Write WAL bytes to data dir

## Problem

Extension pg_stat_statements has function pg_stat_statements_reset().
In vanilla Postgres this function can only be called by superuser role
or other users/roles explicitly granted.
In Neon no end user can use superuser role.
Instead we have neon_superuser role.
We need to grant execute on pg_stat_statements_reset() to neon_superuser

## Summary of changes

Modify the Postgres v14, v15, v16 contrib in our compute docker file to
grant execute on pg_stat_statements_reset() to neon_superuser.
(Modifying it in our docker file is preferable to changes in
neondatabase/postgres because we want to limit the changes in our fork
that we have to carry with each new version of Postgres).

Note that the interface of proc/function pg_stat_statements_reset
changed in pg_stat_statements version 1.7

So for versions up to and including 1.6 we must

`GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO
neon_superuser;`

and for versions starting from 1.7 we must

`GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) TO
neon_superuser;`

If we just use `GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO
neon_superuser;` for all version this results in the following error for
versions 1.7+:

```sql
neondb=> create extension pg_stat_statements;
ERROR:  function pg_stat_statements_reset() does not exist
```



## Checklist before requesting a review

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

## Checklist before merging

- [ ] Do not forget to reformat commit message to not include the above
checklist

## I have run the following test and could now invoke
pg_stat_statements_reset() using default user

```bash
(neon) peterbendel@Peters-MBP neon % kubectl get pods | grep compute-quiet-mud-88416983       
compute-quiet-mud-88416983-74f4bf67db-crl4c            3/3     Running     0          7m26s
(neon) peterbendel@Peters-MBP neon % kubectl set image deploy/compute-quiet-mud-88416983 compute-node=neondatabase/compute-node-v15:7307610371
deployment.apps/compute-quiet-mud-88416983 image updated
(neon) peterbendel@Peters-MBP neon % psql postgresql://peterbendel:<secret>@ep-bitter-sunset-73589702.us-east-2.aws.neon.build/neondb     
psql (16.1, server 15.5)
SSL connection (protocol: TLSv1.3, cipher: TLS_AES_256_GCM_SHA384, compression: off)
Type "help" for help.

neondb=> select version();
                                              version                                              
---------------------------------------------------------------------------------------------------
 PostgreSQL 15.5 on x86_64-pc-linux-gnu, compiled by gcc (Debian 10.2.1-6) 10.2.1 20210110, 64-bit
(1 row)

neondb=> create extension pg_stat_statements;
CREATE EXTENSION

neondb=> select pg_stat_statements_reset();
 pg_stat_statements_reset 
--------------------------
 
(1 row)
```

.github/workflows/benchmarking.yml

@@ -11,7 +11,7 @@ on:
     #          │ │ ┌───────────── day of the month (1 - 31)
     #          │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
     #          │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
-    - cron:  '0 3 * * *' # run once a day, timezone is utc
+    - cron:   '0 3 * * *' # run once a day, timezone is utc
 
   workflow_dispatch: # adds ability to run this manually
     inputs:
@@ -23,6 +23,21 @@ on:
         type: boolean
         description: 'Publish perf report. If not set, the report will be published only for the main branch'
         required: false
+      collect_olap_explain:
+        type: boolean
+        description: 'Collect EXPLAIN ANALYZE for OLAP queries. If not set, EXPLAIN ANALYZE will not be collected'
+        required: false
+        default: false
+      collect_pg_stat_statements:
+        type: boolean
+        description: 'Collect pg_stat_statements for OLAP queries. If not set, pg_stat_statements will not be collected'
+        required: false
+        default: false
+      run_AWS_RDS_AND_AURORA:
+        type: boolean
+        description: 'AWS-RDS and AWS-AURORA normally only run on Saturday. Set this to true to run them on every workflow_dispatch'
+        required: false
+        default: false
 
 defaults:
   run:
@@ -113,6 +128,8 @@ jobs:
     # - neon-captest-reuse: Reusing existing project
     # - rds-aurora: Aurora Postgres Serverless v2 with autoscaling from 0.5 to 2 ACUs
     # - rds-postgres: RDS Postgres db.m5.large instance (2 vCPU, 8 GiB) with gp3 EBS storage
+    env:
+      RUN_AWS_RDS_AND_AURORA: ${{ github.event.inputs.run_AWS_RDS_AND_AURORA || 'false' }}
     runs-on: ubuntu-latest
     outputs:
       pgbench-compare-matrix: ${{ steps.pgbench-compare-matrix.outputs.matrix }}
@@ -152,7 +169,7 @@ jobs:
           ]
         }'
 
-        if [ "$(date +%A)" = "Saturday" ]; then
+        if [ "$(date +%A)" = "Saturday" ] || [ ${RUN_AWS_RDS_AND_AURORA} = "true" ]; then
           matrix=$(echo "$matrix" | jq '.include += [{ "platform": "rds-postgres" },
                                                    { "platform": "rds-aurora"   }]')
         fi
@@ -171,9 +188,9 @@ jobs:
           ]
         }'
 
-        if [ "$(date +%A)" = "Saturday" ]; then
+        if [ "$(date +%A)" = "Saturday" ] || [ ${RUN_AWS_RDS_AND_AURORA} = "true" ]; then
           matrix=$(echo "$matrix" | jq '.include += [{ "platform": "rds-postgres", "scale": "10" },
-                                                   { "platform": "rds-aurora",   "scale": "10" }]')
+                                                    { "platform": "rds-aurora",   "scale": "10" }]')
         fi
 
         echo "matrix=$(echo "$matrix" | jq --compact-output '.')" >> $GITHUB_OUTPUT
@@ -337,6 +354,8 @@ jobs:
       POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
       DEFAULT_PG_VERSION: 14
       TEST_OUTPUT: /tmp/test_output
+      TEST_OLAP_COLLECT_EXPLAIN: ${{ github.event.inputs.collect_olap_explain }}
+      TEST_OLAP_COLLECT_PG_STAT_STATEMENTS: ${{ github.event.inputs.collect_pg_stat_statements }}
       BUILD_TYPE: remote
       SAVE_PERF_REPORT: ${{ github.event.inputs.save_perf_report || ( github.ref_name == 'main' ) }}
       PLATFORM: ${{ matrix.platform }}
@@ -399,6 +418,8 @@ jobs:
       env:
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
+        TEST_OLAP_COLLECT_EXPLAIN: ${{ github.event.inputs.collect_olap_explain || 'false' }}
+        TEST_OLAP_COLLECT_PG_STAT_STATEMENTS: ${{ github.event.inputs.collect_pg_stat_statements || 'false' }}
         BENCHMARK_CONNSTR: ${{ steps.set-up-connstr.outputs.connstr }}
         TEST_OLAP_SCALE: 10
 

.github/workflows/build_and_push_docker_image.yml

@@ -0,0 +1,105 @@
+name: Build and Push Docker Image
+
+on:
+  workflow_call:
+    inputs:
+      dockerfile-path:
+        required: true
+        type: string
+      image-name:
+        required: true
+        type: string
+    outputs:
+      build-tools-tag:
+        description: "tag generated for build tools"
+        value: ${{ jobs.tag.outputs.build-tools-tag }}
+
+jobs:
+  check-if-build-tools-dockerfile-changed:
+    runs-on: ubuntu-latest
+    outputs:
+      docker_file_changed: ${{ steps.dockerfile.outputs.docker_file_changed }}
+    steps:
+      - name: Check if Dockerfile.buildtools has changed
+        id: dockerfile
+        run: |
+          if [[ "$GITHUB_EVENT_NAME" != "pull_request" ]]; then
+            echo "docker_file_changed=false" >> $GITHUB_OUTPUT
+            exit
+          fi
+          updated_files=$(gh pr --repo neondatabase/neon diff ${{ github.event.pull_request.number }} --name-only)
+          if [[ $updated_files == *"Dockerfile.buildtools"* ]]; then
+            echo "docker_file_changed=true" >> $GITHUB_OUTPUT
+          fi
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+  tag:
+    runs-on: ubuntu-latest
+    needs: [ check-if-build-tools-dockerfile-changed ]
+    outputs:
+      build-tools-tag: ${{steps.buildtools-tag.outputs.image_tag}}
+
+    steps:
+      - name: Get buildtools tag
+        env:
+          DOCKERFILE_CHANGED: ${{ needs.check-if-build-tools-dockerfile-changed.outputs.docker_file_changed }}
+        run: |
+          if [[ "$GITHUB_EVENT_NAME" == "pull_request" ]] && [[ "${DOCKERFILE_CHANGED}" == "true" ]]; then
+            IMAGE_TAG=$GITHUB_RUN_ID
+          else
+            IMAGE_TAG=pinned
+          fi
+
+          echo "image_tag=${IMAGE_TAG}" >> $GITHUB_OUTPUT
+        shell: bash
+        id: buildtools-tag
+
+  kaniko:
+    if: needs.check-if-build-tools-dockerfile-changed.outputs.docker_file_changed == 'true'
+    needs: [ tag, check-if-build-tools-dockerfile-changed ]
+    runs-on: [ self-hosted, dev, x64 ]
+    container: gcr.io/kaniko-project/executor:v1.7.0-debug
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v1
+
+      - name: Configure ECR login
+        run: echo "{\"credsStore\":\"ecr-login\"}" > /kaniko/.docker/config.json
+
+      - name: Kaniko build
+        run: /kaniko/executor --reproducible --snapshotMode=redo --skip-unused-stages --dockerfile ${{ inputs.dockerfile-path }} --cache=true --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache  --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}-amd64
+
+  kaniko-arm:
+    if: needs.check-if-build-tools-dockerfile-changed.outputs.docker_file_changed == 'true'
+    needs: [ tag, check-if-build-tools-dockerfile-changed ]
+    runs-on: [ self-hosted, dev, arm64 ]
+    container: gcr.io/kaniko-project/executor:v1.7.0-debug
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v1
+
+      - name: Configure ECR login
+        run: echo "{\"credsStore\":\"ecr-login\"}" > /kaniko/.docker/config.json
+
+      - name: Kaniko build
+        run: /kaniko/executor --reproducible --snapshotMode=redo --skip-unused-stages --dockerfile ${{ inputs.dockerfile-path }} --cache=true --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}-arm64
+
+  manifest:
+    if: needs.check-if-build-tools-dockerfile-changed.outputs.docker_file_changed == 'true'
+    name: 'manifest'
+    runs-on: [ self-hosted, dev, x64 ]
+    needs:
+      - tag
+      - kaniko
+      - kaniko-arm
+      - check-if-build-tools-dockerfile-changed
+
+    steps:
+      - name: Create manifest
+        run: docker manifest create 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }} --amend 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}-amd64 --amend 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}-arm64
+
+      - name: Push manifest
+        run: docker manifest push 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}

.github/workflows/build_and_test.yml

@@ -44,7 +44,6 @@ jobs:
 
         exit 1
 
-
   tag:
     needs: [ check-permissions ]
     runs-on: [ self-hosted, gen3, small ]
@@ -74,11 +73,19 @@ jobs:
         shell: bash
         id: build-tag
 
-  check-codestyle-python:
+  build-buildtools-image:
     needs: [ check-permissions ]
+    uses: ./.github/workflows/build_and_push_docker_image.yml
+    with:
+      dockerfile-path: Dockerfile.buildtools
+      image-name: build-tools
+    secrets: inherit
+
+  check-codestyle-python:
+    needs: [ check-permissions, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, small ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
 
     steps:
@@ -108,10 +115,10 @@ jobs:
         run: poetry run mypy .
 
   check-codestyle-rust:
-    needs: [ check-permissions ]
+    needs: [ check-permissions, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, large ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
 
     steps:
@@ -175,10 +182,10 @@ jobs:
         run: cargo deny check --hide-inclusion-graph
 
   build-neon:
-    needs: [ check-permissions, tag ]
+    needs: [ check-permissions, tag, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, large ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
     strategy:
       fail-fast: false
@@ -338,7 +345,7 @@ jobs:
 
           # Run separate tests for real S3
           export ENABLE_REAL_S3_REMOTE_STORAGE=nonempty
-          export REMOTE_STORAGE_S3_BUCKET=neon-github-public-dev
+          export REMOTE_STORAGE_S3_BUCKET=neon-github-ci-tests
           export REMOTE_STORAGE_S3_REGION=eu-central-1
           # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
           ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_s3
@@ -408,10 +415,10 @@ jobs:
         uses: ./.github/actions/save-coverage-data
 
   regress-tests:
-    needs: [ check-permissions, build-neon, tag ]
+    needs: [ check-permissions, build-neon, build-buildtools-image, tag ]
     runs-on: [ self-hosted, gen3, large ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       # Default shared memory is 64mb
       options: --init --shm-size=512mb
     strategy:
@@ -447,10 +454,10 @@ jobs:
         uses: ./.github/actions/save-coverage-data
 
   benchmarks:
-    needs: [ check-permissions, build-neon ]
+    needs: [ check-permissions, build-neon, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, small ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       # Default shared memory is 64mb
       options: --init --shm-size=512mb
     if: github.ref_name == 'main' || contains(github.event.pull_request.labels.*.name, 'run-benchmarks')
@@ -479,12 +486,12 @@ jobs:
       # while coverage is currently collected for the debug ones
 
   create-test-report:
-    needs: [ check-permissions, regress-tests, coverage-report, benchmarks ]
+    needs: [ check-permissions, regress-tests, coverage-report, benchmarks, build-buildtools-image ]
     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
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
 
     steps:
@@ -526,11 +533,10 @@ jobs:
             })
 
   coverage-report:
-    needs: [ check-permissions, regress-tests ]
-
+    needs: [ check-permissions, regress-tests, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, small ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
     strategy:
       fail-fast: false
@@ -694,7 +700,7 @@ jobs:
             }"
 
   neon-image:
-    needs: [ check-permissions, tag ]
+    needs: [ check-permissions, build-buildtools-image, tag ]
     runs-on: [ self-hosted, gen3, large ]
     container: gcr.io/kaniko-project/executor:v1.9.2-debug
     defaults:
@@ -733,6 +739,7 @@ jobs:
                            --context .
                            --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
                            --build-arg BUILD_TAG=${{ needs.tag.outputs.build-tag }}
+                           --build-arg TAG=${{ needs.build-buildtools-image.outputs.build-tools-tag }}
                            --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
                            --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:${{needs.tag.outputs.build-tag}}
                            --destination neondatabase/neon:${{needs.tag.outputs.build-tag}}
@@ -743,7 +750,7 @@ jobs:
 
   compute-tools-image:
     runs-on: [ self-hosted, gen3, large ]
-    needs: [ check-permissions, tag ]
+    needs: [ check-permissions, build-buildtools-image, tag ]
     container: gcr.io/kaniko-project/executor:v1.9.2-debug
     defaults:
       run:
@@ -778,6 +785,7 @@ jobs:
                            --context .
                            --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
                            --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}}
+                           --build-arg TAG=${{needs.build-buildtools-image.outputs.build-tools-tag}}
                            --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
                            --dockerfile Dockerfile.compute-tools
                            --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}}
@@ -788,7 +796,7 @@ jobs:
         run: rm -rf ~/.ecr
 
   compute-node-image:
-    needs: [ check-permissions, tag ]
+    needs: [ check-permissions, build-buildtools-image, tag ]
     runs-on: [ self-hosted, gen3, large ]
     container:
       image: gcr.io/kaniko-project/executor:v1.9.2-debug
@@ -836,6 +844,7 @@ jobs:
                            --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
                            --build-arg PG_VERSION=${{ matrix.version }}
                            --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}}
+                           --build-arg TAG=${{needs.build-buildtools-image.outputs.build-tools-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/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
@@ -857,7 +866,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.19.0
+      VM_BUILDER_VERSION: v0.21.0
 
     steps:
       - name: Checkout

.github/workflows/neon_extra_builds.yml

@@ -218,7 +218,7 @@ jobs:
 
           # Run separate tests for real S3
           export ENABLE_REAL_S3_REMOTE_STORAGE=nonempty
-          export REMOTE_STORAGE_S3_BUCKET=neon-github-public-dev
+          export REMOTE_STORAGE_S3_BUCKET=neon-github-ci-tests
           export REMOTE_STORAGE_S3_REGION=eu-central-1
           # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
           cargo test $CARGO_FLAGS --package remote_storage --test test_real_s3

.github/workflows/update_build_tools_image.yml

@@ -0,0 +1,130 @@
+name: 'Update build tools image tag'
+
+# This workflow it used to update tag of build tools in ECR.
+# The most common use case is adding/moving `pinned` tag to `${GITHUB_RUN_IT}` image.
+
+on:
+  workflow_dispatch:
+    inputs:
+      from-tag:
+        description: 'Source tag'
+        required: true
+        type: string
+      to-tag:
+        description: 'Destination tag'
+        required: true
+        type: string
+        default: 'pinned'
+
+defaults:
+  run:
+    shell: bash -euo pipefail {0}
+
+env:
+  AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+  AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_KEY_DEV }}
+
+permissions: {}
+
+jobs:
+  tag-image:
+    runs-on: [ self-hosted, gen3, small ]
+    container: golang:1.19-bullseye
+
+    env:
+      IMAGE: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools
+      FROM_TAG: ${{ inputs.from-tag }}
+      TO_TAG: ${{ inputs.to-tag }}
+    outputs:
+      next-digest-buildtools: ${{ steps.next-digest.outputs.next-digest-buildtools }}
+      prev-digest-buildtools: ${{ steps.prev-digest.outputs.prev-digest-buildtools }}
+
+    steps:
+      - name: Install Crane & ECR helper
+        run: |
+          go install github.com/google/go-containerregistry/cmd/crane@a54d64203cffcbf94146e04069aae4a97f228ee2 # v0.16.1
+          go install github.com/awslabs/amazon-ecr-credential-helper/ecr-login/cli/docker-credential-ecr-login@adf1bafd791ae7d4ff098108b1e91f36a4da5404 # v0.7.1
+
+      - name: Configure ECR login
+        run: |
+          mkdir /github/home/.docker/
+          echo "{\"credsStore\":\"ecr-login\"}" > /github/home/.docker/config.json
+
+      - name: Get source image digest
+        id: next-digest
+        run: |
+          NEXT_DIGEST=$(crane digest ${IMAGE}:${FROM_TAG} || true)
+          if [ -z "${NEXT_DIGEST}" ]; then
+            echo >&2 "Image ${IMAGE}:${FROM_TAG} does not exist"
+            exit 1
+          fi
+
+          echo "Current ${IMAGE}@${FROM_TAG} image is ${IMAGE}@${NEXT_DIGEST}"
+          echo "next-digest-buildtools=$NEXT_DIGEST" >> $GITHUB_OUTPUT
+
+      - name: Get destination image digest (if already exists)
+        id: prev-digest
+        run: |
+          PREV_DIGEST=$(crane digest ${IMAGE}:${TO_TAG} || true)
+          if [ -z "${PREV_DIGEST}" ]; then
+            echo >&2 "Image ${IMAGE}:${TO_TAG} does not exist (it's ok)"
+          else
+            echo >&2 "Current ${IMAGE}@${TO_TAG} image is ${IMAGE}@${PREV_DIGEST}"
+
+            echo "prev-digest-buildtools=$PREV_DIGEST" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Tag image
+        run: |
+          crane tag "${IMAGE}:${FROM_TAG}" "${TO_TAG}"
+
+  rollback-tag-image:
+    needs:  tag-image
+    if: ${{ !success() }}
+
+    runs-on: [ self-hosted, gen3, small ]
+    container: golang:1.19-bullseye
+
+    env:
+      IMAGE: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools
+      FROM_TAG: ${{ inputs.from-tag }}
+      TO_TAG: ${{ inputs.to-tag }}
+
+    steps:
+      - name: Install Crane & ECR helper
+        run: |
+          go install github.com/google/go-containerregistry/cmd/crane@a54d64203cffcbf94146e04069aae4a97f228ee2 # v0.16.1
+          go install github.com/awslabs/amazon-ecr-credential-helper/ecr-login/cli/docker-credential-ecr-login@adf1bafd791ae7d4ff098108b1e91f36a4da5404 # v0.7.1
+
+      - name: Configure ECR login
+        run: |
+          mkdir /github/home/.docker/
+          echo "{\"credsStore\":\"ecr-login\"}" > /github/home/.docker/config.json
+
+      - name: Restore previous tag if needed
+        run: |
+          NEXT_DIGEST="${{ needs.tag-image.outputs.next-digest-buildtools }}"
+          PREV_DIGEST="${{ needs.tag-image.outputs.prev-digest-buildtools }}"
+
+          if [ -z "${NEXT_DIGEST}" ]; then
+            echo >&2 "Image ${IMAGE}:${FROM_TAG} does not exist, nothing to rollback"
+            exit 0
+          fi
+
+          if [ -z "${PREV_DIGEST}" ]; then
+            # I guess we should delete the tag here/untag the image, but crane does not support it
+            # - https://github.com/google/go-containerregistry/issues/999
+
+            echo >&2 "Image ${IMAGE}:${TO_TAG} did not exist, but it was created by the job, no need to rollback"
+
+            exit 0
+          fi
+
+          CURRENT_DIGEST=$(crane digest "${IMAGE}:${TO_TAG}")
+          if [ "${CURRENT_DIGEST}" == "${NEXT_DIGEST}" ]; then
+            crane tag "${IMAGE}@${PREV_DIGEST}" "${TO_TAG}"
+
+            echo >&2 "Successfully restored ${TO_TAG} tag from ${IMAGE}@${CURRENT_DIGEST} to ${IMAGE}@${PREV_DIGEST}"
+          else
+            echo >&2 "Image ${IMAGE}:${TO_TAG}@${CURRENT_DIGEST} is not required to be restored"
+          fi

.gitignore

@@ -6,6 +6,7 @@ __pycache__/
 test_output/
 .vscode
 .idea
+neon.iml
 /.neon
 /integration_tests/.neon
 

CONTRIBUTING.md

@@ -70,3 +70,17 @@ We're using the following approach to make it work:
 - 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)
+
+## How do I add the "pinned" tag to an buildtools image?
+We use the `pinned` tag for `Dockerfile.buildtools` build images in our CI/CD setup, currently adding the `pinned` tag is a manual operation.
+
+You can call it from GitHub UI: https://github.com/neondatabase/neon/actions/workflows/update_build_tools_image.yml,
+or using GitHub CLI:
+
+```bash
+gh workflow -R neondatabase/neon run update_build_tools_image.yml \
+            -f from-tag=6254913013 \
+            -f to-tag=pinned \
+
+# Default `-f to-tag` is `pinned`, so the parameter can be omitted.
+```

Cargo.lock

@@ -190,9 +190,9 @@ dependencies = [
 
 [[package]]
 name = "async-compression"
-version = "0.4.0"
+version = "0.4.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5b0122885821398cc923ece939e24d1056a2384ee719432397fa9db87230ff11"
+checksum = "bc2d0cfb2a7388d34f590e76686704c494ed7aaceed62ee1ba35cbf363abc2a5"
 dependencies = [
  "flate2",
  "futures-core",
@@ -233,7 +233,7 @@ checksum = "16e62a023e7c117e27523144c5d2459f4397fcc3cab0085af8e2224f643a0193"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -244,7 +244,7 @@ checksum = "b9ccdd8f2a161be9bd5c023df56f1b2a0bd1d83872ae53b71a84a12c9bf6e842"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -881,7 +881,7 @@ dependencies = [
  "regex",
  "rustc-hash",
  "shlex",
- "syn 2.0.28",
+ "syn 2.0.32",
  "which",
 ]
 
@@ -1095,7 +1095,7 @@ dependencies = [
  "heck",
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -1168,6 +1168,7 @@ dependencies = [
  "regex",
  "remote_storage",
  "reqwest",
+ "rust-ini",
  "serde",
  "serde_json",
  "tar",
@@ -1201,6 +1202,26 @@ version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "28c122c3980598d243d63d9a704629a2d748d101f278052ff068be5a4423ab6f"
 
+[[package]]
+name = "const-random"
+version = "0.1.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5aaf16c9c2c612020bcfd042e170f6e32de9b9d75adb5277cdbbd2e2c8c8299a"
+dependencies = [
+ "const-random-macro",
+]
+
+[[package]]
+name = "const-random-macro"
+version = "0.1.16"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f9d839f2a20b0aee515dc581a6172f2321f96cab76c1a38a4c584a194955390e"
+dependencies = [
+ "getrandom 0.2.11",
+ "once_cell",
+ "tiny-keccak",
+]
+
 [[package]]
 name = "const_fn"
 version = "0.4.9"
@@ -1245,16 +1266,19 @@ name = "control_plane"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "async-trait",
  "camino",
  "clap",
  "comfy-table",
  "compute_api",
+ "futures",
  "git-version",
  "hex",
  "hyper",
  "nix 0.26.2",
  "once_cell",
  "pageserver_api",
+ "pageserver_client",
  "postgres",
  "postgres_backend",
  "postgres_connection",
@@ -1268,6 +1292,8 @@ dependencies = [
  "tar",
  "thiserror",
  "tokio",
+ "tokio-postgres",
+ "tokio-util",
  "toml",
  "tracing",
  "url",
@@ -1428,6 +1454,12 @@ dependencies = [
  "winapi",
 ]
 
+[[package]]
+name = "crunchy"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7a81dae078cea95a014a339291cec439d2f232ebe854a9d672b796c6afafa9b7"
+
 [[package]]
 name = "crypto-bigint"
 version = "0.4.9"
@@ -1481,7 +1513,7 @@ dependencies = [
  "proc-macro2",
  "quote",
  "strsim",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -1492,7 +1524,7 @@ checksum = "29a358ff9f12ec09c3e61fef9b5a9902623a695a46a917b07f269bff1445611a"
 dependencies = [
  "darling_core",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -1567,7 +1599,16 @@ checksum = "487585f4d0c6655fe74905e2504d8ad6908e4db67f744eb140876906c2f3175d"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
+]
+
+[[package]]
+name = "dlv-list"
+version = "0.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "442039f5147480ba31067cb00ada1adae6892028e40e45fc5de7b7df6dcc1b5f"
+dependencies = [
+ "const-random",
 ]
 
 [[package]]
@@ -1661,7 +1702,7 @@ dependencies = [
  "darling",
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -1915,7 +1956,7 @@ checksum = "89ca545a94061b6365f2c7355b4b32bd20df3ff95f02da9329b34ccc3bd6ee72"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -2101,6 +2142,20 @@ dependencies = [
  "hashbrown 0.13.2",
 ]
 
+[[package]]
+name = "hdrhistogram"
+version = "7.5.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "765c9198f173dd59ce26ff9f95ef0aafd0a0fe01fb9d72841bc5066a4c06511d"
+dependencies = [
+ "base64 0.21.1",
+ "byteorder",
+ "crossbeam-channel",
+ "flate2",
+ "nom",
+ "num-traits",
+]
+
 [[package]]
 name = "heapless"
 version = "0.8.0"
@@ -2482,13 +2537,14 @@ dependencies = [
 
 [[package]]
 name = "jsonwebtoken"
-version = "8.3.0"
+version = "9.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6971da4d9c3aa03c3d8f3ff0f4155b534aad021292003895a469716b2a230378"
+checksum = "5c7ea04a7c5c055c175f189b6dc6ba036fd62306b58c66c9f6389036c503a3f4"
 dependencies = [
  "base64 0.21.1",
- "pem 1.1.1",
- "ring 0.16.20",
+ "js-sys",
+ "pem 3.0.3",
+ "ring 0.17.6",
  "serde",
  "serde_json",
  "simple_asn1",
@@ -2901,7 +2957,7 @@ checksum = "a948666b637a0f465e8564c73e89d4dde00d72d4d473cc972f390fc3dcee7d9c"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -3023,6 +3079,16 @@ dependencies = [
  "tokio-stream",
 ]
 
+[[package]]
+name = "ordered-multimap"
+version = "0.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a4d6a8c22fc714f0c2373e6091bf6f5e9b37b1bc0b1184874b7e0a4e303d318f"
+dependencies = [
+ "dlv-list",
+ "hashbrown 0.14.0",
+]
+
 [[package]]
 name = "os_info"
 version = "3.7.0"
@@ -3051,6 +3117,28 @@ dependencies = [
  "sha2",
 ]
 
+[[package]]
+name = "pagebench"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "clap",
+ "futures",
+ "hdrhistogram",
+ "humantime",
+ "humantime-serde",
+ "pageserver",
+ "pageserver_api",
+ "pageserver_client",
+ "rand 0.8.5",
+ "serde",
+ "serde_json",
+ "tokio",
+ "tracing",
+ "utils",
+ "workspace_hack",
+]
+
 [[package]]
 name = "pagectl"
 version = "0.1.0"
@@ -3140,6 +3228,7 @@ dependencies = [
  "tokio",
  "tokio-io-timeout",
  "tokio-postgres",
+ "tokio-stream",
  "tokio-tar",
  "tokio-util",
  "toml_edit",
@@ -3162,6 +3251,7 @@ dependencies = [
  "enum-map",
  "hex",
  "postgres_ffi",
+ "rand 0.8.5",
  "serde",
  "serde_json",
  "serde_with",
@@ -3172,6 +3262,27 @@ dependencies = [
  "workspace_hack",
 ]
 
+[[package]]
+name = "pageserver_client"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "bytes",
+ "futures",
+ "pageserver_api",
+ "postgres",
+ "reqwest",
+ "serde",
+ "thiserror",
+ "tokio",
+ "tokio-postgres",
+ "tokio-stream",
+ "tokio-util",
+ "utils",
+ "workspace_hack",
+]
+
 [[package]]
 name = "parking"
 version = "2.1.1"
@@ -3263,18 +3374,19 @@ checksum = "19b17cddbe7ec3f8bc800887bab5e717348c95ea2ca0b1bf0837fb964dc67099"
 
 [[package]]
 name = "pem"
-version = "1.1.1"
+version = "2.0.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a8835c273a76a90455d7344889b0964598e3316e2a79ede8e36f16bdcf2228b8"
+checksum = "6b13fe415cdf3c8e44518e18a7c95a13431d9bdf6d15367d82b23c377fdd441a"
 dependencies = [
- "base64 0.13.1",
+ "base64 0.21.1",
+ "serde",
 ]
 
 [[package]]
 name = "pem"
-version = "2.0.1"
+version = "3.0.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6b13fe415cdf3c8e44518e18a7c95a13431d9bdf6d15367d82b23c377fdd441a"
+checksum = "1b8fcc794035347fb64beda2d3b462595dd2753e3f268d89c5aae77e8cf2c310"
 dependencies = [
  "base64 0.21.1",
  "serde",
@@ -3331,7 +3443,7 @@ checksum = "39407670928234ebc5e6e580247dd567ad73a3578460c5990f9503df207e8f07"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -3538,7 +3650,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3b69d39aab54d069e7f2fe8cb970493e7834601ca2d8c65fd7bbd183578080d1"
 dependencies = [
  "proc-macro2",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -4146,10 +4258,20 @@ dependencies = [
  "regex",
  "relative-path",
  "rustc_version",
- "syn 2.0.28",
+ "syn 2.0.32",
  "unicode-ident",
 ]
 
+[[package]]
+name = "rust-ini"
+version = "0.20.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3e0698206bcb8882bf2a9ecb4c1e7785db57ff052297085a6efd4fe42302068a"
+dependencies = [
+ "cfg-if",
+ "ordered-multimap",
+]
+
 [[package]]
 name = "rustc-demangle"
 version = "0.1.23"
@@ -4292,6 +4414,7 @@ dependencies = [
  "histogram",
  "itertools",
  "pageserver",
+ "pageserver_api",
  "rand 0.8.5",
  "remote_storage",
  "reqwest",
@@ -4399,12 +4522,12 @@ checksum = "d29ab0c6d3fc0ee92fe66e2d99f700eab17a8d57d1c1d3b748380fb20baa78cd"
 
 [[package]]
 name = "sct"
-version = "0.7.0"
+version = "0.7.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d53dcdb7c9f8158937a7981b48accfd39a43af418591a5d008c7b22b5e1b7ca4"
+checksum = "da046153aa2352493d6cb7da4b6e5c0c057d8a1d0a9aa8560baffdd945acd414"
 dependencies = [
- "ring 0.16.20",
- "untrusted 0.7.1",
+ "ring 0.17.6",
+ "untrusted 0.9.0",
 ]
 
 [[package]]
@@ -4580,7 +4703,7 @@ checksum = "aafe972d60b0b9bee71a91b92fee2d4fb3c9d7e8f6b179aa99f27203d99a4816"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -4661,7 +4784,7 @@ dependencies = [
  "darling",
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -4928,9 +5051,9 @@ dependencies = [
 
 [[package]]
 name = "syn"
-version = "2.0.28"
+version = "2.0.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "04361975b3f5e348b2189d8dc55bc942f278b2d482a6a0365de5bdd62d351567"
+checksum = "239814284fd6f1a4ffe4ca893952cdd93c224b6a1571c9a9eadd670295c0c9e2"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -5060,7 +5183,7 @@ checksum = "f9456a42c5b0d803c8cd86e73dd7cc9edd429499f37a3550d286d5e86720569f"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -5103,6 +5226,15 @@ dependencies = [
  "time-core",
 ]
 
+[[package]]
+name = "tiny-keccak"
+version = "2.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2c9d3793400a45f954c52e73d068316d76b6f4e36977e3fcebb13a2721e80237"
+dependencies = [
+ "crunchy",
+]
+
 [[package]]
 name = "tinytemplate"
 version = "1.2.1"
@@ -5178,7 +5310,7 @@ checksum = "5b8a1e28f2deaa14e508979454cb3a223b10b938b45af148bc0986de36f1923b"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -5479,7 +5611,7 @@ checksum = "0f57e3ca2a01450b1a921183a9c9cbfda207fd822cef4ccb00a65402cbba7a74"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -5924,7 +6056,7 @@ dependencies = [
  "once_cell",
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
  "wasm-bindgen-shared",
 ]
 
@@ -5958,7 +6090,7 @@ checksum = "e128beba882dd1eb6200e1dc92ae6c5dbaa4311aa7bb211ca035779e5efc39f8"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
  "wasm-bindgen-backend",
  "wasm-bindgen-shared",
 ]
@@ -6270,6 +6402,7 @@ dependencies = [
  "futures-io",
  "futures-sink",
  "futures-util",
+ "getrandom 0.2.11",
  "hex",
  "hmac",
  "hyper",
@@ -6281,6 +6414,7 @@ dependencies = [
  "num-bigint",
  "num-integer",
  "num-traits",
+ "once_cell",
  "prost",
  "rand 0.8.5",
  "regex",
@@ -6295,7 +6429,7 @@ dependencies = [
  "smallvec",
  "subtle",
  "syn 1.0.109",
- "syn 2.0.28",
+ "syn 2.0.32",
  "time",
  "time-macros",
  "tokio",
@@ -6357,22 +6491,22 @@ dependencies = [
 
 [[package]]
 name = "zerocopy"
-version = "0.7.3"
+version = "0.7.31"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7a7af71d8643341260a65f89fa60c0eeaa907f34544d8f6d9b0df72f069b5e74"
+checksum = "1c4061bedbb353041c12f413700357bec76df2c7e2ca8e4df8bac24c6bf68e3d"
 dependencies = [
  "zerocopy-derive",
 ]
 
 [[package]]
 name = "zerocopy-derive"
-version = "0.7.3"
+version = "0.7.31"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9731702e2f0617ad526794ae28fbc6f6ca8849b5ba729666c2a5bc4b6ddee2cd"
+checksum = "b3c129550b3e6de3fd0ba67ba5c81818f9805e58b8d7fee80a3a59d2c9fc601a"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.28",
+ "syn 2.0.32",
 ]
 
 [[package]]
@@ -6383,30 +6517,28 @@ checksum = "2a0956f1ba7c7909bfb66c2e9e4124ab6f6482560f6628b5aaeba39207c9aad9"
 
 [[package]]
 name = "zstd"
-version = "0.12.4"
+version = "0.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a27595e173641171fc74a1232b7b1c7a7cb6e18222c11e9dfb9888fa424c53c"
+checksum = "bffb3309596d527cfcba7dfc6ed6052f1d39dfbd7c867aa2e865e4a449c10110"
 dependencies = [
  "zstd-safe",
 ]
 
 [[package]]
 name = "zstd-safe"
-version = "6.0.6"
+version = "7.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ee98ffd0b48ee95e6c5168188e44a54550b1564d9d530ee21d5f0eaed1069581"
+checksum = "43747c7422e2924c11144d5229878b98180ef8b06cca4ab5af37afc8a8d8ea3e"
 dependencies = [
- "libc",
  "zstd-sys",
 ]
 
 [[package]]
 name = "zstd-sys"
-version = "2.0.8+zstd.1.5.5"
+version = "2.0.9+zstd.1.5.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5556e6ee25d32df2586c098bbfa278803692a20d0ab9565e049480d52707ec8c"
+checksum = "9e16efa8a874a0481a574084d34cc26fdb3b99627480f785888deb6386506656"
 dependencies = [
  "cc",
- "libc",
  "pkg-config",
 ]

Cargo.toml

@@ -5,6 +5,8 @@ members = [
     "control_plane",
     "pageserver",
     "pageserver/ctl",
+    "pageserver/client",
+    "pageserver/pagebench",
     "proxy",
     "safekeeper",
     "storage_broker",
@@ -78,6 +80,7 @@ futures-util = "0.3"
 git-version = "0.3"
 hashbrown = "0.13"
 hashlink = "0.8.1"
+hdrhistogram = "7.5.2"
 hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
@@ -90,7 +93,7 @@ hyper-tungstenite = "0.11"
 inotify = "0.10.2"
 ipnet = "2.9.0"
 itertools = "0.10"
-jsonwebtoken = "8"
+jsonwebtoken = "9"
 libc = "0.2"
 md5 = "0.7.0"
 memoffset = "0.8"
@@ -182,6 +185,7 @@ compute_api = { version = "0.1", path = "./libs/compute_api/" }
 consumption_metrics = { version = "0.1", path = "./libs/consumption_metrics/" }
 metrics = { version = "0.1", path = "./libs/metrics/" }
 pageserver_api = { version = "0.1", path = "./libs/pageserver_api/" }
+pageserver_client = { path = "./pageserver/client" }
 postgres_backend = { version = "0.1", path = "./libs/postgres_backend/" }
 postgres_connection = { version = "0.1", path = "./libs/postgres_connection/" }
 postgres_ffi = { version = "0.1", path = "./libs/postgres_ffi/" }
@@ -217,6 +221,9 @@ tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", br
 # This is useful for profiling and, to some extent, debug.
 # Besides, debug info should not affect the performance.
 debug = true
+strip = true  # Automatically strip symbols from the binary.
+opt-level = "z"  # Optimize for size.
+lto = true
 
 # disable debug symbols for all packages except this one to decrease binaries size
 [profile.release.package."*"]

Dockerfile

@@ -3,7 +3,7 @@
 ### By default, the binaries inside the image have some mock parameters and can start, but are not intended to be used
 ### inside this image in the real deployments.
 ARG REPOSITORY=neondatabase
-ARG IMAGE=rust
+ARG IMAGE=build-tools
 ARG TAG=pinned
 
 # Build Postgres

Dockerfile.buildtools

@@ -0,0 +1,165 @@
+FROM debian:bullseye-slim
+
+# Add nonroot user
+RUN useradd -ms /bin/bash nonroot -b /home
+SHELL ["/bin/bash", "-c"]
+
+# System deps
+RUN set -e \
+    && apt update \
+    && apt install -y \
+        autoconf \
+        automake \
+        bison \
+        build-essential \
+        ca-certificates \
+        cmake \
+        curl \
+        flex \
+        git \
+        gnupg \
+        gzip \
+        jq \
+        libcurl4-openssl-dev \
+        libbz2-dev \
+        libffi-dev \
+        liblzma-dev \
+        libncurses5-dev \
+        libncursesw5-dev \
+        libpq-dev \
+        libreadline-dev \
+        libseccomp-dev \
+        libsqlite3-dev \
+        libssl-dev \
+        libstdc++-10-dev \
+        libtool \
+        libxml2-dev \
+        libxmlsec1-dev \
+        libxxhash-dev \
+        lsof \
+        make \
+        netcat \
+        net-tools \
+        openssh-client \
+        parallel \
+        pkg-config \
+        unzip \
+        wget \
+        xz-utils \
+        zlib1g-dev \
+        zstd \
+    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+
+# protobuf-compiler (protoc)
+ENV PROTOC_VERSION 25.1
+RUN curl -fsSL "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-$(uname -m | sed 's/aarch64/aarch_64/g').zip" -o "protoc.zip" \
+    && unzip -q protoc.zip -d protoc \
+    && mv protoc/bin/protoc /usr/local/bin/protoc \
+    && mv protoc/include/google /usr/local/include/google \
+    && rm -rf protoc.zip protoc
+
+# LLVM
+ENV LLVM_VERSION=17
+RUN curl -fsSL 'https://apt.llvm.org/llvm-snapshot.gpg.key' | apt-key add - \
+    && echo "deb http://apt.llvm.org/bullseye/ llvm-toolchain-bullseye-${LLVM_VERSION} main" > /etc/apt/sources.list.d/llvm.stable.list \
+    && apt update \
+    && apt install -y clang-${LLVM_VERSION} llvm-${LLVM_VERSION} \
+    && bash -c 'for f in /usr/bin/clang*-${LLVM_VERSION} /usr/bin/llvm*-${LLVM_VERSION}; do ln -s "${f}" "${f%-${LLVM_VERSION}}"; done' \
+    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+
+# PostgreSQL 14
+RUN curl -fsSL 'https://www.postgresql.org/media/keys/ACCC4CF8.asc' | apt-key add - \
+    && echo 'deb http://apt.postgresql.org/pub/repos/apt bullseye-pgdg main' > /etc/apt/sources.list.d/pgdg.list \
+    && apt update \
+    && apt install -y postgresql-client-14 \
+    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+
+# AWS CLI
+RUN curl "https://awscli.amazonaws.com/awscli-exe-linux-$(uname -m).zip" -o "awscliv2.zip" \
+    && unzip -q awscliv2.zip \
+    && ./aws/install \
+    && rm awscliv2.zip
+
+# Mold: A Modern Linker
+ENV MOLD_VERSION v2.4.0
+RUN set -e \
+    && git clone https://github.com/rui314/mold.git \
+    && mkdir mold/build \
+    && cd mold/build \
+    && git checkout ${MOLD_VERSION} \
+    && cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_COMPILER=clang++ .. \
+    && cmake --build . -j $(nproc) \
+    && cmake --install . \
+    && cd .. \
+    && rm -rf mold
+
+# LCOV
+# Build lcov from a fork:
+# It includes several bug fixes on top on v2.0 release (https://github.com/linux-test-project/lcov/compare/v2.0...master)
+# And patches from us:
+# - Generates json file with code coverage summary (https://github.com/neondatabase/lcov/commit/426e7e7a22f669da54278e9b55e6d8caabd00af0.tar.gz)
+RUN for package in Capture::Tiny DateTime Devel::Cover Digest::MD5 File::Spec JSON::XS Memory::Process Time::HiRes JSON; do yes | perl -MCPAN -e "CPAN::Shell->notest('install', '$package')"; done \
+    && wget https://github.com/neondatabase/lcov/archive/426e7e7a22f669da54278e9b55e6d8caabd00af0.tar.gz -O lcov.tar.gz \
+    && echo "61a22a62e20908b8b9e27d890bd0ea31f567a7b9668065589266371dcbca0992  lcov.tar.gz" | sha256sum --check \
+    && mkdir -p lcov && tar -xzf lcov.tar.gz -C lcov --strip-components=1 \
+    && cd lcov \
+    && make install \
+    && rm -rf ../lcov.tar.gz
+
+# Switch to nonroot user
+USER nonroot:nonroot
+WORKDIR /home/nonroot
+
+# Python
+ENV PYTHON_VERSION=3.9.2 \
+    PYENV_ROOT=/home/nonroot/.pyenv \
+    PATH=/home/nonroot/.pyenv/shims:/home/nonroot/.pyenv/bin:/home/nonroot/.poetry/bin:$PATH
+RUN set -e \
+    && cd $HOME \
+    && curl -sSO https://raw.githubusercontent.com/pyenv/pyenv-installer/master/bin/pyenv-installer \
+    && chmod +x pyenv-installer \
+    && ./pyenv-installer \
+    && export PYENV_ROOT=/home/nonroot/.pyenv \
+    && export PATH="$PYENV_ROOT/bin:$PATH" \
+    && export PATH="$PYENV_ROOT/shims:$PATH" \
+    && pyenv install ${PYTHON_VERSION} \
+    && pyenv global ${PYTHON_VERSION} \
+    && python --version \
+    && pip install --upgrade pip \
+    && pip --version \
+    && pip install pipenv wheel poetry
+
+# Switch to nonroot user (again)
+USER nonroot:nonroot
+WORKDIR /home/nonroot
+
+# Rust
+# Please keep the version of llvm (installed above) in sync with rust llvm (`rustc --version --verbose | grep LLVM`)
+ENV RUSTC_VERSION=1.74.0
+ENV RUSTUP_HOME="/home/nonroot/.rustup"
+ENV PATH="/home/nonroot/.cargo/bin:${PATH}"
+RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && whoami && \
+	chmod +x rustup-init && \
+	./rustup-init -y --default-toolchain ${RUSTC_VERSION} && \
+	rm rustup-init && \
+    export PATH="$HOME/.cargo/bin:$PATH" && \
+    . "$HOME/.cargo/env" && \
+    cargo --version && rustup --version && \
+    rustup component add llvm-tools-preview rustfmt clippy && \
+    cargo install --git https://github.com/paritytech/cachepot && \
+    cargo install rustfilt && \
+    cargo install cargo-hakari && \
+    cargo install cargo-deny && \
+    cargo install cargo-hack && \
+    rm -rf /home/nonroot/.cargo/registry && \
+    rm -rf /home/nonroot/.cargo/git
+ENV RUSTC_WRAPPER=cachepot
+
+# Show versions
+RUN whoami \
+    && python --version \
+    && pip --version \
+    && cargo --version --verbose \
+    && rustup --version --verbose \
+    && rustc --version --verbose \
+    && clang --version

Dockerfile.compute-node

@@ -1,6 +1,6 @@
 ARG PG_VERSION
 ARG REPOSITORY=neondatabase
-ARG IMAGE=rust
+ARG IMAGE=build-tools
 ARG TAG=pinned
 ARG BUILD_TAG
 
@@ -48,7 +48,29 @@ RUN cd postgres && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgrowlocks.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgstattuple.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/refint.control && \
-    echo 'trusted = true' >> /usr/local/pgsql/share/extension/xml2.control
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/xml2.control && \
+    # We need to grant EXECUTE on pg_stat_statements_reset() to neon_superuser.
+    # In vanilla postgres this function is limited to Postgres role superuser.
+    # In neon we have neon_superuser role that is not a superuser but replaces superuser in some cases.
+    # We could add the additional grant statements to the postgres repository but it would be hard to maintain, 
+    # whenever we need to pick up a new postgres version and we want to limit the changes in our postgres fork,
+    # so we do it here.
+    old_list="pg_stat_statements--1.0--1.1.sql pg_stat_statements--1.1--1.2.sql pg_stat_statements--1.2--1.3.sql pg_stat_statements--1.3--1.4.sql pg_stat_statements--1.4--1.5.sql pg_stat_statements--1.4.sql pg_stat_statements--1.5--1.6.sql"; \
+    # the first loop is for pg_stat_statement extension version <= 1.6
+    for file in /usr/local/pgsql/share/extension/pg_stat_statements--*.sql; do \
+        filename=$(basename "$file"); \
+        if echo "$old_list" | grep -q -F "$filename"; then \
+            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO neon_superuser;' >> $file; \
+        fi; \
+    done; \
+    # the second loop is for pg_stat_statement extension versions >= 1.7, 
+    # where pg_stat_statement_reset() got 3 additional arguments
+    for file in /usr/local/pgsql/share/extension/pg_stat_statements--*.sql; do \
+        filename=$(basename "$file"); \
+        if ! echo "$old_list" | grep -q -F "$filename"; then \
+            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) TO neon_superuser;' >> $file; \
+        fi; \
+    done      
 
 #########################################################################################
 #
@@ -569,6 +591,23 @@ RUN wget https://github.com/ChenHuajun/pg_roaringbitmap/archive/refs/tags/v0.5.4
     make -j $(getconf _NPROCESSORS_ONLN) install && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/roaringbitmap.control
 
+#########################################################################################
+#
+# Layer "pg-semver-pg-build"
+# compile pg_semver extension
+#
+#########################################################################################
+FROM build-deps AS pg-semver-pg-build
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+ENV PATH "/usr/local/pgsql/bin/:$PATH"
+RUN wget https://github.com/theory/pg-semver/archive/refs/tags/v0.32.1.tar.gz -O pg_semver.tar.gz && \
+    echo "fbdaf7512026d62eec03fad8687c15ed509b6ba395bff140acd63d2e4fbe25d7 pg_semver.tar.gz" | sha256sum --check && \
+    mkdir pg_semver-src && cd pg_semver-src && tar xvzf ../pg_semver.tar.gz --strip-components=1 -C . && \
+    make -j $(getconf _NPROCESSORS_ONLN) && \
+    make -j $(getconf _NPROCESSORS_ONLN) install && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/semver.control
+
 #########################################################################################
 #
 # Layer "pg-embedding-pg-build"
@@ -768,6 +807,7 @@ COPY --from=pg-pgx-ulid-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=rdkit-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-uuidv7-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-roaringbitmap-pg-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pg-semver-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-embedding-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=wal2json-pg-build /usr/local/pgsql /usr/local/pgsql
 COPY pgxn/ pgxn/

Dockerfile.compute-tools

@@ -1,7 +1,7 @@
 # First transient image to build compute_tools binaries
 # NB: keep in sync with rust image version in .github/workflows/build_and_test.yml
 ARG REPOSITORY=neondatabase
-ARG IMAGE=rust
+ARG IMAGE=build-tools
 ARG TAG=pinned
 ARG BUILD_TAG
 

README.md

@@ -29,13 +29,14 @@ 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 lsof libicu-dev
+libcurl4-openssl-dev openssl python3-poetry lsof libicu-dev
 ```
 * 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 lsof libicu-devel
+  protobuf-devel libcurl-devel openssl poetry lsof libicu-devel libpq-devel python3-devel \
+  libffi-devel
 ```
 * On Arch based systems, these packages are needed:
 ```bash

compute_tools/Cargo.toml

@@ -37,5 +37,6 @@ 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"
+zstd = "0.13"
 bytes = "1.0"
+rust-ini = "0.20.0"

compute_tools/src/bin/compute_ctl.rs

@@ -31,7 +31,9 @@
 //!             -C 'postgresql://cloud_admin@localhost/postgres' \
 //!             -S /var/db/postgres/specs/current.json \
 //!             -b /usr/local/bin/postgres \
-//!             -r http://pg-ext-s3-gateway
+//!             -r http://pg-ext-s3-gateway \
+//!             --pgbouncer-connstr 'host=localhost port=6432 dbname=pgbouncer user=cloud_admin sslmode=disable'
+//!             --pgbouncer-ini-path /etc/pgbouncer.ini \
 //! ```
 //!
 use std::collections::HashMap;
@@ -99,6 +101,9 @@ fn main() -> Result<()> {
     let spec_json = matches.get_one::<String>("spec");
     let spec_path = matches.get_one::<String>("spec-path");
 
+    let pgbouncer_connstr = matches.get_one::<String>("pgbouncer-connstr");
+    let pgbouncer_ini_path = matches.get_one::<String>("pgbouncer-ini-path");
+
     // Extract OpenTelemetry context for the startup actions from the
     // TRACEPARENT and TRACESTATE env variables, and attach it to the current
     // tracing context.
@@ -209,6 +214,8 @@ fn main() -> Result<()> {
         ext_remote_storage: ext_remote_storage.map(|s| s.to_string()),
         ext_download_progress: RwLock::new(HashMap::new()),
         build_tag,
+        pgbouncer_connstr: pgbouncer_connstr.map(|s| s.to_string()),
+        pgbouncer_ini_path: pgbouncer_ini_path.map(|s| s.to_string()),
     };
     let compute = Arc::new(compute_node);
 
@@ -493,6 +500,23 @@ fn cli() -> clap::Command {
                 )
                 .value_name("FILECACHE_CONNSTR"),
         )
+        .arg(
+            Arg::new("pgbouncer-connstr")
+                .long("pgbouncer-connstr")
+                .default_value(
+                    "host=localhost port=6432 dbname=pgbouncer user=cloud_admin sslmode=disable",
+                )
+                .value_name("PGBOUNCER_CONNSTR"),
+        )
+        .arg(
+            Arg::new("pgbouncer-ini-path")
+                .long("pgbouncer-ini-path")
+                // Note: this doesn't match current path for pgbouncer.ini.
+                // Until we fix it, we need to pass the path explicitly
+                // or this will be effectively no-op.
+                .default_value("/etc/pgbouncer.ini")
+                .value_name("PGBOUNCER_INI_PATH"),
+        )
 }
 
 #[test]

compute_tools/src/compute.rs

@@ -7,6 +7,7 @@ use std::path::Path;
 use std::process::{Command, Stdio};
 use std::str::FromStr;
 use std::sync::{Condvar, Mutex, RwLock};
+use std::thread;
 use std::time::Instant;
 
 use anyhow::{Context, Result};
@@ -64,6 +65,10 @@ pub struct ComputeNode {
     // key: ext_archive_name, value: started download time, download_completed?
     pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
     pub build_tag: String,
+    // connection string to pgbouncer to change settings
+    pub pgbouncer_connstr: Option<String>,
+    // path to pgbouncer.ini to change settings
+    pub pgbouncer_ini_path: Option<String>,
 }
 
 // store some metrics about download size that might impact startup time
@@ -737,6 +742,31 @@ impl ComputeNode {
     pub fn reconfigure(&self) -> Result<()> {
         let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
 
+        if let Some(connstr) = &self.pgbouncer_connstr {
+            info!("tuning pgbouncer with connstr: {:?}", connstr);
+
+            let rt = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .expect("failed to create rt");
+
+            // Spawn a thread to do the tuning,
+            // so that we don't block the main thread that starts Postgres.
+            let pgbouncer_settings = spec.pgbouncer_settings.clone();
+            let connstr_clone = connstr.clone();
+            let pgbouncer_ini_path = self.pgbouncer_ini_path.clone();
+            let _handle = thread::spawn(move || {
+                let res = rt.block_on(tune_pgbouncer(
+                    pgbouncer_settings,
+                    &connstr_clone,
+                    pgbouncer_ini_path,
+                ));
+                if let Err(err) = res {
+                    error!("error while tuning pgbouncer: {err:?}");
+                }
+            });
+        }
+
         // Write new config
         let pgdata_path = Path::new(&self.pgdata);
         let postgresql_conf_path = pgdata_path.join("postgresql.conf");
@@ -791,6 +821,32 @@ impl ComputeNode {
             pspec.timeline_id,
         );
 
+        // tune pgbouncer
+        if let Some(connstr) = &self.pgbouncer_connstr {
+            info!("tuning pgbouncer with connstr: {:?}", connstr);
+
+            let rt = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .expect("failed to create rt");
+
+            // Spawn a thread to do the tuning,
+            // so that we don't block the main thread that starts Postgres.
+            let pgbouncer_settings = pspec.spec.pgbouncer_settings.clone();
+            let connstr_clone = connstr.clone();
+            let pgbouncer_ini_path = self.pgbouncer_ini_path.clone();
+            let _handle = thread::spawn(move || {
+                let res = rt.block_on(tune_pgbouncer(
+                    pgbouncer_settings,
+                    &connstr_clone,
+                    pgbouncer_ini_path,
+                ));
+                if let Err(err) = res {
+                    error!("error while tuning pgbouncer: {err:?}");
+                }
+            });
+        }
+
         info!(
             "start_compute spec.remote_extensions {:?}",
             pspec.spec.remote_extensions

compute_tools/src/pg_helpers.rs

@@ -9,9 +9,11 @@ use std::process::Child;
 use std::time::{Duration, Instant};
 
 use anyhow::{bail, Result};
+use ini::Ini;
 use notify::{RecursiveMode, Watcher};
 use postgres::{Client, Transaction};
-use tracing::{debug, instrument};
+use tokio_postgres::NoTls;
+use tracing::{debug, error, info, instrument};
 
 use compute_api::spec::{Database, GenericOption, GenericOptions, PgIdent, Role};
 
@@ -359,3 +361,68 @@ pub fn create_pgdata(pgdata: &str) -> Result<()> {
 
     Ok(())
 }
+
+/// Update pgbouncer.ini with provided options
+pub fn update_pgbouncer_ini(
+    pgbouncer_config: HashMap<String, String>,
+    pgbouncer_ini_path: &str,
+) -> Result<()> {
+    let mut conf = Ini::load_from_file(pgbouncer_ini_path)?;
+    let section = conf.section_mut(Some("pgbouncer")).unwrap();
+
+    for (option_name, value) in pgbouncer_config.iter() {
+        section.insert(option_name, value);
+    }
+
+    conf.write_to_file(pgbouncer_ini_path)?;
+    Ok(())
+}
+
+/// Tune pgbouncer.
+/// 1. Apply new config using pgbouncer admin console
+/// 2. Add new values to pgbouncer.ini to preserve them after restart
+pub async fn tune_pgbouncer(
+    pgbouncer_settings: Option<HashMap<String, String>>,
+    pgbouncer_connstr: &str,
+    pgbouncer_ini_path: Option<String>,
+) -> Result<()> {
+    if let Some(pgbouncer_config) = pgbouncer_settings {
+        // Apply new config
+        let connect_result = tokio_postgres::connect(pgbouncer_connstr, NoTls).await;
+        let (client, connection) = connect_result.unwrap();
+        tokio::spawn(async move {
+            if let Err(e) = connection.await {
+                eprintln!("connection error: {}", e);
+            }
+        });
+
+        for (option_name, value) in pgbouncer_config.iter() {
+            info!(
+                "Applying pgbouncer setting change: {} = {}",
+                option_name, value
+            );
+            let query = format!("SET {} = {}", option_name, value);
+
+            let result = client.simple_query(&query).await;
+
+            info!("Applying pgbouncer setting change: {}", query);
+            info!("pgbouncer setting change result: {:?}", result);
+
+            if let Err(err) = result {
+                // Don't fail on error, just print it into log
+                error!(
+                    "Failed to apply pgbouncer setting change: {},  {}",
+                    query, err
+                );
+            };
+        }
+
+        // save values to pgbouncer.ini
+        // so that they are preserved after pgbouncer restart
+        if let Some(pgbouncer_ini_path) = pgbouncer_ini_path {
+            update_pgbouncer_ini(pgbouncer_config, &pgbouncer_ini_path)?;
+        }
+    }
+
+    Ok(())
+}

compute_tools/src/spec.rs

@@ -298,7 +298,7 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                 // safe to add more permissions here. BYPASSRLS and REPLICATION are inherited
                 // from neon_superuser.
                 let mut query: String = format!(
-                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB IN ROLE neon_superuser",
+                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
                     name.pg_quote()
                 );
                 info!("role create query: '{}'", &query);
@@ -370,33 +370,49 @@ pub fn handle_role_deletions(spec: &ComputeSpec, connstr: &str, client: &mut Cli
     Ok(())
 }
 
+fn reassign_owned_objects_in_one_db(
+    conf: Config,
+    role_name: &PgIdent,
+    db_owner: &PgIdent,
+) -> Result<()> {
+    let mut client = conf.connect(NoTls)?;
+
+    // This will reassign all dependent objects to the db owner
+    let reassign_query = format!(
+        "REASSIGN OWNED BY {} TO {}",
+        role_name.pg_quote(),
+        db_owner.pg_quote()
+    );
+    info!(
+        "reassigning objects owned by '{}' in db '{}' to '{}'",
+        role_name,
+        conf.get_dbname().unwrap_or(""),
+        db_owner
+    );
+    client.simple_query(&reassign_query)?;
+
+    // This now will only drop privileges of the role
+    let drop_query = format!("DROP OWNED BY {}", role_name.pg_quote());
+    client.simple_query(&drop_query)?;
+    Ok(())
+}
+
 // Reassign all owned objects in all databases to the owner of the database.
 fn reassign_owned_objects(spec: &ComputeSpec, connstr: &str, role_name: &PgIdent) -> Result<()> {
     for db in &spec.cluster.databases {
         if db.owner != *role_name {
             let mut conf = Config::from_str(connstr)?;
             conf.dbname(&db.name);
-
-            let mut client = conf.connect(NoTls)?;
-
-            // This will reassign all dependent objects to the db owner
-            let reassign_query = format!(
-                "REASSIGN OWNED BY {} TO {}",
-                role_name.pg_quote(),
-                db.owner.pg_quote()
-            );
-            info!(
-                "reassigning objects owned by '{}' in db '{}' to '{}'",
-                role_name, &db.name, &db.owner
-            );
-            client.simple_query(&reassign_query)?;
-
-            // This now will only drop privileges of the role
-            let drop_query = format!("DROP OWNED BY {}", role_name.pg_quote());
-            client.simple_query(&drop_query)?;
+            reassign_owned_objects_in_one_db(conf, role_name, &db.owner)?;
         }
     }
 
+    // Also handle case when there are no databases in the spec.
+    // In this case we need to reassign objects in the default database.
+    let conf = Config::from_str(connstr)?;
+    let db_owner = PgIdent::from_str("cloud_admin")?;
+    reassign_owned_objects_in_one_db(conf, role_name, &db_owner)?;
+
     Ok(())
 }
 

control_plane/Cargo.toml

@@ -6,9 +6,11 @@ license.workspace = true
 
 [dependencies]
 anyhow.workspace = true
+async-trait.workspace = true
 camino.workspace = true
 clap.workspace = true
 comfy-table.workspace = true
+futures.workspace = true
 git-version.workspace = true
 nix.workspace = true
 once_cell.workspace = true
@@ -24,10 +26,11 @@ tar.workspace = true
 thiserror.workspace = true
 toml.workspace = true
 tokio.workspace = true
+tokio-postgres.workspace = true
+tokio-util.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.
 pageserver_api.workspace = true
+pageserver_client.workspace = true
 postgres_backend.workspace = true
 safekeeper_api.workspace = true
 postgres_connection.workspace = true

control_plane/src/attachment_service.rs

@@ -9,7 +9,7 @@ pub struct AttachmentService {
     env: LocalEnv,
     listen: String,
     path: PathBuf,
-    client: reqwest::blocking::Client,
+    client: reqwest::Client,
 }
 
 const COMMAND: &str = "attachment_service";
@@ -53,7 +53,7 @@ impl AttachmentService {
             env: env.clone(),
             path,
             listen,
-            client: reqwest::blocking::ClientBuilder::new()
+            client: reqwest::ClientBuilder::new()
                 .build()
                 .expect("Failed to construct http client"),
         }
@@ -64,7 +64,7 @@ impl AttachmentService {
             .expect("non-Unicode path")
     }
 
-    pub fn start(&self) -> anyhow::Result<Child> {
+    pub async fn start(&self) -> anyhow::Result<Child> {
         let path_str = self.path.to_string_lossy();
 
         background_process::start_process(
@@ -73,10 +73,11 @@ impl AttachmentService {
             &self.env.attachment_service_bin(),
             ["-l", &self.listen, "-p", &path_str],
             [],
-            background_process::InitialPidFile::Create(&self.pid_file()),
+            background_process::InitialPidFile::Create(self.pid_file()),
             // TODO: a real status check
-            || Ok(true),
+            || async move { anyhow::Ok(true) },
         )
+        .await
     }
 
     pub fn stop(&self, immediate: bool) -> anyhow::Result<()> {
@@ -84,7 +85,7 @@ impl AttachmentService {
     }
 
     /// Call into the attach_hook API, for use before handing out attachments to pageservers
-    pub fn attach_hook(
+    pub async fn attach_hook(
         &self,
         tenant_id: TenantId,
         pageserver_id: NodeId,
@@ -104,16 +105,16 @@ impl AttachmentService {
             node_id: Some(pageserver_id),
         };
 
-        let response = self.client.post(url).json(&request).send()?;
+        let response = self.client.post(url).json(&request).send().await?;
         if response.status() != StatusCode::OK {
             return Err(anyhow!("Unexpected status {}", response.status()));
         }
 
-        let response = response.json::<AttachHookResponse>()?;
+        let response = response.json::<AttachHookResponse>().await?;
         Ok(response.gen)
     }
 
-    pub fn inspect(&self, tenant_id: TenantId) -> anyhow::Result<Option<(u32, NodeId)>> {
+    pub async fn inspect(&self, tenant_id: TenantId) -> anyhow::Result<Option<(u32, NodeId)>> {
         use hyper::StatusCode;
 
         let url = self
@@ -126,12 +127,12 @@ impl AttachmentService {
 
         let request = InspectRequest { tenant_id };
 
-        let response = self.client.post(url).json(&request).send()?;
+        let response = self.client.post(url).json(&request).send().await?;
         if response.status() != StatusCode::OK {
             return Err(anyhow!("Unexpected status {}", response.status()));
         }
 
-        let response = response.json::<InspectResponse>()?;
+        let response = response.json::<InspectResponse>().await?;
         Ok(response.attachment)
     }
 }

control_plane/src/background_process.rs

@@ -44,15 +44,15 @@ const NOTICE_AFTER_RETRIES: u64 = 50;
 
 /// Argument to `start_process`, to indicate whether it should create pidfile or if the process creates
 /// it itself.
-pub enum InitialPidFile<'t> {
+pub enum InitialPidFile {
     /// Create a pidfile, to allow future CLI invocations to manipulate the process.
-    Create(&'t Utf8Path),
+    Create(Utf8PathBuf),
     /// The process will create the pidfile itself, need to wait for that event.
-    Expect(&'t Utf8Path),
+    Expect(Utf8PathBuf),
 }
 
 /// Start a background child process using the parameters given.
-pub fn start_process<F, AI, A, EI>(
+pub async fn start_process<F, Fut, AI, A, EI>(
     process_name: &str,
     datadir: &Path,
     command: &Path,
@@ -62,7 +62,8 @@ pub fn start_process<F, AI, A, EI>(
     process_status_check: F,
 ) -> anyhow::Result<Child>
 where
-    F: Fn() -> anyhow::Result<bool>,
+    F: Fn() -> Fut,
+    Fut: std::future::Future<Output = anyhow::Result<bool>>,
     AI: IntoIterator<Item = A>,
     A: AsRef<OsStr>,
     // Not generic AsRef<OsStr>, otherwise empty `envs` prevents type inference
@@ -89,7 +90,7 @@ where
     let filled_cmd = fill_remote_storage_secrets_vars(fill_rust_env_vars(background_command));
     filled_cmd.envs(envs);
 
-    let pid_file_to_check = match initial_pid_file {
+    let pid_file_to_check = match &initial_pid_file {
         InitialPidFile::Create(path) => {
             pre_exec_create_pidfile(filled_cmd, path);
             path
@@ -107,7 +108,7 @@ where
     );
 
     for retries in 0..RETRIES {
-        match process_started(pid, Some(pid_file_to_check), &process_status_check) {
+        match process_started(pid, pid_file_to_check, &process_status_check).await {
             Ok(true) => {
                 println!("\n{process_name} started, pid: {pid}");
                 return Ok(spawned_process);
@@ -316,22 +317,20 @@ where
     cmd
 }
 
-fn process_started<F>(
+async fn process_started<F, Fut>(
     pid: Pid,
-    pid_file_to_check: Option<&Utf8Path>,
+    pid_file_to_check: &Utf8Path,
     status_check: &F,
 ) -> anyhow::Result<bool>
 where
-    F: Fn() -> anyhow::Result<bool>,
+    F: Fn() -> Fut,
+    Fut: std::future::Future<Output = anyhow::Result<bool>>,
 {
-    match status_check() {
-        Ok(true) => match pid_file_to_check {
-            Some(pid_file_path) => match pid_file::read(pid_file_path)? {
-                PidFileRead::NotExist => Ok(false),
-                PidFileRead::LockedByOtherProcess(pid_in_file) => Ok(pid_in_file == pid),
-                PidFileRead::NotHeldByAnyProcess(_) => Ok(false),
-            },
-            None => Ok(true),
+    match status_check().await {
+        Ok(true) => match pid_file::read(pid_file_to_check)? {
+            PidFileRead::NotExist => Ok(false),
+            PidFileRead::LockedByOtherProcess(pid_in_file) => Ok(pid_in_file == pid),
+            PidFileRead::NotHeldByAnyProcess(_) => Ok(false),
         },
         Ok(false) => Ok(false),
         Err(e) => anyhow::bail!("process failed to start: {e}"),

control_plane/src/bin/neon_local.rs

@@ -120,15 +120,20 @@ fn main() -> Result<()> {
         let mut env = LocalEnv::load_config().context("Error loading config")?;
         let original_env = env.clone();
 
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .unwrap();
+
         let subcommand_result = match sub_name {
-            "tenant" => handle_tenant(sub_args, &mut env),
-            "timeline" => handle_timeline(sub_args, &mut env),
-            "start" => handle_start_all(sub_args, &env),
+            "tenant" => rt.block_on(handle_tenant(sub_args, &mut env)),
+            "timeline" => rt.block_on(handle_timeline(sub_args, &mut env)),
+            "start" => rt.block_on(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),
+            "pageserver" => rt.block_on(handle_pageserver(sub_args, &env)),
+            "attachment_service" => rt.block_on(handle_attachment_service(sub_args, &env)),
+            "safekeeper" => rt.block_on(handle_safekeeper(sub_args, &env)),
+            "endpoint" => rt.block_on(handle_endpoint(sub_args, &env)),
             "mappings" => handle_mappings(sub_args, &mut env),
             "pg" => bail!("'pg' subcommand has been renamed to 'endpoint'"),
             _ => bail!("unexpected subcommand {sub_name}"),
@@ -269,12 +274,13 @@ fn print_timeline(
 
 /// Returns a map of timeline IDs to timeline_id@lsn strings.
 /// Connects to the pageserver to query this information.
-fn get_timeline_infos(
+async fn get_timeline_infos(
     env: &local_env::LocalEnv,
     tenant_id: &TenantId,
 ) -> Result<HashMap<TimelineId, TimelineInfo>> {
     Ok(get_default_pageserver(env)
-        .timeline_list(tenant_id)?
+        .timeline_list(tenant_id)
+        .await?
         .into_iter()
         .map(|timeline_info| (timeline_info.timeline_id, timeline_info))
         .collect())
@@ -373,11 +379,14 @@ fn pageserver_config_overrides(init_match: &ArgMatches) -> Vec<&str> {
         .collect()
 }
 
-fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> anyhow::Result<()> {
+async fn handle_tenant(
+    tenant_match: &ArgMatches,
+    env: &mut local_env::LocalEnv,
+) -> anyhow::Result<()> {
     let pageserver = get_default_pageserver(env);
     match tenant_match.subcommand() {
         Some(("list", _)) => {
-            for t in pageserver.tenant_list()? {
+            for t in pageserver.tenant_list().await? {
                 println!("{} {:?}", t.id, t.state);
             }
         }
@@ -394,12 +403,16 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
                 // 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)?
+                attachment_service
+                    .attach_hook(tenant_id, pageserver.conf.id)
+                    .await?
             } else {
                 None
             };
 
-            pageserver.tenant_create(tenant_id, generation, tenant_conf)?;
+            pageserver
+                .tenant_create(tenant_id, generation, tenant_conf)
+                .await?;
             println!("tenant {tenant_id} successfully created on the pageserver");
 
             // Create an initial timeline for the new tenant
@@ -409,14 +422,16 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
                 .copied()
                 .context("Failed to parse postgres version from the argument string")?;
 
-            let timeline_info = pageserver.timeline_create(
-                tenant_id,
-                new_timeline_id,
-                None,
-                None,
-                Some(pg_version),
-                None,
-            )?;
+            let timeline_info = pageserver
+                .timeline_create(
+                    tenant_id,
+                    new_timeline_id,
+                    None,
+                    None,
+                    Some(pg_version),
+                    None,
+                )
+                .await?;
             let new_timeline_id = timeline_info.timeline_id;
             let last_record_lsn = timeline_info.last_record_lsn;
 
@@ -450,6 +465,7 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
 
             pageserver
                 .tenant_config(tenant_id, tenant_conf)
+                .await
                 .with_context(|| format!("Tenant config failed for tenant with id {tenant_id}"))?;
             println!("tenant {tenant_id} successfully configured on the pageserver");
         }
@@ -458,7 +474,7 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
             let new_pageserver = get_pageserver(env, matches)?;
             let new_pageserver_id = new_pageserver.conf.id;
 
-            migrate_tenant(env, tenant_id, new_pageserver)?;
+            migrate_tenant(env, tenant_id, new_pageserver).await?;
             println!("tenant {tenant_id} migrated to {}", new_pageserver_id);
         }
 
@@ -468,13 +484,13 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
     Ok(())
 }
 
-fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -> Result<()> {
+async fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -> Result<()> {
     let pageserver = get_default_pageserver(env);
 
     match timeline_match.subcommand() {
         Some(("list", list_match)) => {
             let tenant_id = get_tenant_id(list_match, env)?;
-            let timelines = pageserver.timeline_list(&tenant_id)?;
+            let timelines = pageserver.timeline_list(&tenant_id).await?;
             print_timelines_tree(timelines, env.timeline_name_mappings())?;
         }
         Some(("create", create_match)) => {
@@ -490,14 +506,16 @@ fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -
 
             let new_timeline_id_opt = parse_timeline_id(create_match)?;
 
-            let timeline_info = pageserver.timeline_create(
-                tenant_id,
-                new_timeline_id_opt,
-                None,
-                None,
-                Some(pg_version),
-                None,
-            )?;
+            let timeline_info = pageserver
+                .timeline_create(
+                    tenant_id,
+                    new_timeline_id_opt,
+                    None,
+                    None,
+                    Some(pg_version),
+                    None,
+                )
+                .await?;
             let new_timeline_id = timeline_info.timeline_id;
 
             let last_record_lsn = timeline_info.last_record_lsn;
@@ -542,7 +560,9 @@ fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -
 
             let mut cplane = ComputeControlPlane::load(env.clone())?;
             println!("Importing timeline into pageserver ...");
-            pageserver.timeline_import(tenant_id, timeline_id, base, pg_wal, pg_version)?;
+            pageserver
+                .timeline_import(tenant_id, timeline_id, base, pg_wal, pg_version)
+                .await?;
             env.register_branch_mapping(name.to_string(), tenant_id, timeline_id)?;
 
             println!("Creating endpoint for imported timeline ...");
@@ -578,14 +598,16 @@ fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -
                 .map(|lsn_str| Lsn::from_str(lsn_str))
                 .transpose()
                 .context("Failed to parse ancestor start Lsn from the request")?;
-            let timeline_info = pageserver.timeline_create(
-                tenant_id,
-                None,
-                start_lsn,
-                Some(ancestor_timeline_id),
-                None,
-                None,
-            )?;
+            let timeline_info = pageserver
+                .timeline_create(
+                    tenant_id,
+                    None,
+                    start_lsn,
+                    Some(ancestor_timeline_id),
+                    None,
+                    None,
+                )
+                .await?;
             let new_timeline_id = timeline_info.timeline_id;
 
             let last_record_lsn = timeline_info.last_record_lsn;
@@ -604,7 +626,7 @@ fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -
     Ok(())
 }
 
-fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
+async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
     let (sub_name, sub_args) = match ep_match.subcommand() {
         Some(ep_subcommand_data) => ep_subcommand_data,
         None => bail!("no endpoint subcommand provided"),
@@ -614,10 +636,12 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
     match sub_name {
         "list" => {
             let tenant_id = get_tenant_id(sub_args, env)?;
-            let timeline_infos = get_timeline_infos(env, &tenant_id).unwrap_or_else(|e| {
-                eprintln!("Failed to load timeline info: {}", e);
-                HashMap::new()
-            });
+            let timeline_infos = get_timeline_infos(env, &tenant_id)
+                .await
+                .unwrap_or_else(|e| {
+                    eprintln!("Failed to load timeline info: {}", e);
+                    HashMap::new()
+                });
 
             let timeline_name_mappings = env.timeline_name_mappings();
 
@@ -791,7 +815,9 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
             };
 
             println!("Starting existing endpoint {endpoint_id}...");
-            endpoint.start(&auth_token, safekeepers, remote_ext_config)?;
+            endpoint
+                .start(&auth_token, safekeepers, remote_ext_config)
+                .await?;
         }
         "reconfigure" => {
             let endpoint_id = sub_args
@@ -809,7 +835,7 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                 } else {
                     None
                 };
-            endpoint.reconfigure(pageserver_id)?;
+            endpoint.reconfigure(pageserver_id).await?;
         }
         "stop" => {
             let endpoint_id = sub_args
@@ -875,11 +901,12 @@ fn get_pageserver(env: &local_env::LocalEnv, args: &ArgMatches) -> Result<PageSe
     ))
 }
 
-fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
+async fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
     match sub_match.subcommand() {
         Some(("start", subcommand_args)) => {
             if let Err(e) = get_pageserver(env, subcommand_args)?
                 .start(&pageserver_config_overrides(subcommand_args))
+                .await
             {
                 eprintln!("pageserver start failed: {e}");
                 exit(1);
@@ -906,7 +933,10 @@ fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
                 exit(1);
             }
 
-            if let Err(e) = pageserver.start(&pageserver_config_overrides(subcommand_args)) {
+            if let Err(e) = pageserver
+                .start(&pageserver_config_overrides(subcommand_args))
+                .await
+            {
                 eprintln!("pageserver start failed: {e}");
                 exit(1);
             }
@@ -920,14 +950,17 @@ fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
                 exit(1);
             }
 
-            if let Err(e) = pageserver.start(&pageserver_config_overrides(subcommand_args)) {
+            if let Err(e) = pageserver
+                .start(&pageserver_config_overrides(subcommand_args))
+                .await
+            {
                 eprintln!("pageserver start failed: {e}");
                 exit(1);
             }
         }
 
         Some(("status", subcommand_args)) => {
-            match get_pageserver(env, subcommand_args)?.check_status() {
+            match get_pageserver(env, subcommand_args)?.check_status().await {
                 Ok(_) => println!("Page server is up and running"),
                 Err(err) => {
                     eprintln!("Page server is not available: {}", err);
@@ -942,11 +975,14 @@ fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
     Ok(())
 }
 
-fn handle_attachment_service(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
+async 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() {
+            if let Err(e) = svc.start().await {
                 eprintln!("start failed: {e}");
                 exit(1);
             }
@@ -987,7 +1023,7 @@ fn safekeeper_extra_opts(init_match: &ArgMatches) -> Vec<String> {
         .collect()
 }
 
-fn handle_safekeeper(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
+async 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,
         None => bail!("no safekeeper subcommand provided"),
@@ -1005,7 +1041,7 @@ fn handle_safekeeper(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
         "start" => {
             let extra_opts = safekeeper_extra_opts(sub_args);
 
-            if let Err(e) = safekeeper.start(extra_opts) {
+            if let Err(e) = safekeeper.start(extra_opts).await {
                 eprintln!("safekeeper start failed: {}", e);
                 exit(1);
             }
@@ -1031,7 +1067,7 @@ fn handle_safekeeper(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
             }
 
             let extra_opts = safekeeper_extra_opts(sub_args);
-            if let Err(e) = safekeeper.start(extra_opts) {
+            if let Err(e) = safekeeper.start(extra_opts).await {
                 eprintln!("safekeeper start failed: {}", e);
                 exit(1);
             }
@@ -1044,15 +1080,15 @@ fn handle_safekeeper(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
     Ok(())
 }
 
-fn handle_start_all(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> anyhow::Result<()> {
+async fn handle_start_all(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> anyhow::Result<()> {
     // Endpoints are not started automatically
 
-    broker::start_broker_process(env)?;
+    broker::start_broker_process(env).await?;
 
     // 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() {
+        if let Err(e) = attachment_service.start().await {
             eprintln!("attachment_service start failed: {:#}", e);
             try_stop_all(env, true);
             exit(1);
@@ -1061,7 +1097,10 @@ fn handle_start_all(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> anyhow
 
     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)) {
+        if let Err(e) = pageserver
+            .start(&pageserver_config_overrides(sub_match))
+            .await
+        {
             eprintln!("pageserver {} start failed: {:#}", ps_conf.id, e);
             try_stop_all(env, true);
             exit(1);
@@ -1070,7 +1109,7 @@ fn handle_start_all(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> anyhow
 
     for node in env.safekeepers.iter() {
         let safekeeper = SafekeeperNode::from_env(env, node);
-        if let Err(e) = safekeeper.start(vec![]) {
+        if let Err(e) = safekeeper.start(vec![]).await {
             eprintln!("safekeeper {} start failed: {:#}", safekeeper.id, e);
             try_stop_all(env, false);
             exit(1);

control_plane/src/broker.rs

@@ -11,7 +11,7 @@ use camino::Utf8PathBuf;
 
 use crate::{background_process, local_env};
 
-pub fn start_broker_process(env: &local_env::LocalEnv) -> anyhow::Result<()> {
+pub async fn start_broker_process(env: &local_env::LocalEnv) -> anyhow::Result<()> {
     let broker = &env.broker;
     let listen_addr = &broker.listen_addr;
 
@@ -19,15 +19,15 @@ pub fn start_broker_process(env: &local_env::LocalEnv) -> anyhow::Result<()> {
 
     let args = [format!("--listen-addr={listen_addr}")];
 
-    let client = reqwest::blocking::Client::new();
+    let client = reqwest::Client::new();
     background_process::start_process(
         "storage_broker",
         &env.base_data_dir,
         &env.storage_broker_bin(),
         args,
         [],
-        background_process::InitialPidFile::Create(&storage_broker_pid_file_path(env)),
-        || {
+        background_process::InitialPidFile::Create(storage_broker_pid_file_path(env)),
+        || async {
             let url = broker.client_url();
             let status_url = url.join("status").with_context(|| {
                 format!("Failed to append /status path to broker endpoint {url}")
@@ -36,12 +36,13 @@ pub fn start_broker_process(env: &local_env::LocalEnv) -> anyhow::Result<()> {
                 .get(status_url)
                 .build()
                 .with_context(|| format!("Failed to construct request to broker endpoint {url}"))?;
-            match client.execute(request) {
+            match client.execute(request).await {
                 Ok(resp) => Ok(resp.status().is_success()),
                 Err(_) => Ok(false),
             }
         },
     )
+    .await
     .context("Failed to spawn storage_broker subprocess")?;
     Ok(())
 }

control_plane/src/endpoint.rs

@@ -464,7 +464,7 @@ impl Endpoint {
         }
     }
 
-    pub fn start(
+    pub async fn start(
         &self,
         auth_token: &Option<String>,
         safekeepers: Vec<NodeId>,
@@ -537,6 +537,7 @@ impl Endpoint {
             safekeeper_connstrings,
             storage_auth_token: auth_token.clone(),
             remote_extensions,
+            pgbouncer_settings: None,
         };
         let spec_path = self.endpoint_path().join("spec.json");
         std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;
@@ -587,7 +588,7 @@ impl Endpoint {
         const MAX_ATTEMPTS: u32 = 10 * 30; // Wait up to 30 s
         loop {
             attempt += 1;
-            match self.get_status() {
+            match self.get_status().await {
                 Ok(state) => {
                     match state.status {
                         ComputeStatus::Init => {
@@ -629,8 +630,8 @@ impl Endpoint {
     }
 
     // Call the /status HTTP API
-    pub fn get_status(&self) -> Result<ComputeState> {
-        let client = reqwest::blocking::Client::new();
+    pub async fn get_status(&self) -> Result<ComputeState> {
+        let client = reqwest::Client::new();
 
         let response = client
             .request(
@@ -641,16 +642,17 @@ impl Endpoint {
                     self.http_address.port()
                 ),
             )
-            .send()?;
+            .send()
+            .await?;
 
         // Interpret the response
         let status = response.status();
         if !(status.is_client_error() || status.is_server_error()) {
-            Ok(response.json()?)
+            Ok(response.json().await?)
         } else {
             // reqwest does not export its error construction utility functions, so let's craft the message ourselves
             let url = response.url().to_owned();
-            let msg = match response.text() {
+            let msg = match response.text().await {
                 Ok(err_body) => format!("Error: {}", err_body),
                 Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
             };
@@ -658,7 +660,7 @@ impl Endpoint {
         }
     }
 
-    pub fn reconfigure(&self, pageserver_id: Option<NodeId>) -> Result<()> {
+    pub async fn reconfigure(&self, pageserver_id: Option<NodeId>) -> Result<()> {
         let mut spec: ComputeSpec = {
             let spec_path = self.endpoint_path().join("spec.json");
             let file = std::fs::File::open(spec_path)?;
@@ -687,7 +689,7 @@ impl Endpoint {
             spec.pageserver_connstring = Some(format!("postgresql://no_user@{host}:{port}"));
         }
 
-        let client = reqwest::blocking::Client::new();
+        let client = reqwest::Client::new();
         let response = client
             .post(format!(
                 "http://{}:{}/configure",
@@ -698,14 +700,15 @@ impl Endpoint {
                 "{{\"spec\":{}}}",
                 serde_json::to_string_pretty(&spec)?
             ))
-            .send()?;
+            .send()
+            .await?;
 
         let status = response.status();
         if !(status.is_client_error() || status.is_server_error()) {
             Ok(())
         } else {
             let url = response.url().to_owned();
-            let msg = match response.text() {
+            let msg = match response.text().await {
                 Ok(err_body) => format!("Error: {}", err_body),
                 Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
             };

control_plane/src/pageserver.rs

@@ -6,28 +6,24 @@
 //!
 use std::borrow::Cow;
 use std::collections::HashMap;
-use std::fs::File;
-use std::io::{BufReader, Write};
+
+use std::io;
+use std::io::Write;
 use std::num::NonZeroU64;
 use std::path::PathBuf;
 use std::process::{Child, Command};
 use std::time::Duration;
-use std::{io, result};
 
 use anyhow::{bail, Context};
 use camino::Utf8PathBuf;
-use pageserver_api::models::{
-    self, LocationConfig, TenantInfo, TenantLocationConfigRequest, TimelineInfo,
-};
+use futures::SinkExt;
+use pageserver_api::models::{self, LocationConfig, TenantInfo, TimelineInfo};
 use pageserver_api::shard::TenantShardId;
+use pageserver_client::mgmt_api;
 use postgres_backend::AuthType;
 use postgres_connection::{parse_host_port, PgConnectionConfig};
-use reqwest::blocking::{Client, RequestBuilder, Response};
-use reqwest::{IntoUrl, Method};
-use thiserror::Error;
 use utils::auth::{Claims, Scope};
 use utils::{
-    http::error::HttpErrorBody,
     id::{TenantId, TimelineId},
     lsn::Lsn,
 };
@@ -38,45 +34,6 @@ use crate::{background_process, local_env::LocalEnv};
 /// Directory within .neon which will be used by default for LocalFs remote storage.
 pub const PAGESERVER_REMOTE_STORAGE_DIR: &str = "local_fs_remote_storage/pageserver";
 
-#[derive(Error, Debug)]
-pub enum PageserverHttpError {
-    #[error("Reqwest error: {0}")]
-    Transport(#[from] reqwest::Error),
-
-    #[error("Error: {0}")]
-    Response(String),
-}
-
-impl From<anyhow::Error> for PageserverHttpError {
-    fn from(e: anyhow::Error) -> Self {
-        Self::Response(e.to_string())
-    }
-}
-
-type Result<T> = result::Result<T, PageserverHttpError>;
-
-pub trait ResponseErrorMessageExt: Sized {
-    fn error_from_body(self) -> Result<Self>;
-}
-
-impl ResponseErrorMessageExt for Response {
-    fn error_from_body(self) -> Result<Self> {
-        let status = self.status();
-        if !(status.is_client_error() || status.is_server_error()) {
-            return Ok(self);
-        }
-
-        // reqwest does not export its error construction utility functions, so let's craft the message ourselves
-        let url = self.url().to_owned();
-        Err(PageserverHttpError::Response(
-            match self.json::<HttpErrorBody>() {
-                Ok(err_body) => format!("Error: {}", err_body.msg),
-                Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
-            },
-        ))
-    }
-}
-
 //
 // Control routines for pageserver.
 //
@@ -87,8 +44,7 @@ pub struct PageServerNode {
     pub pg_connection_config: PgConnectionConfig,
     pub conf: PageServerConf,
     pub env: LocalEnv,
-    pub http_client: Client,
-    pub http_base_url: String,
+    pub http_client: mgmt_api::Client,
 }
 
 impl PageServerNode {
@@ -100,8 +56,19 @@ impl PageServerNode {
             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", conf.listen_http_addr),
+            http_client: mgmt_api::Client::new(
+                format!("http://{}", conf.listen_http_addr),
+                {
+                    match conf.http_auth_type {
+                        AuthType::Trust => None,
+                        AuthType::NeonJWT => Some(
+                            env.generate_auth_token(&Claims::new(None, Scope::PageServerApi))
+                                .unwrap(),
+                        ),
+                    }
+                }
+                .as_deref(),
+            ),
         }
     }
 
@@ -182,8 +149,8 @@ impl PageServerNode {
             .expect("non-Unicode path")
     }
 
-    pub fn start(&self, config_overrides: &[&str]) -> anyhow::Result<Child> {
-        self.start_node(config_overrides, false)
+    pub async fn start(&self, config_overrides: &[&str]) -> anyhow::Result<Child> {
+        self.start_node(config_overrides, false).await
     }
 
     fn pageserver_init(&self, config_overrides: &[&str]) -> anyhow::Result<()> {
@@ -224,7 +191,12 @@ impl PageServerNode {
         Ok(())
     }
 
-    fn start_node(&self, config_overrides: &[&str], update_config: bool) -> anyhow::Result<Child> {
+    async fn start_node(
+        &self,
+        config_overrides: &[&str],
+        update_config: bool,
+    ) -> anyhow::Result<Child> {
+        // TODO: using a thread here because start_process() is not async but we need to call check_status()
         let datadir = self.repo_path();
         print!(
             "Starting pageserver node {} at '{}' in {:?}",
@@ -232,7 +204,7 @@ impl PageServerNode {
             self.pg_connection_config.raw_address(),
             datadir
         );
-        io::stdout().flush()?;
+        io::stdout().flush().context("flush stdout")?;
 
         let datadir_path_str = datadir.to_str().with_context(|| {
             format!(
@@ -244,20 +216,23 @@ impl PageServerNode {
         if update_config {
             args.push(Cow::Borrowed("--update-config"));
         }
-
         background_process::start_process(
             "pageserver",
             &datadir,
             &self.env.pageserver_bin(),
             args.iter().map(Cow::as_ref),
             self.pageserver_env_variables()?,
-            background_process::InitialPidFile::Expect(&self.pid_file()),
-            || match self.check_status() {
-                Ok(()) => Ok(true),
-                Err(PageserverHttpError::Transport(_)) => Ok(false),
-                Err(e) => Err(anyhow::anyhow!("Failed to check node status: {e}")),
+            background_process::InitialPidFile::Expect(self.pid_file()),
+            || async {
+                let st = self.check_status().await;
+                match st {
+                    Ok(()) => Ok(true),
+                    Err(mgmt_api::Error::ReceiveBody(_)) => Ok(false),
+                    Err(e) => Err(anyhow::anyhow!("Failed to check node status: {e}")),
+                }
             },
         )
+        .await
     }
 
     fn pageserver_basic_args<'a>(
@@ -303,7 +278,12 @@ impl PageServerNode {
         background_process::stop_process(immediate, "pageserver", &self.pid_file())
     }
 
-    pub fn page_server_psql_client(&self) -> anyhow::Result<postgres::Client> {
+    pub async fn page_server_psql_client(
+        &self,
+    ) -> anyhow::Result<(
+        tokio_postgres::Client,
+        tokio_postgres::Connection<tokio_postgres::Socket, tokio_postgres::tls::NoTlsStream>,
+    )> {
         let mut config = self.pg_connection_config.clone();
         if self.conf.pg_auth_type == AuthType::NeonJWT {
             let token = self
@@ -311,36 +291,18 @@ impl PageServerNode {
                 .generate_auth_token(&Claims::new(None, Scope::PageServerApi))?;
             config = config.set_password(Some(token));
         }
-        Ok(config.connect_no_tls()?)
+        Ok(config.connect_no_tls().await?)
     }
 
-    fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> anyhow::Result<RequestBuilder> {
-        let mut builder = self.http_client.request(method, url);
-        if self.conf.http_auth_type == AuthType::NeonJWT {
-            let token = self
-                .env
-                .generate_auth_token(&Claims::new(None, Scope::PageServerApi))?;
-            builder = builder.bearer_auth(token)
-        }
-        Ok(builder)
+    pub async fn check_status(&self) -> mgmt_api::Result<()> {
+        self.http_client.status().await
     }
 
-    pub fn check_status(&self) -> Result<()> {
-        self.http_request(Method::GET, format!("{}/status", self.http_base_url))?
-            .send()?
-            .error_from_body()?;
-        Ok(())
+    pub async fn tenant_list(&self) -> mgmt_api::Result<Vec<TenantInfo>> {
+        self.http_client.list_tenants().await
     }
 
-    pub fn tenant_list(&self) -> Result<Vec<TenantInfo>> {
-        Ok(self
-            .http_request(Method::GET, format!("{}/tenant", self.http_base_url))?
-            .send()?
-            .error_from_body()?
-            .json()?)
-    }
-
-    pub fn tenant_create(
+    pub async fn tenant_create(
         &self,
         new_tenant_id: TenantId,
         generation: Option<u32>,
@@ -418,23 +380,10 @@ impl PageServerNode {
         if !settings.is_empty() {
             bail!("Unrecognized tenant settings: {settings:?}")
         }
-        self.http_request(Method::POST, format!("{}/tenant", self.http_base_url))?
-            .json(&request)
-            .send()?
-            .error_from_body()?
-            .json::<Option<String>>()
-            .with_context(|| {
-                format!("Failed to parse tenant creation response for tenant id: {new_tenant_id:?}")
-            })?
-            .context("No tenant id was found in the tenant creation response")
-            .and_then(|tenant_id_string| {
-                tenant_id_string.parse().with_context(|| {
-                    format!("Failed to parse response string as tenant id: '{tenant_id_string}'")
-                })
-            })
+        Ok(self.http_client.tenant_create(&request).await?)
     }
 
-    pub fn tenant_config(
+    pub async fn tenant_config(
         &self,
         tenant_id: TenantId,
         mut settings: HashMap<&str, &str>,
@@ -513,54 +462,30 @@ impl PageServerNode {
             bail!("Unrecognized tenant settings: {settings:?}")
         }
 
-        self.http_request(Method::PUT, format!("{}/tenant/config", self.http_base_url))?
-            .json(&models::TenantConfigRequest { tenant_id, config })
-            .send()?
-            .error_from_body()?;
+        self.http_client
+            .tenant_config(&models::TenantConfigRequest { tenant_id, config })
+            .await?;
 
         Ok(())
     }
 
-    pub fn location_config(
+    pub async fn location_config(
         &self,
         tenant_id: TenantId,
         config: LocationConfig,
         flush_ms: Option<Duration>,
     ) -> anyhow::Result<()> {
-        let req_body = TenantLocationConfigRequest { tenant_id, config };
-
-        let path = format!(
-            "{}/tenant/{}/location_config",
-            self.http_base_url, tenant_id
-        );
-        let path = if let Some(flush_ms) = flush_ms {
-            format!("{}?flush_ms={}", path, flush_ms.as_millis())
-        } else {
-            path
-        };
-
-        self.http_request(Method::PUT, path)?
-            .json(&req_body)
-            .send()?
-            .error_from_body()?;
-
-        Ok(())
+        Ok(self
+            .http_client
+            .location_config(tenant_id, config, flush_ms)
+            .await?)
     }
 
-    pub fn timeline_list(&self, tenant_id: &TenantId) -> anyhow::Result<Vec<TimelineInfo>> {
-        let timeline_infos: Vec<TimelineInfo> = self
-            .http_request(
-                Method::GET,
-                format!("{}/tenant/{}/timeline", self.http_base_url, tenant_id),
-            )?
-            .send()?
-            .error_from_body()?
-            .json()?;
-
-        Ok(timeline_infos)
+    pub async fn timeline_list(&self, tenant_id: &TenantId) -> anyhow::Result<Vec<TimelineInfo>> {
+        Ok(self.http_client.list_timelines(*tenant_id).await?)
     }
 
-    pub fn timeline_create(
+    pub async fn timeline_create(
         &self,
         tenant_id: TenantId,
         new_timeline_id: Option<TimelineId>,
@@ -571,29 +496,14 @@ impl PageServerNode {
     ) -> anyhow::Result<TimelineInfo> {
         // If timeline ID was not specified, generate one
         let new_timeline_id = new_timeline_id.unwrap_or(TimelineId::generate());
-
-        self.http_request(
-            Method::POST,
-            format!("{}/tenant/{}/timeline", self.http_base_url, tenant_id),
-        )?
-        .json(&models::TimelineCreateRequest {
+        let req = models::TimelineCreateRequest {
             new_timeline_id,
             ancestor_start_lsn,
             ancestor_timeline_id,
             pg_version,
             existing_initdb_timeline_id,
-        })
-        .send()?
-        .error_from_body()?
-        .json::<Option<TimelineInfo>>()
-        .with_context(|| {
-            format!("Failed to parse timeline creation response for tenant id: {tenant_id}")
-        })?
-        .with_context(|| {
-            format!(
-                "No timeline id was found in the timeline creation response for tenant {tenant_id}"
-            )
-        })
+        };
+        Ok(self.http_client.timeline_create(tenant_id, &req).await?)
     }
 
     /// Import a basebackup prepared using either:
@@ -605,7 +515,7 @@ impl PageServerNode {
     /// * `timeline_id` - id to assign to imported timeline
     /// * `base` - (start lsn of basebackup, path to `base.tar` file)
     /// * `pg_wal` - if there's any wal to import: (end lsn, path to `pg_wal.tar`)
-    pub fn timeline_import(
+    pub async fn timeline_import(
         &self,
         tenant_id: TenantId,
         timeline_id: TimelineId,
@@ -613,36 +523,60 @@ impl PageServerNode {
         pg_wal: Option<(Lsn, PathBuf)>,
         pg_version: u32,
     ) -> anyhow::Result<()> {
-        let mut client = self.page_server_psql_client()?;
+        let (client, conn) = self.page_server_psql_client().await?;
+        // The connection object performs the actual communication with the database,
+        // so spawn it off to run on its own.
+        tokio::spawn(async move {
+            if let Err(e) = conn.await {
+                eprintln!("connection error: {}", e);
+            }
+        });
+        tokio::pin!(client);
 
         // Init base reader
         let (start_lsn, base_tarfile_path) = base;
-        let base_tarfile = File::open(base_tarfile_path)?;
-        let mut base_reader = BufReader::new(base_tarfile);
+        let base_tarfile = tokio::fs::File::open(base_tarfile_path).await?;
+        let base_tarfile = tokio_util::io::ReaderStream::new(base_tarfile);
 
         // Init wal reader if necessary
         let (end_lsn, wal_reader) = if let Some((end_lsn, wal_tarfile_path)) = pg_wal {
-            let wal_tarfile = File::open(wal_tarfile_path)?;
-            let wal_reader = BufReader::new(wal_tarfile);
+            let wal_tarfile = tokio::fs::File::open(wal_tarfile_path).await?;
+            let wal_reader = tokio_util::io::ReaderStream::new(wal_tarfile);
             (end_lsn, Some(wal_reader))
         } else {
             (start_lsn, None)
         };
 
-        // Import base
-        let import_cmd = format!(
-            "import basebackup {tenant_id} {timeline_id} {start_lsn} {end_lsn} {pg_version}"
-        );
-        let mut writer = client.copy_in(&import_cmd)?;
-        io::copy(&mut base_reader, &mut writer)?;
-        writer.finish()?;
+        let copy_in = |reader, cmd| {
+            let client = &client;
+            async move {
+                let writer = client.copy_in(&cmd).await?;
+                let writer = std::pin::pin!(writer);
+                let mut writer = writer.sink_map_err(|e| {
+                    std::io::Error::new(std::io::ErrorKind::Other, format!("{e}"))
+                });
+                let mut reader = std::pin::pin!(reader);
+                writer.send_all(&mut reader).await?;
+                writer.into_inner().finish().await?;
+                anyhow::Ok(())
+            }
+        };
 
+        // Import base
+        copy_in(
+            base_tarfile,
+            format!(
+                "import basebackup {tenant_id} {timeline_id} {start_lsn} {end_lsn} {pg_version}"
+            ),
+        )
+        .await?;
         // Import wal if necessary
-        if let Some(mut wal_reader) = wal_reader {
-            let import_cmd = format!("import wal {tenant_id} {timeline_id} {start_lsn} {end_lsn}");
-            let mut writer = client.copy_in(&import_cmd)?;
-            io::copy(&mut wal_reader, &mut writer)?;
-            writer.finish()?;
+        if let Some(wal_reader) = wal_reader {
+            copy_in(
+                wal_reader,
+                format!("import wal {tenant_id} {timeline_id} {start_lsn} {end_lsn}"),
+            )
+            .await?;
         }
 
         Ok(())

control_plane/src/safekeeper.rs

@@ -13,7 +13,6 @@ use std::{io, result};
 use anyhow::Context;
 use camino::Utf8PathBuf;
 use postgres_connection::PgConnectionConfig;
-use reqwest::blocking::{Client, RequestBuilder, Response};
 use reqwest::{IntoUrl, Method};
 use thiserror::Error;
 use utils::{http::error::HttpErrorBody, id::NodeId};
@@ -34,12 +33,14 @@ pub enum SafekeeperHttpError {
 
 type Result<T> = result::Result<T, SafekeeperHttpError>;
 
+#[async_trait::async_trait]
 pub trait ResponseErrorMessageExt: Sized {
-    fn error_from_body(self) -> Result<Self>;
+    async fn error_from_body(self) -> Result<Self>;
 }
 
-impl ResponseErrorMessageExt for Response {
-    fn error_from_body(self) -> Result<Self> {
+#[async_trait::async_trait]
+impl ResponseErrorMessageExt for reqwest::Response {
+    async fn error_from_body(self) -> Result<Self> {
         let status = self.status();
         if !(status.is_client_error() || status.is_server_error()) {
             return Ok(self);
@@ -48,7 +49,7 @@ impl ResponseErrorMessageExt for Response {
         // reqwest does not export its error construction utility functions, so let's craft the message ourselves
         let url = self.url().to_owned();
         Err(SafekeeperHttpError::Response(
-            match self.json::<HttpErrorBody>() {
+            match self.json::<HttpErrorBody>().await {
                 Ok(err_body) => format!("Error: {}", err_body.msg),
                 Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
             },
@@ -69,7 +70,7 @@ pub struct SafekeeperNode {
 
     pub pg_connection_config: PgConnectionConfig,
     pub env: LocalEnv,
-    pub http_client: Client,
+    pub http_client: reqwest::Client,
     pub http_base_url: String,
 }
 
@@ -80,7 +81,7 @@ impl SafekeeperNode {
             conf: conf.clone(),
             pg_connection_config: Self::safekeeper_connection_config(conf.pg_port),
             env: env.clone(),
-            http_client: Client::new(),
+            http_client: reqwest::Client::new(),
             http_base_url: format!("http://127.0.0.1:{}/v1", conf.http_port),
         }
     }
@@ -103,7 +104,7 @@ impl SafekeeperNode {
             .expect("non-Unicode path")
     }
 
-    pub fn start(&self, extra_opts: Vec<String>) -> anyhow::Result<Child> {
+    pub async fn start(&self, extra_opts: Vec<String>) -> anyhow::Result<Child> {
         print!(
             "Starting safekeeper at '{}' in '{}'",
             self.pg_connection_config.raw_address(),
@@ -191,13 +192,16 @@ impl SafekeeperNode {
             &self.env.safekeeper_bin(),
             &args,
             [],
-            background_process::InitialPidFile::Expect(&self.pid_file()),
-            || match self.check_status() {
-                Ok(()) => Ok(true),
-                Err(SafekeeperHttpError::Transport(_)) => Ok(false),
-                Err(e) => Err(anyhow::anyhow!("Failed to check node status: {e}")),
+            background_process::InitialPidFile::Expect(self.pid_file()),
+            || async {
+                match self.check_status().await {
+                    Ok(()) => Ok(true),
+                    Err(SafekeeperHttpError::Transport(_)) => Ok(false),
+                    Err(e) => Err(anyhow::anyhow!("Failed to check node status: {e}")),
+                }
             },
         )
+        .await
     }
 
     ///
@@ -216,7 +220,7 @@ impl SafekeeperNode {
         )
     }
 
-    fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> RequestBuilder {
+    fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> reqwest::RequestBuilder {
         // TODO: authentication
         //if self.env.auth_type == AuthType::NeonJWT {
         //    builder = builder.bearer_auth(&self.env.safekeeper_auth_token)
@@ -224,10 +228,12 @@ impl SafekeeperNode {
         self.http_client.request(method, url)
     }
 
-    pub fn check_status(&self) -> Result<()> {
+    pub async fn check_status(&self) -> Result<()> {
         self.http_request(Method::GET, format!("{}/{}", self.http_base_url, "status"))
-            .send()?
-            .error_from_body()?;
+            .send()
+            .await?
+            .error_from_body()
+            .await?;
         Ok(())
     }
 }

control_plane/src/tenant_migration.rs

@@ -19,11 +19,11 @@ use utils::{
 };
 
 /// Given an attached pageserver, retrieve the LSN for all timelines
-fn get_lsns(
+async fn get_lsns(
     tenant_id: TenantId,
     pageserver: &PageServerNode,
 ) -> anyhow::Result<HashMap<TimelineId, Lsn>> {
-    let timelines = pageserver.timeline_list(&tenant_id)?;
+    let timelines = pageserver.timeline_list(&tenant_id).await?;
     Ok(timelines
         .into_iter()
         .map(|t| (t.timeline_id, t.last_record_lsn))
@@ -32,13 +32,13 @@ fn get_lsns(
 
 /// Wait for the timeline LSNs on `pageserver` to catch up with or overtake
 /// `baseline`.
-fn await_lsn(
+async fn await_lsn(
     tenant_id: TenantId,
     pageserver: &PageServerNode,
     baseline: HashMap<TimelineId, Lsn>,
 ) -> anyhow::Result<()> {
     loop {
-        let latest = match get_lsns(tenant_id, pageserver) {
+        let latest = match get_lsns(tenant_id, pageserver).await {
             Ok(l) => l,
             Err(e) => {
                 println!(
@@ -84,7 +84,7 @@ fn await_lsn(
 ///  - Coordinate attach/secondary/detach on pageservers
 ///  - call into attachment_service for generations
 ///  - reconfigure compute endpoints to point to new attached pageserver
-pub fn migrate_tenant(
+pub async fn migrate_tenant(
     env: &LocalEnv,
     tenant_id: TenantId,
     dest_ps: PageServerNode,
@@ -108,16 +108,18 @@ pub fn migrate_tenant(
         }
     }
 
-    let previous = attachment_service.inspect(tenant_id)?;
+    let previous = attachment_service.inspect(tenant_id).await?;
     let mut baseline_lsns = None;
     if let Some((generation, origin_ps_id)) = &previous {
         let origin_ps = PageServerNode::from_env(env, env.get_pageserver_conf(*origin_ps_id)?);
 
         if origin_ps_id == &dest_ps.conf.id {
             println!("🔁 Already attached to {origin_ps_id}, freshening...");
-            let gen = attachment_service.attach_hook(tenant_id, dest_ps.conf.id)?;
+            let gen = attachment_service
+                .attach_hook(tenant_id, dest_ps.conf.id)
+                .await?;
             let dest_conf = build_location_config(LocationConfigMode::AttachedSingle, gen, None);
-            dest_ps.location_config(tenant_id, dest_conf, None)?;
+            dest_ps.location_config(tenant_id, dest_conf, None).await?;
             println!("✅ Migration complete");
             return Ok(());
         }
@@ -126,20 +128,24 @@ pub fn migrate_tenant(
 
         let stale_conf =
             build_location_config(LocationConfigMode::AttachedStale, Some(*generation), None);
-        origin_ps.location_config(tenant_id, stale_conf, Some(Duration::from_secs(10)))?;
+        origin_ps
+            .location_config(tenant_id, stale_conf, Some(Duration::from_secs(10)))
+            .await?;
 
-        baseline_lsns = Some(get_lsns(tenant_id, &origin_ps)?);
+        baseline_lsns = Some(get_lsns(tenant_id, &origin_ps).await?);
     }
 
-    let gen = attachment_service.attach_hook(tenant_id, dest_ps.conf.id)?;
+    let gen = attachment_service
+        .attach_hook(tenant_id, dest_ps.conf.id)
+        .await?;
     let dest_conf = build_location_config(LocationConfigMode::AttachedMulti, gen, None);
 
     println!("🔁 Attaching to pageserver {}", dest_ps.conf.id);
-    dest_ps.location_config(tenant_id, dest_conf, None)?;
+    dest_ps.location_config(tenant_id, dest_conf, None).await?;
 
     if let Some(baseline) = baseline_lsns {
         println!("🕑 Waiting for LSN to catch up...");
-        await_lsn(tenant_id, &dest_ps, baseline)?;
+        await_lsn(tenant_id, &dest_ps, baseline).await?;
     }
 
     let cplane = ComputeControlPlane::load(env.clone())?;
@@ -149,7 +155,7 @@ pub fn migrate_tenant(
                 "🔁 Reconfiguring endpoint {} to use pageserver {}",
                 endpoint_name, dest_ps.conf.id
             );
-            endpoint.reconfigure(Some(dest_ps.conf.id))?;
+            endpoint.reconfigure(Some(dest_ps.conf.id)).await?;
         }
     }
 
@@ -159,7 +165,7 @@ pub fn migrate_tenant(
         }
 
         let other_ps = PageServerNode::from_env(env, other_ps_conf);
-        let other_ps_tenants = other_ps.tenant_list()?;
+        let other_ps_tenants = other_ps.tenant_list().await?;
 
         // Check if this tenant is attached
         let found = other_ps_tenants
@@ -181,7 +187,9 @@ pub fn migrate_tenant(
             "💤 Switching to secondary mode on pageserver {}",
             other_ps.conf.id
         );
-        other_ps.location_config(tenant_id, secondary_conf, None)?;
+        other_ps
+            .location_config(tenant_id, secondary_conf, None)
+            .await?;
     }
 
     println!(
@@ -189,7 +197,7 @@ pub fn migrate_tenant(
         dest_ps.conf.id
     );
     let dest_conf = build_location_config(LocationConfigMode::AttachedSingle, gen, None);
-    dest_ps.location_config(tenant_id, dest_conf, None)?;
+    dest_ps.location_config(tenant_id, dest_conf, None).await?;
 
     println!("✅ Migration complete");
 

deny.toml

@@ -35,6 +35,7 @@ allow = [
     "Artistic-2.0",
     "BSD-2-Clause",
     "BSD-3-Clause",
+    "CC0-1.0",
     "ISC",
     "MIT",
     "MPL-2.0",

docs/rfcs/029-getpage-throttling.md

@@ -0,0 +1,197 @@
+# Per-Tenant GetPage@LSN Throttling
+
+Author: Christian Schwarz
+Date: Oct 24, 2023
+
+## Summary
+
+This RFC proposes per-tenant throttling of GetPage@LSN requests inside Pageserver
+and the interactions with its client, i.e., the neon_smgr component in Compute.
+
+The result of implementing & executing this RFC will be a fleet-wide upper limit for
+**"the highest GetPage/second that Pageserver can support for a single tenant/shard"**.
+
+## Background
+
+### GetPage@LSN Request Flow
+
+Pageserver exposes its `page_service.rs` as a libpq listener.
+The Computes' `neon_smgr` module connects to that libpq listener.
+Once a connection is established, the protocol allows Compute to request page images at a given LSN.
+We call these requests GetPage@LSN requests, or GetPage requests for short.
+Other request types can be sent, but these are low traffic compared to GetPage requests
+and are not the concern of this RFC.
+
+Pageserver associates one libpq connection with one tokio task.
+
+Per connection/task, the pq protocol is handled by the common `postgres_backend` crate.
+Its `run_message_loop` function invokes the `page_service` specific `impl<IO> postgres_backend::Handler<IO> for PageServerHandler`.
+Requests are processed in the order in which they arrive via the TCP-based pq protocol.
+So, there is no concurrent request processing within one connection/task.
+
+There is a degree of natural pipelining:
+Compute can "fill the pipe" by sending more than one GetPage request into the libpq TCP stream.
+And Pageserver can fill the pipe with responses in the other direction.
+Both directions are subject to the limit of tx/rx buffers, nodelay, TCP flow control, etc.
+
+### GetPage@LSN Access Pattern
+
+The Compute has its own hierarchy of caches, specifically `shared_buffers` and the `local file cache` (LFC).
+Compute only issues GetPage requests to Pageserver if it encounters a miss in these caches.
+
+If the working set stops fitting into Compute's caches, requests to Pageserver increase sharply -- the Compute starts *thrashing*.
+
+## Motivation
+
+In INC-69, a tenant issued 155k GetPage/second for a period of 10 minutes and 60k GetPage/second for a period of 3h,
+then dropping to ca 18k GetPage/second for a period of 9h.
+
+We noticed this because of an internal GetPage latency SLO burn rate alert, i.e.,
+the request latency profile during this period significantly exceeded what was acceptable according to the internal SLO.
+
+Sadly, we do not have the observability data to determine the impact of this tenant on other tenants on the same tenants.
+
+However, here are some illustrative data points for the 155k period:
+The tenant was responsible for >= 99% of the GetPage traffic and, frankly, the overall activity on this Pageserver instance.
+We were serving pages at 10 Gb/s (`155k x 8 kbyte (PAGE_SZ) per second is 1.12GiB/s = 9.4Gb/s.`)
+The CPU utilization of the instance was 75% user+system.
+Pageserver page cache served 1.75M accesses/second at a hit rate of ca 90%.
+The hit rate for materialized pages was ca. 40%.
+Curiously, IOPS to the Instance Store NVMe were very low, rarely exceeding 100.
+
+The fact that the IOPS were so low / the materialized page cache hit rate was so high suggests that **this tenant's compute's caches were thrashing**.
+The compute was of type `k8s-pod`; hence, auto-scaling could/would not have helped remediate the thrashing by provisioning more RAM.
+The consequence was that the **thrashing translated into excessive GetPage requests against Pageserver**.
+
+My claim is that it was **unhealthy to serve this workload at the pace we did**:
+* it is likely that other tenants were/would have experienced high latencies (again, we sadly don't have per-tenant latency data to confirm this)
+* more importantly, it was **unsustainable** to serve traffic at this pace for multiple reasons:
+    * **predictability of performance**: when the working set grows, the pageserver materialized page cache hit rate drops.
+      At some point, we're bound by the EC2 Instance Store NVMe drive's IOPS limit.
+      The result is an **uneven** performance profile from the Compute perspective.
+
+    * **economics**: Neon currently does not charge for IOPS, only capacity.
+      **We cannot afford to undercut the market in IOPS/$ this drastically; it leads to adverse selection and perverse incentives.**
+      For example, the 155k IOPS, which we served for 10min, would cost ca. 6.5k$/month when provisioned as an io2 EBS volume.
+      Even the 18k IOPS, which we served for 9h, would cost ca. 1.1k$/month when provisioned as an io2 EBS volume.
+      We charge 0$.
+      It could be economically advantageous to keep using a low-DRAM compute because Pageserver IOPS are fast enough and free.
+
+
+Note: It is helpful to think of Pageserver as a disk, because it's precisely where `neon_smgr` sits:
+vanilla Postgres gets its pages from disk, Neon Postgres gets them from Pageserver.
+So, regarding the above performance & economic arguments, it is fair to say that we currently provide an "as-fast-as-possible-IOPS" disk that we charge for only by capacity.
+
+## Solution: Throttling GetPage Requests
+
+**The consequence of the above analysis must be that Pageserver throttles GetPage@LSN requests**.
+That is, unless we want to start charging for provisioned GetPage@LSN/second.
+Throttling sets the correct incentive for a thrashing Compute to scale up its DRAM to the working set size.
+Neon Autoscaling will make this easy, [eventually](https://github.com/neondatabase/neon/pull/3913).
+
+## The Design Space
+
+What that remains is the question about *policy* and *mechanism*:
+
+**Policy** concerns itself with the question of what limit applies to a given connection|timeline|tenant.
+Candidates are:
+
+* hard limit, same limit value per connection|timeline|tenant
+    * Per-tenant will provide an upper bound for the impact of a tenant on a given Pageserver instance.
+      This is a major operational pain point / risk right now.
+* hard limit, configurable per connection|timeline|tenant
+    * This outsources policy to console/control plane, with obvious advantages for flexible structuring of what service we offer to customers.
+    * Note that this is not a mechanism to guarantee a minium provisioned rate, i.e., this is not a mechanism to guarantee a certain QoS for a tenant.
+* fair share among active connections|timelines|tenants per instance
+    * example: each connection|timeline|tenant gets a fair fraction of the machine's GetPage/second capacity
+    * NB: needs definition of "active", and knowledge of available GetPage/second capacity in advance
+* ...
+
+
+Regarding **mechanism**, it's clear that **backpressure** is the way to go.
+However, we must choose between
+* **implicit** backpressure through pq/TCP and
+* **explicit** rejection of requests + retries with exponential backoff
+
+Further, there is the question of how throttling GetPage@LSN will affect the **internal GetPage latency SLO**:
+where do we measure the SLI for Pageserver's internal getpage latency SLO? Before or after the throttling?
+
+And when we eventually move the measurement point into the Computes (to avoid coordinated omission),
+how do we avoid counting throttling-induced latency toward the internal getpage latency SLI/SLO?
+
+## Scope Of This RFC
+
+**This RFC proposes introducing a hard GetPage@LSN/second limit per tenant, with the same value applying to each tenant on a Pageserver**.
+
+This proposal is easy to implement and significantly de-risks operating large Pageservers,
+based on the assumption that extremely-high-GetPage-rate-episodes like the one from the "Motivation" section are uncorrelated between tenants.
+
+For example, suppose we pick a limit that allows up to 10 tenants to go at limit rate.
+Suppose our Pageserver can serve 100k GetPage/second total at a 100% page cache miss rate.
+If each tenant gets a hard limit of 10k GetPage/second, we can serve up to 10 tenants at limit speed without latency degradation.
+
+The mechanism for backpressure will be TCP-based implicit backpressure.
+The compute team isn't concerned about prefetch queue depth.
+Pageserver will implement it by delaying the reading of requests from the libpq connection(s).
+
+The rate limit will be implemented using a per-tenant token bucket.
+The bucket will be be shared among all connections to the tenant.
+The bucket implementation supports starvation-preventing `await`ing.
+The current candidate for the implementation is [`leaky_bucket`](https://docs.rs/leaky-bucket/).
+The getpage@lsn benchmark that's being added in https://github.com/neondatabase/neon/issues/5771
+can be used to evaluate the overhead of sharing the bucket among connections of a tenant.
+A possible technique to mitigate the impact of sharing the bucket would be to maintain a buffer of a few tokens per connection handler.
+
+Regarding metrics / the internal GetPage latency SLO:
+we will measure the GetPage latency SLO _after_ the throttler and introduce a new metric to measure the amount of throttling, quantified by:
+- histogram that records the tenants' observations of queue depth before they start waiting (one such histogram per pageserver)
+- histogram that records the tenants' observations of time spent waiting (one such histogram per pageserver)
+
+Further observability measures:
+- an INFO log message at frequency 1/min if the tenant/timeline/connection was throttled in that last minute.
+  The message will identify the tenant/timeline/connection to allow correlation with compute logs/stats.
+
+Rollout will happen as follows:
+- deploy 1: implementation + config: disabled by default, ability to enable it per tenant through tenant_conf
+- experimentation in staging and later production to study impact & interaction with auto-scaling
+- determination of a sensible global default value
+  - the value will be chosen as high as possible ...
+  - ... but low enough to work towards this RFC's goal that one tenant should not be able to dominate a pageserver instance.
+- deploy 2: implementation fixes if any + config: enabled by default with the aforementioned global default
+- reset of the experimental per-tenant overrides
+- gain experience & lower the limit over time
+  - we stop lowering the limit as soon as this RFC's goal is achieved, i.e.,
+    once we decide that in practice the chosen value sufficiently de-risks operating large pageservers
+
+The per-tenant override will remain for emergencies and testing.
+But since Console doesn't preserve it during tenant migrations, it isn't durably configurable for the tenant.
+
+Toward the upper layers of the Neon stack, the resulting limit will be
+**"the highest GetPage/second that Pageserver can support for a single tenant"**.
+
+### Rationale
+
+We decided against error + retry because of worries about starvation.
+
+## Future Work
+
+Enable per-tenant emergency override of the limit via Console.
+Should be part of a more general framework to specify tenant config overrides.
+**NB:** this is **not** the right mechanism to _sell_ different max GetPage/second levels to users,
+or _auto-scale_ the GetPage/second levels. Such functionality will require a separate RFC that
+concerns itself with GetPage/second capacity planning.
+
+Compute-side metrics for GetPage latency.
+
+Back-channel to inform Compute/Autoscaling/ControlPlane that the project is being throttled.
+
+Compute-side neon_smgr improvements to avoid sending the same GetPage request multiple times if multiple backends experience a cache miss.
+
+Dealing with read-only endpoints: users use read-only endpoints to scale reads for a single tenant.
+Possibly there are also assumptions around read-only endpoints not affecting the primary read-write endpoint's performance.
+With per-tenant rate limiting, we will not meet that expectation.
+However, we can currently only scale per tenant.
+Soon, we will have sharding (#5505), which will apply the throttling on a per-shard basis.
+But, that's orthogonal to scaling reads: if many endpoints hit one shard, they share the same throttling limit.
+To solve this properly, I think we'll need replicas for tenants / shard.
+To performance-isolate a tenant's endpoints from each other, we'd then route them to different replicas.

libs/compute_api/src/spec.rs

@@ -73,6 +73,8 @@ pub struct ComputeSpec {
 
     // information about available remote extensions
     pub remote_extensions: Option<RemoteExtSpec>,
+
+    pub pgbouncer_settings: Option<HashMap<String, String>>,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.

libs/compute_api/tests/cluster_spec.json

@@ -243,5 +243,9 @@
         "public_extensions": [
           "postgis"
         ]
+      },
+      "pgbouncer_settings": {
+        "default_pool_size": "42",
+        "pool_mode": "session"
       }
 }

libs/pageserver_api/Cargo.toml

@@ -24,3 +24,4 @@ workspace_hack.workspace = true
 
 [dev-dependencies]
 bincode.workspace = true
+rand.workspace = true

libs/pageserver_api/src/key.rs

@@ -144,3 +144,37 @@ impl Key {
 pub fn is_rel_block_key(key: &Key) -> bool {
     key.field1 == 0x00 && key.field4 != 0
 }
+
+impl std::str::FromStr for Key {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> std::result::Result<Self, Self::Err> {
+        Self::from_hex(s)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::str::FromStr;
+
+    use crate::key::Key;
+
+    use rand::Rng;
+    use rand::SeedableRng;
+
+    #[test]
+    fn display_fromstr_bijection() {
+        let mut rng = rand::rngs::StdRng::seed_from_u64(42);
+
+        let key = Key {
+            field1: rng.gen(),
+            field2: rng.gen(),
+            field3: rng.gen(),
+            field4: rng.gen(),
+            field5: rng.gen(),
+            field6: rng.gen(),
+        };
+
+        assert_eq!(key, Key::from_str(&format!("{key}")).unwrap());
+    }
+}

libs/pageserver_api/src/keyspace.rs

@@ -1,11 +1,12 @@
-use crate::repository::{key_range_size, singleton_range, Key};
 use postgres_ffi::BLCKSZ;
 use std::ops::Range;
 
+use crate::key::Key;
+
 ///
 /// Represents a set of Keys, in a compact form.
 ///
-#[derive(Clone, Debug, Default)]
+#[derive(Clone, Debug, Default, PartialEq, Eq)]
 pub struct KeySpace {
     /// Contiguous ranges of keys that belong to the key space. In key order,
     /// and with no overlap.
@@ -186,6 +187,33 @@ impl KeySpaceRandomAccum {
     }
 }
 
+pub fn key_range_size(key_range: &Range<Key>) -> u32 {
+    let start = key_range.start;
+    let end = key_range.end;
+
+    if end.field1 != start.field1
+        || end.field2 != start.field2
+        || end.field3 != start.field3
+        || end.field4 != start.field4
+    {
+        return u32::MAX;
+    }
+
+    let start = (start.field5 as u64) << 32 | start.field6 as u64;
+    let end = (end.field5 as u64) << 32 | end.field6 as u64;
+
+    let diff = end - start;
+    if diff > u32::MAX as u64 {
+        u32::MAX
+    } else {
+        diff as u32
+    }
+}
+
+pub fn singleton_range(key: Key) -> Range<Key> {
+    key..key.next()
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

libs/pageserver_api/src/lib.rs

@@ -5,6 +5,7 @@ use const_format::formatcp;
 /// Public API types
 pub mod control_api;
 pub mod key;
+pub mod keyspace;
 pub mod models;
 pub mod reltag;
 pub mod shard;

libs/pageserver_api/src/models.rs

@@ -1,5 +1,8 @@
+pub mod partitioning;
+
 use std::{
     collections::HashMap,
+    io::Read,
     num::{NonZeroU64, NonZeroUsize},
     time::SystemTime,
 };
@@ -17,7 +20,7 @@ use utils::{
 
 use crate::{reltag::RelTag, shard::TenantShardId};
 use anyhow::bail;
-use bytes::{BufMut, Bytes, BytesMut};
+use bytes::{Buf, BufMut, Bytes, BytesMut};
 
 /// The state of a tenant in this pageserver.
 ///
@@ -367,6 +370,14 @@ pub struct TenantInfo {
     pub attachment_status: TenantAttachmentStatus,
 }
 
+#[derive(Serialize, Deserialize, Clone)]
+pub struct TenantDetails {
+    #[serde(flatten)]
+    pub tenant_info: TenantInfo,
+
+    pub timelines: Vec<TimelineId>,
+}
+
 /// This represents the output of the "timeline_detail" and "timeline_list" API calls.
 #[derive(Debug, Serialize, Deserialize, Clone)]
 pub struct TimelineInfo {
@@ -574,6 +585,7 @@ pub enum PagestreamFeMessage {
 }
 
 // Wrapped in libpq CopyData
+#[derive(strum_macros::EnumProperty)]
 pub enum PagestreamBeMessage {
     Exists(PagestreamExistsResponse),
     Nblocks(PagestreamNblocksResponse),
@@ -582,6 +594,29 @@ pub enum PagestreamBeMessage {
     DbSize(PagestreamDbSizeResponse),
 }
 
+// Keep in sync with `pagestore_client.h`
+#[repr(u8)]
+enum PagestreamBeMessageTag {
+    Exists = 100,
+    Nblocks = 101,
+    GetPage = 102,
+    Error = 103,
+    DbSize = 104,
+}
+impl TryFrom<u8> for PagestreamBeMessageTag {
+    type Error = u8;
+    fn try_from(value: u8) -> Result<Self, u8> {
+        match value {
+            100 => Ok(PagestreamBeMessageTag::Exists),
+            101 => Ok(PagestreamBeMessageTag::Nblocks),
+            102 => Ok(PagestreamBeMessageTag::GetPage),
+            103 => Ok(PagestreamBeMessageTag::Error),
+            104 => Ok(PagestreamBeMessageTag::DbSize),
+            _ => Err(value),
+        }
+    }
+}
+
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamExistsRequest {
     pub latest: bool,
@@ -737,35 +772,91 @@ impl PagestreamBeMessage {
     pub fn serialize(&self) -> Bytes {
         let mut bytes = BytesMut::new();
 
+        use PagestreamBeMessageTag as Tag;
         match self {
             Self::Exists(resp) => {
-                bytes.put_u8(100); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::Exists as u8);
                 bytes.put_u8(resp.exists as u8);
             }
 
             Self::Nblocks(resp) => {
-                bytes.put_u8(101); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::Nblocks as u8);
                 bytes.put_u32(resp.n_blocks);
             }
 
             Self::GetPage(resp) => {
-                bytes.put_u8(102); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::GetPage as u8);
                 bytes.put(&resp.page[..]);
             }
 
             Self::Error(resp) => {
-                bytes.put_u8(103); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::Error as u8);
                 bytes.put(resp.message.as_bytes());
                 bytes.put_u8(0); // null terminator
             }
             Self::DbSize(resp) => {
-                bytes.put_u8(104); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::DbSize as u8);
                 bytes.put_i64(resp.db_size);
             }
         }
 
         bytes.into()
     }
+
+    pub fn deserialize(buf: Bytes) -> anyhow::Result<Self> {
+        let mut buf = buf.reader();
+        let msg_tag = buf.read_u8()?;
+
+        use PagestreamBeMessageTag as Tag;
+        let ok =
+            match Tag::try_from(msg_tag).map_err(|tag: u8| anyhow::anyhow!("invalid tag {tag}"))? {
+                Tag::Exists => {
+                    let exists = buf.read_u8()?;
+                    Self::Exists(PagestreamExistsResponse {
+                        exists: exists != 0,
+                    })
+                }
+                Tag::Nblocks => {
+                    let n_blocks = buf.read_u32::<BigEndian>()?;
+                    Self::Nblocks(PagestreamNblocksResponse { n_blocks })
+                }
+                Tag::GetPage => {
+                    let mut page = vec![0; 8192]; // TODO: use MaybeUninit
+                    buf.read_exact(&mut page)?;
+                    PagestreamBeMessage::GetPage(PagestreamGetPageResponse { page: page.into() })
+                }
+                Tag::Error => {
+                    let buf = buf.get_ref();
+                    let cstr = std::ffi::CStr::from_bytes_until_nul(buf)?;
+                    let rust_str = cstr.to_str()?;
+                    PagestreamBeMessage::Error(PagestreamErrorResponse {
+                        message: rust_str.to_owned(),
+                    })
+                }
+                Tag::DbSize => {
+                    let db_size = buf.read_i64::<BigEndian>()?;
+                    Self::DbSize(PagestreamDbSizeResponse { db_size })
+                }
+            };
+        let remaining = buf.into_inner();
+        if !remaining.is_empty() {
+            anyhow::bail!(
+                "remaining bytes in msg with tag={msg_tag}: {}",
+                remaining.len()
+            );
+        }
+        Ok(ok)
+    }
+
+    pub fn kind(&self) -> &'static str {
+        match self {
+            Self::Exists(_) => "Exists",
+            Self::Nblocks(_) => "Nblocks",
+            Self::GetPage(_) => "GetPage",
+            Self::Error(_) => "Error",
+            Self::DbSize(_) => "DbSize",
+        }
+    }
 }
 
 #[cfg(test)]

libs/pageserver_api/src/models/partitioning.rs

@@ -0,0 +1,151 @@
+use utils::lsn::Lsn;
+
+#[derive(Debug, PartialEq, Eq)]
+pub struct Partitioning {
+    pub keys: crate::keyspace::KeySpace,
+
+    pub at_lsn: Lsn,
+}
+
+impl serde::Serialize for Partitioning {
+    fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        pub struct KeySpace<'a>(&'a crate::keyspace::KeySpace);
+
+        impl<'a> serde::Serialize for KeySpace<'a> {
+            fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+            where
+                S: serde::Serializer,
+            {
+                use serde::ser::SerializeSeq;
+                let mut seq = serializer.serialize_seq(Some(self.0.ranges.len()))?;
+                for kr in &self.0.ranges {
+                    seq.serialize_element(&KeyRange(kr))?;
+                }
+                seq.end()
+            }
+        }
+
+        use serde::ser::SerializeMap;
+        let mut map = serializer.serialize_map(Some(2))?;
+        map.serialize_key("keys")?;
+        map.serialize_value(&KeySpace(&self.keys))?;
+        map.serialize_key("at_lsn")?;
+        map.serialize_value(&WithDisplay(&self.at_lsn))?;
+        map.end()
+    }
+}
+
+pub struct WithDisplay<'a, T>(&'a T);
+
+impl<'a, T: std::fmt::Display> serde::Serialize for WithDisplay<'a, T> {
+    fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        serializer.collect_str(&self.0)
+    }
+}
+
+pub struct KeyRange<'a>(&'a std::ops::Range<crate::key::Key>);
+
+impl<'a> serde::Serialize for KeyRange<'a> {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        use serde::ser::SerializeTuple;
+        let mut t = serializer.serialize_tuple(2)?;
+        t.serialize_element(&WithDisplay(&self.0.start))?;
+        t.serialize_element(&WithDisplay(&self.0.end))?;
+        t.end()
+    }
+}
+
+impl<'a> serde::Deserialize<'a> for Partitioning {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'a>,
+    {
+        pub struct KeySpace(crate::keyspace::KeySpace);
+
+        impl<'de> serde::Deserialize<'de> for KeySpace {
+            fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+            where
+                D: serde::Deserializer<'de>,
+            {
+                #[serde_with::serde_as]
+                #[derive(serde::Deserialize)]
+                #[serde(transparent)]
+                struct Key(#[serde_as(as = "serde_with::DisplayFromStr")] crate::key::Key);
+
+                #[serde_with::serde_as]
+                #[derive(serde::Deserialize)]
+                struct Range(Key, Key);
+
+                let ranges: Vec<Range> = serde::Deserialize::deserialize(deserializer)?;
+                Ok(Self(crate::keyspace::KeySpace {
+                    ranges: ranges
+                        .into_iter()
+                        .map(|Range(start, end)| (start.0..end.0))
+                        .collect(),
+                }))
+            }
+        }
+
+        #[serde_with::serde_as]
+        #[derive(serde::Deserialize)]
+        struct De {
+            keys: KeySpace,
+            #[serde_as(as = "serde_with::DisplayFromStr")]
+            at_lsn: Lsn,
+        }
+
+        let de: De = serde::Deserialize::deserialize(deserializer)?;
+        Ok(Self {
+            at_lsn: de.at_lsn,
+            keys: de.keys.0,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_serialization_roundtrip() {
+        let reference = r#"
+        {
+            "keys": [
+              [
+                "000000000000000000000000000000000000",
+                "000000000000000000000000000000000001"
+              ],
+              [
+                "000000067F00000001000000000000000000",
+                "000000067F00000001000000000000000002"
+              ],
+              [
+                "030000000000000000000000000000000000",
+                "030000000000000000000000000000000003"
+              ]
+            ],
+            "at_lsn": "0/2240160"
+        }
+        "#;
+
+        let de: Partitioning = serde_json::from_str(reference).unwrap();
+
+        let ser = serde_json::to_string(&de).unwrap();
+
+        let ser_de: serde_json::Value = serde_json::from_str(&ser).unwrap();
+
+        assert_eq!(
+            ser_de,
+            serde_json::from_str::<'_, serde_json::Value>(reference).unwrap()
+        );
+    }
+}

libs/pageserver_api/src/shard.rs

@@ -81,6 +81,10 @@ impl TenantShardId {
     pub fn is_zero(&self) -> bool {
         self.shard_number == ShardNumber(0)
     }
+
+    pub fn is_unsharded(&self) -> bool {
+        self.shard_number == ShardNumber(0) && self.shard_count == ShardCount(0)
+    }
 }
 
 /// Formatting helper
@@ -159,7 +163,7 @@ impl From<[u8; 18]> for TenantShardId {
 /// shard we're dealing with, but do not need to know the full ShardIdentity (because
 /// we won't be doing any page->shard mapping), and do not need to know the fully qualified
 /// TenantShardId.
-#[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy)]
+#[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy, Hash)]
 pub struct ShardIndex {
     pub shard_number: ShardNumber,
     pub shard_count: ShardCount,

libs/postgres_connection/src/lib.rs

@@ -163,8 +163,18 @@ impl PgConnectionConfig {
     }
 
     /// Connect using postgres protocol with TLS disabled.
-    pub fn connect_no_tls(&self) -> Result<postgres::Client, postgres::Error> {
-        postgres::Config::from(self.to_tokio_postgres_config()).connect(postgres::NoTls)
+    pub async fn connect_no_tls(
+        &self,
+    ) -> Result<
+        (
+            tokio_postgres::Client,
+            tokio_postgres::Connection<tokio_postgres::Socket, tokio_postgres::tls::NoTlsStream>,
+        ),
+        postgres::Error,
+    > {
+        self.to_tokio_postgres_config()
+            .connect(postgres::NoTls)
+            .await
     }
 }
 

libs/remote_storage/src/azure_blob.rs

@@ -117,6 +117,8 @@ impl AzureBlobStorage {
     ) -> Result<Download, DownloadError> {
         let mut response = builder.into_stream();
 
+        let mut etag = None;
+        let mut last_modified = None;
         let mut metadata = HashMap::new();
         // TODO give proper streaming response instead of buffering into RAM
         // https://github.com/neondatabase/neon/issues/5563
@@ -124,6 +126,13 @@ impl AzureBlobStorage {
         let mut bufs = Vec::new();
         while let Some(part) = response.next().await {
             let part = part.map_err(to_download_error)?;
+            let etag_str: &str = part.blob.properties.etag.as_ref();
+            if etag.is_none() {
+                etag = Some(etag.unwrap_or_else(|| etag_str.to_owned()));
+            }
+            if last_modified.is_none() {
+                last_modified = Some(part.blob.properties.last_modified.into());
+            }
             if let Some(blob_meta) = part.blob.metadata {
                 metadata.extend(blob_meta.iter().map(|(k, v)| (k.to_owned(), v.to_owned())));
             }
@@ -136,6 +145,8 @@ impl AzureBlobStorage {
         }
         Ok(Download {
             download_stream: Box::pin(futures::stream::iter(bufs.into_iter().map(Ok))),
+            etag,
+            last_modified,
             metadata: Some(StorageMetadata(metadata)),
         })
     }

libs/remote_storage/src/lib.rs

@@ -14,7 +14,9 @@ mod local_fs;
 mod s3_bucket;
 mod simulate_failures;
 
-use std::{collections::HashMap, fmt::Debug, num::NonZeroUsize, pin::Pin, sync::Arc};
+use std::{
+    collections::HashMap, fmt::Debug, num::NonZeroUsize, pin::Pin, sync::Arc, time::SystemTime,
+};
 
 use anyhow::{bail, Context};
 use camino::{Utf8Path, Utf8PathBuf};
@@ -207,8 +209,13 @@ pub trait RemoteStorage: Send + Sync + 'static {
     async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()>;
 }
 
+pub type DownloadStream = Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync>>;
 pub struct Download {
-    pub download_stream: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync>>,
+    pub download_stream: DownloadStream,
+    /// The last time the file was modified (`last-modified` HTTP header)
+    pub last_modified: Option<SystemTime>,
+    /// A way to identify this specific version of the resource (`etag` HTTP header)
+    pub etag: Option<String>,
     /// Extra key-value data, associated with the current remote file.
     pub metadata: Option<StorageMetadata>,
 }

libs/remote_storage/src/local_fs.rs

@@ -18,7 +18,7 @@ use tokio_util::io::ReaderStream;
 use tracing::*;
 use utils::{crashsafe::path_with_suffix_extension, fs_ext::is_directory_empty};
 
-use crate::{Download, DownloadError, Listing, ListingMode, RemotePath};
+use crate::{Download, DownloadError, DownloadStream, Listing, ListingMode, RemotePath};
 
 use super::{RemoteStorage, StorageMetadata};
 
@@ -331,6 +331,8 @@ impl RemoteStorage for LocalFs {
                 .map_err(DownloadError::Other)?;
             Ok(Download {
                 metadata,
+                last_modified: None,
+                etag: None,
                 download_stream: Box::pin(source),
             })
         } else {
@@ -372,17 +374,17 @@ impl RemoteStorage for LocalFs {
                 .await
                 .map_err(DownloadError::Other)?;
 
-            Ok(match end_exclusive {
-                Some(end_exclusive) => Download {
-                    metadata,
-                    download_stream: Box::pin(ReaderStream::new(
-                        source.take(end_exclusive - start_inclusive),
-                    )),
-                },
-                None => Download {
-                    metadata,
-                    download_stream: Box::pin(ReaderStream::new(source)),
-                },
+            let download_stream: DownloadStream = match end_exclusive {
+                Some(end_exclusive) => Box::pin(ReaderStream::new(
+                    source.take(end_exclusive - start_inclusive),
+                )),
+                None => Box::pin(ReaderStream::new(source)),
+            };
+            Ok(Download {
+                metadata,
+                last_modified: None,
+                etag: None,
+                download_stream,
             })
         } else {
             Err(DownloadError::NotFound)

libs/remote_storage/src/s3_bucket.rs

@@ -16,6 +16,7 @@ use aws_config::{
     environment::credentials::EnvironmentVariableCredentialsProvider,
     imds::credentials::ImdsCredentialsProvider,
     meta::credentials::CredentialsProviderChain,
+    profile::ProfileFileCredentialsProvider,
     provider_config::ProviderConfig,
     retry::{RetryConfigBuilder, RetryMode},
     web_identity_token::WebIdentityTokenCredentialsProvider,
@@ -74,20 +75,29 @@ impl S3Bucket {
 
         let region = Some(Region::new(aws_config.bucket_region.clone()));
 
+        let provider_conf = ProviderConfig::without_region().with_region(region.clone());
+
         let credentials_provider = {
             // uses "AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY"
             CredentialsProviderChain::first_try(
                 "env",
                 EnvironmentVariableCredentialsProvider::new(),
             )
+            // uses "AWS_PROFILE" / `aws sso login --profile <profile>`
+            .or_else(
+                "profile-sso",
+                ProfileFileCredentialsProvider::builder()
+                    .configure(&provider_conf)
+                    .build(),
+            )
             // uses "AWS_WEB_IDENTITY_TOKEN_FILE", "AWS_ROLE_ARN", "AWS_ROLE_SESSION_NAME"
             // needed to access remote extensions bucket
-            .or_else("token", {
-                let provider_conf = ProviderConfig::without_region().with_region(region.clone());
+            .or_else(
+                "token",
                 WebIdentityTokenCredentialsProvider::builder()
                     .configure(&provider_conf)
-                    .build()
-            })
+                    .build(),
+            )
             // uses imds v2
             .or_else("imds", ImdsCredentialsProvider::builder().build())
         };
@@ -218,17 +228,11 @@ impl S3Bucket {
 
         let started_at = ScopeGuard::into_inner(started_at);
 
-        if get_object.is_err() {
-            metrics::BUCKET_METRICS.req_seconds.observe_elapsed(
-                kind,
-                AttemptOutcome::Err,
-                started_at,
-            );
-        }
-
         match get_object {
             Ok(object_output) => {
                 let metadata = object_output.metadata().cloned().map(StorageMetadata);
+                let etag = object_output.e_tag.clone();
+                let last_modified = object_output.last_modified.and_then(|t| t.try_into().ok());
 
                 let body = object_output.body;
                 let body = ByteStreamAsStream::from(body);
@@ -237,15 +241,33 @@ impl S3Bucket {
 
                 Ok(Download {
                     metadata,
+                    etag,
+                    last_modified,
                     download_stream: Box::pin(body),
                 })
             }
             Err(SdkError::ServiceError(e)) if matches!(e.err(), GetObjectError::NoSuchKey(_)) => {
+                // Count this in the AttemptOutcome::Ok bucket, because 404 is not
+                // an error: we expect to sometimes fetch an object and find it missing,
+                // e.g. when probing for timeline indices.
+                metrics::BUCKET_METRICS.req_seconds.observe_elapsed(
+                    kind,
+                    AttemptOutcome::Ok,
+                    started_at,
+                );
                 Err(DownloadError::NotFound)
             }
-            Err(e) => Err(DownloadError::Other(
-                anyhow::Error::new(e).context("download s3 object"),
-            )),
+            Err(e) => {
+                metrics::BUCKET_METRICS.req_seconds.observe_elapsed(
+                    kind,
+                    AttemptOutcome::Err,
+                    started_at,
+                );
+
+                Err(DownloadError::Other(
+                    anyhow::Error::new(e).context("download s3 object"),
+                ))
+            }
         }
     }
 }

libs/remote_storage/tests/common/mod.rs

@@ -0,0 +1,200 @@
+use std::collections::HashSet;
+use std::ops::ControlFlow;
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use anyhow::Context;
+use bytes::Bytes;
+use camino::Utf8Path;
+use futures::stream::Stream;
+use once_cell::sync::OnceCell;
+use remote_storage::{Download, GenericRemoteStorage, RemotePath};
+use tokio::task::JoinSet;
+use tracing::{debug, error, info};
+
+static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+
+pub(crate) fn upload_stream(
+    content: std::borrow::Cow<'static, [u8]>,
+) -> (
+    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
+    usize,
+) {
+    use std::borrow::Cow;
+
+    let content = match content {
+        Cow::Borrowed(x) => Bytes::from_static(x),
+        Cow::Owned(vec) => Bytes::from(vec),
+    };
+    wrap_stream(content)
+}
+
+pub(crate) fn wrap_stream(
+    content: bytes::Bytes,
+) -> (
+    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
+    usize,
+) {
+    let len = content.len();
+    let content = futures::future::ready(Ok(content));
+
+    (futures::stream::once(content), len)
+}
+
+pub(crate) async fn download_to_vec(dl: Download) -> anyhow::Result<Vec<u8>> {
+    let mut buf = Vec::new();
+    tokio::io::copy_buf(
+        &mut tokio_util::io::StreamReader::new(dl.download_stream),
+        &mut buf,
+    )
+    .await?;
+    Ok(buf)
+}
+
+// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
+pub(crate) async fn upload_simple_remote_data(
+    client: &Arc<GenericRemoteStorage>,
+    upload_tasks_count: usize,
+) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
+    info!("Creating {upload_tasks_count} remote files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
+            let blob_path = RemotePath::new(
+                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
+            )
+            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
+            task_client.upload(data, len, &blob_path, None).await?;
+
+            Ok::<_, anyhow::Error>(blob_path)
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok(upload_path) => {
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    if upload_tasks_failed {
+        ControlFlow::Break(uploaded_blobs)
+    } else {
+        ControlFlow::Continue(uploaded_blobs)
+    }
+}
+
+pub(crate) async fn cleanup(
+    client: &Arc<GenericRemoteStorage>,
+    objects_to_delete: HashSet<RemotePath>,
+) {
+    info!(
+        "Removing {} objects from the remote storage during cleanup",
+        objects_to_delete.len()
+    );
+    let mut delete_tasks = JoinSet::new();
+    for object_to_delete in objects_to_delete {
+        let task_client = Arc::clone(client);
+        delete_tasks.spawn(async move {
+            debug!("Deleting remote item at path {object_to_delete:?}");
+            task_client
+                .delete(&object_to_delete)
+                .await
+                .with_context(|| format!("{object_to_delete:?} removal"))
+        });
+    }
+
+    while let Some(task_run_result) = delete_tasks.join_next().await {
+        match task_run_result {
+            Ok(task_result) => match task_result {
+                Ok(()) => {}
+                Err(e) => error!("Delete task failed: {e:?}"),
+            },
+            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
+        }
+    }
+}
+pub(crate) struct Uploads {
+    pub(crate) prefixes: HashSet<RemotePath>,
+    pub(crate) blobs: HashSet<RemotePath>,
+}
+
+pub(crate) async fn upload_remote_data(
+    client: &Arc<GenericRemoteStorage>,
+    base_prefix_str: &'static str,
+    upload_tasks_count: usize,
+) -> ControlFlow<Uploads, Uploads> {
+    info!("Creating {upload_tasks_count} remote files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
+            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
+                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
+            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let (data, data_len) =
+                upload_stream(format!("remote blob data {i}").into_bytes().into());
+            task_client.upload(data, data_len, &blob_path, None).await?;
+
+            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok((upload_prefix, upload_path)) => {
+                uploaded_prefixes.insert(upload_prefix);
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    let uploads = Uploads {
+        prefixes: uploaded_prefixes,
+        blobs: uploaded_blobs,
+    };
+    if upload_tasks_failed {
+        ControlFlow::Break(uploads)
+    } else {
+        ControlFlow::Continue(uploads)
+    }
+}
+
+pub(crate) fn ensure_logging_ready() {
+    LOGGING_DONE.get_or_init(|| {
+        utils::logging::init(
+            utils::logging::LogFormat::Test,
+            utils::logging::TracingErrorLayerEnablement::Disabled,
+            utils::logging::Output::Stdout,
+        )
+        .expect("logging init failed");
+    });
+}

libs/remote_storage/tests/test_real_azure.rs

@@ -2,23 +2,23 @@ use std::collections::HashSet;
 use std::env;
 use std::num::NonZeroUsize;
 use std::ops::ControlFlow;
-use std::path::PathBuf;
 use std::sync::Arc;
 use std::time::UNIX_EPOCH;
 
 use anyhow::Context;
-use bytes::Bytes;
 use camino::Utf8Path;
-use futures::stream::Stream;
-use once_cell::sync::OnceCell;
 use remote_storage::{
-    AzureConfig, Download, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
+    AzureConfig, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
 };
 use test_context::{test_context, AsyncTestContext};
-use tokio::task::JoinSet;
-use tracing::{debug, error, info};
+use tracing::{debug, info};
 
-static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+mod common;
+
+use common::{
+    cleanup, download_to_vec, ensure_logging_ready, upload_remote_data, upload_simple_remote_data,
+    upload_stream, wrap_stream,
+};
 
 const ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_AZURE_REMOTE_STORAGE";
 
@@ -30,7 +30,7 @@ const BASE_PREFIX: &str = "test";
 /// If real Azure tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
 /// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
 ///
-/// First, the test creates a set of Azure blobs with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_azure_data`]
+/// First, the test creates a set of Azure blobs with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_remote_data`]
 /// where
 /// * `random_prefix_part` is set for the entire Azure client during the Azure client creation in [`create_azure_client`], to avoid multiple test runs interference
 /// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
@@ -97,7 +97,7 @@ async fn azure_pagination_should_work(
 /// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified. Test will skip real code and pass if env vars not set.
 /// See `Azure_pagination_should_work` for more information.
 ///
-/// First, create a set of Azure objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_azure_data`]
+/// First, create a set of Azure objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_remote_data`]
 /// Then performs the following queries:
 ///    1. `list_files(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
 ///    2. `list_files("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
@@ -218,18 +218,9 @@ async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Res
 
     ctx.client.upload(data, len, &path, None).await?;
 
-    async fn download_and_compare(dl: Download) -> anyhow::Result<Vec<u8>> {
-        let mut buf = Vec::new();
-        tokio::io::copy_buf(
-            &mut tokio_util::io::StreamReader::new(dl.download_stream),
-            &mut buf,
-        )
-        .await?;
-        Ok(buf)
-    }
     // Normal download request
     let dl = ctx.client.download(&path).await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     // Full range (end specified)
@@ -237,12 +228,12 @@ async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Res
         .client
         .download_byte_range(&path, 0, Some(len as u64))
         .await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     // partial range (end specified)
     let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[4..10]);
 
     // partial range (end beyond real end)
@@ -250,17 +241,17 @@ async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Res
         .client
         .download_byte_range(&path, 8, Some(len as u64 * 100))
         .await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[8..]);
 
     // Partial range (end unspecified)
     let dl = ctx.client.download_byte_range(&path, 4, None).await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[4..]);
 
     // Full range (end unspecified)
     let dl = ctx.client.download_byte_range(&path, 0, None).await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     debug!("Cleanup: deleting file at path {path:?}");
@@ -272,17 +263,6 @@ async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Res
     Ok(())
 }
 
-fn ensure_logging_ready() {
-    LOGGING_DONE.get_or_init(|| {
-        utils::logging::init(
-            utils::logging::LogFormat::Test,
-            utils::logging::TracingErrorLayerEnablement::Disabled,
-            utils::logging::Output::Stdout,
-        )
-        .expect("logging init failed");
-    });
-}
-
 struct EnabledAzure {
     client: Arc<GenericRemoteStorage>,
     base_prefix: &'static str,
@@ -352,7 +332,7 @@ impl AsyncTestContext for MaybeEnabledAzureWithTestBlobs {
 
         let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
 
-        match upload_azure_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
+        match upload_remote_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
             ControlFlow::Continue(uploads) => {
                 info!("Remote objects created successfully");
 
@@ -414,7 +394,7 @@ impl AsyncTestContext for MaybeEnabledAzureWithSimpleTestBlobs {
 
         let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
 
-        match upload_simple_azure_data(&enabled.client, upload_tasks_count).await {
+        match upload_simple_remote_data(&enabled.client, upload_tasks_count).await {
             ControlFlow::Continue(uploads) => {
                 info!("Remote objects created successfully");
 
@@ -478,166 +458,3 @@ fn create_azure_client(
         GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
     ))
 }
-
-struct Uploads {
-    prefixes: HashSet<RemotePath>,
-    blobs: HashSet<RemotePath>,
-}
-
-async fn upload_azure_data(
-    client: &Arc<GenericRemoteStorage>,
-    base_prefix_str: &'static str,
-    upload_tasks_count: usize,
-) -> ControlFlow<Uploads, Uploads> {
-    info!("Creating {upload_tasks_count} Azure files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
-            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
-                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
-            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, len, &blob_path, None).await?;
-
-            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok((upload_prefix, upload_path)) => {
-                uploaded_prefixes.insert(upload_prefix);
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    let uploads = Uploads {
-        prefixes: uploaded_prefixes,
-        blobs: uploaded_blobs,
-    };
-    if upload_tasks_failed {
-        ControlFlow::Break(uploads)
-    } else {
-        ControlFlow::Continue(uploads)
-    }
-}
-
-async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
-    info!(
-        "Removing {} objects from the remote storage during cleanup",
-        objects_to_delete.len()
-    );
-    let mut delete_tasks = JoinSet::new();
-    for object_to_delete in objects_to_delete {
-        let task_client = Arc::clone(client);
-        delete_tasks.spawn(async move {
-            debug!("Deleting remote item at path {object_to_delete:?}");
-            task_client
-                .delete(&object_to_delete)
-                .await
-                .with_context(|| format!("{object_to_delete:?} removal"))
-        });
-    }
-
-    while let Some(task_run_result) = delete_tasks.join_next().await {
-        match task_run_result {
-            Ok(task_result) => match task_result {
-                Ok(()) => {}
-                Err(e) => error!("Delete task failed: {e:?}"),
-            },
-            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
-        }
-    }
-}
-
-// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
-async fn upload_simple_azure_data(
-    client: &Arc<GenericRemoteStorage>,
-    upload_tasks_count: usize,
-) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
-    info!("Creating {upload_tasks_count} Azure files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
-            let blob_path = RemotePath::new(
-                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
-            )
-            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, len, &blob_path, None).await?;
-
-            Ok::<_, anyhow::Error>(blob_path)
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok(upload_path) => {
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    if upload_tasks_failed {
-        ControlFlow::Break(uploaded_blobs)
-    } else {
-        ControlFlow::Continue(uploaded_blobs)
-    }
-}
-
-// FIXME: copypasted from test_real_s3, can't remember how to share a module which is not compiled
-// to binary
-fn upload_stream(
-    content: std::borrow::Cow<'static, [u8]>,
-) -> (
-    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
-    usize,
-) {
-    use std::borrow::Cow;
-
-    let content = match content {
-        Cow::Borrowed(x) => Bytes::from_static(x),
-        Cow::Owned(vec) => Bytes::from(vec),
-    };
-    wrap_stream(content)
-}
-
-fn wrap_stream(
-    content: bytes::Bytes,
-) -> (
-    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
-    usize,
-) {
-    let len = content.len();
-    let content = futures::future::ready(Ok(content));
-
-    (futures::stream::once(content), len)
-}

libs/remote_storage/tests/test_real_s3.rs

@@ -2,23 +2,23 @@ use std::collections::HashSet;
 use std::env;
 use std::num::NonZeroUsize;
 use std::ops::ControlFlow;
-use std::path::PathBuf;
 use std::sync::Arc;
 use std::time::UNIX_EPOCH;
 
 use anyhow::Context;
-use bytes::Bytes;
 use camino::Utf8Path;
-use futures::stream::Stream;
-use once_cell::sync::OnceCell;
 use remote_storage::{
     GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
 };
 use test_context::{test_context, AsyncTestContext};
-use tokio::task::JoinSet;
-use tracing::{debug, error, info};
+use tracing::{debug, info};
 
-static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+mod common;
+
+use common::{
+    cleanup, download_to_vec, ensure_logging_ready, upload_remote_data, upload_simple_remote_data,
+    upload_stream, wrap_stream,
+};
 
 const ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
 
@@ -30,7 +30,7 @@ const BASE_PREFIX: &str = "test";
 /// If real S3 tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
 /// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
 ///
-/// First, the test creates a set of S3 objects with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_s3_data`]
+/// First, the test creates a set of S3 objects with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_remote_data`]
 /// where
 /// * `random_prefix_part` is set for the entire S3 client during the S3 client creation in [`create_s3_client`], to avoid multiple test runs interference
 /// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
@@ -95,7 +95,7 @@ async fn s3_pagination_should_work(ctx: &mut MaybeEnabledS3WithTestBlobs) -> any
 /// Uses real S3 and requires [`ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME`] and related S3 cred env vars specified. Test will skip real code and pass if env vars not set.
 /// See `s3_pagination_should_work` for more information.
 ///
-/// First, create a set of S3 objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_s3_data`]
+/// First, create a set of S3 objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_remote_data`]
 /// Then performs the following queries:
 ///    1. `list_files(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
 ///    2. `list_files("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
@@ -198,15 +198,65 @@ async fn s3_delete_objects_works(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()>
     Ok(())
 }
 
-fn ensure_logging_ready() {
-    LOGGING_DONE.get_or_init(|| {
-        utils::logging::init(
-            utils::logging::LogFormat::Test,
-            utils::logging::TracingErrorLayerEnablement::Disabled,
-            utils::logging::Output::Stdout,
-        )
-        .expect("logging init failed");
-    });
+#[test_context(MaybeEnabledS3)]
+#[tokio::test]
+async fn s3_upload_download_works(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()> {
+    let MaybeEnabledS3::Enabled(ctx) = ctx else {
+        return Ok(());
+    };
+
+    let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let orig = bytes::Bytes::from_static("remote blob data here".as_bytes());
+
+    let (data, len) = wrap_stream(orig.clone());
+
+    ctx.client.upload(data, len, &path, None).await?;
+
+    // Normal download request
+    let dl = ctx.client.download(&path).await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig);
+
+    // Full range (end specified)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 0, Some(len as u64))
+        .await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig);
+
+    // partial range (end specified)
+    let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig[4..10]);
+
+    // partial range (end beyond real end)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 8, Some(len as u64 * 100))
+        .await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig[8..]);
+
+    // Partial range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 4, None).await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig[4..]);
+
+    // Full range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 0, None).await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig);
+
+    debug!("Cleanup: deleting file at path {path:?}");
+    ctx.client
+        .delete(&path)
+        .await
+        .with_context(|| format!("{path:?} removal"))?;
+
+    Ok(())
 }
 
 struct EnabledS3 {
@@ -278,7 +328,7 @@ impl AsyncTestContext for MaybeEnabledS3WithTestBlobs {
 
         let enabled = EnabledS3::setup(Some(max_keys_in_list_response)).await;
 
-        match upload_s3_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
+        match upload_remote_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
             ControlFlow::Continue(uploads) => {
                 info!("Remote objects created successfully");
 
@@ -340,7 +390,7 @@ impl AsyncTestContext for MaybeEnabledS3WithSimpleTestBlobs {
 
         let enabled = EnabledS3::setup(Some(max_keys_in_list_response)).await;
 
-        match upload_simple_s3_data(&enabled.client, upload_tasks_count).await {
+        match upload_simple_remote_data(&enabled.client, upload_tasks_count).await {
             ControlFlow::Continue(uploads) => {
                 info!("Remote objects created successfully");
 
@@ -403,166 +453,3 @@ fn create_s3_client(
         GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
     ))
 }
-
-struct Uploads {
-    prefixes: HashSet<RemotePath>,
-    blobs: HashSet<RemotePath>,
-}
-
-async fn upload_s3_data(
-    client: &Arc<GenericRemoteStorage>,
-    base_prefix_str: &'static str,
-    upload_tasks_count: usize,
-) -> ControlFlow<Uploads, Uploads> {
-    info!("Creating {upload_tasks_count} S3 files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
-            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
-                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
-            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let (data, data_len) =
-                upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, data_len, &blob_path, None).await?;
-
-            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok((upload_prefix, upload_path)) => {
-                uploaded_prefixes.insert(upload_prefix);
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    let uploads = Uploads {
-        prefixes: uploaded_prefixes,
-        blobs: uploaded_blobs,
-    };
-    if upload_tasks_failed {
-        ControlFlow::Break(uploads)
-    } else {
-        ControlFlow::Continue(uploads)
-    }
-}
-
-async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
-    info!(
-        "Removing {} objects from the remote storage during cleanup",
-        objects_to_delete.len()
-    );
-    let mut delete_tasks = JoinSet::new();
-    for object_to_delete in objects_to_delete {
-        let task_client = Arc::clone(client);
-        delete_tasks.spawn(async move {
-            debug!("Deleting remote item at path {object_to_delete:?}");
-            task_client
-                .delete(&object_to_delete)
-                .await
-                .with_context(|| format!("{object_to_delete:?} removal"))
-        });
-    }
-
-    while let Some(task_run_result) = delete_tasks.join_next().await {
-        match task_run_result {
-            Ok(task_result) => match task_result {
-                Ok(()) => {}
-                Err(e) => error!("Delete task failed: {e:?}"),
-            },
-            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
-        }
-    }
-}
-
-// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
-async fn upload_simple_s3_data(
-    client: &Arc<GenericRemoteStorage>,
-    upload_tasks_count: usize,
-) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
-    info!("Creating {upload_tasks_count} S3 files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
-            let blob_path = RemotePath::new(
-                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
-            )
-            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let (data, data_len) =
-                upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, data_len, &blob_path, None).await?;
-
-            Ok::<_, anyhow::Error>(blob_path)
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok(upload_path) => {
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    if upload_tasks_failed {
-        ControlFlow::Break(uploaded_blobs)
-    } else {
-        ControlFlow::Continue(uploaded_blobs)
-    }
-}
-
-fn upload_stream(
-    content: std::borrow::Cow<'static, [u8]>,
-) -> (
-    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
-    usize,
-) {
-    use std::borrow::Cow;
-
-    let content = match content {
-        Cow::Borrowed(x) => Bytes::from_static(x),
-        Cow::Owned(vec) => Bytes::from(vec),
-    };
-    wrap_stream(content)
-}
-
-fn wrap_stream(
-    content: bytes::Bytes,
-) -> (
-    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
-    usize,
-) {
-    let len = content.len();
-    let content = futures::future::ready(Ok(content));
-
-    (futures::stream::once(content), len)
-}

libs/utils/src/lsn.rs

@@ -366,6 +366,49 @@ impl MonotonicCounter<Lsn> for RecordLsn {
     }
 }
 
+/// Implements  [`rand::distributions::uniform::UniformSampler`] so we can sample [`Lsn`]s.
+///
+/// This is used by the `pagebench` pageserver benchmarking tool.
+pub struct LsnSampler(<u64 as rand::distributions::uniform::SampleUniform>::Sampler);
+
+impl rand::distributions::uniform::SampleUniform for Lsn {
+    type Sampler = LsnSampler;
+}
+
+impl rand::distributions::uniform::UniformSampler for LsnSampler {
+    type X = Lsn;
+
+    fn new<B1, B2>(low: B1, high: B2) -> Self
+    where
+        B1: rand::distributions::uniform::SampleBorrow<Self::X> + Sized,
+        B2: rand::distributions::uniform::SampleBorrow<Self::X> + Sized,
+    {
+        Self(
+            <u64 as rand::distributions::uniform::SampleUniform>::Sampler::new(
+                low.borrow().0,
+                high.borrow().0,
+            ),
+        )
+    }
+
+    fn new_inclusive<B1, B2>(low: B1, high: B2) -> Self
+    where
+        B1: rand::distributions::uniform::SampleBorrow<Self::X> + Sized,
+        B2: rand::distributions::uniform::SampleBorrow<Self::X> + Sized,
+    {
+        Self(
+            <u64 as rand::distributions::uniform::SampleUniform>::Sampler::new_inclusive(
+                low.borrow().0,
+                high.borrow().0,
+            ),
+        )
+    }
+
+    fn sample<R: rand::prelude::Rng + ?Sized>(&self, rng: &mut R) -> Self::X {
+        Lsn(self.0.sample(rng))
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use crate::bin_ser::BeSer;

libs/utils/src/timeout.rs

@@ -2,8 +2,11 @@ use std::time::Duration;
 
 use tokio_util::sync::CancellationToken;
 
+#[derive(thiserror::Error, Debug)]
 pub enum TimeoutCancellableError {
+    #[error("Timed out")]
     Timeout,
+    #[error("Cancelled")]
     Cancelled,
 }
 

libs/walproposer/bindgen_deps.h

@@ -1 +1,2 @@
+#include "postgres.h"
 #include "walproposer.h"

libs/walproposer/build.rs

@@ -1,3 +1,6 @@
+//! Links with walproposer, pgcommon, pgport and runs bindgen on walproposer.h
+//! to generate Rust bindings for it.
+
 use std::{env, path::PathBuf, process::Command};
 
 use anyhow::{anyhow, Context};

libs/walproposer/src/api_bindings.rs

@@ -1,3 +1,6 @@
+//! A C-Rust shim: defines implementation of C walproposer API, assuming wp
+//! callback_data stores Box to some Rust implementation.
+
 #![allow(dead_code)]
 
 use std::ffi::CStr;
@@ -5,12 +8,12 @@ use std::ffi::CString;
 
 use crate::bindings::uint32;
 use crate::bindings::walproposer_api;
+use crate::bindings::NeonWALReadResult;
 use crate::bindings::PGAsyncReadResult;
 use crate::bindings::PGAsyncWriteResult;
 use crate::bindings::Safekeeper;
 use crate::bindings::Size;
 use crate::bindings::StringInfoData;
-use crate::bindings::TimeLineID;
 use crate::bindings::TimestampTz;
 use crate::bindings::WalProposer;
 use crate::bindings::WalProposerConnStatusType;
@@ -175,31 +178,11 @@ extern "C" fn conn_blocking_write(
     }
 }
 
-extern "C" fn recovery_download(
-    sk: *mut Safekeeper,
-    _timeline: TimeLineID,
-    startpos: XLogRecPtr,
-    endpos: XLogRecPtr,
-) -> bool {
+extern "C" fn recovery_download(wp: *mut WalProposer, sk: *mut Safekeeper) -> bool {
     unsafe {
         let callback_data = (*(*(*sk).wp).config).callback_data;
         let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).recovery_download(&mut (*sk), startpos, endpos)
-    }
-}
-
-#[allow(clippy::unnecessary_cast)]
-extern "C" fn wal_read(
-    sk: *mut Safekeeper,
-    buf: *mut ::std::os::raw::c_char,
-    startptr: XLogRecPtr,
-    count: Size,
-) {
-    unsafe {
-        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, count);
-        let callback_data = (*(*(*sk).wp).config).callback_data;
-        let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).wal_read(&mut (*sk), buf, startptr)
+        (*api).recovery_download(&mut (*wp), &mut (*sk))
     }
 }
 
@@ -211,11 +194,28 @@ extern "C" fn wal_reader_allocate(sk: *mut Safekeeper) {
     }
 }
 
-extern "C" fn free_event_set(wp: *mut WalProposer) {
+#[allow(clippy::unnecessary_cast)]
+extern "C" fn wal_read(
+    sk: *mut Safekeeper,
+    buf: *mut ::std::os::raw::c_char,
+    startptr: XLogRecPtr,
+    count: Size,
+    _errmsg: *mut *mut ::std::os::raw::c_char,
+) -> NeonWALReadResult {
     unsafe {
-        let callback_data = (*(*wp).config).callback_data;
+        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, count);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
         let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).free_event_set(&mut (*wp));
+        // TODO: errmsg is not forwarded
+        (*api).wal_read(&mut (*sk), buf, startptr)
+    }
+}
+
+extern "C" fn wal_reader_events(sk: *mut Safekeeper) -> uint32 {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).wal_reader_events(&mut (*sk))
     }
 }
 
@@ -235,6 +235,14 @@ extern "C" fn update_event_set(sk: *mut Safekeeper, events: uint32) {
     }
 }
 
+extern "C" fn active_state_update_event_set(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).active_state_update_event_set(&mut (*sk));
+    }
+}
+
 extern "C" fn add_safekeeper_event_set(sk: *mut Safekeeper, events: uint32) {
     unsafe {
         let callback_data = (*(*(*sk).wp).config).callback_data;
@@ -243,6 +251,14 @@ extern "C" fn add_safekeeper_event_set(sk: *mut Safekeeper, events: uint32) {
     }
 }
 
+extern "C" fn rm_safekeeper_event_set(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).rm_safekeeper_event_set(&mut (*sk));
+    }
+}
+
 extern "C" fn wait_event_set(
     wp: *mut WalProposer,
     timeout: ::std::os::raw::c_long,
@@ -310,14 +326,6 @@ extern "C" fn process_safekeeper_feedback(wp: *mut WalProposer, commit_lsn: XLog
     }
 }
 
-extern "C" fn confirm_wal_streamed(wp: *mut WalProposer, lsn: XLogRecPtr) {
-    unsafe {
-        let callback_data = (*(*wp).config).callback_data;
-        let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).confirm_wal_streamed(&mut (*wp), lsn)
-    }
-}
-
 extern "C" fn log_internal(
     wp: *mut WalProposer,
     level: ::std::os::raw::c_int,
@@ -332,14 +340,6 @@ extern "C" fn log_internal(
     }
 }
 
-extern "C" fn after_election(wp: *mut WalProposer) {
-    unsafe {
-        let callback_data = (*(*wp).config).callback_data;
-        let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).after_election(&mut (*wp))
-    }
-}
-
 #[derive(Debug)]
 pub enum Level {
     Debug5,
@@ -398,20 +398,20 @@ pub(crate) fn create_api() -> walproposer_api {
         conn_async_write: Some(conn_async_write),
         conn_blocking_write: Some(conn_blocking_write),
         recovery_download: Some(recovery_download),
-        wal_read: Some(wal_read),
         wal_reader_allocate: Some(wal_reader_allocate),
-        free_event_set: Some(free_event_set),
+        wal_read: Some(wal_read),
+        wal_reader_events: Some(wal_reader_events),
         init_event_set: Some(init_event_set),
         update_event_set: Some(update_event_set),
+        active_state_update_event_set: Some(active_state_update_event_set),
         add_safekeeper_event_set: Some(add_safekeeper_event_set),
+        rm_safekeeper_event_set: Some(rm_safekeeper_event_set),
         wait_event_set: Some(wait_event_set),
         strong_random: Some(strong_random),
         get_redo_start_lsn: Some(get_redo_start_lsn),
         finish_sync_safekeepers: Some(finish_sync_safekeepers),
         process_safekeeper_feedback: Some(process_safekeeper_feedback),
-        confirm_wal_streamed: Some(confirm_wal_streamed),
         log_internal: Some(log_internal),
-        after_election: Some(after_election),
     }
 }
 

libs/walproposer/src/walproposer.rs

@@ -6,8 +6,8 @@ use utils::id::TenantTimelineId;
 use crate::{
     api_bindings::{create_api, take_vec_u8, Level},
     bindings::{
-        Safekeeper, WalProposer, WalProposerConfig, WalProposerCreate, WalProposerFree,
-        WalProposerStart,
+        NeonWALReadResult, Safekeeper, WalProposer, WalProposerConfig, WalProposerCreate,
+        WalProposerFree, WalProposerStart,
     },
 };
 
@@ -86,19 +86,19 @@ pub trait ApiImpl {
         todo!()
     }
 
-    fn recovery_download(&self, _sk: &mut Safekeeper, _startpos: u64, _endpos: u64) -> bool {
+    fn recovery_download(&self, _wp: &mut WalProposer, _sk: &mut Safekeeper) -> bool {
         todo!()
     }
 
-    fn wal_read(&self, _sk: &mut Safekeeper, _buf: &mut [u8], _startpos: u64) {
+    fn wal_reader_allocate(&self, _sk: &mut Safekeeper) -> NeonWALReadResult {
         todo!()
     }
 
-    fn wal_reader_allocate(&self, _sk: &mut Safekeeper) {
+    fn wal_read(&self, _sk: &mut Safekeeper, _buf: &mut [u8], _startpos: u64) -> NeonWALReadResult {
         todo!()
     }
 
-    fn free_event_set(&self, _wp: &mut WalProposer) {
+    fn wal_reader_events(&self, _sk: &mut Safekeeper) -> u32 {
         todo!()
     }
 
@@ -110,10 +110,18 @@ pub trait ApiImpl {
         todo!()
     }
 
+    fn active_state_update_event_set(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
     fn add_safekeeper_event_set(&self, _sk: &mut Safekeeper, _events_mask: u32) {
         todo!()
     }
 
+    fn rm_safekeeper_event_set(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
     fn wait_event_set(&self, _wp: &mut WalProposer, _timeout_millis: i64) -> WaitResult {
         todo!()
     }
@@ -134,10 +142,6 @@ pub trait ApiImpl {
         todo!()
     }
 
-    fn confirm_wal_streamed(&self, _wp: &mut WalProposer, _lsn: u64) {
-        todo!()
-    }
-
     fn log_internal(&self, _wp: &mut WalProposer, _level: Level, _msg: &str) {
         todo!()
     }
@@ -240,6 +244,7 @@ impl Drop for Wrapper {
 
 #[cfg(test)]
 mod tests {
+    use core::panic;
     use std::{
         cell::Cell,
         sync::{atomic::AtomicUsize, mpsc::sync_channel},
@@ -247,7 +252,7 @@ mod tests {
 
     use utils::id::TenantTimelineId;
 
-    use crate::{api_bindings::Level, walproposer::Wrapper};
+    use crate::{api_bindings::Level, bindings::NeonWALReadResult, walproposer::Wrapper};
 
     use super::ApiImpl;
 
@@ -355,12 +360,17 @@ mod tests {
             true
         }
 
-        fn wal_reader_allocate(&self, _: &mut crate::bindings::Safekeeper) {
-            println!("wal_reader_allocate")
+        fn recovery_download(
+            &self,
+            _wp: &mut crate::bindings::WalProposer,
+            _sk: &mut crate::bindings::Safekeeper,
+        ) -> bool {
+            true
         }
 
-        fn free_event_set(&self, _: &mut crate::bindings::WalProposer) {
-            println!("free_event_set")
+        fn wal_reader_allocate(&self, _: &mut crate::bindings::Safekeeper) -> NeonWALReadResult {
+            println!("wal_reader_allocate");
+            crate::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
         }
 
         fn init_event_set(&self, _: &mut crate::bindings::WalProposer) {
@@ -383,6 +393,13 @@ mod tests {
             self.wait_events.set(WaitEventsData { sk, event_mask });
         }
 
+        fn rm_safekeeper_event_set(&self, sk: &mut crate::bindings::Safekeeper) {
+            println!(
+                "rm_safekeeper_event_set, sk={:?}",
+                sk as *mut crate::bindings::Safekeeper
+            );
+        }
+
         fn wait_event_set(
             &self,
             _: &mut crate::bindings::WalProposer,

pageserver/Cargo.toml

@@ -63,6 +63,7 @@ thiserror.workspace = true
 tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time"] }
 tokio-io-timeout.workspace = true
 tokio-postgres.workspace = true
+tokio-stream.workspace = true
 tokio-util.workspace = true
 toml_edit = { workspace = true, features = [ "serde" ] }
 tracing.workspace = true

pageserver/client/Cargo.toml

@@ -0,0 +1,22 @@
+[package]
+name = "pageserver_client"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+pageserver_api.workspace = true
+thiserror.workspace = true
+async-trait.workspace = true
+reqwest.workspace = true
+utils.workspace = true
+serde.workspace = true
+workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+tokio-postgres.workspace = true
+tokio-stream.workspace = true
+tokio.workspace = true
+futures.workspace = true
+tokio-util.workspace = true
+anyhow.workspace = true
+postgres.workspace = true
+bytes.workspace = true

pageserver/client/src/lib.rs

@@ -0,0 +1,2 @@
+pub mod mgmt_api;
+pub mod page_service;

pageserver/client/src/mgmt_api.rs

@@ -0,0 +1,202 @@
+use pageserver_api::models::*;
+use reqwest::{IntoUrl, Method};
+use utils::{
+    http::error::HttpErrorBody,
+    id::{TenantId, TimelineId},
+};
+
+pub mod util;
+
+#[derive(Debug)]
+pub struct Client {
+    mgmt_api_endpoint: String,
+    authorization_header: Option<String>,
+    client: reqwest::Client,
+}
+
+#[derive(thiserror::Error, Debug)]
+pub enum Error {
+    #[error("receive body: {0}")]
+    ReceiveBody(reqwest::Error),
+
+    #[error("receive error body: {0}")]
+    ReceiveErrorBody(String),
+
+    #[error("pageserver API: {0}")]
+    ApiError(String),
+}
+
+pub type Result<T> = std::result::Result<T, Error>;
+
+#[async_trait::async_trait]
+pub trait ResponseErrorMessageExt: Sized {
+    async fn error_from_body(self) -> Result<Self>;
+}
+
+#[async_trait::async_trait]
+impl ResponseErrorMessageExt for reqwest::Response {
+    async fn error_from_body(mut self) -> Result<Self> {
+        let status = self.status();
+        if !(status.is_client_error() || status.is_server_error()) {
+            return Ok(self);
+        }
+
+        let url = self.url().to_owned();
+        Err(match self.json::<HttpErrorBody>().await {
+            Ok(HttpErrorBody { msg }) => Error::ApiError(msg),
+            Err(_) => {
+                Error::ReceiveErrorBody(format!("Http error ({}) at {}.", status.as_u16(), url))
+            }
+        })
+    }
+}
+
+impl Client {
+    pub fn new(mgmt_api_endpoint: String, jwt: Option<&str>) -> Self {
+        Self {
+            mgmt_api_endpoint,
+            authorization_header: jwt.map(|jwt| format!("Bearer {jwt}")),
+            client: reqwest::Client::new(),
+        }
+    }
+
+    pub async fn list_tenants(&self) -> Result<Vec<pageserver_api::models::TenantInfo>> {
+        let uri = format!("{}/v1/tenant", self.mgmt_api_endpoint);
+        let resp = self.get(&uri).await?;
+        resp.json().await.map_err(Error::ReceiveBody)
+    }
+
+    pub async fn tenant_details(
+        &self,
+        tenant_id: TenantId,
+    ) -> Result<pageserver_api::models::TenantDetails> {
+        let uri = format!("{}/v1/tenant/{tenant_id}", self.mgmt_api_endpoint);
+        self.get(uri)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
+    pub async fn list_timelines(
+        &self,
+        tenant_id: TenantId,
+    ) -> Result<Vec<pageserver_api::models::TimelineInfo>> {
+        let uri = format!("{}/v1/tenant/{tenant_id}/timeline", self.mgmt_api_endpoint);
+        self.get(&uri)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
+    pub async fn timeline_info(
+        &self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> Result<pageserver_api::models::TimelineInfo> {
+        let uri = format!(
+            "{}/v1/tenant/{tenant_id}/timeline/{timeline_id}",
+            self.mgmt_api_endpoint
+        );
+        self.get(&uri)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
+    pub async fn keyspace(
+        &self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> Result<pageserver_api::models::partitioning::Partitioning> {
+        let uri = format!(
+            "{}/v1/tenant/{tenant_id}/timeline/{timeline_id}/keyspace",
+            self.mgmt_api_endpoint
+        );
+        self.get(&uri)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
+    async fn get<U: IntoUrl>(&self, uri: U) -> Result<reqwest::Response> {
+        self.request(Method::GET, uri, ()).await
+    }
+
+    async fn request<B: serde::Serialize, U: reqwest::IntoUrl>(
+        &self,
+        method: Method,
+        uri: U,
+        body: B,
+    ) -> Result<reqwest::Response> {
+        let req = self.client.request(method, uri);
+        let req = if let Some(value) = &self.authorization_header {
+            req.header(reqwest::header::AUTHORIZATION, value)
+        } else {
+            req
+        };
+        let res = req.json(&body).send().await.map_err(Error::ReceiveBody)?;
+        let response = res.error_from_body().await?;
+        Ok(response)
+    }
+
+    pub async fn status(&self) -> Result<()> {
+        let uri = format!("{}/v1/status", self.mgmt_api_endpoint);
+        self.get(&uri).await?;
+        Ok(())
+    }
+
+    pub async fn tenant_create(&self, req: &TenantCreateRequest) -> Result<TenantId> {
+        let uri = format!("{}/v1/tenant", self.mgmt_api_endpoint);
+        self.request(Method::POST, &uri, req)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
+    pub async fn tenant_config(&self, req: &TenantConfigRequest) -> Result<()> {
+        let uri = format!("{}/v1/tenant/config", self.mgmt_api_endpoint);
+        self.request(Method::PUT, &uri, req).await?;
+        Ok(())
+    }
+
+    pub async fn location_config(
+        &self,
+        tenant_id: TenantId,
+        config: LocationConfig,
+        flush_ms: Option<std::time::Duration>,
+    ) -> Result<()> {
+        let req_body = TenantLocationConfigRequest { tenant_id, config };
+        let path = format!(
+            "{}/v1/tenant/{}/location_config",
+            self.mgmt_api_endpoint, tenant_id
+        );
+        let path = if let Some(flush_ms) = flush_ms {
+            format!("{}?flush_ms={}", path, flush_ms.as_millis())
+        } else {
+            path
+        };
+        self.request(Method::PUT, &path, &req_body).await?;
+        Ok(())
+    }
+
+    pub async fn timeline_create(
+        &self,
+        tenant_id: TenantId,
+        req: &TimelineCreateRequest,
+    ) -> Result<TimelineInfo> {
+        let uri = format!(
+            "{}/v1/tenant/{}/timeline",
+            self.mgmt_api_endpoint, tenant_id
+        );
+        self.request(Method::POST, &uri, req)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+}

pageserver/client/src/mgmt_api/util.rs

@@ -0,0 +1,49 @@
+//! Helpers to do common higher-level tasks with the [`Client`].
+
+use std::sync::Arc;
+
+use tokio::task::JoinSet;
+use utils::id::{TenantId, TenantTimelineId};
+
+use super::Client;
+
+/// Retrieve a list of all of the pageserver's timelines.
+///
+/// Fails if there are sharded tenants present on the pageserver.
+pub async fn get_pageserver_tenant_timelines_unsharded(
+    api_client: &Arc<Client>,
+) -> anyhow::Result<Vec<TenantTimelineId>> {
+    let mut timelines: Vec<TenantTimelineId> = Vec::new();
+    let mut tenants: Vec<TenantId> = Vec::new();
+    for ti in api_client.list_tenants().await? {
+        if !ti.id.is_unsharded() {
+            anyhow::bail!(
+                "only unsharded tenants are supported at this time: {}",
+                ti.id
+            );
+        }
+        tenants.push(ti.id.tenant_id)
+    }
+    let mut js = JoinSet::new();
+    for tenant_id in tenants {
+        js.spawn({
+            let mgmt_api_client = Arc::clone(api_client);
+            async move {
+                (
+                    tenant_id,
+                    mgmt_api_client.tenant_details(tenant_id).await.unwrap(),
+                )
+            }
+        });
+    }
+    while let Some(res) = js.join_next().await {
+        let (tenant_id, details) = res.unwrap();
+        for timeline_id in details.timelines {
+            timelines.push(TenantTimelineId {
+                tenant_id,
+                timeline_id,
+            });
+        }
+    }
+    Ok(timelines)
+}

pageserver/client/src/page_service.rs

@@ -0,0 +1,151 @@
+use std::pin::Pin;
+
+use futures::SinkExt;
+use pageserver_api::{
+    models::{
+        PagestreamBeMessage, PagestreamFeMessage, PagestreamGetPageRequest,
+        PagestreamGetPageResponse,
+    },
+    reltag::RelTag,
+};
+use tokio::task::JoinHandle;
+use tokio_postgres::CopyOutStream;
+use tokio_stream::StreamExt;
+use tokio_util::sync::CancellationToken;
+use utils::{
+    id::{TenantId, TimelineId},
+    lsn::Lsn,
+};
+
+pub struct Client {
+    client: tokio_postgres::Client,
+    cancel_on_client_drop: Option<tokio_util::sync::DropGuard>,
+    conn_task: JoinHandle<()>,
+}
+
+pub struct BasebackupRequest {
+    pub tenant_id: TenantId,
+    pub timeline_id: TimelineId,
+    pub lsn: Option<Lsn>,
+    pub gzip: bool,
+}
+
+impl Client {
+    pub async fn new(connstring: String) -> anyhow::Result<Self> {
+        let (client, connection) = tokio_postgres::connect(&connstring, postgres::NoTls).await?;
+
+        let conn_task_cancel = CancellationToken::new();
+        let conn_task = tokio::spawn({
+            let conn_task_cancel = conn_task_cancel.clone();
+            async move {
+                tokio::select! {
+                    _ = conn_task_cancel.cancelled() => { }
+                    res = connection => {
+                        res.unwrap();
+                    }
+                }
+            }
+        });
+        Ok(Self {
+            cancel_on_client_drop: Some(conn_task_cancel.drop_guard()),
+            conn_task,
+            client,
+        })
+    }
+
+    pub async fn pagestream(
+        self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> anyhow::Result<PagestreamClient> {
+        let copy_both: tokio_postgres::CopyBothDuplex<bytes::Bytes> = self
+            .client
+            .copy_both_simple(&format!("pagestream {tenant_id} {timeline_id}"))
+            .await?;
+        let Client {
+            cancel_on_client_drop,
+            conn_task,
+            client: _,
+        } = self;
+        Ok(PagestreamClient {
+            copy_both: Box::pin(copy_both),
+            conn_task,
+            cancel_on_client_drop,
+        })
+    }
+
+    pub async fn basebackup(&self, req: &BasebackupRequest) -> anyhow::Result<CopyOutStream> {
+        let BasebackupRequest {
+            tenant_id,
+            timeline_id,
+            lsn,
+            gzip,
+        } = req;
+        let mut args = Vec::with_capacity(5);
+        args.push("basebackup".to_string());
+        args.push(format!("{tenant_id}"));
+        args.push(format!("{timeline_id}"));
+        if let Some(lsn) = lsn {
+            args.push(format!("{lsn}"));
+        }
+        if *gzip {
+            args.push("--gzip".to_string())
+        }
+        Ok(self.client.copy_out(&args.join(" ")).await?)
+    }
+}
+
+/// Create using [`Client::pagestream`].
+pub struct PagestreamClient {
+    copy_both: Pin<Box<tokio_postgres::CopyBothDuplex<bytes::Bytes>>>,
+    cancel_on_client_drop: Option<tokio_util::sync::DropGuard>,
+    conn_task: JoinHandle<()>,
+}
+
+pub struct RelTagBlockNo {
+    pub rel_tag: RelTag,
+    pub block_no: u32,
+}
+
+impl PagestreamClient {
+    pub async fn shutdown(mut self) {
+        let _ = self.cancel_on_client_drop.take();
+        self.conn_task.await.unwrap();
+    }
+
+    pub async fn getpage(
+        &mut self,
+        key: RelTagBlockNo,
+        lsn: Lsn,
+    ) -> anyhow::Result<PagestreamGetPageResponse> {
+        let req = PagestreamGetPageRequest {
+            latest: false,
+            rel: key.rel_tag,
+            blkno: key.block_no,
+            lsn,
+        };
+        let req = PagestreamFeMessage::GetPage(req);
+        let req: bytes::Bytes = req.serialize();
+        // let mut req = tokio_util::io::ReaderStream::new(&req);
+        let mut req = tokio_stream::once(Ok(req));
+
+        self.copy_both.send_all(&mut req).await?;
+
+        let next: Option<Result<bytes::Bytes, _>> = self.copy_both.next().await;
+        let next: bytes::Bytes = next.unwrap()?;
+
+        let msg = PagestreamBeMessage::deserialize(next)?;
+        match msg {
+            PagestreamBeMessage::GetPage(p) => Ok(p),
+            PagestreamBeMessage::Error(e) => anyhow::bail!("Error: {:?}", e),
+            PagestreamBeMessage::Exists(_)
+            | PagestreamBeMessage::Nblocks(_)
+            | PagestreamBeMessage::DbSize(_) => {
+                anyhow::bail!(
+                    "unexpected be message kind in response to getpage request: {}",
+                    msg.kind()
+                )
+            }
+        }
+    }
+}

pageserver/pagebench/Cargo.toml

@@ -0,0 +1,26 @@
+[package]
+name = "pagebench"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+anyhow.workspace = true
+clap.workspace = true
+futures.workspace = true
+hdrhistogram.workspace = true
+humantime.workspace = true
+humantime-serde.workspace = true
+rand.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+tracing.workspace = true
+tokio.workspace = true
+
+pageserver = { path = ".." }
+pageserver_client.workspace = true
+pageserver_api.workspace = true
+utils = { path = "../../libs/utils/" }
+workspace_hack = { version = "0.1", path = "../../workspace_hack" }

pageserver/pagebench/src/cmd/basebackup.rs

@@ -0,0 +1,272 @@
+use anyhow::Context;
+use pageserver_client::page_service::BasebackupRequest;
+
+use utils::id::TenantTimelineId;
+use utils::lsn::Lsn;
+
+use rand::prelude::*;
+use tokio::sync::Barrier;
+use tokio::task::JoinSet;
+use tracing::{debug, info, instrument};
+
+use std::collections::HashMap;
+use std::num::NonZeroUsize;
+use std::ops::Range;
+use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::Instant;
+
+use crate::util::tokio_thread_local_stats::AllThreadLocalStats;
+use crate::util::{request_stats, tokio_thread_local_stats};
+
+/// basebackup@LatestLSN
+#[derive(clap::Parser)]
+pub(crate) struct Args {
+    #[clap(long, default_value = "http://localhost:9898")]
+    mgmt_api_endpoint: String,
+    #[clap(long, default_value = "localhost:64000")]
+    page_service_host_port: String,
+    #[clap(long)]
+    pageserver_jwt: Option<String>,
+    #[clap(long, default_value = "1")]
+    num_clients: NonZeroUsize,
+    #[clap(long, default_value = "1.0")]
+    gzip_probability: f64,
+    #[clap(long)]
+    runtime: Option<humantime::Duration>,
+    #[clap(long)]
+    limit_to_first_n_targets: Option<usize>,
+    targets: Option<Vec<TenantTimelineId>>,
+}
+
+#[derive(Debug, Default)]
+struct LiveStats {
+    completed_requests: AtomicU64,
+}
+
+impl LiveStats {
+    fn inc(&self) {
+        self.completed_requests.fetch_add(1, Ordering::Relaxed);
+    }
+}
+
+struct Target {
+    timeline: TenantTimelineId,
+    lsn_range: Option<Range<Lsn>>,
+}
+
+#[derive(serde::Serialize)]
+struct Output {
+    total: request_stats::Output,
+}
+
+tokio_thread_local_stats::declare!(STATS: request_stats::Stats);
+
+pub(crate) fn main(args: Args) -> anyhow::Result<()> {
+    tokio_thread_local_stats::main!(STATS, move |thread_local_stats| {
+        main_impl(args, thread_local_stats)
+    })
+}
+
+async fn main_impl(
+    args: Args,
+    all_thread_local_stats: AllThreadLocalStats<request_stats::Stats>,
+) -> anyhow::Result<()> {
+    let args: &'static Args = Box::leak(Box::new(args));
+
+    let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
+        args.mgmt_api_endpoint.clone(),
+        args.pageserver_jwt.as_deref(),
+    ));
+
+    // discover targets
+    let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
+        &mgmt_api_client,
+        crate::util::cli::targets::Spec {
+            limit_to_first_n_targets: args.limit_to_first_n_targets,
+            targets: args.targets.clone(),
+        },
+    )
+    .await?;
+    let mut js = JoinSet::new();
+    for timeline in &timelines {
+        js.spawn({
+            let timeline = *timeline;
+            // FIXME: this triggers initial logical size calculation
+            // https://github.com/neondatabase/neon/issues/6168
+            let info = mgmt_api_client
+                .timeline_info(timeline.tenant_id, timeline.timeline_id)
+                .await
+                .unwrap();
+            async move {
+                anyhow::Ok(Target {
+                    timeline,
+                    // TODO: support lsn_range != latest LSN
+                    lsn_range: Some(info.last_record_lsn..(info.last_record_lsn + 1)),
+                })
+            }
+        });
+    }
+    let mut all_targets: Vec<Target> = Vec::new();
+    while let Some(res) = js.join_next().await {
+        all_targets.push(res.unwrap().unwrap());
+    }
+
+    let live_stats = Arc::new(LiveStats::default());
+
+    let num_client_tasks = timelines.len();
+    let num_live_stats_dump = 1;
+    let num_work_sender_tasks = 1;
+
+    let start_work_barrier = Arc::new(tokio::sync::Barrier::new(
+        num_client_tasks + num_live_stats_dump + num_work_sender_tasks,
+    ));
+    let all_work_done_barrier = Arc::new(tokio::sync::Barrier::new(num_client_tasks));
+
+    tokio::spawn({
+        let stats = Arc::clone(&live_stats);
+        let start_work_barrier = Arc::clone(&start_work_barrier);
+        async move {
+            start_work_barrier.wait().await;
+            loop {
+                let start = std::time::Instant::now();
+                tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+                let completed_requests = stats.completed_requests.swap(0, Ordering::Relaxed);
+                let elapsed = start.elapsed();
+                info!(
+                    "RPS: {:.0}",
+                    completed_requests as f64 / elapsed.as_secs_f64()
+                );
+            }
+        }
+    });
+
+    let mut work_senders = HashMap::new();
+    let mut tasks = Vec::new();
+    for tl in &timelines {
+        let (sender, receiver) = tokio::sync::mpsc::channel(1); // TODO: not sure what the implications of this are
+        work_senders.insert(tl, sender);
+        tasks.push(tokio::spawn(client(
+            args,
+            *tl,
+            Arc::clone(&start_work_barrier),
+            receiver,
+            Arc::clone(&all_work_done_barrier),
+            Arc::clone(&live_stats),
+        )));
+    }
+
+    let work_sender = async move {
+        start_work_barrier.wait().await;
+        loop {
+            let (timeline, work) = {
+                let mut rng = rand::thread_rng();
+                let target = all_targets.choose(&mut rng).unwrap();
+                let lsn = target.lsn_range.clone().map(|r| rng.gen_range(r));
+                (
+                    target.timeline,
+                    Work {
+                        lsn,
+                        gzip: rng.gen_bool(args.gzip_probability),
+                    },
+                )
+            };
+            let sender = work_senders.get(&timeline).unwrap();
+            // TODO: what if this blocks?
+            sender.send(work).await.ok().unwrap();
+        }
+    };
+
+    if let Some(runtime) = args.runtime {
+        match tokio::time::timeout(runtime.into(), work_sender).await {
+            Ok(()) => unreachable!("work sender never terminates"),
+            Err(_timeout) => {
+                // this implicitly drops the work_senders, making all the clients exit
+            }
+        }
+    } else {
+        work_sender.await;
+        unreachable!("work sender never terminates");
+    }
+
+    for t in tasks {
+        t.await.unwrap();
+    }
+
+    let output = Output {
+        total: {
+            let mut agg_stats = request_stats::Stats::new();
+            for stats in all_thread_local_stats.lock().unwrap().iter() {
+                let stats = stats.lock().unwrap();
+                agg_stats.add(&stats);
+            }
+            agg_stats.output()
+        },
+    };
+
+    let output = serde_json::to_string_pretty(&output).unwrap();
+    println!("{output}");
+
+    anyhow::Ok(())
+}
+
+#[derive(Copy, Clone)]
+struct Work {
+    lsn: Option<Lsn>,
+    gzip: bool,
+}
+
+#[instrument(skip_all)]
+async fn client(
+    args: &'static Args,
+    timeline: TenantTimelineId,
+    start_work_barrier: Arc<Barrier>,
+    mut work: tokio::sync::mpsc::Receiver<Work>,
+    all_work_done_barrier: Arc<Barrier>,
+    live_stats: Arc<LiveStats>,
+) {
+    start_work_barrier.wait().await;
+
+    let client = pageserver_client::page_service::Client::new(crate::util::connstring::connstring(
+        &args.page_service_host_port,
+        args.pageserver_jwt.as_deref(),
+    ))
+    .await
+    .unwrap();
+
+    while let Some(Work { lsn, gzip }) = work.recv().await {
+        let start = Instant::now();
+        let copy_out_stream = client
+            .basebackup(&BasebackupRequest {
+                tenant_id: timeline.tenant_id,
+                timeline_id: timeline.timeline_id,
+                lsn,
+                gzip,
+            })
+            .await
+            .with_context(|| format!("start basebackup for {timeline}"))
+            .unwrap();
+
+        use futures::StreamExt;
+        let size = Arc::new(AtomicUsize::new(0));
+        copy_out_stream
+            .for_each({
+                |r| {
+                    let size = Arc::clone(&size);
+                    async move {
+                        let size = Arc::clone(&size);
+                        size.fetch_add(r.unwrap().len(), Ordering::Relaxed);
+                    }
+                }
+            })
+            .await;
+        debug!("basebackup size is {} bytes", size.load(Ordering::Relaxed));
+        let elapsed = start.elapsed();
+        live_stats.inc();
+        STATS.with(|stats| {
+            stats.borrow().lock().unwrap().observe(elapsed).unwrap();
+        });
+    }
+
+    all_work_done_barrier.wait().await;
+}

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -0,0 +1,335 @@
+use anyhow::Context;
+use futures::future::join_all;
+use pageserver::pgdatadir_mapping::key_to_rel_block;
+use pageserver::repository;
+use pageserver_api::key::is_rel_block_key;
+use pageserver_client::page_service::RelTagBlockNo;
+
+use utils::id::TenantTimelineId;
+use utils::lsn::Lsn;
+
+use rand::prelude::*;
+use tokio::sync::Barrier;
+use tokio::task::JoinSet;
+use tracing::{info, instrument};
+
+use std::collections::HashMap;
+use std::future::Future;
+use std::num::NonZeroUsize;
+use std::pin::Pin;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+
+use crate::util::tokio_thread_local_stats::AllThreadLocalStats;
+use crate::util::{request_stats, tokio_thread_local_stats};
+
+/// GetPage@LatestLSN, uniformly distributed across the compute-accessible keyspace.
+#[derive(clap::Parser)]
+pub(crate) struct Args {
+    #[clap(long, default_value = "http://localhost:9898")]
+    mgmt_api_endpoint: String,
+    #[clap(long, default_value = "postgres://postgres@localhost:64000")]
+    page_service_connstring: String,
+    #[clap(long)]
+    pageserver_jwt: Option<String>,
+    #[clap(long, default_value = "1")]
+    num_clients: NonZeroUsize,
+    #[clap(long)]
+    runtime: Option<humantime::Duration>,
+    #[clap(long)]
+    per_target_rate_limit: Option<usize>,
+    #[clap(long)]
+    limit_to_first_n_targets: Option<usize>,
+    targets: Option<Vec<TenantTimelineId>>,
+}
+
+#[derive(Debug, Default)]
+struct LiveStats {
+    completed_requests: AtomicU64,
+}
+
+impl LiveStats {
+    fn inc(&self) {
+        self.completed_requests.fetch_add(1, Ordering::Relaxed);
+    }
+}
+
+#[derive(Clone)]
+struct KeyRange {
+    timeline: TenantTimelineId,
+    timeline_lsn: Lsn,
+    start: i128,
+    end: i128,
+}
+
+impl KeyRange {
+    fn len(&self) -> i128 {
+        self.end - self.start
+    }
+}
+
+#[derive(serde::Serialize)]
+struct Output {
+    total: request_stats::Output,
+}
+
+tokio_thread_local_stats::declare!(STATS: request_stats::Stats);
+
+pub(crate) fn main(args: Args) -> anyhow::Result<()> {
+    tokio_thread_local_stats::main!(STATS, move |thread_local_stats| {
+        main_impl(args, thread_local_stats)
+    })
+}
+
+async fn main_impl(
+    args: Args,
+    all_thread_local_stats: AllThreadLocalStats<request_stats::Stats>,
+) -> anyhow::Result<()> {
+    let args: &'static Args = Box::leak(Box::new(args));
+
+    let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
+        args.mgmt_api_endpoint.clone(),
+        args.pageserver_jwt.as_deref(),
+    ));
+
+    // discover targets
+    let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
+        &mgmt_api_client,
+        crate::util::cli::targets::Spec {
+            limit_to_first_n_targets: args.limit_to_first_n_targets,
+            targets: args.targets.clone(),
+        },
+    )
+    .await?;
+
+    let mut js = JoinSet::new();
+    for timeline in &timelines {
+        js.spawn({
+            let mgmt_api_client = Arc::clone(&mgmt_api_client);
+            let timeline = *timeline;
+            async move {
+                let partitioning = mgmt_api_client
+                    .keyspace(timeline.tenant_id, timeline.timeline_id)
+                    .await?;
+                let lsn = partitioning.at_lsn;
+
+                let ranges = partitioning
+                    .keys
+                    .ranges
+                    .iter()
+                    .filter_map(|r| {
+                        let start = r.start;
+                        let end = r.end;
+                        // filter out non-relblock keys
+                        match (is_rel_block_key(&start), is_rel_block_key(&end)) {
+                            (true, true) => Some(KeyRange {
+                                timeline,
+                                timeline_lsn: lsn,
+                                start: start.to_i128(),
+                                end: end.to_i128(),
+                            }),
+                            (true, false) | (false, true) => {
+                                unimplemented!("split up range")
+                            }
+                            (false, false) => None,
+                        }
+                    })
+                    .collect::<Vec<_>>();
+
+                anyhow::Ok(ranges)
+            }
+        });
+    }
+    let mut all_ranges: Vec<KeyRange> = Vec::new();
+    while let Some(res) = js.join_next().await {
+        all_ranges.extend(res.unwrap().unwrap());
+    }
+
+    let live_stats = Arc::new(LiveStats::default());
+
+    let num_client_tasks = timelines.len();
+    let num_live_stats_dump = 1;
+    let num_work_sender_tasks = 1;
+
+    let start_work_barrier = Arc::new(tokio::sync::Barrier::new(
+        num_client_tasks + num_live_stats_dump + num_work_sender_tasks,
+    ));
+    let all_work_done_barrier = Arc::new(tokio::sync::Barrier::new(num_client_tasks));
+
+    tokio::spawn({
+        let stats = Arc::clone(&live_stats);
+        let start_work_barrier = Arc::clone(&start_work_barrier);
+        async move {
+            start_work_barrier.wait().await;
+            loop {
+                let start = std::time::Instant::now();
+                tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+                let completed_requests = stats.completed_requests.swap(0, Ordering::Relaxed);
+                let elapsed = start.elapsed();
+                info!(
+                    "RPS: {:.0}",
+                    completed_requests as f64 / elapsed.as_secs_f64()
+                );
+            }
+        }
+    });
+
+    let mut work_senders = HashMap::new();
+    let mut tasks = Vec::new();
+    for tl in &timelines {
+        let (sender, receiver) = tokio::sync::mpsc::channel(10); // TODO: not sure what the implications of this are
+        work_senders.insert(tl, sender);
+        tasks.push(tokio::spawn(client(
+            args,
+            *tl,
+            Arc::clone(&start_work_barrier),
+            receiver,
+            Arc::clone(&all_work_done_barrier),
+            Arc::clone(&live_stats),
+        )));
+    }
+
+    let work_sender: Pin<Box<dyn Send + Future<Output = ()>>> = match args.per_target_rate_limit {
+        None => Box::pin(async move {
+            let weights = rand::distributions::weighted::WeightedIndex::new(
+                all_ranges.iter().map(|v| v.len()),
+            )
+            .unwrap();
+
+            start_work_barrier.wait().await;
+
+            loop {
+                let (range, key) = {
+                    let mut rng = rand::thread_rng();
+                    let r = &all_ranges[weights.sample(&mut rng)];
+                    let key: i128 = rng.gen_range(r.start..r.end);
+                    let key = repository::Key::from_i128(key);
+                    let (rel_tag, block_no) =
+                        key_to_rel_block(key).expect("we filter non-rel-block keys out above");
+                    (r, RelTagBlockNo { rel_tag, block_no })
+                };
+                let sender = work_senders.get(&range.timeline).unwrap();
+                // TODO: what if this blocks?
+                sender.send((key, range.timeline_lsn)).await.ok().unwrap();
+            }
+        }),
+        Some(rps_limit) => Box::pin(async move {
+            let period = Duration::from_secs_f64(1.0 / (rps_limit as f64));
+
+            let make_timeline_task: &dyn Fn(
+                TenantTimelineId,
+            )
+                -> Pin<Box<dyn Send + Future<Output = ()>>> = &|timeline| {
+                let sender = work_senders.get(&timeline).unwrap();
+                let ranges: Vec<KeyRange> = all_ranges
+                    .iter()
+                    .filter(|r| r.timeline == timeline)
+                    .cloned()
+                    .collect();
+                let weights = rand::distributions::weighted::WeightedIndex::new(
+                    ranges.iter().map(|v| v.len()),
+                )
+                .unwrap();
+
+                Box::pin(async move {
+                    let mut ticker = tokio::time::interval(period);
+                    ticker.set_missed_tick_behavior(
+                        /* TODO review this choice */
+                        tokio::time::MissedTickBehavior::Burst,
+                    );
+                    loop {
+                        ticker.tick().await;
+                        let (range, key) = {
+                            let mut rng = rand::thread_rng();
+                            let r = &ranges[weights.sample(&mut rng)];
+                            let key: i128 = rng.gen_range(r.start..r.end);
+                            let key = repository::Key::from_i128(key);
+                            let (rel_tag, block_no) = key_to_rel_block(key)
+                                .expect("we filter non-rel-block keys out above");
+                            (r, RelTagBlockNo { rel_tag, block_no })
+                        };
+                        sender.send((key, range.timeline_lsn)).await.ok().unwrap();
+                    }
+                })
+            };
+
+            let tasks: Vec<_> = work_senders
+                .keys()
+                .map(|tl| make_timeline_task(**tl))
+                .collect();
+
+            start_work_barrier.wait().await;
+
+            join_all(tasks).await;
+        }),
+    };
+
+    if let Some(runtime) = args.runtime {
+        match tokio::time::timeout(runtime.into(), work_sender).await {
+            Ok(()) => unreachable!("work sender never terminates"),
+            Err(_timeout) => {
+                // this implicitly drops the work_senders, making all the clients exit
+            }
+        }
+    } else {
+        work_sender.await;
+        unreachable!("work sender never terminates");
+    }
+
+    for t in tasks {
+        t.await.unwrap();
+    }
+
+    let output = Output {
+        total: {
+            let mut agg_stats = request_stats::Stats::new();
+            for stats in all_thread_local_stats.lock().unwrap().iter() {
+                let stats = stats.lock().unwrap();
+                agg_stats.add(&stats);
+            }
+            agg_stats.output()
+        },
+    };
+
+    let output = serde_json::to_string_pretty(&output).unwrap();
+    println!("{output}");
+
+    anyhow::Ok(())
+}
+
+#[instrument(skip_all)]
+async fn client(
+    args: &'static Args,
+    timeline: TenantTimelineId,
+    start_work_barrier: Arc<Barrier>,
+    mut work: tokio::sync::mpsc::Receiver<(RelTagBlockNo, Lsn)>,
+    all_work_done_barrier: Arc<Barrier>,
+    live_stats: Arc<LiveStats>,
+) {
+    start_work_barrier.wait().await;
+
+    let client = pageserver_client::page_service::Client::new(args.page_service_connstring.clone())
+        .await
+        .unwrap();
+    let mut client = client
+        .pagestream(timeline.tenant_id, timeline.timeline_id)
+        .await
+        .unwrap();
+
+    while let Some((key, lsn)) = work.recv().await {
+        let start = Instant::now();
+        client
+            .getpage(key, lsn)
+            .await
+            .with_context(|| format!("getpage for {timeline}"))
+            .unwrap();
+        let elapsed = start.elapsed();
+        live_stats.inc();
+        STATS.with(|stats| {
+            stats.borrow().lock().unwrap().observe(elapsed).unwrap();
+        });
+    }
+
+    all_work_done_barrier.wait().await;
+}

pageserver/pagebench/src/cmd/trigger_initial_size_calculation.rs

@@ -0,0 +1,85 @@
+use std::sync::Arc;
+
+use humantime::Duration;
+use tokio::task::JoinSet;
+use utils::id::TenantTimelineId;
+
+#[derive(clap::Parser)]
+pub(crate) struct Args {
+    #[clap(long, default_value = "http://localhost:9898")]
+    mgmt_api_endpoint: String,
+    #[clap(long, default_value = "localhost:64000")]
+    page_service_host_port: String,
+    #[clap(long)]
+    pageserver_jwt: Option<String>,
+    #[clap(
+        long,
+        help = "if specified, poll mgmt api to check whether init logical size calculation has completed"
+    )]
+    poll_for_completion: Option<Duration>,
+    #[clap(long)]
+    limit_to_first_n_targets: Option<usize>,
+    targets: Option<Vec<TenantTimelineId>>,
+}
+
+pub(crate) fn main(args: Args) -> anyhow::Result<()> {
+    let rt = tokio::runtime::Builder::new_multi_thread()
+        .enable_all()
+        .build()
+        .unwrap();
+
+    let main_task = rt.spawn(main_impl(args));
+    rt.block_on(main_task).unwrap()
+}
+
+async fn main_impl(args: Args) -> anyhow::Result<()> {
+    let args: &'static Args = Box::leak(Box::new(args));
+
+    let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
+        args.mgmt_api_endpoint.clone(),
+        args.pageserver_jwt.as_deref(),
+    ));
+
+    // discover targets
+    let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
+        &mgmt_api_client,
+        crate::util::cli::targets::Spec {
+            limit_to_first_n_targets: args.limit_to_first_n_targets,
+            targets: args.targets.clone(),
+        },
+    )
+    .await?;
+
+    // kick it off
+
+    let mut js = JoinSet::new();
+    for tl in timelines {
+        let mgmt_api_client = Arc::clone(&mgmt_api_client);
+        js.spawn(async move {
+            // TODO: API to explicitly trigger initial logical size computation.
+            // Should probably also avoid making it a side effect of timeline details to trigger initial logical size calculation.
+            // => https://github.com/neondatabase/neon/issues/6168
+            let info = mgmt_api_client
+                .timeline_info(tl.tenant_id, tl.timeline_id)
+                .await
+                .unwrap();
+
+            if let Some(period) = args.poll_for_completion {
+                let mut ticker = tokio::time::interval(period.into());
+                ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
+                let mut info = info;
+                while !info.current_logical_size_is_accurate {
+                    ticker.tick().await;
+                    info = mgmt_api_client
+                        .timeline_info(tl.tenant_id, tl.timeline_id)
+                        .await
+                        .unwrap();
+                }
+            }
+        });
+    }
+    while let Some(res) = js.join_next().await {
+        let _: () = res.unwrap();
+    }
+    Ok(())
+}

pageserver/pagebench/src/main.rs

@@ -0,0 +1,48 @@
+use clap::Parser;
+use utils::logging;
+
+/// Re-usable pieces of code that aren't CLI-specific.
+mod util {
+    pub(crate) mod connstring;
+    pub(crate) mod request_stats;
+    #[macro_use]
+    pub(crate) mod tokio_thread_local_stats;
+    /// Re-usable pieces of CLI-specific code.
+    pub(crate) mod cli {
+        pub(crate) mod targets;
+    }
+}
+
+/// The pagebench CLI sub-commands, dispatched in [`main`] below.
+mod cmd {
+    pub(super) mod basebackup;
+    pub(super) mod getpage_latest_lsn;
+    pub(super) mod trigger_initial_size_calculation;
+}
+
+/// Component-level performance test for pageserver.
+#[derive(clap::Parser)]
+enum Args {
+    Basebackup(cmd::basebackup::Args),
+    GetPageLatestLsn(cmd::getpage_latest_lsn::Args),
+    TriggerInitialSizeCalculation(cmd::trigger_initial_size_calculation::Args),
+}
+
+fn main() {
+    logging::init(
+        logging::LogFormat::Plain,
+        logging::TracingErrorLayerEnablement::Disabled,
+        logging::Output::Stderr,
+    )
+    .unwrap();
+
+    let args = Args::parse();
+    match args {
+        Args::Basebackup(args) => cmd::basebackup::main(args),
+        Args::GetPageLatestLsn(args) => cmd::getpage_latest_lsn::main(args),
+        Args::TriggerInitialSizeCalculation(args) => {
+            cmd::trigger_initial_size_calculation::main(args)
+        }
+    }
+    .unwrap()
+}

pageserver/pagebench/src/util/cli/targets.rs

@@ -0,0 +1,34 @@
+use std::sync::Arc;
+
+use pageserver_client::mgmt_api;
+use tracing::info;
+use utils::id::TenantTimelineId;
+
+pub(crate) struct Spec {
+    pub(crate) limit_to_first_n_targets: Option<usize>,
+    pub(crate) targets: Option<Vec<TenantTimelineId>>,
+}
+
+pub(crate) async fn discover(
+    api_client: &Arc<mgmt_api::Client>,
+    spec: Spec,
+) -> anyhow::Result<Vec<TenantTimelineId>> {
+    let mut timelines = if let Some(targets) = spec.targets {
+        targets
+    } else {
+        mgmt_api::util::get_pageserver_tenant_timelines_unsharded(api_client).await?
+    };
+
+    if let Some(limit) = spec.limit_to_first_n_targets {
+        timelines.sort(); // for determinism
+        timelines.truncate(limit);
+        if timelines.len() < limit {
+            anyhow::bail!("pageserver has less than limit_to_first_n_targets={limit} tenants");
+        }
+    }
+
+    info!("timelines:\n{:?}", timelines);
+    info!("number of timelines:\n{:?}", timelines.len());
+
+    Ok(timelines)
+}

pageserver/pagebench/src/util/connstring.rs

@@ -0,0 +1,8 @@
+pub(crate) fn connstring(host_port: &str, jwt: Option<&str>) -> String {
+    let colon_and_jwt = if let Some(jwt) = jwt {
+        format!(":{jwt}") // TODO: urlescape
+    } else {
+        String::new()
+    };
+    format!("postgres://postgres{colon_and_jwt}@{host_port}")
+}

pageserver/pagebench/src/util/request_stats.rs

@@ -0,0 +1,88 @@
+use std::time::Duration;
+
+use anyhow::Context;
+
+pub(crate) struct Stats {
+    latency_histo: hdrhistogram::Histogram<u64>,
+}
+
+impl Stats {
+    pub(crate) fn new() -> Self {
+        Self {
+            // Initialize with fixed bounds so that we panic at runtime instead of resizing the histogram,
+            // which would skew the benchmark results.
+            latency_histo: hdrhistogram::Histogram::new_with_bounds(1, 1_000_000_000, 3).unwrap(),
+        }
+    }
+    pub(crate) fn observe(&mut self, latency: Duration) -> anyhow::Result<()> {
+        let micros: u64 = latency
+            .as_micros()
+            .try_into()
+            .context("latency greater than u64")?;
+        self.latency_histo
+            .record(micros)
+            .context("add to histogram")?;
+        Ok(())
+    }
+    pub(crate) fn output(&self) -> Output {
+        let latency_percentiles = std::array::from_fn(|idx| {
+            let micros = self
+                .latency_histo
+                .value_at_percentile(LATENCY_PERCENTILES[idx]);
+            Duration::from_micros(micros)
+        });
+        Output {
+            request_count: self.latency_histo.len(),
+            latency_mean: Duration::from_micros(self.latency_histo.mean() as u64),
+            latency_percentiles: LatencyPercentiles {
+                latency_percentiles,
+            },
+        }
+    }
+    pub(crate) fn add(&mut self, other: &Self) {
+        let Self {
+            ref mut latency_histo,
+        } = self;
+        latency_histo.add(&other.latency_histo).unwrap();
+    }
+}
+
+impl Default for Stats {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+const LATENCY_PERCENTILES: [f64; 4] = [95.0, 99.00, 99.90, 99.99];
+
+struct LatencyPercentiles {
+    latency_percentiles: [Duration; 4],
+}
+
+impl serde::Serialize for LatencyPercentiles {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        use serde::ser::SerializeMap;
+        let mut ser = serializer.serialize_map(Some(LATENCY_PERCENTILES.len()))?;
+        for p in LATENCY_PERCENTILES {
+            ser.serialize_entry(
+                &format!("p{p}"),
+                &format!(
+                    "{}",
+                    &humantime::format_duration(self.latency_percentiles[0])
+                ),
+            )?;
+        }
+        ser.end()
+    }
+}
+
+#[derive(serde::Serialize)]
+pub(crate) struct Output {
+    request_count: u64,
+    #[serde(with = "humantime_serde")]
+    latency_mean: Duration,
+    latency_percentiles: LatencyPercentiles,
+}

pageserver/pagebench/src/util/tokio_thread_local_stats.rs

@@ -0,0 +1,45 @@
+pub(crate) type ThreadLocalStats<T> = Arc<Mutex<T>>;
+pub(crate) type AllThreadLocalStats<T> = Arc<Mutex<Vec<ThreadLocalStats<T>>>>;
+
+macro_rules! declare {
+    ($THREAD_LOCAL_NAME:ident: $T:ty) => {
+        thread_local! {
+            pub static $THREAD_LOCAL_NAME: std::cell::RefCell<crate::util::tokio_thread_local_stats::ThreadLocalStats<$T>> = std::cell::RefCell::new(
+                std::sync::Arc::new(std::sync::Mutex::new(Default::default()))
+            );
+        }
+    };
+}
+
+use std::sync::{Arc, Mutex};
+
+pub(crate) use declare;
+
+macro_rules! main {
+    ($THREAD_LOCAL_NAME:ident, $main_impl:expr) => {{
+        let main_impl = $main_impl;
+        let all = Arc::new(Mutex::new(Vec::new()));
+
+        let rt = tokio::runtime::Builder::new_multi_thread()
+            .on_thread_start({
+                let all = Arc::clone(&all);
+                move || {
+                    // pre-initialize the thread local stats by accessesing them
+                    // (some stats like requests_stats::Stats are quite costly to initialize,
+                    //  we don't want to pay that cost during the measurement period)
+                    $THREAD_LOCAL_NAME.with(|stats| {
+                        let stats: Arc<_> = Arc::clone(&*stats.borrow());
+                        all.lock().unwrap().push(stats);
+                    });
+                }
+            })
+            .enable_all()
+            .build()
+            .unwrap();
+
+        let main_task = rt.spawn(main_impl(all));
+        rt.block_on(main_task).unwrap()
+    }};
+}
+
+pub(crate) use main;

pageserver/src/config.rs

@@ -41,6 +41,8 @@ use crate::{
     TIMELINE_DELETE_MARK_SUFFIX, TIMELINE_UNINIT_MARK_SUFFIX,
 };
 
+use self::defaults::DEFAULT_CONCURRENT_TENANT_WARMUP;
+
 pub mod defaults {
     use crate::tenant::config::defaults::*;
     use const_format::formatcp;
@@ -61,6 +63,8 @@ pub mod defaults {
 
     pub const DEFAULT_LOG_FORMAT: &str = "plain";
 
+    pub const DEFAULT_CONCURRENT_TENANT_WARMUP: usize = 8;
+
     pub const DEFAULT_CONCURRENT_TENANT_SIZE_LOGICAL_SIZE_QUERIES: usize =
         super::ConfigurableSemaphore::DEFAULT_INITIAL.get();
 
@@ -94,6 +98,7 @@ pub mod defaults {
 #log_format = '{DEFAULT_LOG_FORMAT}'
 
 #concurrent_tenant_size_logical_size_queries = '{DEFAULT_CONCURRENT_TENANT_SIZE_LOGICAL_SIZE_QUERIES}'
+#concurrent_tenant_warmup = '{DEFAULT_CONCURRENT_TENANT_WARMUP}'
 
 #metric_collection_interval = '{DEFAULT_METRIC_COLLECTION_INTERVAL}'
 #cached_metric_collection_interval = '{DEFAULT_CACHED_METRIC_COLLECTION_INTERVAL}'
@@ -180,6 +185,11 @@ pub struct PageServerConf {
 
     pub log_format: LogFormat,
 
+    /// Number of tenants which will be concurrently loaded from remote storage proactively on startup,
+    /// does not limit tenants loaded in response to client I/O.  A lower value implicitly deprioritizes
+    /// loading such tenants, vs. other work in the system.
+    pub concurrent_tenant_warmup: ConfigurableSemaphore,
+
     /// Number of concurrent [`Tenant::gather_size_inputs`](crate::tenant::Tenant::gather_size_inputs) allowed.
     pub concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore,
     /// Limit of concurrent [`Tenant::gather_size_inputs`] issued by module `eviction_task`.
@@ -283,6 +293,7 @@ struct PageServerConfigBuilder {
 
     log_format: BuilderValue<LogFormat>,
 
+    concurrent_tenant_warmup: BuilderValue<NonZeroUsize>,
     concurrent_tenant_size_logical_size_queries: BuilderValue<NonZeroUsize>,
 
     metric_collection_interval: BuilderValue<Duration>,
@@ -340,6 +351,8 @@ impl Default for PageServerConfigBuilder {
             .expect("cannot parse default keepalive interval")),
             log_format: Set(LogFormat::from_str(DEFAULT_LOG_FORMAT).unwrap()),
 
+            concurrent_tenant_warmup: Set(NonZeroUsize::new(DEFAULT_CONCURRENT_TENANT_WARMUP)
+                .expect("Invalid default constant")),
             concurrent_tenant_size_logical_size_queries: Set(
                 ConfigurableSemaphore::DEFAULT_INITIAL,
             ),
@@ -453,6 +466,10 @@ impl PageServerConfigBuilder {
         self.log_format = BuilderValue::Set(log_format)
     }
 
+    pub fn concurrent_tenant_warmup(&mut self, u: NonZeroUsize) {
+        self.concurrent_tenant_warmup = BuilderValue::Set(u);
+    }
+
     pub fn concurrent_tenant_size_logical_size_queries(&mut self, u: NonZeroUsize) {
         self.concurrent_tenant_size_logical_size_queries = BuilderValue::Set(u);
     }
@@ -518,6 +535,9 @@ impl PageServerConfigBuilder {
     }
 
     pub fn build(self) -> anyhow::Result<PageServerConf> {
+        let concurrent_tenant_warmup = self
+            .concurrent_tenant_warmup
+            .ok_or(anyhow!("missing concurrent_tenant_warmup"))?;
         let concurrent_tenant_size_logical_size_queries = self
             .concurrent_tenant_size_logical_size_queries
             .ok_or(anyhow!(
@@ -570,6 +590,7 @@ impl PageServerConfigBuilder {
                 .broker_keepalive_interval
                 .ok_or(anyhow!("No broker keepalive interval provided"))?,
             log_format: self.log_format.ok_or(anyhow!("missing log_format"))?,
+            concurrent_tenant_warmup: ConfigurableSemaphore::new(concurrent_tenant_warmup),
             concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::new(
                 concurrent_tenant_size_logical_size_queries,
             ),
@@ -807,6 +828,11 @@ impl PageServerConf {
                 "log_format" => builder.log_format(
                     LogFormat::from_config(&parse_toml_string(key, item)?)?
                 ),
+                "concurrent_tenant_warmup" => builder.concurrent_tenant_warmup({
+                    let input = parse_toml_string(key, item)?;
+                    let permits = input.parse::<usize>().context("expected a number of initial permits, not {s:?}")?;
+                    NonZeroUsize::new(permits).context("initial semaphore permits out of range: 0, use other configuration to disable a feature")?
+                }),
                 "concurrent_tenant_size_logical_size_queries" => builder.concurrent_tenant_size_logical_size_queries({
                     let input = parse_toml_string(key, item)?;
                     let permits = input.parse::<usize>().context("expected a number of initial permits, not {s:?}")?;
@@ -904,6 +930,10 @@ impl PageServerConf {
             broker_endpoint: storage_broker::DEFAULT_ENDPOINT.parse().unwrap(),
             broker_keepalive_interval: Duration::from_secs(5000),
             log_format: LogFormat::from_str(defaults::DEFAULT_LOG_FORMAT).unwrap(),
+            concurrent_tenant_warmup: ConfigurableSemaphore::new(
+                NonZeroUsize::new(DEFAULT_CONCURRENT_TENANT_WARMUP)
+                    .expect("Invalid default constant"),
+            ),
             concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::default(),
             eviction_task_immitated_concurrent_logical_size_queries: ConfigurableSemaphore::default(
             ),
@@ -1122,6 +1152,9 @@ background_task_maximum_delay = '334 s'
                     storage_broker::DEFAULT_KEEPALIVE_INTERVAL
                 )?,
                 log_format: LogFormat::from_str(defaults::DEFAULT_LOG_FORMAT).unwrap(),
+                concurrent_tenant_warmup: ConfigurableSemaphore::new(
+                    NonZeroUsize::new(DEFAULT_CONCURRENT_TENANT_WARMUP).unwrap()
+                ),
                 concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::default(),
                 eviction_task_immitated_concurrent_logical_size_queries:
                     ConfigurableSemaphore::default(),
@@ -1188,6 +1221,9 @@ background_task_maximum_delay = '334 s'
                 broker_endpoint: storage_broker::DEFAULT_ENDPOINT.parse().unwrap(),
                 broker_keepalive_interval: Duration::from_secs(5),
                 log_format: LogFormat::Json,
+                concurrent_tenant_warmup: ConfigurableSemaphore::new(
+                    NonZeroUsize::new(DEFAULT_CONCURRENT_TENANT_WARMUP).unwrap()
+                ),
                 concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::default(),
                 eviction_task_immitated_concurrent_logical_size_queries:
                     ConfigurableSemaphore::default(),
@@ -1432,6 +1468,7 @@ threshold = "20m"
                 period: Duration::from_secs(10),
                 #[cfg(feature = "testing")]
                 mock_statvfs: None,
+                eviction_order: crate::disk_usage_eviction_task::EvictionOrder::AbsoluteAccessed,
             })
         );
         match &conf.default_tenant_conf.eviction_policy {

pageserver/src/disk_usage_eviction_task.rs

@@ -74,6 +74,45 @@ pub struct DiskUsageEvictionTaskConfig {
     pub period: Duration,
     #[cfg(feature = "testing")]
     pub mock_statvfs: Option<crate::statvfs::mock::Behavior>,
+    /// Select sorting for evicted layers
+    #[serde(default)]
+    pub eviction_order: EvictionOrder,
+}
+
+/// Selects the sort order for eviction candidates *after* per tenant `min_resident_size`
+/// partitioning.
+#[derive(Default, Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "type", content = "args")]
+pub enum EvictionOrder {
+    /// Order the layers to be evicted by how recently they have been accessed in absolute
+    /// time.
+    ///
+    /// This strategy is unfair when some tenants grow faster than others towards the slower
+    /// growing.
+    #[default]
+    AbsoluteAccessed,
+
+    /// Order the layers to be evicted by how recently they have been accessed relatively within
+    /// the set of resident layers of a tenant.
+    ///
+    /// This strategy will evict layers more fairly but is untested.
+    RelativeAccessed {
+        #[serde(default)]
+        highest_layer_count_loses_first: bool,
+    },
+}
+
+impl EvictionOrder {
+    /// Return true, if with [`Self::RelativeAccessed`] order the tenants with the highest layer
+    /// counts should be the first ones to have their layers evicted.
+    fn highest_layer_count_loses_first(&self) -> bool {
+        match self {
+            EvictionOrder::AbsoluteAccessed => false,
+            EvictionOrder::RelativeAccessed {
+                highest_layer_count_loses_first,
+            } => *highest_layer_count_loses_first,
+        }
+    }
 }
 
 #[derive(Default)]
@@ -192,7 +231,14 @@ async fn disk_usage_eviction_task_iteration(
 ) -> anyhow::Result<()> {
     let usage_pre = filesystem_level_usage::get(tenants_dir, task_config)
         .context("get filesystem-level disk usage before evictions")?;
-    let res = disk_usage_eviction_task_iteration_impl(state, storage, usage_pre, cancel).await;
+    let res = disk_usage_eviction_task_iteration_impl(
+        state,
+        storage,
+        usage_pre,
+        task_config.eviction_order,
+        cancel,
+    )
+    .await;
     match res {
         Ok(outcome) => {
             debug!(?outcome, "disk_usage_eviction_iteration finished");
@@ -278,6 +324,7 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     state: &State,
     _storage: &GenericRemoteStorage,
     usage_pre: U,
+    eviction_order: EvictionOrder,
     cancel: &CancellationToken,
 ) -> anyhow::Result<IterationOutcome<U>> {
     // use tokio's mutex to get a Sync guard (instead of std::sync::Mutex)
@@ -297,7 +344,7 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
         "running disk usage based eviction due to pressure"
     );
 
-    let candidates = match collect_eviction_candidates(cancel).await? {
+    let candidates = match collect_eviction_candidates(eviction_order, cancel).await? {
         EvictionCandidates::Cancelled => {
             return Ok(IterationOutcome::Cancelled);
         }
@@ -307,16 +354,16 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     // Debug-log the list of candidates
     let now = SystemTime::now();
     for (i, (partition, candidate)) in candidates.iter().enumerate() {
+        let nth = i + 1;
         let desc = candidate.layer.layer_desc();
+        let total_candidates = candidates.len();
+        let size = desc.file_size;
+        let rel = candidate.relative_last_activity;
         debug!(
-            "cand {}/{}: size={}, no_access_for={}us, partition={:?}, {}/{}/{}",
-            i + 1,
-            candidates.len(),
-            desc.file_size,
+            "cand {nth}/{total_candidates}: size={size}, rel_last_activity={rel}, no_access_for={}us, partition={partition:?}, {}/{}/{}",
             now.duration_since(candidate.last_activity_ts)
                 .unwrap()
                 .as_micros(),
-            partition,
             desc.tenant_shard_id,
             desc.timeline_id,
             candidate.layer,
@@ -459,6 +506,7 @@ struct EvictionCandidate {
     timeline: Arc<Timeline>,
     layer: Layer,
     last_activity_ts: SystemTime,
+    relative_last_activity: finite_f32::FiniteF32,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
@@ -478,24 +526,24 @@ enum EvictionCandidates {
 /// order. A caller that evicts in that order, until pressure is relieved, implements
 /// the eviction policy outlined in the module comment.
 ///
-/// # Example
+/// # Example with EvictionOrder::AbsoluteAccessed
 ///
 /// Imagine that there are two tenants, A and B, with five layers each, a-e.
 /// Each layer has size 100, and both tenant's min_resident_size is 150.
 /// The eviction order would be
 ///
 /// ```text
-/// partition last_activity_ts    tenant/layer
-/// Above     18:30               A/c
-/// Above     19:00               A/b
-/// Above     18:29               B/c
-/// Above     19:05               B/b
-/// Above     20:00               B/a
-/// Above     20:03               A/a
-/// Below     20:30               A/d
-/// Below     20:40               B/d
-/// Below     20:45               B/e
-/// Below     20:58               A/e
+/// partition last_activity_ts tenant/layer
+/// Above     18:30            A/c
+/// Above     19:00            A/b
+/// Above     18:29            B/c
+/// Above     19:05            B/b
+/// Above     20:00            B/a
+/// Above     20:03            A/a
+/// Below     20:30            A/d
+/// Below     20:40            B/d
+/// Below     20:45            B/e
+/// Below     20:58            A/e
 /// ```
 ///
 /// Now, if we need to evict 300 bytes to relieve pressure, we'd evict `A/c, A/b, B/c`.
@@ -505,7 +553,77 @@ enum EvictionCandidates {
 /// `A/c, A/b, B/c, B/b, B/a, A/a, A/d, B/d, B/e`, reaching into the `Below` partition
 /// after exhauting the `Above` partition.
 /// So, we did not respect each tenant's min_resident_size.
+///
+/// # Example with EvictionOrder::RelativeAccessed
+///
+/// ```text
+/// partition relative_age last_activity_ts tenant/layer
+/// Above     0/4          18:30            A/c
+/// Above     0/4          18:29            B/c
+/// Above     1/4          19:00            A/b
+/// Above     1/4          19:05            B/b
+/// Above     2/4          20:00            B/a
+/// Above     2/4          20:03            A/a
+/// Below     3/4          20:30            A/d
+/// Below     3/4          20:40            B/d
+/// Below     4/4          20:45            B/e
+/// Below     4/4          20:58            A/e
+/// ```
+///
+/// With tenants having the same number of layers the picture does not change much. The same with
+/// A having many more layers **resident** (not all of them listed):
+///
+/// ```text
+/// Above       0/100      18:30            A/c
+/// Above       0/4        18:29            B/c
+/// Above       1/100      19:00            A/b
+/// Above       2/100      20:03            A/a
+/// Above       3/100      20:03            A/nth_3
+/// Above       4/100      20:03            A/nth_4
+///             ...
+/// Above       1/4        19:05            B/b
+/// Above      25/100      20:04            A/nth_25
+///             ...
+/// Above       2/4        20:00            B/a
+/// Above      50/100      20:10            A/nth_50
+///             ...
+/// Below       3/4        20:40            B/d
+/// Below      99/100      20:30            A/nth_99
+/// Below       4/4        20:45            B/e
+/// Below     100/100      20:58            A/nth_100
+/// ```
+///
+/// Now it's easier to see that because A has grown fast it has more layers to get evicted. What is
+/// difficult to see is what happens on the next round assuming the evicting 23 from the above list
+/// relieves the pressure (22 A layers gone, 1 B layers gone) but a new fast growing tenant C has
+/// appeared:
+///
+/// ```text
+/// Above       0/87       20:04            A/nth_23
+/// Above       0/3        19:05            B/b
+/// Above       0/50       20:59            C/nth_0
+/// Above       1/87       20:04            A/nth_24
+/// Above       1/50       21:00            C/nth_1
+/// Above       2/87       20:04            A/nth_25
+///             ...
+/// Above      16/50       21:02            C/nth_16
+/// Above       1/3        20:00            B/a
+/// Above      27/87       20:10            A/nth_50
+///             ...
+/// Below       2/3        20:40            B/d
+/// Below      49/50       21:05            C/nth_49
+/// Below      86/87       20:30            A/nth_99
+/// Below       3/3        20:45            B/e
+/// Below      50/50       21:05            C/nth_50
+/// Below      87/87       20:58            A/nth_100
+/// ```
+///
+/// Now relieving pressure with 23 layers would cost:
+/// - tenant A 14 layers
+/// - tenant B 1 layer
+/// - tenant C 8 layers
 async fn collect_eviction_candidates(
+    eviction_order: EvictionOrder,
     cancel: &CancellationToken,
 ) -> anyhow::Result<EvictionCandidates> {
     // get a snapshot of the list of tenants
@@ -591,12 +709,63 @@ async fn collect_eviction_candidates(
         tenant_candidates
             .sort_unstable_by_key(|(_, layer_info)| std::cmp::Reverse(layer_info.last_activity_ts));
         let mut cumsum: i128 = 0;
-        for (timeline, layer_info) in tenant_candidates.into_iter() {
+
+        // keeping the -1 or not decides if every tenant should lose their least recently accessed
+        // layer OR if this should happen in the order of having highest layer count:
+        let fudge = if eviction_order.highest_layer_count_loses_first() {
+            // relative_age vs. tenant layer count:
+            // - 0.1..=1.0 (10 layers)
+            // - 0.01..=1.0 (100 layers)
+            // - 0.001..=1.0 (1000 layers)
+            //
+            // leading to evicting less of the smallest tenants.
+            0
+        } else {
+            // use full 0.0..=1.0 range, which means even the smallest tenants could always lose a
+            // layer. the actual ordering is unspecified: for 10k tenants on a pageserver it could
+            // be that less than 10k layer evictions is enough, so we would not need to evict from
+            // all tenants.
+            //
+            // as the tenant ordering is now deterministic this could hit the same tenants
+            // disproportionetly on multiple invocations. alternative could be to remember how many
+            // layers did we evict last time from this tenant, and inject that as an additional
+            // fudge here.
+            1
+        };
+
+        let total = tenant_candidates
+            .len()
+            .checked_sub(fudge)
+            .filter(|&x| x > 0)
+            // support 0 or 1 resident layer tenants as well
+            .unwrap_or(1);
+        let divider = total as f32;
+
+        for (i, (timeline, layer_info)) in tenant_candidates.into_iter().enumerate() {
             let file_size = layer_info.file_size();
+
+            // as we iterate this reverse sorted list, the most recently accessed layer will always
+            // be 1.0; this is for us to evict it last.
+            let relative_last_activity = if matches!(
+                eviction_order,
+                EvictionOrder::RelativeAccessed { .. }
+            ) {
+                // another possibility: use buckets, like (256.0 * relative_last_activity) as u8 or
+                // similarly for u16. unsure how it would help.
+                finite_f32::FiniteF32::try_from_normalized((total - i) as f32 / divider)
+                    .unwrap_or_else(|val| {
+                        tracing::warn!(%fudge, "calculated invalid relative_last_activity for i={i}, total={total}: {val}");
+                        finite_f32::FiniteF32::ZERO
+                    })
+            } else {
+                finite_f32::FiniteF32::ZERO
+            };
+
             let candidate = EvictionCandidate {
                 timeline,
                 last_activity_ts: layer_info.last_activity_ts,
                 layer: layer_info.layer,
+                relative_last_activity,
             };
             let partition = if cumsum > min_resident_size as i128 {
                 MinResidentSizePartition::Above
@@ -610,8 +779,19 @@ async fn collect_eviction_candidates(
 
     debug_assert!(MinResidentSizePartition::Above < MinResidentSizePartition::Below,
         "as explained in the function's doc comment, layers that aren't in the tenant's min_resident_size are evicted first");
-    candidates
-        .sort_unstable_by_key(|(partition, candidate)| (*partition, candidate.last_activity_ts));
+
+    match eviction_order {
+        EvictionOrder::AbsoluteAccessed => {
+            candidates.sort_unstable_by_key(|(partition, candidate)| {
+                (*partition, candidate.last_activity_ts)
+            });
+        }
+        EvictionOrder::RelativeAccessed { .. } => {
+            candidates.sort_unstable_by_key(|(partition, candidate)| {
+                (*partition, candidate.relative_last_activity)
+            });
+        }
+    }
 
     Ok(EvictionCandidates::Finished(candidates))
 }
@@ -640,6 +820,66 @@ impl std::ops::Deref for TimelineKey {
     }
 }
 
+/// A totally ordered f32 subset we can use with sorting functions.
+mod finite_f32 {
+
+    /// A totally ordered f32 subset we can use with sorting functions.
+    #[derive(Clone, Copy, PartialEq)]
+    pub struct FiniteF32(f32);
+
+    impl std::fmt::Debug for FiniteF32 {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            std::fmt::Debug::fmt(&self.0, f)
+        }
+    }
+
+    impl std::fmt::Display for FiniteF32 {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            std::fmt::Display::fmt(&self.0, f)
+        }
+    }
+
+    impl std::cmp::Eq for FiniteF32 {}
+
+    impl std::cmp::PartialOrd for FiniteF32 {
+        fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
+            Some(self.cmp(other))
+        }
+    }
+
+    impl std::cmp::Ord for FiniteF32 {
+        fn cmp(&self, other: &Self) -> std::cmp::Ordering {
+            self.0.total_cmp(&other.0)
+        }
+    }
+
+    impl TryFrom<f32> for FiniteF32 {
+        type Error = f32;
+
+        fn try_from(value: f32) -> Result<Self, Self::Error> {
+            if value.is_finite() {
+                Ok(FiniteF32(value))
+            } else {
+                Err(value)
+            }
+        }
+    }
+
+    impl FiniteF32 {
+        pub const ZERO: FiniteF32 = FiniteF32(0.0);
+
+        pub fn try_from_normalized(value: f32) -> Result<Self, f32> {
+            if (0.0..=1.0).contains(&value) {
+                // -0.0 is within the range, make sure it is assumed 0.0..=1.0
+                let value = value.abs();
+                Ok(FiniteF32(value))
+            } else {
+                Err(value)
+            }
+        }
+    }
+}
+
 mod filesystem_level_usage {
     use anyhow::Context;
     use camino::Utf8Path;
@@ -721,6 +961,7 @@ mod filesystem_level_usage {
 
     #[test]
     fn max_usage_pct_pressure() {
+        use super::EvictionOrder;
         use super::Usage as _;
         use std::time::Duration;
         use utils::serde_percent::Percent;
@@ -732,6 +973,7 @@ mod filesystem_level_usage {
                 period: Duration::MAX,
                 #[cfg(feature = "testing")]
                 mock_statvfs: None,
+                eviction_order: EvictionOrder::default(),
             },
             total_bytes: 100_000,
             avail_bytes: 0,

pageserver/src/http/mod.rs

@@ -1,4 +1,2 @@
 pub mod routes;
 pub use routes::make_router;
-
-pub use pageserver_api::models;

pageserver/src/http/openapi_spec.yml

@@ -159,6 +159,12 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/ConflictError"
+        "412":
+          description: Deletion may not proceed, tenant is not in Active state
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/PreconditionFailedError"
         "500":
           description: Generic operation error
           content:

pageserver/src/http/routes.rs

@@ -14,6 +14,7 @@ use hyper::header;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use metrics::launch_timestamp::LaunchTimestamp;
+use pageserver_api::models::TenantDetails;
 use pageserver_api::models::{
     DownloadRemoteLayersTaskSpawnRequest, LocationConfigMode, TenantAttachRequest,
     TenantLoadRequest, TenantLocationConfigRequest,
@@ -28,16 +29,13 @@ use utils::http::endpoint::request_span;
 use utils::http::json::json_request_or_empty_body;
 use utils::http::request::{get_request_param, must_get_query_param, parse_query_param};
 
-use super::models::{
-    StatusResponse, TenantConfigRequest, TenantCreateRequest, TenantCreateResponse, TenantInfo,
-    TimelineCreateRequest, TimelineGcRequest, TimelineInfo,
-};
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::deletion_queue::DeletionQueueClient;
 use crate::metrics::{StorageTimeOperation, STORAGE_TIME_GLOBAL};
 use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::task_mgr::TaskKind;
 use crate::tenant::config::{LocationConf, TenantConfOpt};
+use crate::tenant::mgr::GetActiveTenantError;
 use crate::tenant::mgr::{
     GetTenantError, SetNewTenantConfigError, TenantManager, TenantMapError, TenantMapInsertError,
     TenantSlotError, TenantSlotUpsertError, TenantStateError,
@@ -50,6 +48,10 @@ use crate::tenant::timeline::Timeline;
 use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError, TenantSharedResources};
 use crate::{config::PageServerConf, tenant::mgr};
 use crate::{disk_usage_eviction_task, tenant};
+use pageserver_api::models::{
+    StatusResponse, TenantConfigRequest, TenantCreateRequest, TenantCreateResponse, TenantInfo,
+    TimelineCreateRequest, TimelineGcRequest, TimelineInfo,
+};
 use utils::{
     auth::SwappableJwtAuth,
     generation::Generation,
@@ -65,7 +67,12 @@ use utils::{
 };
 
 // Imports only used for testing APIs
-use super::models::ConfigureFailpointsRequest;
+use pageserver_api::models::ConfigureFailpointsRequest;
+
+// For APIs that require an Active tenant, how long should we block waiting for that state?
+// This is not functionally necessary (clients will retry), but avoids generating a lot of
+// failed API calls while tenants are activating.
+const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(5000);
 
 pub struct State {
     conf: &'static PageServerConf,
@@ -233,6 +240,19 @@ impl From<GetTenantError> for ApiError {
     }
 }
 
+impl From<GetActiveTenantError> for ApiError {
+    fn from(e: GetActiveTenantError) -> ApiError {
+        match e {
+            GetActiveTenantError::WillNotBecomeActive(_) => ApiError::Conflict(format!("{}", e)),
+            GetActiveTenantError::Cancelled => ApiError::ShuttingDown,
+            GetActiveTenantError::NotFound(gte) => gte.into(),
+            GetActiveTenantError::WaitForActiveTimeout { .. } => {
+                ApiError::ResourceUnavailable(format!("{}", e).into())
+            }
+        }
+    }
+}
+
 impl From<SetNewTenantConfigError> for ApiError {
     fn from(e: SetNewTenantConfigError) -> ApiError {
         match e {
@@ -288,6 +308,7 @@ impl From<crate::tenant::delete::DeleteTenantError> for ApiError {
             SlotUpsertError(e) => e.into(),
             Other(o) => ApiError::InternalServerError(o),
             e @ InvalidState(_) => ApiError::PreconditionFailed(e.to_string().into_boxed_str()),
+            Cancelled => ApiError::ShuttingDown,
         }
     }
 }
@@ -435,7 +456,10 @@ async fn timeline_create_handler(
     let state = get_state(&request);
 
     async {
-        let tenant = state.tenant_manager.get_attached_tenant_shard(tenant_shard_id, true)?;
+        let tenant = state.tenant_manager.get_attached_tenant_shard(tenant_shard_id, false)?;
+
+        tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+
         match tenant.create_timeline(
             new_timeline_id,
             request_data.ancestor_timeline_id.map(TimelineId::from),
@@ -570,8 +594,6 @@ async fn get_lsn_by_timestamp_handler(
         )));
     }
 
-    let version: Option<u8> = parse_query_param(&request, "version")?;
-
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     let timestamp_raw = must_get_query_param(&request, "timestamp")?;
     let timestamp = humantime::parse_rfc3339(&timestamp_raw)
@@ -584,31 +606,18 @@ async fn get_lsn_by_timestamp_handler(
     let result = timeline
         .find_lsn_for_timestamp(timestamp_pg, &cancel, &ctx)
         .await?;
-
-    if version.unwrap_or(0) > 1 {
-        #[derive(serde::Serialize)]
-        struct Result {
-            lsn: Lsn,
-            kind: &'static str,
-        }
-        let (lsn, kind) = match result {
-            LsnForTimestamp::Present(lsn) => (lsn, "present"),
-            LsnForTimestamp::Future(lsn) => (lsn, "future"),
-            LsnForTimestamp::Past(lsn) => (lsn, "past"),
-            LsnForTimestamp::NoData(lsn) => (lsn, "nodata"),
-        };
-        json_response(StatusCode::OK, Result { lsn, kind })
-    } else {
-        // FIXME: this is a temporary crutch not to break backwards compatibility
-        // See https://github.com/neondatabase/neon/pull/5608
-        let result = match result {
-            LsnForTimestamp::Present(lsn) => format!("{lsn}"),
-            LsnForTimestamp::Future(_lsn) => "future".into(),
-            LsnForTimestamp::Past(_lsn) => "past".into(),
-            LsnForTimestamp::NoData(_lsn) => "nodata".into(),
-        };
-        json_response(StatusCode::OK, result)
+    #[derive(serde::Serialize)]
+    struct Result {
+        lsn: Lsn,
+        kind: &'static str,
     }
+    let (lsn, kind) = match result {
+        LsnForTimestamp::Present(lsn) => (lsn, "present"),
+        LsnForTimestamp::Future(lsn) => (lsn, "future"),
+        LsnForTimestamp::Past(lsn) => (lsn, "past"),
+        LsnForTimestamp::NoData(lsn) => (lsn, "nodata"),
+    };
+    json_response(StatusCode::OK, Result { lsn, kind })
 }
 
 async fn get_timestamp_of_lsn_handler(
@@ -694,11 +703,23 @@ async fn timeline_delete_handler(
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
-    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
     let state = get_state(&request);
 
-    state.tenant_manager.delete_timeline(tenant_shard_id, timeline_id, &ctx)
-        .instrument(info_span!("timeline_delete", tenant_id=%tenant_shard_id.tenant_id, shard=%tenant_shard_id.shard_slug(), %timeline_id))
+    let tenant = state
+        .tenant_manager
+        .get_attached_tenant_shard(tenant_shard_id, false)
+        .map_err(|e| {
+            match e {
+                // GetTenantError has a built-in conversion to ApiError, but in this context we don't
+                // want to treat missing tenants as 404, to avoid ambiguity with successful deletions.
+                GetTenantError::NotFound(_) => ApiError::PreconditionFailed(
+                    "Requested tenant is missing".to_string().into_boxed_str(),
+                ),
+                e => e.into(),
+            }
+        })?;
+    tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+    tenant.delete_timeline(timeline_id).instrument(info_span!("timeline_delete", tenant_id=%tenant_shard_id.tenant_id, shard=%tenant_shard_id.shard_slug(), %timeline_id))
         .await?;
 
     json_response(StatusCode::ACCEPTED, ())
@@ -838,11 +859,14 @@ async fn tenant_status(
         }
 
         let state = tenant.current_state();
-        Result::<_, ApiError>::Ok(TenantInfo {
-            id: tenant_shard_id,
-            state: state.clone(),
-            current_physical_size: Some(current_physical_size),
-            attachment_status: state.attachment_status(),
+        Result::<_, ApiError>::Ok(TenantDetails {
+            tenant_info: TenantInfo {
+                id: tenant_shard_id,
+                state: state.clone(),
+                current_physical_size: Some(current_physical_size),
+                attachment_status: state.attachment_status(),
+            },
+            timelines: tenant.list_timeline_ids(),
         })
     }
     .instrument(info_span!("tenant_status_handler",
@@ -863,7 +887,9 @@ async fn tenant_delete_handler(
 
     let state = get_state(&request);
 
-    mgr::delete_tenant(state.conf, state.remote_storage.clone(), tenant_shard_id)
+    state
+        .tenant_manager
+        .delete_tenant(tenant_shard_id, ACTIVE_TENANT_TIMEOUT)
         .instrument(info_span!("tenant_delete_handler",
             tenant_id = %tenant_shard_id.tenant_id,
             shard = %tenant_shard_id.shard_slug()
@@ -1136,7 +1162,10 @@ async fn tenant_create_handler(
 
     // We created the tenant. Existing API semantics are that the tenant
     // is Active when this function returns.
-    if let res @ Err(_) = new_tenant.wait_to_become_active().await {
+    if let res @ Err(_) = new_tenant
+        .wait_to_become_active(ACTIVE_TENANT_TIMEOUT)
+        .await
+    {
         // This shouldn't happen because we just created the tenant directory
         // in tenant::mgr::create_tenant, and there aren't any remote timelines
         // to load, so, nothing can really fail during load.
@@ -1487,69 +1516,6 @@ async fn timeline_collect_keyspace(
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
-    struct Partitioning {
-        keys: crate::keyspace::KeySpace,
-
-        at_lsn: Lsn,
-    }
-
-    impl serde::Serialize for Partitioning {
-        fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
-        where
-            S: serde::Serializer,
-        {
-            use serde::ser::SerializeMap;
-            let mut map = serializer.serialize_map(Some(2))?;
-            map.serialize_key("keys")?;
-            map.serialize_value(&KeySpace(&self.keys))?;
-            map.serialize_key("at_lsn")?;
-            map.serialize_value(&WithDisplay(&self.at_lsn))?;
-            map.end()
-        }
-    }
-
-    struct WithDisplay<'a, T>(&'a T);
-
-    impl<'a, T: std::fmt::Display> serde::Serialize for WithDisplay<'a, T> {
-        fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
-        where
-            S: serde::Serializer,
-        {
-            serializer.collect_str(&self.0)
-        }
-    }
-
-    struct KeySpace<'a>(&'a crate::keyspace::KeySpace);
-
-    impl<'a> serde::Serialize for KeySpace<'a> {
-        fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
-        where
-            S: serde::Serializer,
-        {
-            use serde::ser::SerializeSeq;
-            let mut seq = serializer.serialize_seq(Some(self.0.ranges.len()))?;
-            for kr in &self.0.ranges {
-                seq.serialize_element(&KeyRange(kr))?;
-            }
-            seq.end()
-        }
-    }
-
-    struct KeyRange<'a>(&'a std::ops::Range<crate::repository::Key>);
-
-    impl<'a> serde::Serialize for KeyRange<'a> {
-        fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
-        where
-            S: serde::Serializer,
-        {
-            use serde::ser::SerializeTuple;
-            let mut t = serializer.serialize_tuple(2)?;
-            t.serialize_element(&WithDisplay(&self.0.start))?;
-            t.serialize_element(&WithDisplay(&self.0.end))?;
-            t.end()
-        }
-    }
-
     let at_lsn: Option<Lsn> = parse_query_param(&request, "at_lsn")?;
 
     async {
@@ -1561,7 +1527,9 @@ async fn timeline_collect_keyspace(
             .await
             .map_err(|e| ApiError::InternalServerError(e.into()))?;
 
-        json_response(StatusCode::OK, Partitioning { keys, at_lsn })
+        let res = pageserver_api::models::partitioning::Partitioning { keys, at_lsn };
+
+        json_response(StatusCode::OK, res)
     }
     .instrument(info_span!("timeline_collect_keyspace", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
     .await
@@ -1601,19 +1569,22 @@ async fn disk_usage_eviction_run(
     struct Config {
         /// How many bytes to evict before reporting that pressure is relieved.
         evict_bytes: u64,
+
+        #[serde(default)]
+        eviction_order: crate::disk_usage_eviction_task::EvictionOrder,
     }
 
     #[derive(Debug, Clone, Copy, serde::Serialize)]
     struct Usage {
         // remains unchanged after instantiation of the struct
-        config: Config,
+        evict_bytes: u64,
         // updated by `add_available_bytes`
         freed_bytes: u64,
     }
 
     impl crate::disk_usage_eviction_task::Usage for Usage {
         fn has_pressure(&self) -> bool {
-            self.config.evict_bytes > self.freed_bytes
+            self.evict_bytes > self.freed_bytes
         }
 
         fn add_available_bytes(&mut self, bytes: u64) {
@@ -1624,7 +1595,7 @@ async fn disk_usage_eviction_run(
     let config = json_request::<Config>(&mut r).await?;
 
     let usage = Usage {
-        config,
+        evict_bytes: config.evict_bytes,
         freed_bytes: 0,
     };
 
@@ -1639,7 +1610,11 @@ async fn disk_usage_eviction_run(
     let state = state.disk_usage_eviction_state.clone();
 
     let res = crate::disk_usage_eviction_task::disk_usage_eviction_task_iteration_impl(
-        &state, storage, usage, &cancel,
+        &state,
+        storage,
+        usage,
+        config.eviction_order,
+        &cancel,
     )
     .await;
 

pageserver/src/lib.rs

@@ -10,7 +10,7 @@ pub mod deletion_queue;
 pub mod disk_usage_eviction_task;
 pub mod http;
 pub mod import_datadir;
-pub mod keyspace;
+pub use pageserver_api::keyspace;
 pub mod metrics;
 pub mod page_cache;
 pub mod page_service;

pageserver/src/metrics.rs

@@ -522,14 +522,18 @@ pub(crate) mod initial_logical_size {
     impl StartCalculation {
         pub(crate) fn first(&self, circumstances: StartCircumstances) -> OngoingCalculationGuard {
             let circumstances_label: &'static str = circumstances.into();
-            self.0.with_label_values(&["first", circumstances_label]);
+            self.0
+                .with_label_values(&["first", circumstances_label])
+                .inc();
             OngoingCalculationGuard {
                 inc_drop_calculation: Some(DROP_CALCULATION.first.clone()),
             }
         }
         pub(crate) fn retry(&self, circumstances: StartCircumstances) -> OngoingCalculationGuard {
             let circumstances_label: &'static str = circumstances.into();
-            self.0.with_label_values(&["retry", circumstances_label]);
+            self.0
+                .with_label_values(&["retry", circumstances_label])
+                .inc();
             OngoingCalculationGuard {
                 inc_drop_calculation: Some(DROP_CALCULATION.retry.clone()),
             }
@@ -684,14 +688,54 @@ pub static STARTUP_IS_LOADING: Lazy<UIntGauge> = Lazy::new(|| {
     .expect("Failed to register pageserver_startup_is_loading")
 });
 
-/// How long did tenants take to go from construction to active state?
-pub(crate) static TENANT_ACTIVATION: Lazy<Histogram> = Lazy::new(|| {
-    register_histogram!(
+/// Metrics related to the lifecycle of a [`crate::tenant::Tenant`] object: things
+/// like how long it took to load.
+///
+/// Note that these are process-global metrics, _not_ per-tenant metrics.  Per-tenant
+/// metrics are rather expensive, and usually fine grained stuff makes more sense
+/// at a timeline level than tenant level.
+pub(crate) struct TenantMetrics {
+    /// How long did tenants take to go from construction to active state?
+    pub(crate) activation: Histogram,
+    pub(crate) preload: Histogram,
+    pub(crate) attach: Histogram,
+
+    /// How many tenants are included in the initial startup of the pagesrever?
+    pub(crate) startup_scheduled: IntCounter,
+    pub(crate) startup_complete: IntCounter,
+}
+
+pub(crate) static TENANT: Lazy<TenantMetrics> = Lazy::new(|| {
+    TenantMetrics {
+    activation: register_histogram!(
         "pageserver_tenant_activation_seconds",
         "Time taken by tenants to activate, in seconds",
         CRITICAL_OP_BUCKETS.into()
     )
-    .expect("Failed to register pageserver_tenant_activation_seconds metric")
+    .expect("Failed to register metric"),
+    preload: register_histogram!(
+        "pageserver_tenant_preload_seconds",
+        "Time taken by tenants to load remote metadata on startup/attach, in seconds",
+        CRITICAL_OP_BUCKETS.into()
+    )
+    .expect("Failed to register metric"),
+    attach: register_histogram!(
+        "pageserver_tenant_attach_seconds",
+        "Time taken by tenants to intialize, after remote metadata is already loaded",
+        CRITICAL_OP_BUCKETS.into()
+    )
+    .expect("Failed to register metric"),
+    startup_scheduled: register_int_counter!(
+        "pageserver_tenant_startup_scheduled",
+        "Number of tenants included in pageserver startup (doesn't count tenants attached later)"
+    ).expect("Failed to register metric"),
+    startup_complete: register_int_counter!(
+        "pageserver_tenant_startup_complete",
+        "Number of tenants that have completed warm-up, or activated on-demand during initial startup: \
+         should eventually reach `pageserver_tenant_startup_scheduled_total`.  Does not include broken \
+         tenants: such cases will lead to this metric never reaching the scheduled count."
+    ).expect("Failed to register metric"),
+}
 });
 
 /// Each `Timeline`'s  [`EVICTIONS_WITH_LOW_RESIDENCE_DURATION`] metric.
@@ -979,12 +1023,62 @@ static SMGR_QUERY_TIME_PER_TENANT_TIMELINE: Lazy<HistogramVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+static SMGR_QUERY_TIME_GLOBAL_BUCKETS: Lazy<Vec<f64>> = Lazy::new(|| {
+    [
+        1,
+        10,
+        20,
+        40,
+        60,
+        80,
+        100,
+        200,
+        300,
+        400,
+        500,
+        600,
+        700,
+        800,
+        900,
+        1_000, // 1ms
+        2_000,
+        4_000,
+        6_000,
+        8_000,
+        10_000, // 10ms
+        20_000,
+        40_000,
+        60_000,
+        80_000,
+        100_000,
+        200_000,
+        400_000,
+        600_000,
+        800_000,
+        1_000_000, // 1s
+        2_000_000,
+        4_000_000,
+        6_000_000,
+        8_000_000,
+        10_000_000, // 10s
+        20_000_000,
+        50_000_000,
+        100_000_000,
+        200_000_000,
+        1_000_000_000, // 1000s
+    ]
+    .into_iter()
+    .map(Duration::from_micros)
+    .map(|d| d.as_secs_f64())
+    .collect()
+});
+
 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(),
+        SMGR_QUERY_TIME_GLOBAL_BUCKETS.clone(),
     )
     .expect("failed to define a metric")
 });
@@ -2213,6 +2307,9 @@ pub fn preinitialize_metrics() {
     // Deletion queue stats
     Lazy::force(&DELETION_QUEUE);
 
+    // Tenant stats
+    Lazy::force(&TENANT);
+
     // Tenant manager stats
     Lazy::force(&TENANT_MANAGER);
 

pageserver/src/pgdatadir_mapping.rs

@@ -1776,6 +1776,7 @@ pub fn is_inherited_key(key: Key) -> bool {
     key != AUX_FILES_KEY
 }
 
+/// Guaranteed to return `Ok()` if [[is_rel_block_key]] returns `true` for `key`.
 pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
     Ok(match key.field1 {
         0x00 => (
@@ -1790,7 +1791,6 @@ pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
         _ => anyhow::bail!("unexpected value kind 0x{:02x}", key.field1),
     })
 }
-
 pub fn is_rel_fsm_block_key(key: Key) -> bool {
     key.field1 == 0x00 && key.field4 != 0 && key.field5 == FSM_FORKNUM && key.field6 != 0xffffffff
 }

pageserver/src/repository.rs

@@ -2,38 +2,11 @@ use crate::walrecord::NeonWalRecord;
 use anyhow::Result;
 use bytes::Bytes;
 use serde::{Deserialize, Serialize};
-use std::ops::{AddAssign, Range};
+use std::ops::AddAssign;
 use std::time::Duration;
 
 pub use pageserver_api::key::{Key, KEY_SIZE};
 
-pub fn key_range_size(key_range: &Range<Key>) -> u32 {
-    let start = key_range.start;
-    let end = key_range.end;
-
-    if end.field1 != start.field1
-        || end.field2 != start.field2
-        || end.field3 != start.field3
-        || end.field4 != start.field4
-    {
-        return u32::MAX;
-    }
-
-    let start = (start.field5 as u64) << 32 | start.field6 as u64;
-    let end = (end.field5 as u64) << 32 | end.field6 as u64;
-
-    let diff = end - start;
-    if diff > u32::MAX as u64 {
-        u32::MAX
-    } else {
-        diff as u32
-    }
-}
-
-pub fn singleton_range(key: Key) -> Range<Key> {
-    key..key.next()
-}
-
 /// A 'value' stored for a one Key.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[cfg_attr(test, derive(PartialEq))]

pageserver/src/tenant.rs

@@ -36,6 +36,8 @@ use utils::crashsafe::path_with_suffix_extension;
 use utils::fs_ext;
 use utils::sync::gate::Gate;
 use utils::sync::gate::GateGuard;
+use utils::timeout::timeout_cancellable;
+use utils::timeout::TimeoutCancellableError;
 
 use self::config::AttachedLocationConfig;
 use self::config::AttachmentMode;
@@ -59,7 +61,7 @@ use crate::deletion_queue::DeletionQueueClient;
 use crate::deletion_queue::DeletionQueueError;
 use crate::import_datadir;
 use crate::is_uninit_mark;
-use crate::metrics::TENANT_ACTIVATION;
+use crate::metrics::TENANT;
 use crate::metrics::{remove_tenant_metrics, TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC};
 use crate::repository::GcResult;
 use crate::task_mgr;
@@ -226,7 +228,7 @@ pub struct Tenant {
 
     /// The value creation timestamp, used to measure activation delay, see:
     /// <https://github.com/neondatabase/neon/issues/4025>
-    loading_started_at: Instant,
+    constructed_at: Instant,
 
     state: watch::Sender<TenantState>,
 
@@ -276,6 +278,11 @@ pub struct Tenant {
 
     eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
 
+    /// If the tenant is in Activating state, notify this to encourage it
+    /// to proceed to Active as soon as possible, rather than waiting for lazy
+    /// background warmup.
+    pub(crate) activate_now_sem: tokio::sync::Semaphore,
+
     pub(crate) delete_progress: Arc<tokio::sync::Mutex<DeleteTenantFlow>>,
 
     // Cancellation token fires when we have entered shutdown().  This is a parent of
@@ -622,6 +629,14 @@ impl Tenant {
             "attach tenant",
             false,
             async move {
+                // Is this tenant being spawned as part of process startup?
+                let starting_up = init_order.is_some();
+                scopeguard::defer! {
+                    if starting_up {
+                        TENANT.startup_complete.inc();
+                    }
+                }
+
                 // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
                 let make_broken =
                     |t: &Tenant, err: anyhow::Error| {
@@ -648,8 +663,62 @@ impl Tenant {
                     .as_mut()
                     .and_then(|x| x.initial_tenant_load_remote.take());
 
+                enum AttachType<'a> {
+                    // During pageserver startup, we are attaching this tenant lazily in the background
+                    Warmup(tokio::sync::SemaphorePermit<'a>),
+                    // During pageserver startup, we are attaching this tenant as soon as we can,
+                    // because a client tried to access it.
+                    OnDemand,
+                    // During normal operations after startup, we are attaching a tenant.
+                    Normal,
+                }
+
+                // Before doing any I/O, wait for either or:
+                // - A client to attempt to access to this tenant (on-demand loading)
+                // - A permit to become available in the warmup semaphore (background warmup)
+                //
+                // Some-ness of init_order is how we know if we're attaching during startup or later
+                // in process lifetime.
+                let attach_type = if init_order.is_some() {
+                    tokio::select!(
+                        _ = tenant_clone.activate_now_sem.acquire() => {
+                            tracing::info!("Activating tenant (on-demand)");
+                            AttachType::OnDemand
+                        },
+                        permit_result = conf.concurrent_tenant_warmup.inner().acquire() => {
+                            match permit_result {
+                                Ok(p) => {
+                                    tracing::info!("Activating tenant (warmup)");
+                                    AttachType::Warmup(p)
+                                }
+                                Err(_) => {
+                                    // This is unexpected: the warmup semaphore should stay alive
+                                    // for the lifetime of init_order.  Log a warning and proceed.
+                                    tracing::warn!("warmup_limit semaphore unexpectedly closed");
+                                    AttachType::Normal
+                                }
+                            }
+
+                        }
+                        _ = tenant_clone.cancel.cancelled() => {
+                            // This is safe, but should be pretty rare: it is interesting if a tenant
+                            // stayed in Activating for such a long time that shutdown found it in
+                            // that state.
+                            tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
+                            return Ok(());
+                        },
+                    )
+                } else {
+                    AttachType::Normal
+                };
+
+                let preload_timer = TENANT.preload.start_timer();
                 let preload = match mode {
-                    SpawnMode::Create => {None},
+                    SpawnMode::Create => {
+                        // Don't count the skipped preload into the histogram of preload durations
+                        preload_timer.stop_and_discard();
+                        None
+                    },
                     SpawnMode::Normal => {
                         match &remote_storage {
                             Some(remote_storage) => Some(
@@ -659,7 +728,11 @@ impl Tenant {
                                         tracing::info_span!(parent: None, "attach_preload", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()),
                                     )
                                     .await {
-                                        Ok(p) => p,
+                                        Ok(p) => {
+                                            preload_timer.observe_duration();
+                                            p
+                                        }
+                                            ,
                                         Err(e) => {
                                             make_broken(&tenant_clone, anyhow::anyhow!(e));
                                                 return Ok(());
@@ -721,15 +794,43 @@ impl Tenant {
                     }
                 }
 
+                // We will time the duration of the attach phase unless this is a creation (attach will do no work)
+                let attach_timer = match mode {
+                    SpawnMode::Create => None,
+                    SpawnMode::Normal => {Some(TENANT.attach.start_timer())}
+                };
                 match tenant_clone.attach(preload, &ctx).await {
                     Ok(()) => {
                         info!("attach finished, activating");
+                        if let Some(t)=  attach_timer {t.observe_duration();}
                         tenant_clone.activate(broker_client, None, &ctx);
                     }
                     Err(e) => {
+                        if let Some(t)=  attach_timer {t.observe_duration();}
                         make_broken(&tenant_clone, anyhow::anyhow!(e));
                     }
                 }
+
+                // If we are doing an opportunistic warmup attachment at startup, initialize
+                // logical size at the same time.  This is better than starting a bunch of idle tenants
+                // with cold caches and then coming back later to initialize their logical sizes.
+                //
+                // It also prevents the warmup proccess competing with the concurrency limit on
+                // logical size calculations: if logical size calculation semaphore is saturated,
+                // then warmup will wait for that before proceeding to the next tenant.
+                if let AttachType::Warmup(_permit) = attach_type {
+                    let mut futs = FuturesUnordered::new();
+                    let timelines: Vec<_> = tenant_clone.timelines.lock().unwrap().values().cloned().collect();
+                    for t in timelines {
+                        futs.push(t.await_initial_logical_size())
+                    }
+                    tracing::info!("Waiting for initial logical sizes while warming up...");
+                    while futs.next().await.is_some() {
+
+                    }
+                    tracing::info!("Warm-up complete");
+                }
+
                 Ok(())
             }
             .instrument({
@@ -1451,6 +1552,10 @@ impl Tenant {
             .collect()
     }
 
+    pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
+        self.timelines.lock().unwrap().keys().cloned().collect()
+    }
+
     /// This is used to create the initial 'main' timeline during bootstrapping,
     /// or when importing a new base backup. The caller is expected to load an
     /// initial image of the datadir to the new timeline after this.
@@ -1696,6 +1801,15 @@ impl Tenant {
         Ok(loaded_timeline)
     }
 
+    pub(crate) async fn delete_timeline(
+        self: Arc<Self>,
+        timeline_id: TimelineId,
+    ) -> Result<(), DeleteTimelineError> {
+        DeleteTimelineFlow::run(&self, timeline_id, false).await?;
+
+        Ok(())
+    }
+
     /// perform one garbage collection iteration, removing old data files from disk.
     /// this function is periodically called by gc task.
     /// also it can be explicitly requested through page server api 'do_gc' command.
@@ -1857,7 +1971,7 @@ impl Tenant {
                 );
                 *current_state = TenantState::Active;
 
-                let elapsed = self.loading_started_at.elapsed();
+                let elapsed = self.constructed_at.elapsed();
                 let total_timelines = timelines_accessor.len();
 
                 // log a lot of stuff, because some tenants sometimes suffer from user-visible
@@ -1872,7 +1986,7 @@ impl Tenant {
                     "activation attempt finished"
                 );
 
-                TENANT_ACTIVATION.observe(elapsed.as_secs_f64());
+                TENANT.activation.observe(elapsed.as_secs_f64());
             });
         }
     }
@@ -2127,18 +2241,41 @@ impl Tenant {
         self.state.subscribe()
     }
 
-    pub(crate) async fn wait_to_become_active(&self) -> Result<(), GetActiveTenantError> {
+    /// The activate_now semaphore is initialized with zero units.  As soon as
+    /// we add a unit, waiters will be able to acquire a unit and proceed.
+    pub(crate) fn activate_now(&self) {
+        self.activate_now_sem.add_permits(1);
+    }
+
+    pub(crate) async fn wait_to_become_active(
+        &self,
+        timeout: Duration,
+    ) -> Result<(), GetActiveTenantError> {
         let mut receiver = self.state.subscribe();
         loop {
             let current_state = receiver.borrow_and_update().clone();
             match current_state {
                 TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
                     // in these states, there's a chance that we can reach ::Active
-                    receiver.changed().await.map_err(
-                        |_e: tokio::sync::watch::error::RecvError|
-                            // Tenant existed but was dropped: report it as non-existent
-                            GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
-                    )?;
+                    self.activate_now();
+                    match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
+                        Ok(r) => {
+                            r.map_err(
+                            |_e: tokio::sync::watch::error::RecvError|
+                                // Tenant existed but was dropped: report it as non-existent
+                                GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
+                        )?
+                        }
+                        Err(TimeoutCancellableError::Cancelled) => {
+                            return Err(GetActiveTenantError::Cancelled);
+                        }
+                        Err(TimeoutCancellableError::Timeout) => {
+                            return Err(GetActiveTenantError::WaitForActiveTimeout {
+                                latest_state: Some(self.current_state()),
+                                wait_time: timeout,
+                            });
+                        }
+                    }
                 }
                 TenantState::Active { .. } => {
                     return Ok(());
@@ -2463,7 +2600,7 @@ impl Tenant {
             conf,
             // using now here is good enough approximation to catch tenants with really long
             // activation times.
-            loading_started_at: Instant::now(),
+            constructed_at: Instant::now(),
             tenant_conf: Arc::new(RwLock::new(attached_conf)),
             timelines: Mutex::new(HashMap::new()),
             timelines_creating: Mutex::new(HashSet::new()),
@@ -2475,6 +2612,7 @@ impl Tenant {
             cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
             cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
             eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
+            activate_now_sem: tokio::sync::Semaphore::new(0),
             delete_progress: Arc::new(tokio::sync::Mutex::new(DeleteTenantFlow::default())),
             cancel: CancellationToken::default(),
             gate: Gate::new(format!("Tenant<{tenant_shard_id}>")),
@@ -2996,6 +3134,7 @@ impl Tenant {
 
     /// For unit tests, make this visible so that other modules can directly create timelines
     #[cfg(test)]
+    #[tracing::instrument(fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
     pub(crate) async fn bootstrap_timeline_test(
         &self,
         timeline_id: TimelineId,
@@ -3059,6 +3198,7 @@ impl Tenant {
                     storage,
                     &self.tenant_shard_id,
                     &existing_initdb_timeline_id,
+                    &self.cancel,
                 )
                 .await
                 .context("download initdb tar")?;
@@ -3099,6 +3239,7 @@ impl Tenant {
                             &timeline_id,
                             pgdata_zstd.try_clone().await?,
                             tar_zst_size,
+                            &self.cancel,
                         )
                         .await
                     },
@@ -3106,9 +3247,7 @@ impl Tenant {
                     3,
                     u32::MAX,
                     "persist_initdb_tar_zst",
-                    backoff::Cancel::new(self.cancel.clone(), || {
-                        anyhow::anyhow!("initdb upload cancelled")
-                    }),
+                    backoff::Cancel::new(self.cancel.clone(), || anyhow::anyhow!("Cancelled")),
                 )
                 .await?;
 

pageserver/src/tenant/delete.rs

@@ -48,6 +48,9 @@ pub(crate) enum DeleteTenantError {
     #[error("Timeline {0}")]
     Timeline(#[from] DeleteTimelineError),
 
+    #[error("Cancelled")]
+    Cancelled,
+
     #[error(transparent)]
     Other(#[from] anyhow::Error),
 }
@@ -71,6 +74,7 @@ async fn create_remote_delete_mark(
     conf: &PageServerConf,
     remote_storage: &GenericRemoteStorage,
     tenant_shard_id: &TenantShardId,
+    cancel: &CancellationToken,
 ) -> Result<(), DeleteTenantError> {
     let remote_mark_path = remote_tenant_delete_mark_path(conf, tenant_shard_id)?;
 
@@ -87,8 +91,7 @@ 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!()),
+        backoff::Cancel::new(cancel.clone(), || anyhow::anyhow!("Cancelled")),
     )
     .await
     .context("mark_upload")?;
@@ -170,6 +173,7 @@ async fn remove_tenant_remote_delete_mark(
     conf: &PageServerConf,
     remote_storage: Option<&GenericRemoteStorage>,
     tenant_shard_id: &TenantShardId,
+    cancel: &CancellationToken,
 ) -> Result<(), DeleteTenantError> {
     if let Some(remote_storage) = remote_storage {
         let path = remote_tenant_delete_mark_path(conf, tenant_shard_id)?;
@@ -179,8 +183,7 @@ 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!()),
+            backoff::Cancel::new(cancel.clone(), || anyhow::anyhow!("Cancelled")),
         )
         .await
         .context("remove_tenant_remote_delete_mark")?;
@@ -322,9 +325,15 @@ impl DeleteTenantFlow {
         // Though sounds scary, different mark name?
         // Detach currently uses remove_dir_all so in case of a crash we can end up in a weird state.
         if let Some(remote_storage) = &remote_storage {
-            create_remote_delete_mark(conf, remote_storage, &tenant.tenant_shard_id)
-                .await
-                .context("remote_mark")?
+            create_remote_delete_mark(
+                conf,
+                remote_storage,
+                &tenant.tenant_shard_id,
+                // Can't use tenant.cancel, it's already shut down.  TODO: wire in an appropriate token
+                &CancellationToken::new(),
+            )
+            .await
+            .context("remote_mark")?
         }
 
         fail::fail_point!("tenant-delete-before-create-local-mark", |_| {
@@ -524,8 +533,14 @@ impl DeleteTenantFlow {
                 .context("timelines dir not empty")?;
         }
 
-        remove_tenant_remote_delete_mark(conf, remote_storage.as_ref(), &tenant.tenant_shard_id)
-            .await?;
+        remove_tenant_remote_delete_mark(
+            conf,
+            remote_storage.as_ref(),
+            &tenant.tenant_shard_id,
+            // Can't use tenant.cancel, it's already shut down.  TODO: wire in an appropriate token
+            &CancellationToken::new(),
+        )
+        .await?;
 
         fail::fail_point!("tenant-delete-before-cleanup-remaining-fs-traces", |_| {
             Err(anyhow::anyhow!(

pageserver/src/tenant/mgr.rs

@@ -28,7 +28,7 @@ use crate::control_plane_client::{
     ControlPlaneClient, ControlPlaneGenerationsApi, RetryForeverError,
 };
 use crate::deletion_queue::DeletionQueueClient;
-use crate::metrics::TENANT_MANAGER as METRICS;
+use crate::metrics::{TENANT, TENANT_MANAGER as METRICS};
 use crate::task_mgr::{self, TaskKind};
 use crate::tenant::config::{
     AttachedLocationConfig, AttachmentMode, LocationConf, LocationMode, TenantConfOpt,
@@ -44,7 +44,6 @@ use utils::generation::Generation;
 use utils::id::{TenantId, TimelineId};
 
 use super::delete::DeleteTenantError;
-use super::timeline::delete::DeleteTimelineFlow;
 use super::TenantSharedResources;
 
 /// For a tenant that appears in TenantsMap, it may either be
@@ -430,6 +429,13 @@ pub async fn init_tenant_mgr(
     let tenant_generations =
         init_load_generations(conf, &tenant_configs, &resources, &cancel).await?;
 
+    tracing::info!(
+        "Attaching {} tenants at startup, warming up {} at a time",
+        tenant_configs.len(),
+        conf.concurrent_tenant_warmup.initial_permits()
+    );
+    TENANT.startup_scheduled.inc_by(tenant_configs.len() as u64);
+
     // Construct `Tenant` objects and start them running
     for (tenant_shard_id, location_conf) in tenant_configs {
         let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
@@ -508,10 +514,7 @@ pub async fn init_tenant_mgr(
             &ctx,
         ) {
             Ok(tenant) => {
-                tenants.insert(
-                    TenantShardId::unsharded(tenant.tenant_id()),
-                    TenantSlot::Attached(tenant),
-                );
+                tenants.insert(tenant_shard_id, TenantSlot::Attached(tenant));
             }
             Err(e) => {
                 error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to start tenant: {e:#}");
@@ -848,17 +851,6 @@ impl TenantManager {
         }
     }
 
-    pub(crate) async fn delete_timeline(
-        &self,
-        tenant_shard_id: TenantShardId,
-        timeline_id: TimelineId,
-        _ctx: &RequestContext,
-    ) -> Result<(), DeleteTimelineError> {
-        let tenant = self.get_attached_tenant_shard(tenant_shard_id, true)?;
-        DeleteTimelineFlow::run(&tenant, timeline_id, false).await?;
-        Ok(())
-    }
-
     #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
     pub(crate) async fn upsert_location(
         &self,
@@ -967,35 +959,27 @@ impl TenantManager {
         }
 
         let tenant_path = self.conf.tenant_path(&tenant_shard_id);
+        let timelines_path = self.conf.timelines_path(&tenant_shard_id);
+
+        // Directory structure is the same for attached and secondary modes:
+        // create it if it doesn't exist.  Timeline load/creation expects the
+        // timelines/ subdir to already exist.
+        //
+        // Does not need to be fsync'd because local storage is just a cache.
+        tokio::fs::create_dir_all(&timelines_path)
+            .await
+            .with_context(|| format!("Creating {timelines_path}"))?;
+
+        // Before activating either secondary or attached mode, persist the
+        // configuration, so that on restart we will re-attach (or re-start
+        // secondary) on the tenant.
+        Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
+            .await
+            .map_err(SetNewTenantConfigError::Persist)?;
 
         let new_slot = match &new_location_config.mode {
-            LocationMode::Secondary(_) => {
-                // Directory doesn't need to be fsync'd because if we crash it can
-                // safely be recreated next time this tenant location is configured.
-                tokio::fs::create_dir_all(&tenant_path)
-                    .await
-                    .with_context(|| format!("Creating {tenant_path}"))?;
-
-                Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
-                    .await
-                    .map_err(SetNewTenantConfigError::Persist)?;
-
-                TenantSlot::Secondary
-            }
+            LocationMode::Secondary(_) => TenantSlot::Secondary,
             LocationMode::Attached(_attach_config) => {
-                let timelines_path = self.conf.timelines_path(&tenant_shard_id);
-
-                // Directory doesn't need to be fsync'd because we do not depend on
-                // it to exist after crashes: it may be recreated when tenant is
-                // re-attached, see https://github.com/neondatabase/neon/issues/5550
-                tokio::fs::create_dir_all(&tenant_path)
-                    .await
-                    .with_context(|| format!("Creating {timelines_path}"))?;
-
-                Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
-                    .await
-                    .map_err(SetNewTenantConfigError::Persist)?;
-
                 let shard_identity = new_location_config.shard;
                 let tenant = tenant_spawn(
                     self.conf,
@@ -1107,6 +1091,71 @@ impl TenantManager {
                 .collect(),
         }
     }
+
+    pub(crate) async fn delete_tenant(
+        &self,
+        tenant_shard_id: TenantShardId,
+        activation_timeout: Duration,
+    ) -> Result<(), DeleteTenantError> {
+        // We acquire a SlotGuard during this function to protect against concurrent
+        // changes while the ::prepare phase of DeleteTenantFlow executes, but then
+        // have to return the Tenant to the map while the background deletion runs.
+        //
+        // TODO: refactor deletion to happen outside the lifetime of a Tenant.
+        // Currently, deletion requires a reference to the tenants map in order to
+        // keep the Tenant in the map until deletion is complete, and then remove
+        // it at the end.
+        //
+        // See https://github.com/neondatabase/neon/issues/5080
+
+        let slot_guard =
+            tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist)?;
+
+        // unwrap is safe because we used MustExist mode when acquiring
+        let tenant = match slot_guard.get_old_value().as_ref().unwrap() {
+            TenantSlot::Attached(tenant) => tenant.clone(),
+            _ => {
+                // Express "not attached" as equivalent to "not found"
+                return Err(DeleteTenantError::NotAttached);
+            }
+        };
+
+        match tenant.current_state() {
+            TenantState::Broken { .. } | TenantState::Stopping { .. } => {
+                // If a tenant is broken or stopping, DeleteTenantFlow can
+                // handle it: broken tenants proceed to delete, stopping tenants
+                // are checked for deletion already in progress.
+            }
+            _ => {
+                tenant
+                    .wait_to_become_active(activation_timeout)
+                    .await
+                    .map_err(|e| match e {
+                        GetActiveTenantError::WillNotBecomeActive(_) => {
+                            DeleteTenantError::InvalidState(tenant.current_state())
+                        }
+                        GetActiveTenantError::Cancelled => DeleteTenantError::Cancelled,
+                        GetActiveTenantError::NotFound(_) => DeleteTenantError::NotAttached,
+                        GetActiveTenantError::WaitForActiveTimeout {
+                            latest_state: _latest_state,
+                            wait_time: _wait_time,
+                        } => DeleteTenantError::InvalidState(tenant.current_state()),
+                    })?;
+            }
+        }
+
+        let result = DeleteTenantFlow::run(
+            self.conf,
+            self.resources.remote_storage.clone(),
+            &TENANTS,
+            tenant,
+        )
+        .await;
+
+        // The Tenant goes back into the map in Stopping state, it will eventually be removed by DeleteTenantFLow
+        slot_guard.revert();
+        result
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -1221,7 +1270,10 @@ pub(crate) async fn get_active_tenant_with_timeout(
                         // Fast path: we don't need to do any async waiting.
                         return Ok(tenant.clone());
                     }
-                    _ => (WaitFor::Tenant(tenant.clone()), tenant_shard_id),
+                    _ => {
+                        tenant.activate_now();
+                        (WaitFor::Tenant(tenant.clone()), tenant_shard_id)
+                    }
                 }
             }
             Some(TenantSlot::Secondary) => {
@@ -1275,63 +1327,10 @@ pub(crate) async fn get_active_tenant_with_timeout(
     };
 
     tracing::debug!("Waiting for tenant to enter active state...");
-    match timeout_cancellable(
-        deadline.duration_since(Instant::now()),
-        cancel,
-        tenant.wait_to_become_active(),
-    )
-    .await
-    {
-        Ok(Ok(())) => Ok(tenant),
-        Ok(Err(e)) => Err(e),
-        Err(TimeoutCancellableError::Timeout) => {
-            let latest_state = tenant.current_state();
-            if latest_state == TenantState::Active {
-                Ok(tenant)
-            } else {
-                Err(GetActiveTenantError::WaitForActiveTimeout {
-                    latest_state: Some(latest_state),
-                    wait_time: timeout,
-                })
-            }
-        }
-        Err(TimeoutCancellableError::Cancelled) => Err(GetActiveTenantError::Cancelled),
-    }
-}
-
-pub(crate) async fn delete_tenant(
-    conf: &'static PageServerConf,
-    remote_storage: Option<GenericRemoteStorage>,
-    tenant_shard_id: TenantShardId,
-) -> Result<(), DeleteTenantError> {
-    // We acquire a SlotGuard during this function to protect against concurrent
-    // changes while the ::prepare phase of DeleteTenantFlow executes, but then
-    // have to return the Tenant to the map while the background deletion runs.
-    //
-    // TODO: refactor deletion to happen outside the lifetime of a Tenant.
-    // Currently, deletion requires a reference to the tenants map in order to
-    // keep the Tenant in the map until deletion is complete, and then remove
-    // it at the end.
-    //
-    // See https://github.com/neondatabase/neon/issues/5080
-
-    // TODO(sharding): make delete API sharding-aware
-    let slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist)?;
-
-    // unwrap is safe because we used MustExist mode when acquiring
-    let tenant = match slot_guard.get_old_value().as_ref().unwrap() {
-        TenantSlot::Attached(tenant) => tenant.clone(),
-        _ => {
-            // Express "not attached" as equivalent to "not found"
-            return Err(DeleteTenantError::NotAttached);
-        }
-    };
-
-    let result = DeleteTenantFlow::run(conf, remote_storage, &TENANTS, tenant).await;
-
-    // The Tenant goes back into the map in Stopping state, it will eventually be removed by DeleteTenantFLow
-    slot_guard.revert();
-    result
+    tenant
+        .wait_to_become_active(deadline.duration_since(Instant::now()))
+        .await?;
+    Ok(tenant)
 }
 
 #[derive(Debug, thiserror::Error)]

pageserver/src/tenant/remote_timeline_client.rs

@@ -196,10 +196,12 @@ pub(crate) use upload::upload_initdb_dir;
 use utils::backoff::{
     self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
 };
+use utils::timeout::{timeout_cancellable, TimeoutCancellableError};
 
 use std::collections::{HashMap, VecDeque};
 use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, Mutex};
+use std::time::Duration;
 
 use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
 use std::ops::DerefMut;
@@ -316,6 +318,47 @@ pub struct RemoteTimelineClient {
     storage_impl: GenericRemoteStorage,
 
     deletion_queue_client: DeletionQueueClient,
+
+    cancel: CancellationToken,
+}
+
+/// This timeout is intended to deal with hangs in lower layers, e.g. stuck TCP flows.  It is not
+/// intended to be snappy enough for prompt shutdown, as we have a CancellationToken for that.
+const UPLOAD_TIMEOUT: Duration = Duration::from_secs(120);
+const DOWNLOAD_TIMEOUT: Duration = Duration::from_secs(120);
+
+/// Wrapper for timeout_cancellable that flattens result and converts TimeoutCancellableError to anyhow.
+///
+/// This is a convenience for the various upload functions.  In future
+/// the anyhow::Error result should be replaced with a more structured type that
+/// enables callers to avoid handling shutdown as an error.
+async fn upload_cancellable<F>(cancel: &CancellationToken, future: F) -> anyhow::Result<()>
+where
+    F: std::future::Future<Output = anyhow::Result<()>>,
+{
+    match timeout_cancellable(UPLOAD_TIMEOUT, cancel, future).await {
+        Ok(Ok(())) => Ok(()),
+        Ok(Err(e)) => Err(e),
+        Err(TimeoutCancellableError::Timeout) => Err(anyhow::anyhow!("Timeout")),
+        Err(TimeoutCancellableError::Cancelled) => Err(anyhow::anyhow!("Shutting down")),
+    }
+}
+/// Wrapper for timeout_cancellable that flattens result and converts TimeoutCancellableError to DownloaDError.
+async fn download_cancellable<F, R>(
+    cancel: &CancellationToken,
+    future: F,
+) -> Result<R, DownloadError>
+where
+    F: std::future::Future<Output = Result<R, DownloadError>>,
+{
+    match timeout_cancellable(DOWNLOAD_TIMEOUT, cancel, future).await {
+        Ok(Ok(r)) => Ok(r),
+        Ok(Err(e)) => Err(e),
+        Err(TimeoutCancellableError::Timeout) => {
+            Err(DownloadError::Other(anyhow::anyhow!("Timed out")))
+        }
+        Err(TimeoutCancellableError::Cancelled) => Err(DownloadError::Cancelled),
+    }
 }
 
 impl RemoteTimelineClient {
@@ -351,6 +394,7 @@ impl RemoteTimelineClient {
                 &tenant_shard_id,
                 &timeline_id,
             )),
+            cancel: CancellationToken::new(),
         }
     }
 
@@ -501,6 +545,7 @@ impl RemoteTimelineClient {
         &self,
         layer_file_name: &LayerFileName,
         layer_metadata: &LayerFileMetadata,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<u64> {
         let downloaded_size = {
             let _unfinished_gauge_guard = self.metrics.call_begin(
@@ -517,6 +562,7 @@ impl RemoteTimelineClient {
                 self.timeline_id,
                 layer_file_name,
                 layer_metadata,
+                cancel,
             )
             .measure_remote_op(
                 self.tenant_shard_id.tenant_id,
@@ -971,6 +1017,7 @@ impl RemoteTimelineClient {
                     &self.timeline_id,
                     self.generation,
                     &index_part_with_deleted_at,
+                    &self.cancel,
                 )
             },
             |_e| false,
@@ -980,8 +1027,7 @@ impl RemoteTimelineClient {
             // when executed as part of tenant deletion this happens in the background
             2,
             "persist_index_part_with_deleted_flag",
-            // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
-            backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
+            backoff::Cancel::new(self.cancel.clone(), || anyhow::anyhow!("Cancelled")),
         )
         .await?;
 
@@ -1281,6 +1327,7 @@ impl RemoteTimelineClient {
                         path,
                         layer_metadata,
                         self.generation,
+                        &self.cancel,
                     )
                     .measure_remote_op(
                         self.tenant_shard_id.tenant_id,
@@ -1307,6 +1354,7 @@ impl RemoteTimelineClient {
                         &self.timeline_id,
                         self.generation,
                         index_part,
+                        &self.cancel,
                     )
                     .measure_remote_op(
                         self.tenant_shard_id.tenant_id,
@@ -1828,6 +1876,7 @@ mod tests {
                     &self.harness.tenant_shard_id,
                     &TIMELINE_ID,
                 )),
+                cancel: CancellationToken::new(),
             })
         }
 
@@ -2143,15 +2192,6 @@ mod tests {
 
         let index_part_bytes = serde_json::to_vec(&example_index_part).unwrap();
 
-        let timeline_path = test_state.harness.timeline_path(&TIMELINE_ID);
-        let remote_timeline_dir = test_state.harness.remote_fs_dir.join(
-            timeline_path
-                .strip_prefix(&test_state.harness.conf.workdir)
-                .unwrap(),
-        );
-
-        std::fs::create_dir_all(remote_timeline_dir).expect("creating test dir should work");
-
         let index_path = test_state.harness.remote_fs_dir.join(
             remote_index_path(
                 &test_state.harness.tenant_shard_id,
@@ -2160,6 +2200,10 @@ mod tests {
             )
             .get_path(),
         );
+
+        std::fs::create_dir_all(index_path.parent().unwrap())
+            .expect("creating test dir should work");
+
         eprintln!("Writing {index_path}");
         std::fs::write(&index_path, index_part_bytes).unwrap();
         example_index_part

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -5,7 +5,6 @@
 
 use std::collections::HashSet;
 use std::future::Future;
-use std::time::Duration;
 
 use anyhow::{anyhow, Context};
 use camino::{Utf8Path, Utf8PathBuf};
@@ -14,13 +13,17 @@ use tokio::fs::{self, File, OpenOptions};
 use tokio::io::{AsyncSeekExt, AsyncWriteExt};
 use tokio_util::sync::CancellationToken;
 use tracing::warn;
+use utils::timeout::timeout_cancellable;
 use utils::{backoff, crashsafe};
 
 use crate::config::PageServerConf;
-use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
+use crate::tenant::remote_timeline_client::{
+    download_cancellable, remote_layer_path, remote_timelines_path, DOWNLOAD_TIMEOUT,
+};
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::timeline::span::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::Generation;
+use crate::virtual_file::on_fatal_io_error;
 use crate::TEMP_FILE_SUFFIX;
 use remote_storage::{DownloadError, GenericRemoteStorage, ListingMode};
 use utils::crashsafe::path_with_suffix_extension;
@@ -32,8 +35,6 @@ use super::{
     FAILED_DOWNLOAD_WARN_THRESHOLD, FAILED_REMOTE_OP_RETRIES, INITDB_PATH,
 };
 
-static MAX_DOWNLOAD_DURATION: Duration = Duration::from_secs(120);
-
 ///
 /// If 'metadata' is given, we will validate that the downloaded file's size matches that
 /// in the metadata. (In the future, we might do more cross-checks, like CRC validation)
@@ -46,6 +47,7 @@ pub async fn download_layer_file<'a>(
     timeline_id: TimelineId,
     layer_file_name: &'a LayerFileName,
     layer_metadata: &'a LayerFileMetadata,
+    cancel: &CancellationToken,
 ) -> Result<u64, DownloadError> {
     debug_assert_current_span_has_tenant_and_timeline_id();
 
@@ -73,14 +75,18 @@ pub async fn download_layer_file<'a>(
     // If pageserver crashes the temp file will be deleted on startup and re-downloaded.
     let temp_file_path = path_with_suffix_extension(&local_path, TEMP_DOWNLOAD_EXTENSION);
 
+    let cancel_inner = cancel.clone();
     let (mut destination_file, bytes_amount) = download_retry(
         || async {
             let destination_file = tokio::fs::File::create(&temp_file_path)
                 .await
                 .with_context(|| format!("create a destination file for layer '{temp_file_path}'"))
                 .map_err(DownloadError::Other)?;
-            let download = storage
-                .download(&remote_path)
+
+            // Cancellation safety: it is safe to cancel this future, because it isn't writing to a local
+            // file: the write to local file doesn't start until after the request header is returned
+            // and we start draining the body stream below
+            let download = download_cancellable(&cancel_inner, storage.download(&remote_path))
                 .await
                 .with_context(|| {
                     format!(
@@ -94,12 +100,33 @@ pub async fn download_layer_file<'a>(
 
             let mut reader = tokio_util::io::StreamReader::new(download.download_stream);
 
-            let bytes_amount = tokio::time::timeout(
-                MAX_DOWNLOAD_DURATION,
+            // Cancellation safety: it is safe to cancel this future because it is writing into a temporary file,
+            // and we will unlink the temporary file if there is an error.  This unlink is important because we
+            // are in a retry loop, and we wouldn't want to leave behind a rogue write I/O to a file that
+            // we will imminiently try and write to again.
+            let bytes_amount: u64 = match timeout_cancellable(
+                DOWNLOAD_TIMEOUT,
+                &cancel_inner,
                 tokio::io::copy_buf(&mut reader, &mut destination_file),
             )
             .await
-            .map_err(|e| DownloadError::Other(anyhow::anyhow!("Timed out  {:?}", e)))?
+            .with_context(|| {
+                format!(
+                    "download layer at remote path '{remote_path:?}' into file {temp_file_path:?}"
+                )
+            })
+            .map_err(DownloadError::Other)?
+            {
+                Ok(b) => Ok(b),
+                Err(e) => {
+                    // Remove incomplete files: on restart Timeline would do this anyway, but we must
+                    // do it here for the retry case.
+                    if let Err(e) = tokio::fs::remove_file(&temp_file_path).await {
+                        on_fatal_io_error(&e, &format!("Removing temporary file {temp_file_path}"));
+                    }
+                    Err(e)
+                }
+            }
             .with_context(|| {
                 format!(
                     "download layer at remote path '{remote_path:?}' into file {temp_file_path:?}"
@@ -112,6 +139,7 @@ pub async fn download_layer_file<'a>(
             Ok((destination_file, bytes_amount))
         },
         &format!("download {remote_path:?}"),
+        cancel,
     )
     .await?;
 
@@ -188,8 +216,14 @@ pub async fn list_remote_timelines(
         anyhow::bail!("storage-sync-list-remote-timelines");
     });
 
+    let cancel_inner = cancel.clone();
     let listing = download_retry_forever(
-        || storage.list(Some(&remote_path), ListingMode::WithDelimiter),
+        || {
+            download_cancellable(
+                &cancel_inner,
+                storage.list(Some(&remote_path), ListingMode::WithDelimiter),
+            )
+        },
         &format!("list timelines for {tenant_shard_id}"),
         cancel,
     )
@@ -230,9 +264,13 @@ async fn do_download_index_part(
 
     let remote_path = remote_index_path(tenant_shard_id, timeline_id, index_generation);
 
+    let cancel_inner = cancel.clone();
     let index_part_bytes = download_retry_forever(
         || async {
-            let index_part_download = storage.download(&remote_path).await?;
+            // Cancellation: if is safe to cancel this future because we're just downloading into
+            // a memory buffer, not touching local disk.
+            let index_part_download =
+                download_cancellable(&cancel_inner, storage.download(&remote_path)).await?;
 
             let mut index_part_bytes = Vec::new();
             let mut stream = std::pin::pin!(index_part_download.download_stream);
@@ -347,10 +385,7 @@ pub(super) async fn download_index_part(
         FAILED_DOWNLOAD_WARN_THRESHOLD,
         FAILED_REMOTE_OP_RETRIES,
         "listing index_part files",
-        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
-        backoff::Cancel::new(CancellationToken::new(), || -> anyhow::Error {
-            unreachable!()
-        }),
+        backoff::Cancel::new(cancel.clone(), || anyhow::anyhow!("Cancelled")),
     )
     .await
     .map_err(DownloadError::Other)?;
@@ -389,6 +424,7 @@ pub(crate) async fn download_initdb_tar_zst(
     storage: &GenericRemoteStorage,
     tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
+    cancel: &CancellationToken,
 ) -> Result<(Utf8PathBuf, File), DownloadError> {
     debug_assert_current_span_has_tenant_and_timeline_id();
 
@@ -406,6 +442,8 @@ pub(crate) async fn download_initdb_tar_zst(
         "{INITDB_PATH}.download-{timeline_id}.{TEMP_FILE_SUFFIX}"
     ));
 
+    let cancel_inner = cancel.clone();
+
     let file = download_retry(
         || async {
             let file = OpenOptions::new()
@@ -418,10 +456,14 @@ pub(crate) async fn download_initdb_tar_zst(
                 .with_context(|| format!("tempfile creation {temp_path}"))
                 .map_err(DownloadError::Other)?;
 
-            let download = storage.download(&remote_path).await?;
+            let download =
+                download_cancellable(&cancel_inner, storage.download(&remote_path)).await?;
             let mut download = tokio_util::io::StreamReader::new(download.download_stream);
             let mut writer = tokio::io::BufWriter::with_capacity(8 * 1024, file);
 
+            // TODO: this consumption of the response body should be subject to timeout + cancellation, but
+            // not without thinking carefully about how to recover safely from cancelling a write to
+            // local storage (e.g. by writing into a temp file as we do in download_layer)
             tokio::io::copy_buf(&mut download, &mut writer)
                 .await
                 .with_context(|| format!("download initdb.tar.zst at {remote_path:?}"))
@@ -437,6 +479,7 @@ pub(crate) async fn download_initdb_tar_zst(
             Ok(file)
         },
         &format!("download {remote_path}"),
+        cancel,
     )
     .await
     .map_err(|e| {
@@ -460,7 +503,11 @@ pub(crate) async fn download_initdb_tar_zst(
 /// with backoff.
 ///
 /// (See similar logic for uploads in `perform_upload_task`)
-async fn download_retry<T, O, F>(op: O, description: &str) -> Result<T, DownloadError>
+async fn download_retry<T, O, F>(
+    op: O,
+    description: &str,
+    cancel: &CancellationToken,
+) -> Result<T, DownloadError>
 where
     O: FnMut() -> F,
     F: Future<Output = Result<T, DownloadError>>,
@@ -471,10 +518,7 @@ where
         FAILED_DOWNLOAD_WARN_THRESHOLD,
         FAILED_REMOTE_OP_RETRIES,
         description,
-        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
-        backoff::Cancel::new(CancellationToken::new(), || -> DownloadError {
-            unreachable!()
-        }),
+        backoff::Cancel::new(cancel.clone(), || DownloadError::Cancelled),
     )
     .await
 }

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -7,12 +7,14 @@ use pageserver_api::shard::TenantShardId;
 use std::io::{ErrorKind, SeekFrom};
 use tokio::fs::{self, File};
 use tokio::io::AsyncSeekExt;
+use tokio_util::sync::CancellationToken;
 
 use super::Generation;
 use crate::{
     config::PageServerConf,
     tenant::remote_timeline_client::{
         index::IndexPart, remote_index_path, remote_initdb_archive_path, remote_path,
+        upload_cancellable,
     },
 };
 use remote_storage::GenericRemoteStorage;
@@ -29,6 +31,7 @@ pub(super) async fn upload_index_part<'a>(
     timeline_id: &TimelineId,
     generation: Generation,
     index_part: &'a IndexPart,
+    cancel: &CancellationToken,
 ) -> anyhow::Result<()> {
     tracing::trace!("uploading new index part");
 
@@ -44,14 +47,16 @@ pub(super) async fn upload_index_part<'a>(
     let index_part_bytes = bytes::Bytes::from(index_part_bytes);
 
     let remote_path = remote_index_path(tenant_shard_id, timeline_id, generation);
-    storage
-        .upload_storage_object(
+    upload_cancellable(
+        cancel,
+        storage.upload_storage_object(
             futures::stream::once(futures::future::ready(Ok(index_part_bytes))),
             index_part_size,
             &remote_path,
-        )
-        .await
-        .with_context(|| format!("upload index part for '{tenant_shard_id} / {timeline_id}'"))
+        ),
+    )
+    .await
+    .with_context(|| format!("upload index part for '{tenant_shard_id} / {timeline_id}'"))
 }
 
 /// Attempts to upload given layer files.
@@ -64,6 +69,7 @@ pub(super) async fn upload_timeline_layer<'a>(
     source_path: &'a Utf8Path,
     known_metadata: &'a LayerFileMetadata,
     generation: Generation,
+    cancel: &CancellationToken,
 ) -> anyhow::Result<()> {
     fail_point!("before-upload-layer", |_| {
         bail!("failpoint before-upload-layer")
@@ -107,8 +113,7 @@ pub(super) async fn upload_timeline_layer<'a>(
 
     let reader = tokio_util::io::ReaderStream::with_capacity(source_file, super::BUFFER_SIZE);
 
-    storage
-        .upload(reader, fs_size, &storage_path, None)
+    upload_cancellable(cancel, storage.upload(reader, fs_size, &storage_path, None))
         .await
         .with_context(|| format!("upload layer from local path '{source_path}'"))?;
 
@@ -122,6 +127,7 @@ pub(crate) async fn upload_initdb_dir(
     timeline_id: &TimelineId,
     mut initdb_tar_zst: File,
     size: u64,
+    cancel: &CancellationToken,
 ) -> anyhow::Result<()> {
     tracing::trace!("uploading initdb dir");
 
@@ -131,8 +137,10 @@ pub(crate) async fn upload_initdb_dir(
     let file = tokio_util::io::ReaderStream::with_capacity(initdb_tar_zst, super::BUFFER_SIZE);
 
     let remote_path = remote_initdb_archive_path(tenant_id, timeline_id);
-    storage
-        .upload_storage_object(file, size as usize, &remote_path)
-        .await
-        .with_context(|| format!("upload initdb dir for '{tenant_id} / {timeline_id}'"))
+    upload_cancellable(
+        cancel,
+        storage.upload_storage_object(file, size as usize, &remote_path),
+    )
+    .await
+    .with_context(|| format!("upload initdb dir for '{tenant_id} / {timeline_id}'"))
 }

pageserver/src/tenant/storage_layer/layer.rs

@@ -259,8 +259,9 @@ impl Layer {
 
         layer
             .get_value_reconstruct_data(key, lsn_range, reconstruct_data, &self.0, ctx)
-            .instrument(tracing::info_span!("get_value_reconstruct_data", layer=%self))
+            .instrument(tracing::debug_span!("get_value_reconstruct_data", layer=%self))
             .await
+            .with_context(|| format!("get_value_reconstruct_data for layer {self}"))
     }
 
     /// Download the layer if evicted.
@@ -654,7 +655,6 @@ impl LayerInner {
     }
 
     /// Cancellation safe.
-    #[tracing::instrument(skip_all, fields(layer=%self))]
     async fn get_or_maybe_download(
         self: &Arc<Self>,
         allow_download: bool,
@@ -663,95 +663,101 @@ impl LayerInner {
         let mut init_permit = None;
 
         loop {
-            let download = move |permit| async move {
-                // disable any scheduled but not yet running eviction deletions for this
-                let next_version = 1 + self.version.fetch_add(1, Ordering::Relaxed);
+            let download = move |permit| {
+                async move {
+                    // disable any scheduled but not yet running eviction deletions for this
+                    let next_version = 1 + self.version.fetch_add(1, Ordering::Relaxed);
 
-                // count cancellations, which currently remain largely unexpected
-                let init_cancelled =
-                    scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
+                    // count cancellations, which currently remain largely unexpected
+                    let init_cancelled =
+                        scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
 
-                // no need to make the evict_and_wait wait for the actual download to complete
-                drop(self.status.send(Status::Downloaded));
+                    // no need to make the evict_and_wait wait for the actual download to complete
+                    drop(self.status.send(Status::Downloaded));
 
-                let timeline = self
-                    .timeline
-                    .upgrade()
-                    .ok_or_else(|| DownloadError::TimelineShutdown)?;
+                    let timeline = self
+                        .timeline
+                        .upgrade()
+                        .ok_or_else(|| DownloadError::TimelineShutdown)?;
 
-                // FIXME: grab a gate
+                    // FIXME: grab a gate
 
-                let can_ever_evict = timeline.remote_client.as_ref().is_some();
+                    let can_ever_evict = timeline.remote_client.as_ref().is_some();
 
-                // check if we really need to be downloaded; could have been already downloaded by a
-                // cancelled previous attempt.
-                let needs_download = self
-                    .needs_download()
-                    .await
-                    .map_err(DownloadError::PreStatFailed)?;
+                    // check if we really need to be downloaded; could have been already downloaded by a
+                    // cancelled previous attempt.
+                    let needs_download = self
+                        .needs_download()
+                        .await
+                        .map_err(DownloadError::PreStatFailed)?;
 
-                let permit = if let Some(reason) = needs_download {
-                    if let NeedsDownload::NotFile(ft) = reason {
-                        return Err(DownloadError::NotFile(ft));
+                    let permit = if let Some(reason) = needs_download {
+                        if let NeedsDownload::NotFile(ft) = reason {
+                            return Err(DownloadError::NotFile(ft));
+                        }
+
+                        // only reset this after we've decided we really need to download. otherwise it'd
+                        // be impossible to mark cancelled downloads for eviction, like one could imagine
+                        // we would like to do for prefetching which was not needed.
+                        self.wanted_evicted.store(false, Ordering::Release);
+
+                        if !can_ever_evict {
+                            return Err(DownloadError::NoRemoteStorage);
+                        }
+
+                        if let Some(ctx) = ctx {
+                            self.check_expected_download(ctx)?;
+                        }
+
+                        if !allow_download {
+                            // this does look weird, but for LayerInner the "downloading" means also changing
+                            // internal once related state ...
+                            return Err(DownloadError::DownloadRequired);
+                        }
+
+                        tracing::info!(%reason, "downloading on-demand");
+
+                        self.spawn_download_and_wait(timeline, permit).await?
+                    } else {
+                        // the file is present locally, probably by a previous but cancelled call to
+                        // get_or_maybe_download. alternatively we might be running without remote storage.
+                        LAYER_IMPL_METRICS.inc_init_needed_no_download();
+
+                        permit
+                    };
+
+                    let since_last_eviction =
+                        self.last_evicted_at.lock().unwrap().map(|ts| ts.elapsed());
+                    if let Some(since_last_eviction) = since_last_eviction {
+                        // FIXME: this will not always be recorded correctly until #6028 (the no
+                        // download needed branch above)
+                        LAYER_IMPL_METRICS.record_redownloaded_after(since_last_eviction);
                     }
 
-                    // only reset this after we've decided we really need to download. otherwise it'd
-                    // be impossible to mark cancelled downloads for eviction, like one could imagine
-                    // we would like to do for prefetching which was not needed.
-                    self.wanted_evicted.store(false, Ordering::Release);
+                    let res = Arc::new(DownloadedLayer {
+                        owner: Arc::downgrade(self),
+                        kind: tokio::sync::OnceCell::default(),
+                        version: next_version,
+                    });
 
-                    if !can_ever_evict {
-                        return Err(DownloadError::NoRemoteStorage);
+                    self.access_stats.record_residence_event(
+                        LayerResidenceStatus::Resident,
+                        LayerResidenceEventReason::ResidenceChange,
+                    );
+
+                    let waiters = self.inner.initializer_count();
+                    if waiters > 0 {
+                        tracing::info!(
+                            waiters,
+                            "completing the on-demand download for other tasks"
+                        );
                     }
 
-                    if let Some(ctx) = ctx {
-                        self.check_expected_download(ctx)?;
-                    }
+                    scopeguard::ScopeGuard::into_inner(init_cancelled);
 
-                    if !allow_download {
-                        // this does look weird, but for LayerInner the "downloading" means also changing
-                        // internal once related state ...
-                        return Err(DownloadError::DownloadRequired);
-                    }
-
-                    tracing::info!(%reason, "downloading on-demand");
-
-                    self.spawn_download_and_wait(timeline, permit).await?
-                } else {
-                    // the file is present locally, probably by a previous but cancelled call to
-                    // get_or_maybe_download. alternatively we might be running without remote storage.
-                    LAYER_IMPL_METRICS.inc_init_needed_no_download();
-
-                    permit
-                };
-
-                let since_last_eviction =
-                    self.last_evicted_at.lock().unwrap().map(|ts| ts.elapsed());
-                if let Some(since_last_eviction) = since_last_eviction {
-                    // FIXME: this will not always be recorded correctly until #6028 (the no
-                    // download needed branch above)
-                    LAYER_IMPL_METRICS.record_redownloaded_after(since_last_eviction);
+                    Ok((ResidentOrWantedEvicted::Resident(res), permit))
                 }
-
-                let res = Arc::new(DownloadedLayer {
-                    owner: Arc::downgrade(self),
-                    kind: tokio::sync::OnceCell::default(),
-                    version: next_version,
-                });
-
-                self.access_stats.record_residence_event(
-                    LayerResidenceStatus::Resident,
-                    LayerResidenceEventReason::ResidenceChange,
-                );
-
-                let waiters = self.inner.initializer_count();
-                if waiters > 0 {
-                    tracing::info!(waiters, "completing the on-demand download for other tasks");
-                }
-
-                scopeguard::ScopeGuard::into_inner(init_cancelled);
-
-                Ok((ResidentOrWantedEvicted::Resident(res), permit))
+                .instrument(tracing::info_span!("get_or_maybe_download", layer=%self))
             };
 
             if let Some(init_permit) = init_permit.take() {
@@ -862,6 +868,7 @@ impl LayerInner {
                 let result = client.download_layer_file(
                     &this.desc.filename(),
                     &this.metadata(),
+                    &crate::task_mgr::shutdown_token()
                 )
                 .await;
 
@@ -871,6 +878,23 @@ impl LayerInner {
                         Ok(())
                     }
                     Err(e) => {
+                        let consecutive_failures =
+                            this.consecutive_failures.fetch_add(1, Ordering::Relaxed);
+
+                        let backoff = utils::backoff::exponential_backoff_duration_seconds(
+                            consecutive_failures.min(u32::MAX as usize) as u32,
+                            1.5,
+                            60.0,
+                        );
+
+                        let backoff = std::time::Duration::from_secs_f64(backoff);
+
+                        tokio::select! {
+                            _ = tokio::time::sleep(backoff) => {},
+                            _ = crate::task_mgr::shutdown_token().cancelled_owned() => {},
+                            _ = timeline.cancel.cancelled() => {},
+                        };
+
                         Err(e)
                     }
                 };
@@ -919,21 +943,9 @@ impl LayerInner {
                 Ok(permit)
             }
             Ok((Err(e), _permit)) => {
-                // FIXME: this should be with the spawned task and be cancellation sensitive
-                //
-                // while we should not need this, this backoff has turned out to be useful with
-                // a bug of unexpectedly deleted remote layer file (#5787).
-                let consecutive_failures =
-                    self.consecutive_failures.fetch_add(1, Ordering::Relaxed);
+                // sleep already happened in the spawned task, if it was not cancelled
+                let consecutive_failures = self.consecutive_failures.load(Ordering::Relaxed);
                 tracing::error!(consecutive_failures, "layer file download failed: {e:#}");
-                let backoff = utils::backoff::exponential_backoff_duration_seconds(
-                    consecutive_failures.min(u32::MAX as usize) as u32,
-                    1.5,
-                    60.0,
-                );
-                let backoff = std::time::Duration::from_secs_f64(backoff);
-
-                tokio::time::sleep(backoff).await;
                 Err(DownloadError::DownloadFailed)
             }
             Err(_gone) => Err(DownloadError::DownloadCancelled),

pageserver/src/tenant/timeline.rs

@@ -1734,6 +1734,7 @@ impl Timeline {
                 self.current_logical_size.current_size().accuracy(),
                 logical_size::Accuracy::Exact,
             );
+            self.current_logical_size.initialized.add_permits(1);
             return;
         };
 
@@ -1779,6 +1780,11 @@ impl Timeline {
         cancel: CancellationToken,
         background_ctx: RequestContext,
     ) {
+        scopeguard::defer! {
+            // Irrespective of the outcome of this operation, we should unblock anyone waiting for it.
+            self.current_logical_size.initialized.add_permits(1);
+        }
+
         enum BackgroundCalculationError {
             Cancelled,
             Other(anyhow::Error),
@@ -3104,6 +3110,32 @@ impl Timeline {
 
         Ok(image_layers)
     }
+
+    /// Wait until the background initial logical size calculation is complete, or
+    /// this Timeline is shut down.  Calling this function will cause the initial
+    /// logical size calculation to skip waiting for the background jobs barrier.
+    pub(crate) async fn await_initial_logical_size(self: Arc<Self>) {
+        if let Some(await_bg_cancel) = self
+            .current_logical_size
+            .cancel_wait_for_background_loop_concurrency_limit_semaphore
+            .get()
+        {
+            await_bg_cancel.cancel();
+        } else {
+            // We should not wait if we were not able to explicitly instruct
+            // the logical size cancellation to skip the concurrency limit semaphore.
+            // TODO: this is an unexpected case.  We should restructure so that it
+            // can't happen.
+            tracing::info!(
+                "await_initial_logical_size: can't get semaphore cancel token, skipping"
+            );
+        }
+
+        tokio::select!(
+            _ = self.current_logical_size.initialized.acquire() => {},
+            _ = self.cancel.cancelled() => {}
+        )
+    }
 }
 
 #[derive(Default)]

pageserver/src/tenant/timeline/logical_size.rs

@@ -34,6 +34,9 @@ pub(super) struct LogicalSize {
     pub(crate) cancel_wait_for_background_loop_concurrency_limit_semaphore:
         OnceCell<CancellationToken>,
 
+    /// Once the initial logical size is initialized, this is notified.
+    pub(crate) initialized: tokio::sync::Semaphore,
+
     /// Latest Lsn that has its size uncalculated, could be absent for freshly created timelines.
     pub initial_part_end: Option<Lsn>,
 
@@ -125,6 +128,7 @@ impl LogicalSize {
             initial_part_end: None,
             size_added_after_initial: AtomicI64::new(0),
             did_return_approximate_to_walreceiver: AtomicBool::new(false),
+            initialized: tokio::sync::Semaphore::new(0),
         }
     }
 
@@ -135,6 +139,7 @@ impl LogicalSize {
             initial_part_end: Some(compute_to),
             size_added_after_initial: AtomicI64::new(0),
             did_return_approximate_to_walreceiver: AtomicBool::new(false),
+            initialized: tokio::sync::Semaphore::new(0),
         }
     }
 

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

@@ -138,7 +138,7 @@ pub(super) async fn connection_manager_loop_step(
                     Ok(Some(broker_update)) => connection_manager_state.register_timeline_update(broker_update),
                     Err(status) => {
                         match status.code() {
-                            Code::Unknown if status.message().contains("stream closed because of a broken pipe") => {
+                            Code::Unknown if status.message().contains("stream closed because of a broken pipe") || status.message().contains("connection reset") => {
                                 // tonic's error handling doesn't provide a clear code for disconnections: we get
                                 // "h2 protocol error: error reading a body from connection: stream closed because of a broken pipe"
                                 info!("broker disconnected: {status}");

pageserver/src/walingest.rs

@@ -1612,6 +1612,7 @@ impl<'a> WalIngest<'a> {
 mod tests {
     use super::*;
     use crate::tenant::harness::*;
+    use crate::tenant::remote_timeline_client::{remote_initdb_archive_path, INITDB_PATH};
     use crate::tenant::Timeline;
     use postgres_ffi::v14::xlog_utils::SIZEOF_CHECKPOINT;
     use postgres_ffi::RELSEG_SIZE;
@@ -2177,21 +2178,25 @@ mod tests {
         let pg_version = 15; // The test data was generated by pg15
         let path = "test_data/sk_wal_segment_from_pgbench";
         let wal_segment_path = format!("{path}/000000010000000000000001.zst");
+        let source_initdb_path = format!("{path}/{INITDB_PATH}");
         let startpoint = Lsn::from_hex("14AEC08").unwrap();
         let endpoint = Lsn::from_hex("1FFFF98").unwrap();
 
+        let harness = TenantHarness::create("test_ingest_real_wal").unwrap();
+        let (tenant, ctx) = harness.load().await;
+
+        let remote_initdb_path = remote_initdb_archive_path(&tenant.tenant_id(), &TIMELINE_ID);
+        let initdb_path = harness.remote_fs_dir.join(remote_initdb_path.get_path());
+
+        std::fs::create_dir_all(initdb_path.parent().unwrap())
+            .expect("creating test dir should work");
+        std::fs::copy(source_initdb_path, initdb_path).expect("copying the initdb.tar.zst works");
+
         // Bootstrap a real timeline. We can't use create_test_timeline because
         // it doesn't create a real checkpoint, and Walingest::new tries to parse
         // the garbage data.
-        //
-        // TODO use the initdb.tar.zst file stored with the test data to avoid
-        //      problems with inconsistent initdb results after pg minor version bumps.
-        let (tenant, ctx) = TenantHarness::create("test_ingest_real_wal")
-            .unwrap()
-            .load()
-            .await;
         let tline = tenant
-            .bootstrap_timeline_test(TIMELINE_ID, pg_version, None, &ctx)
+            .bootstrap_timeline_test(TIMELINE_ID, pg_version, Some(TIMELINE_ID), &ctx)
             .await
             .unwrap();
 

pgxn/neon/Makefile

@@ -9,6 +9,7 @@ OBJS = \
 	libpagestore.o \
 	neon.o \
 	neon_utils.o \
+	neon_walreader.o \
 	pagestore_smgr.o \
 	relsize_cache.o \
 	walproposer.o \

pgxn/neon/control_plane_connector.c

@@ -19,20 +19,21 @@
  *-------------------------------------------------------------------------
  */
 #include "postgres.h"
+
+#include <curl/curl.h>
+
+#include "access/xact.h"
+#include "commands/defrem.h"
+#include "fmgr.h"
+#include "libpq/crypt.h"
+#include "miscadmin.h"
 #include "tcop/pquery.h"
 #include "tcop/utility.h"
-#include "access/xact.h"
+#include "utils/acl.h"
+#include "utils/guc.h"
 #include "utils/hsearch.h"
 #include "utils/memutils.h"
-#include "commands/defrem.h"
-#include "miscadmin.h"
-#include "utils/acl.h"
-#include "fmgr.h"
-#include "utils/guc.h"
-#include "port.h"
-#include <curl/curl.h>
 #include "utils/jsonb.h"
-#include "libpq/crypt.h"
 
 static ProcessUtility_hook_type PreviousProcessUtilityHook = NULL;
 

pgxn/neon/extension_server.c

@@ -1,4 +1,3 @@
-
 /*-------------------------------------------------------------------------
  *
  * extension_server.c
@@ -10,21 +9,11 @@
  *-------------------------------------------------------------------------
  */
 #include "postgres.h"
-#include "tcop/pquery.h"
-#include "tcop/utility.h"
-#include "access/xact.h"
-#include "utils/hsearch.h"
-#include "utils/memutils.h"
-#include "commands/defrem.h"
-#include "miscadmin.h"
-#include "utils/acl.h"
-#include "fmgr.h"
-#include "utils/guc.h"
-#include "port.h"
-#include "fmgr.h"
 
 #include <curl/curl.h>
 
+#include "utils/guc.h"
+
 static int	extension_server_port = 0;
 
 static download_extension_file_hook_type prev_download_extension_file_hook = NULL;

pgxn/neon/file_cache.c

@@ -13,32 +13,30 @@
  *-------------------------------------------------------------------------
  */
 
+#include "postgres.h"
+
 #include <sys/file.h>
 #include <unistd.h>
 #include <fcntl.h>
 
-#include "postgres.h"
-
 #include "neon_pgversioncompat.h"
 
+#include "access/parallel.h"
 #include "funcapi.h"
 #include "miscadmin.h"
-#include "pgstat.h"
 #include "pagestore_client.h"
-#include "access/parallel.h"
+#include "pgstat.h"
 #include "postmaster/bgworker.h"
 #include RELFILEINFO_HDR
 #include "storage/buf_internals.h"
-#include "storage/latch.h"
+#include "storage/fd.h"
 #include "storage/ipc.h"
+#include "storage/latch.h"
 #include "storage/lwlock.h"
+#include "storage/pg_shmem.h"
 #include "utils/builtins.h"
 #include "utils/dynahash.h"
 #include "utils/guc.h"
-#include "storage/fd.h"
-#include "storage/pg_shmem.h"
-#include "storage/buf_internals.h"
-#include "pgstat.h"
 
 /*
  * Local file cache is used to temporary store relations pages in local file system.
@@ -102,8 +100,6 @@ static shmem_request_hook_type prev_shmem_request_hook;
 
 #define LFC_ENABLED() (lfc_ctl->limit != 0)
 
-void		PGDLLEXPORT FileCacheMonitorMain(Datum main_arg);
-
 /*
  * Local file cache is optional and Neon can work without it.
  * In case of any any errors with this cache, we should disable it but to not throw error.

pgxn/neon/libpagestore.c

@@ -14,28 +14,24 @@
  */
 #include "postgres.h"
 
-#include "pagestore_client.h"
-#include "fmgr.h"
 #include "access/xlog.h"
-#include "access/xlogutils.h"
-#include "storage/buf_internals.h"
-#include "storage/lwlock.h"
-#include "storage/ipc.h"
-#include "storage/pg_shmem.h"
-#include "c.h"
-#include "postmaster/interrupt.h"
-
+#include "fmgr.h"
 #include "libpq-fe.h"
-#include "libpq/pqformat.h"
 #include "libpq/libpq.h"
-
+#include "libpq/pqformat.h"
 #include "miscadmin.h"
 #include "pgstat.h"
+#include "postmaster/interrupt.h"
+#include "storage/buf_internals.h"
+#include "storage/ipc.h"
+#include "storage/lwlock.h"
+#include "storage/pg_shmem.h"
 #include "utils/guc.h"
 
 #include "neon.h"
-#include "walproposer.h"
 #include "neon_utils.h"
+#include "pagestore_client.h"
+#include "walproposer.h"
 
 #define PageStoreTrace DEBUG5
 
@@ -62,8 +58,8 @@ char	   *neon_auth_token;
 int			readahead_buffer_size = 128;
 int			flush_every_n_requests = 8;
 
-int			n_reconnect_attempts = 0;
-int			max_reconnect_attempts = 60;
+static int n_reconnect_attempts = 0;
+static int max_reconnect_attempts = 60;
 
 #define MAX_PAGESERVER_CONNSTRING_SIZE 256
 
@@ -83,8 +79,6 @@ static PagestoreShmemState *pagestore_shared;
 static uint64 pagestore_local_counter = 0;
 static char local_pageserver_connstring[MAX_PAGESERVER_CONNSTRING_SIZE];
 
-bool		(*old_redo_read_buffer_filter) (XLogReaderState *record, uint8 block_id) = NULL;
-
 static bool pageserver_flush(void);
 static void pageserver_disconnect(void);
 
@@ -627,8 +621,6 @@ pg_init_libpagestore(void)
 		smgr_hook = smgr_neon;
 		smgr_init_hook = smgr_init_neon;
 		dbsize_hook = neon_dbsize;
-		old_redo_read_buffer_filter = redo_read_buffer_filter;
-		redo_read_buffer_filter = neon_redo_read_buffer_filter;
 	}
 
 	lfc_init();

pgxn/neon/libpqwalproposer.h

@@ -0,0 +1,96 @@
+/*
+ * Interface to set of libpq wrappers walproposer and neon_walreader need.
+ * Similar to libpqwalreceiver, but it has blocking connection establishment and
+ * pqexec which don't fit us. Implementation is at walproposer_pg.c.
+ */
+#ifndef ___LIBPQWALPROPOSER_H__
+#define ___LIBPQWALPROPOSER_H__
+
+/* Re-exported and modified ExecStatusType */
+typedef enum
+{
+	/* We received a single CopyBoth result */
+	WP_EXEC_SUCCESS_COPYBOTH,
+
+	/*
+	 * Any success result other than a single CopyBoth was received. The
+	 * specifics of the result were already logged, but it may be useful to
+	 * provide an error message indicating which safekeeper messed up.
+	 *
+	 * Do not expect PQerrorMessage to be appropriately set.
+	 */
+	WP_EXEC_UNEXPECTED_SUCCESS,
+
+	/*
+	 * No result available at this time. Wait until read-ready, then call
+	 * again. Internally, this is returned when PQisBusy indicates that
+	 * PQgetResult would block.
+	 */
+	WP_EXEC_NEEDS_INPUT,
+	/* Catch-all failure. Check PQerrorMessage. */
+	WP_EXEC_FAILED,
+} WalProposerExecStatusType;
+
+/* Possible return values from walprop_async_read */
+typedef enum
+{
+	/* The full read was successful. buf now points to the data */
+	PG_ASYNC_READ_SUCCESS,
+
+	/*
+	 * The read is ongoing. Wait until the connection is read-ready, then try
+	 * again.
+	 */
+	PG_ASYNC_READ_TRY_AGAIN,
+	/* Reading failed. Check PQerrorMessage(conn) */
+	PG_ASYNC_READ_FAIL,
+} PGAsyncReadResult;
+
+/* Possible return values from walprop_async_write */
+typedef enum
+{
+	/* The write fully completed */
+	PG_ASYNC_WRITE_SUCCESS,
+
+	/*
+	 * The write started, but you'll need to call PQflush some more times to
+	 * finish it off. We just tried, so it's best to wait until the connection
+	 * is read- or write-ready to try again.
+	 *
+	 * If it becomes read-ready, call PQconsumeInput and flush again. If it
+	 * becomes write-ready, just call PQflush.
+	 */
+	PG_ASYNC_WRITE_TRY_FLUSH,
+	/* Writing failed. Check PQerrorMessage(conn) */
+	PG_ASYNC_WRITE_FAIL,
+} PGAsyncWriteResult;
+
+/*
+ * This header is included by walproposer.h to define walproposer_api; if we're
+ * building walproposer without pg, ignore libpq part, leaving only interface
+ * types.
+ */
+#ifndef WALPROPOSER_LIB
+
+#include "libpq-fe.h"
+
+/*
+ * Sometimes working directly with underlying PGconn is simpler, export the
+ * whole thing for simplicity.
+ */
+typedef struct WalProposerConn
+{
+	PGconn	   *pg_conn;
+	bool		is_nonblocking; /* whether the connection is non-blocking */
+	char	   *recvbuf;		/* last received CopyData message from
+								 * walprop_async_read */
+} WalProposerConn;
+
+extern WalProposerConn *libpqwp_connect_start(char *conninfo);
+extern bool libpqwp_send_query(WalProposerConn *conn, char *query);
+extern WalProposerExecStatusType libpqwp_get_query_result(WalProposerConn *conn);
+extern PGAsyncReadResult libpqwp_async_read(WalProposerConn *conn, char **buf, int *amount);
+extern void libpqwp_disconnect(WalProposerConn *conn);
+
+#endif							/* WALPROPOSER_LIB */
+#endif							/* ___LIBPQWALPROPOSER_H__ */

pgxn/neon/neon.h

@@ -27,13 +27,6 @@ extern void pg_init_walproposer(void);
 
 extern void pg_init_extension_server(void);
 
-/*
- * Returns true if we shouldn't do REDO on that block in record indicated by
- * block_id; false otherwise.
- */
-extern bool neon_redo_read_buffer_filter(XLogReaderState *record, uint8 block_id);
-extern bool (*old_redo_read_buffer_filter) (XLogReaderState *record, uint8 block_id);
-
 extern uint64 BackpressureThrottlingTime(void);
 extern void replication_feedback_get_lsns(XLogRecPtr *writeLsn, XLogRecPtr *flushLsn, XLogRecPtr *applyLsn);
 

pgxn/neon/neon_utils.c

@@ -1,32 +1,10 @@
+
+#include <sys/resource.h>
+
 #include "postgres.h"
 
-#include "access/timeline.h"
-#include "access/xlogutils.h"
-#include "common/logging.h"
-#include "common/ip.h"
-#include "funcapi.h"
-#include "libpq/libpq.h"
+#include "lib/stringinfo.h"
 #include "libpq/pqformat.h"
-#include "miscadmin.h"
-#include "postmaster/interrupt.h"
-#include "replication/slot.h"
-#include "replication/walsender_private.h"
-
-#include "storage/ipc.h"
-#include "utils/builtins.h"
-#include "utils/ps_status.h"
-
-#include "libpq-fe.h"
-#include <netinet/tcp.h>
-#include <unistd.h>
-
-#if PG_VERSION_NUM >= 150000
-#include "access/xlogutils.h"
-#include "access/xlogrecovery.h"
-#endif
-#if PG_MAJORVERSION_NUM >= 16
-#include "utils/guc.h"
-#endif
 
 /*
  * Convert a character which represents a hexadecimal digit to an integer.
@@ -114,3 +92,25 @@ pq_sendint64_le(StringInfo buf, uint64 i)
 	memcpy(buf->data + buf->len, &i, sizeof(uint64));
 	buf->len += sizeof(uint64);
 }
+
+/*
+ * Disables core dump for the current process.
+ */
+void
+disable_core_dump()
+{
+	struct rlimit rlim;
+
+#ifdef WALPROPOSER_LIB			/* skip in simulation mode */
+	return;
+#endif
+
+	rlim.rlim_cur = 0;
+	rlim.rlim_max = 0;
+	if (setrlimit(RLIMIT_CORE, &rlim))
+	{
+		int			save_errno = errno;
+
+		fprintf(stderr, "WARNING: disable cores setrlimit failed: %s", strerror(save_errno));
+	}
+}

pgxn/neon/neon_utils.h

@@ -1,12 +1,11 @@
 #ifndef __NEON_UTILS_H__
 #define __NEON_UTILS_H__
 
-#include "postgres.h"
-
 bool		HexDecodeString(uint8 *result, char *input, int nbytes);
 uint32		pq_getmsgint32_le(StringInfo msg);
 uint64		pq_getmsgint64_le(StringInfo msg);
 void		pq_sendint32_le(StringInfo buf, uint32 i);
 void		pq_sendint64_le(StringInfo buf, uint64 i);
+extern void disable_core_dump();
 
 #endif							/* __NEON_UTILS_H__ */

pgxn/neon/neon_walreader.c

@@ -0,0 +1,742 @@
+/*
+ * Like WALRead, but when WAL segment doesn't exist locally instead of throwing
+ * ERROR asynchronously tries to fetch it from the most advanced safekeeper.
+ *
+ * We can't use libpqwalreceiver as it blocks during connection establishment
+ * (and waiting for PQExec result), so use libpqwalproposer instead.
+ *
+ * TODO: keepalives are currently never sent, so the other side can close the
+ * connection prematurely.
+ *
+ * TODO: close conn if reading takes too long to prevent stuck connections.
+ */
+#include "postgres.h"
+
+#include <sys/stat.h>
+#include <unistd.h>
+
+#include "access/xlog_internal.h"
+#include "access/xlogdefs.h"
+#include "access/xlogreader.h"
+#include "libpq/pqformat.h"
+#include "storage/fd.h"
+#include "utils/wait_event.h"
+
+#include "libpq-fe.h"
+
+#include "neon_walreader.h"
+#include "walproposer.h"
+
+#define NEON_WALREADER_ERR_MSG_LEN 512
+
+/*
+ * Can be called where NeonWALReader *state is available in the context, adds log_prefix.
+ */
+#define nwr_log(elevel, fmt, ...) elog(elevel, "%s" fmt, state->log_prefix, ## __VA_ARGS__)
+
+static NeonWALReadResult NeonWALReadRemote(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli);
+static NeonWALReadResult NeonWALReaderReadMsg(NeonWALReader *state);
+static void NeonWALReaderResetRemote(NeonWALReader *state);
+static bool NeonWALReadLocal(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli);
+static bool neon_wal_segment_open(NeonWALReader *state, XLogSegNo nextSegNo, TimeLineID *tli_p);
+static void neon_wal_segment_close(NeonWALReader *state);
+static bool is_wal_segment_exists(XLogSegNo segno, int segsize,
+								  TimeLineID tli);
+
+/*
+ * State of connection to donor safekeeper.
+ */
+typedef enum
+{
+	/* no remote connection */
+	RS_NONE,
+	/* doing PQconnectPoll, need readable socket */
+	RS_CONNECTING_READ,
+	/* doing PQconnectPoll, need writable socket */
+	RS_CONNECTING_WRITE,
+	/* Waiting for START_REPLICATION result */
+	RS_WAIT_EXEC_RESULT,
+	/* replication stream established */
+	RS_ESTABLISHED,
+} NeonWALReaderRemoteState;
+
+struct NeonWALReader
+{
+	/*
+	 * LSN before which we assume WAL is not available locally. Exists because
+	 * though first segment after startup always exists, part before
+	 * basebackup LSN is filled with zeros.
+	 */
+	XLogRecPtr	available_lsn;
+	WALSegmentContext segcxt;
+	WALOpenSegment seg;
+	int			wre_errno;
+	/* Explains failure to read, static for simplicity. */
+	char		err_msg[NEON_WALREADER_ERR_MSG_LEN];
+
+	/*
+	 * Saved info about request in progress, used to check validity of
+	 * arguments after resume and remember how far we accomplished it. req_lsn
+	 * is 0 if there is no request in progress.
+	 */
+	XLogRecPtr	req_lsn;
+	Size		req_len;
+	Size		req_progress;
+	WalProposer *wp;			/* we learn donor through walproposer */
+	char		donor_name[64]; /* saved donor safekeeper name for logging */
+	/* state of connection to safekeeper */
+	NeonWALReaderRemoteState rem_state;
+	WalProposerConn *wp_conn;
+
+	/*
+	 * position in wp_conn recvbuf from which we'll copy WAL next time, or
+	 * NULL if there is no unprocessed message
+	 */
+	char	   *wal_ptr;
+	Size		wal_rem_len;	/* how many unprocessed bytes left in recvbuf */
+
+	/*
+	 * LSN of wal_ptr position according to walsender to cross check against
+	 * read request
+	 */
+	XLogRecPtr	rem_lsn;
+
+	/* prepended to lines logged by neon_walreader, if provided */
+	char		log_prefix[64];
+};
+
+/* palloc and initialize NeonWALReader */
+NeonWALReader *
+NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, WalProposer *wp, char *log_prefix)
+{
+	NeonWALReader *reader;
+
+	reader = (NeonWALReader *)
+		palloc_extended(sizeof(NeonWALReader),
+						MCXT_ALLOC_NO_OOM | MCXT_ALLOC_ZERO);
+	if (!reader)
+		return NULL;
+
+	reader->available_lsn = available_lsn;
+	reader->seg.ws_file = -1;
+	reader->seg.ws_segno = 0;
+	reader->seg.ws_tli = 0;
+	reader->segcxt.ws_segsize = wal_segment_size;
+
+	reader->wp = wp;
+
+	reader->rem_state = RS_NONE;
+
+	if (log_prefix)
+		strlcpy(reader->log_prefix, log_prefix, sizeof(reader->log_prefix));
+
+	return reader;
+}
+
+void
+NeonWALReaderFree(NeonWALReader *state)
+{
+	if (state->seg.ws_file != -1)
+		neon_wal_segment_close(state);
+	if (state->wp_conn)
+		libpqwp_disconnect(state->wp_conn);
+	pfree(state);
+}
+
+/*
+ * Like vanilla WALRead, but if requested position is before available_lsn or
+ * WAL segment doesn't exist on disk, it tries to fetch needed segment from the
+ * advanced safekeeper.
+ *
+ * Read 'count' bytes into 'buf', starting at location 'startptr', from WAL
+ * fetched from timeline 'tli'.
+ *
+ * Returns NEON_WALREAD_SUCCESS if succeeded, NEON_WALREAD_ERROR if an error
+ * occurs, in which case 'err' has the desciption. Error always closes remote
+ * connection, if there was any, so socket subscription should be removed.
+ *
+ * NEON_WALREAD_WOULDBLOCK means caller should obtain socket to wait for with
+ * NeonWALReaderSocket and call NeonWALRead again with exactly the same
+ * arguments when NeonWALReaderEvents happen on the socket. Note that per libpq
+ * docs during connection establishment (before first successful read) socket
+ * underneath might change.
+ *
+ * Also, eventually walreader should switch from remote to local read; caller
+ * should remove subscription to socket then by checking NeonWALReaderEvents
+ * after successful read (otherwise next read might reopen the connection with
+ * different socket).
+ *
+ * Reading not monotonically is not supported and will result in error.
+ *
+ * Caller should be sure that WAL up to requested LSN exists, otherwise
+ * NEON_WALREAD_WOULDBLOCK might be always returned.
+ */
+NeonWALReadResult
+NeonWALRead(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli)
+{
+	/*
+	 * If requested data is before known available basebackup lsn or there is
+	 * already active remote state, do remote read.
+	 */
+	if (startptr < state->available_lsn || state->rem_state != RS_NONE)
+	{
+		return NeonWALReadRemote(state, buf, startptr, count, tli);
+	}
+	if (NeonWALReadLocal(state, buf, startptr, count, tli))
+	{
+		return NEON_WALREAD_SUCCESS;
+	}
+	else if (state->wre_errno == ENOENT)
+	{
+		nwr_log(LOG, "local read failed as segment at %X/%X doesn't exist, attempting remote",
+				LSN_FORMAT_ARGS(startptr));
+		return NeonWALReadRemote(state, buf, startptr, count, tli);
+	}
+	else
+	{
+		return NEON_WALREAD_ERROR;
+	}
+}
+
+/* Do the read from remote safekeeper. */
+static NeonWALReadResult
+NeonWALReadRemote(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli)
+{
+	if (state->rem_state == RS_NONE)
+	{
+		XLogRecPtr	donor_lsn;
+
+		/* no connection yet; start one */
+		Safekeeper *donor = GetDonor(state->wp, &donor_lsn);
+
+		if (donor == NULL)
+		{
+			snprintf(state->err_msg, sizeof(state->err_msg),
+					 "failed to establish remote connection to fetch WAL: no donor available");
+			return NEON_WALREAD_ERROR;
+		}
+		snprintf(state->donor_name, sizeof(state->donor_name), "%s:%s", donor->host, donor->port);
+		nwr_log(LOG, "establishing connection to %s, flush_lsn %X/%X to fetch WAL",
+				state->donor_name, LSN_FORMAT_ARGS(donor_lsn));
+		state->wp_conn = libpqwp_connect_start(donor->conninfo);
+		if (PQstatus(state->wp_conn->pg_conn) == CONNECTION_BAD)
+		{
+			snprintf(state->err_msg, sizeof(state->err_msg),
+					 "failed to connect to %s to fetch WAL: immediately failed with %s",
+					 state->donor_name, PQerrorMessage(state->wp_conn->pg_conn));
+			NeonWALReaderResetRemote(state);
+			return NEON_WALREAD_ERROR;
+		}
+		/* we'll poll immediately */
+		state->rem_state = RS_CONNECTING_READ;
+	}
+
+	if (state->rem_state == RS_CONNECTING_READ || state->rem_state == RS_CONNECTING_WRITE)
+	{
+		switch (PQconnectPoll(state->wp_conn->pg_conn))
+		{
+			case PGRES_POLLING_FAILED:
+				snprintf(state->err_msg, sizeof(state->err_msg),
+						 "failed to connect to %s to fetch WAL: poll error: %s",
+						 state->donor_name, PQerrorMessage(state->wp_conn->pg_conn));
+				NeonWALReaderResetRemote(state);
+				return NEON_WALREAD_ERROR;
+			case PGRES_POLLING_READING:
+				state->rem_state = RS_CONNECTING_READ;
+				return NEON_WALREAD_WOULDBLOCK;
+			case PGRES_POLLING_WRITING:
+				state->rem_state = RS_CONNECTING_WRITE;
+				return NEON_WALREAD_WOULDBLOCK;
+			case PGRES_POLLING_OK:
+				{
+					/* connection successfully established */
+					char		start_repl_query[128];
+
+					snprintf(start_repl_query, sizeof(start_repl_query),
+							 "START_REPLICATION PHYSICAL %X/%X (term='" UINT64_FORMAT "')",
+							 LSN_FORMAT_ARGS(startptr), state->wp->propTerm);
+					nwr_log(LOG, "connection to %s to fetch WAL succeeded, running %s",
+							state->donor_name, start_repl_query);
+					if (!libpqwp_send_query(state->wp_conn, start_repl_query))
+					{
+						snprintf(state->err_msg, sizeof(state->err_msg),
+								 "failed to send %s query to %s: %s",
+								 start_repl_query, state->donor_name, PQerrorMessage(state->wp_conn->pg_conn));
+						NeonWALReaderResetRemote(state);
+						return NEON_WALREAD_ERROR;
+					}
+					state->rem_state = RS_WAIT_EXEC_RESULT;
+					break;
+				}
+
+			default:			/* there is unused PGRES_POLLING_ACTIVE */
+				Assert(false);
+				return NEON_WALREAD_ERROR;	/* keep the compiler quiet */
+		}
+	}
+
+	if (state->rem_state == RS_WAIT_EXEC_RESULT)
+	{
+		switch (libpqwp_get_query_result(state->wp_conn))
+		{
+			case WP_EXEC_SUCCESS_COPYBOTH:
+				state->rem_state = RS_ESTABLISHED;
+				break;
+			case WP_EXEC_NEEDS_INPUT:
+				return NEON_WALREAD_WOULDBLOCK;
+			case WP_EXEC_FAILED:
+				snprintf(state->err_msg, sizeof(state->err_msg),
+						 "get START_REPLICATION result from %s failed: %s",
+						 state->donor_name, PQerrorMessage(state->wp_conn->pg_conn));
+				NeonWALReaderResetRemote(state);
+				return NEON_WALREAD_ERROR;
+			default:			/* can't happen */
+				snprintf(state->err_msg, sizeof(state->err_msg),
+						 "get START_REPLICATION result from %s: unexpected result",
+						 state->donor_name);
+				NeonWALReaderResetRemote(state);
+				return NEON_WALREAD_ERROR;
+		}
+	}
+
+	Assert(state->rem_state == RS_ESTABLISHED);
+
+	/*
+	 * If we had the request before, verify args are the same and advance the
+	 * result ptr according to the progress; otherwise register the request.
+	 */
+	if (state->req_lsn != InvalidXLogRecPtr)
+	{
+		if (state->req_lsn != startptr || state->req_len != count)
+		{
+			snprintf(state->err_msg, sizeof(state->err_msg),
+					 "args changed during request, was %X/%X %zu, now %X/%X %zu",
+					 LSN_FORMAT_ARGS(state->req_lsn), state->req_len, LSN_FORMAT_ARGS(startptr), count);
+			NeonWALReaderResetRemote(state);
+			return NEON_WALREAD_ERROR;
+		}
+		nwr_log(DEBUG5, "continuing remote read at req_lsn=%X/%X len=%zu, req_progress=%zu",
+				LSN_FORMAT_ARGS(startptr),
+				count,
+				state->req_progress);
+		buf += state->req_progress;
+	}
+	else
+	{
+		state->req_lsn = startptr;
+		state->req_len = count;
+		state->req_progress = 0;
+		nwr_log(DEBUG5, "starting remote read req_lsn=%X/%X len=%zu",
+				LSN_FORMAT_ARGS(startptr),
+				count);
+	}
+
+	while (true)
+	{
+		Size		to_copy;
+
+		/*
+		 * If we have no ready data, receive new message.
+		 */
+		if (state->wal_rem_len == 0 &&
+
+		/*
+		 * check for the sake of 0 length reads; walproposer does these for
+		 * heartbeats, though generally they shouldn't hit remote source.
+		 */
+			state->req_len - state->req_progress > 0)
+		{
+			NeonWALReadResult read_msg_res = NeonWALReaderReadMsg(state);
+
+			if (read_msg_res != NEON_WALREAD_SUCCESS)
+				return read_msg_res;
+		}
+
+		if (state->req_lsn + state->req_progress != state->rem_lsn)
+		{
+			snprintf(state->err_msg, sizeof(state->err_msg),
+					 "expected remote WAL at %X/%X but got %X/%X. Non monotonic read requests could have caused this. req_lsn=%X/%X len=%zu",
+					 LSN_FORMAT_ARGS(state->req_lsn + state->req_progress),
+					 LSN_FORMAT_ARGS(state->rem_lsn),
+					 LSN_FORMAT_ARGS(state->req_lsn),
+					 state->req_len);
+			NeonWALReaderResetRemote(state);
+			return NEON_WALREAD_ERROR;
+		}
+
+		/* We can copy min of (available, requested) bytes. */
+		to_copy =
+			Min(state->req_len - state->req_progress, state->wal_rem_len);
+		memcpy(buf, state->wal_ptr, to_copy);
+		state->wal_ptr += to_copy;
+		state->wal_rem_len -= to_copy;
+		state->rem_lsn += to_copy;
+		if (state->wal_rem_len == 0)
+			state->wal_ptr = NULL;	/* freed by libpqwalproposer */
+		buf += to_copy;
+		state->req_progress += to_copy;
+		if (state->req_progress == state->req_len)
+		{
+			XLogSegNo	next_segno;
+			XLogSegNo	req_segno;
+
+			XLByteToSeg(state->req_lsn, req_segno, state->segcxt.ws_segsize);
+			XLByteToSeg(state->rem_lsn, next_segno, state->segcxt.ws_segsize);
+
+			/*
+			 * Request completed. If there is a chance of serving next one
+			 * locally, close the connection.
+			 */
+			if (state->req_lsn < state->available_lsn &&
+				state->rem_lsn >= state->available_lsn)
+			{
+				nwr_log(LOG, "closing remote connection as available_lsn %X/%X crossed and next read at %X/%X is likely to be served locally",
+						LSN_FORMAT_ARGS(state->available_lsn), LSN_FORMAT_ARGS(state->rem_lsn));
+				NeonWALReaderResetRemote(state);
+			}
+			else if (state->rem_lsn >= state->available_lsn && next_segno > req_segno &&
+					 is_wal_segment_exists(next_segno, state->segcxt.ws_segsize, tli))
+			{
+				nwr_log(LOG, "closing remote connection as WAL file at next lsn %X/%X exists",
+						LSN_FORMAT_ARGS(state->rem_lsn));
+				NeonWALReaderResetRemote(state);
+			}
+			state->req_lsn = InvalidXLogRecPtr;
+			state->req_len = 0;
+			state->req_progress = 0;
+			return NEON_WALREAD_SUCCESS;
+		}
+	}
+}
+
+/*
+ * Read one WAL message from the stream, sets state->wal_ptr in case of success.
+ * Resets remote state in case of failure.
+ */
+static NeonWALReadResult
+NeonWALReaderReadMsg(NeonWALReader *state)
+{
+	while (true)				/* loop until we get 'w' */
+	{
+		char	   *copydata_ptr;
+		int			copydata_size;
+		StringInfoData s;
+		char		msg_type;
+		int			hdrlen;
+
+		Assert(state->rem_state == RS_ESTABLISHED);
+		Assert(state->wal_ptr == NULL && state->wal_rem_len == 0);
+
+		switch (libpqwp_async_read(state->wp_conn,
+								   &copydata_ptr,
+								   &copydata_size))
+		{
+			case PG_ASYNC_READ_SUCCESS:
+				break;
+			case PG_ASYNC_READ_TRY_AGAIN:
+				return NEON_WALREAD_WOULDBLOCK;
+			case PG_ASYNC_READ_FAIL:
+				snprintf(state->err_msg,
+						 sizeof(state->err_msg),
+						 "req_lsn=%X/%X, req_len=%zu, req_progress=%zu, get copydata failed: %s",
+						 LSN_FORMAT_ARGS(state->req_lsn),
+						 state->req_len,
+						 state->req_progress,
+						 PQerrorMessage(state->wp_conn->pg_conn));
+				goto err;
+		}
+
+		/* put data on StringInfo to parse */
+		s.data = copydata_ptr;
+		s.len = copydata_size;
+		s.cursor = 0;
+		s.maxlen = -1;
+
+		if (copydata_size == 0)
+		{
+			snprintf(state->err_msg,
+					 sizeof(state->err_msg),
+					 "zero length copydata received");
+			goto err;
+		}
+		msg_type = pq_getmsgbyte(&s);
+		switch (msg_type)
+		{
+			case 'w':
+				{
+					XLogRecPtr	start_lsn;
+
+					hdrlen = sizeof(int64) + sizeof(int64) + sizeof(int64);
+					if (s.len - s.cursor < hdrlen)
+					{
+						snprintf(state->err_msg,
+								 sizeof(state->err_msg),
+								 "invalid WAL message received from primary");
+						goto err;
+					}
+
+					start_lsn = pq_getmsgint64(&s);
+					pq_getmsgint64(&s); /* XLogRecPtr	end_lsn; */
+					pq_getmsgint64(&s); /* TimestampTz send_time */
+
+					state->rem_lsn = start_lsn;
+					state->wal_rem_len = (Size) (s.len - s.cursor);
+					state->wal_ptr = (char *) pq_getmsgbytes(&s, s.len - s.cursor);
+					nwr_log(DEBUG5, "received WAL msg at %X/%X len %zu",
+							LSN_FORMAT_ARGS(state->rem_lsn), state->wal_rem_len);
+
+					return NEON_WALREAD_SUCCESS;
+				}
+			case 'k':
+				{
+					XLogRecPtr	end_lsn;
+					bool		reply_requested;
+
+					hdrlen = sizeof(int64) + sizeof(int64) + sizeof(char);
+					if (s.len - s.cursor < hdrlen)
+					{
+						snprintf(state->err_msg, sizeof(state->err_msg),
+								 "invalid keepalive message received from primary");
+						goto err;
+					}
+
+					end_lsn = pq_getmsgint64(&s);
+					pq_getmsgint64(&s); /* TimestampTz timestamp; */
+					reply_requested = pq_getmsgbyte(&s);
+					nwr_log(DEBUG5, "received keepalive end_lsn=%X/%X reply_requested=%d",
+							LSN_FORMAT_ARGS(end_lsn),
+							reply_requested);
+					if (end_lsn < state->req_lsn + state->req_len)
+					{
+						snprintf(state->err_msg, sizeof(state->err_msg),
+								 "closing remote connection: requested WAL up to %X/%X, but current donor %s has only up to %X/%X",
+								 LSN_FORMAT_ARGS(state->req_lsn + state->req_len), state->donor_name, LSN_FORMAT_ARGS(end_lsn));
+						goto err;
+					}
+					continue;
+				}
+			default:
+				nwr_log(WARNING, "invalid replication message type %d", msg_type);
+				continue;
+		}
+	}
+err:
+	NeonWALReaderResetRemote(state);
+	return NEON_WALREAD_ERROR;
+}
+
+/* reset remote connection and request in progress */
+static void
+NeonWALReaderResetRemote(NeonWALReader *state)
+{
+	state->req_lsn = InvalidXLogRecPtr;
+	state->req_len = 0;
+	state->req_progress = 0;
+	state->rem_state = RS_NONE;
+	if (state->wp_conn)
+	{
+		libpqwp_disconnect(state->wp_conn);
+		state->wp_conn = NULL;
+	}
+	state->donor_name[0] = '\0';
+	state->wal_ptr = NULL;
+	state->wal_rem_len = 0;
+	state->rem_lsn = InvalidXLogRecPtr;
+}
+
+/*
+ * Return socket of connection to remote source. Must be called only when
+ * connection exists (NeonWALReaderEvents returns non zero).
+ */
+pgsocket
+NeonWALReaderSocket(NeonWALReader *state)
+{
+	if (!state->wp_conn)
+		nwr_log(FATAL, "NeonWALReaderSocket is called without active remote connection");
+	return PQsocket(state->wp_conn->pg_conn);
+}
+
+/*
+ * Whether remote connection is established. Once this is done, until successful
+ * local read or error socket is stable and user can update socket events
+ * instead of readding it each time.
+ */
+bool
+NeonWALReaderIsRemConnEstablished(NeonWALReader *state)
+{
+	return state->rem_state == RS_ESTABLISHED;
+}
+
+/*
+ * Returns events user should wait on connection socket or 0 if remote
+ * connection is not active.
+ */
+extern uint32
+NeonWALReaderEvents(NeonWALReader *state)
+{
+	switch (state->rem_state)
+	{
+		case RS_NONE:
+			return 0;
+		case RS_CONNECTING_READ:
+			return WL_SOCKET_READABLE;
+		case RS_CONNECTING_WRITE:
+			return WL_SOCKET_WRITEABLE;
+		case RS_WAIT_EXEC_RESULT:
+		case RS_ESTABLISHED:
+			return WL_SOCKET_READABLE;
+		default:
+			Assert(false);
+			return 0;			/* make compiler happy */
+	}
+}
+
+static bool
+NeonWALReadLocal(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli)
+{
+	char	   *p;
+	XLogRecPtr	recptr;
+	Size		nbytes;
+
+	p = buf;
+	recptr = startptr;
+	nbytes = count;
+
+	while (nbytes > 0)
+	{
+		uint32		startoff;
+		int			segbytes;
+		int			readbytes;
+
+		startoff = XLogSegmentOffset(recptr, state->segcxt.ws_segsize);
+
+		/*
+		 * If the data we want is not in a segment we have open, close what we
+		 * have (if anything) and open the next one, using the caller's
+		 * provided openSegment callback.
+		 */
+		if (state->seg.ws_file < 0 ||
+			!XLByteInSeg(recptr, state->seg.ws_segno, state->segcxt.ws_segsize) ||
+			tli != state->seg.ws_tli)
+		{
+			XLogSegNo	nextSegNo;
+
+			neon_wal_segment_close(state);
+
+			XLByteToSeg(recptr, nextSegNo, state->segcxt.ws_segsize);
+			if (!neon_wal_segment_open(state, nextSegNo, &tli))
+			{
+				char		fname[MAXFNAMELEN];
+
+				state->wre_errno = errno;
+
+				XLogFileName(fname, tli, nextSegNo, state->segcxt.ws_segsize);
+				snprintf(state->err_msg, sizeof(state->err_msg), "failed to open WAL segment %s while reading at %X/%X: %s",
+						 fname, LSN_FORMAT_ARGS(recptr), strerror(state->wre_errno));
+				return false;
+			}
+
+			/* This shouldn't happen -- indicates a bug in segment_open */
+			Assert(state->seg.ws_file >= 0);
+
+			/* Update the current segment info. */
+			state->seg.ws_tli = tli;
+			state->seg.ws_segno = nextSegNo;
+		}
+
+		/* How many bytes are within this segment? */
+		if (nbytes > (state->segcxt.ws_segsize - startoff))
+			segbytes = state->segcxt.ws_segsize - startoff;
+		else
+			segbytes = nbytes;
+
+#ifndef FRONTEND
+		pgstat_report_wait_start(WAIT_EVENT_WAL_READ);
+#endif
+
+		/* Reset errno first; eases reporting non-errno-affecting errors */
+		errno = 0;
+		readbytes = pg_pread(state->seg.ws_file, p, segbytes, (off_t) startoff);
+
+#ifndef FRONTEND
+		pgstat_report_wait_end();
+#endif
+
+		if (readbytes <= 0)
+		{
+			char		fname[MAXFNAMELEN];
+
+			XLogFileName(fname, state->seg.ws_tli, state->seg.ws_segno, state->segcxt.ws_segsize);
+
+			if (readbytes < 0)
+			{
+				state->wre_errno = errno;
+				snprintf(state->err_msg, sizeof(state->err_msg), "could not read from log segment %s, offset %d: %m: %s",
+						 fname, startoff, strerror(state->wre_errno));
+			}
+			else
+			{
+				snprintf(state->err_msg, sizeof(state->err_msg), "could not read from log segment %s, offset %d: %m: unexpected EOF",
+						 fname, startoff);
+			}
+			return false;
+		}
+
+		/* Update state for read */
+		recptr += readbytes;
+		nbytes -= readbytes;
+		p += readbytes;
+	}
+
+	return true;
+}
+
+/*
+ * Copy of vanilla wal_segment_open, but returns false in case of error instead
+ * of ERROR, with errno set.
+ *
+ * XLogReaderRoutine->segment_open callback for local pg_wal files
+ */
+static bool
+neon_wal_segment_open(NeonWALReader *state, XLogSegNo nextSegNo,
+					  TimeLineID *tli_p)
+{
+	TimeLineID	tli = *tli_p;
+	char		path[MAXPGPATH];
+
+	XLogFilePath(path, tli, nextSegNo, state->segcxt.ws_segsize);
+	nwr_log(DEBUG5, "opening %s", path);
+	state->seg.ws_file = BasicOpenFile(path, O_RDONLY | PG_BINARY);
+	if (state->seg.ws_file >= 0)
+		return true;
+
+	return false;
+}
+
+static bool
+is_wal_segment_exists(XLogSegNo segno, int segsize, TimeLineID tli)
+{
+	struct stat stat_buffer;
+	char		path[MAXPGPATH];
+
+	XLogFilePath(path, tli, segno, segsize);
+	return stat(path, &stat_buffer) == 0;
+}
+
+/* copy of vanilla wal_segment_close with NeonWALReader */
+static void
+neon_wal_segment_close(NeonWALReader *state)
+{
+	if (state->seg.ws_file >= 0)
+	{
+		close(state->seg.ws_file);
+		/* need to check errno? */
+		state->seg.ws_file = -1;
+	}
+}
+
+char *
+NeonWALReaderErrMsg(NeonWALReader *state)
+{
+	return state->err_msg;
+}

pgxn/neon/neon_walreader.h

@@ -0,0 +1,30 @@
+#ifndef __NEON_WALREADER_H__
+#define __NEON_WALREADER_H__
+
+#include "access/xlogdefs.h"
+
+/* forward declare so we don't have to expose the struct to the public */
+struct NeonWALReader;
+typedef struct NeonWALReader NeonWALReader;
+
+/* avoid including walproposer.h as it includes us */
+struct WalProposer;
+typedef struct WalProposer WalProposer;
+
+/* NeonWALRead return value */
+typedef enum
+{
+	NEON_WALREAD_SUCCESS,
+	NEON_WALREAD_WOULDBLOCK,
+	NEON_WALREAD_ERROR,
+} NeonWALReadResult;
+
+extern NeonWALReader *NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, WalProposer *wp, char *log_prefix);
+extern void NeonWALReaderFree(NeonWALReader *state);
+extern NeonWALReadResult NeonWALRead(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli);
+extern pgsocket NeonWALReaderSocket(NeonWALReader *state);
+extern uint32 NeonWALReaderEvents(NeonWALReader *state);
+extern bool NeonWALReaderIsRemConnEstablished(NeonWALReader *state);
+extern char *NeonWALReaderErrMsg(NeonWALReader *state);
+
+#endif							/* __NEON_WALREADER_H__ */

pgxn/neon/pagestore_client.h

@@ -13,19 +13,16 @@
 #ifndef pageserver_h
 #define pageserver_h
 
-#include "postgres.h"
 #include "neon_pgversioncompat.h"
 
 #include "access/xlogdefs.h"
 #include RELFILEINFO_HDR
-#include "storage/block.h"
-#include "storage/smgr.h"
 #include "lib/stringinfo.h"
 #include "libpq/pqformat.h"
+#include "storage/block.h"
+#include "storage/smgr.h"
 #include "utils/memutils.h"
 
-#include "pg_config.h"
-
 typedef enum
 {
 	/* pagestore_client -> pagestore */
@@ -158,11 +155,8 @@ extern page_server_api *page_server;
 extern char *page_server_connstring;
 extern int	flush_every_n_requests;
 extern int	readahead_buffer_size;
-extern bool seqscan_prefetch_enabled;
-extern int	seqscan_prefetch_distance;
 extern char *neon_timeline;
 extern char *neon_tenant;
-extern bool wal_redo;
 extern int32 max_cluster_size;
 
 extern const f_smgr *smgr_neon(BackendId backend, NRelFileInfo rinfo);