fmt

## Problem

Usually, the connection itself is quite fast (bellow 10ms for p999:
https://neonprod.grafana.net/goto/aOyn8vYIg?orgId=1).

It doesn't make a lot of sense to wait for a lot of time for the lock,
if it takes a lot of time to acquire it, probably, something goes wrong.

We also spawn a lot of retries, but they are not super helpful (0 means
that it was connected successfully, 1, most probably, that it was
re-request of the compute node address
https://neonprod.grafana.net/goto/J_8VQvLIR?orgId=1). Let's try to keep
a small number of retries.

.dockerignore

@@ -22,6 +22,7 @@
 !s3_scrubber/
 !safekeeper/
 !storage_broker/
+!storage_controller/
 !trace/
 !vendor/postgres-*/
 !workspace_hack/

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

@@ -150,7 +150,7 @@ runs:
 
         # Use aws s3 cp (instead of aws s3 sync) to keep files from previous runs to make old URLs work,
         # and to keep files on the host to upload them to the database
-        time aws s3 cp --recursive --only-show-errors "${WORKDIR}/report" "s3://${BUCKET}/${REPORT_PREFIX}/${GITHUB_RUN_ID}"
+        time s5cmd --log error cp "${WORKDIR}/report/*" "s3://${BUCKET}/${REPORT_PREFIX}/${GITHUB_RUN_ID}/"
 
         # Generate redirect
         cat <<EOF > ${WORKDIR}/index.html

.github/actions/neon-branch-create/action.yml

@@ -10,7 +10,7 @@ inputs:
     required: true
   api_host:
     desctiption: 'Neon API host'
-    default: console.stage.neon.tech
+    default: console-stage.neon.build
 outputs:
   dsn:
     description: 'Created Branch DSN (for main database)'

.github/actions/neon-branch-delete/action.yml

@@ -13,7 +13,7 @@ inputs:
     required: true
   api_host:
     desctiption: 'Neon API host'
-    default: console.stage.neon.tech
+    default: console-stage.neon.build
 
 runs:
   using: "composite"

.github/actions/neon-project-create/action.yml

@@ -13,7 +13,7 @@ inputs:
     default: 15
   api_host:
     desctiption: 'Neon API host'
-    default: console.stage.neon.tech
+    default: console-stage.neon.build
   provisioner:
     desctiption: 'k8s-pod or k8s-neonvm'
     default: 'k8s-pod'

.github/actions/neon-project-delete/action.yml

@@ -10,7 +10,7 @@ inputs:
     required: true
   api_host:
     desctiption: 'Neon API host'
-    default: console.stage.neon.tech
+    default: console-stage.neon.build
 
 runs:
   using: "composite"

.github/workflows/approved-for-ci-run.yml

@@ -18,6 +18,7 @@ on:
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number }}
+  cancel-in-progress: false
 
 env:
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/benchmarking.yml

@@ -147,15 +147,16 @@ jobs:
             "neonvm-captest-new"
           ],
           "db_size": [ "10gb" ],
-          "include": [{ "platform": "neon-captest-freetier",   "db_size": "3gb"  },
-                      { "platform": "neon-captest-new",        "db_size": "50gb" },
-                      { "platform": "neonvm-captest-freetier", "db_size": "3gb"  },
-                      { "platform": "neonvm-captest-new",      "db_size": "50gb" }]
+          "include": [{ "platform": "neon-captest-freetier",         "db_size": "3gb"  },
+                      { "platform": "neon-captest-new",              "db_size": "50gb" },
+                      { "platform": "neonvm-captest-freetier",       "db_size": "3gb"  },
+                      { "platform": "neonvm-captest-new",            "db_size": "50gb" },
+                      { "platform": "neonvm-captest-sharding-reuse", "db_size": "50gb" }]
         }'
 
         if [ "$(date +%A)" = "Saturday" ]; then
           matrix=$(echo "$matrix" | jq '.include += [{ "platform": "rds-postgres", "db_size": "10gb"},
-                                                   { "platform": "rds-aurora",   "db_size": "50gb"}]')
+                                                     { "platform": "rds-aurora",   "db_size": "50gb"}]')
         fi
 
         echo "matrix=$(echo "$matrix" | jq --compact-output '.')" >> $GITHUB_OUTPUT
@@ -171,7 +172,7 @@ jobs:
 
         if [ "$(date +%A)" = "Saturday" ] || [ ${RUN_AWS_RDS_AND_AURORA} = "true" ]; then
           matrix=$(echo "$matrix" | jq '.include += [{ "platform": "rds-postgres" },
-                                                   { "platform": "rds-aurora"   }]')
+                                                     { "platform": "rds-aurora"   }]')
         fi
 
         echo "matrix=$(echo "$matrix" | jq --compact-output '.')" >> $GITHUB_OUTPUT
@@ -190,7 +191,7 @@ jobs:
 
         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 '.')" >> $GITHUB_OUTPUT
@@ -253,6 +254,9 @@ jobs:
           neon-captest-reuse)
             CONNSTR=${{ secrets.BENCHMARK_CAPTEST_CONNSTR }}
             ;;
+          neonvm-captest-sharding-reuse)
+            CONNSTR=${{ secrets.BENCHMARK_CAPTEST_SHARDING_CONNSTR }}
+            ;;
           neon-captest-new | neon-captest-freetier | neonvm-captest-new | neonvm-captest-freetier)
             CONNSTR=${{ steps.create-neon-project.outputs.dsn }}
             ;;
@@ -270,11 +274,15 @@ jobs:
 
         echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT
 
-        QUERY="SELECT version();"
+        QUERIES=("SELECT version()")
         if [[ "${PLATFORM}" = "neon"* ]]; then
-          QUERY="${QUERY} SHOW neon.tenant_id; SHOW neon.timeline_id;"
+          QUERIES+=("SHOW neon.tenant_id")
+          QUERIES+=("SHOW neon.timeline_id")
         fi
-        psql ${CONNSTR} -c "${QUERY}"
+
+        for q in "${QUERIES[@]}"; do
+          psql ${CONNSTR} -c "${q}"
+        done
 
     - name: Benchmark init
       uses: ./.github/actions/run-python-test-set
@@ -401,11 +409,15 @@ jobs:
 
         echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT
 
-        QUERY="SELECT version();"
+        QUERIES=("SELECT version()")
         if [[ "${PLATFORM}" = "neon"* ]]; then
-          QUERY="${QUERY} SHOW neon.tenant_id; SHOW neon.timeline_id;"
+          QUERIES+=("SHOW neon.tenant_id")
+          QUERIES+=("SHOW neon.timeline_id")
         fi
-        psql ${CONNSTR} -c "${QUERY}"
+
+        for q in "${QUERIES[@]}"; do
+          psql ${CONNSTR} -c "${q}"
+        done
 
     - name: ClickBench benchmark
       uses: ./.github/actions/run-python-test-set
@@ -507,11 +519,15 @@ jobs:
 
         echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT
 
-        QUERY="SELECT version();"
+        QUERIES=("SELECT version()")
         if [[ "${PLATFORM}" = "neon"* ]]; then
-          QUERY="${QUERY} SHOW neon.tenant_id; SHOW neon.timeline_id;"
+          QUERIES+=("SHOW neon.tenant_id")
+          QUERIES+=("SHOW neon.timeline_id")
         fi
-        psql ${CONNSTR} -c "${QUERY}"
+
+        for q in "${QUERIES[@]}"; do
+          psql ${CONNSTR} -c "${q}"
+        done
 
     - name: Run TPC-H benchmark
       uses: ./.github/actions/run-python-test-set
@@ -597,11 +613,15 @@ jobs:
 
         echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT
 
-        QUERY="SELECT version();"
+        QUERIES=("SELECT version()")
         if [[ "${PLATFORM}" = "neon"* ]]; then
-          QUERY="${QUERY} SHOW neon.tenant_id; SHOW neon.timeline_id;"
+          QUERIES+=("SHOW neon.tenant_id")
+          QUERIES+=("SHOW neon.timeline_id")
         fi
-        psql ${CONNSTR} -c "${QUERY}"
+
+        for q in "${QUERIES[@]}"; do
+          psql ${CONNSTR} -c "${q}"
+        done
 
     - name: Run user examples
       uses: ./.github/actions/run-python-test-set

.github/workflows/build-build-tools-image.yml

@@ -21,6 +21,7 @@ defaults:
 
 concurrency:
   group: build-build-tools-image-${{ inputs.image-tag }}
+  cancel-in-progress: false
 
 # No permission for GITHUB_TOKEN by default; the **minimal required** set of permissions should be granted in each job.
 permissions: {}

.github/workflows/build_and_test.yml

@@ -461,6 +461,7 @@ jobs:
 
       - name: Pytest regression tests
         uses: ./.github/actions/run-python-test-set
+        timeout-minutes: 60
         with:
           build_type: ${{ matrix.build_type }}
           test_selection: regress
@@ -476,6 +477,8 @@ jobs:
           BUILD_TAG: ${{ needs.tag.outputs.build-tag }}
           PAGESERVER_VIRTUAL_FILE_IO_ENGINE: tokio-epoll-uring
           PAGESERVER_GET_VECTORED_IMPL: vectored
+          PAGESERVER_GET_IMPL: vectored
+          PAGESERVER_VALIDATE_VEC_GET: true
 
       # Temporary disable this step until we figure out why it's so flaky
       # Ref https://github.com/neondatabase/neon/issues/4540
@@ -555,6 +558,9 @@ jobs:
           PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
           TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}"
           PAGESERVER_VIRTUAL_FILE_IO_ENGINE: tokio-epoll-uring
+          PAGESERVER_GET_VECTORED_IMPL: vectored
+          PAGESERVER_GET_IMPL: vectored
+          PAGESERVER_VALIDATE_VEC_GET: false
       # XXX: no coverage data handling here, since benchmarks are run on release builds,
       # while coverage is currently collected for the debug ones
 
@@ -734,7 +740,7 @@ jobs:
         run: |
           mkdir -p .docker-custom
           echo DOCKER_CONFIG=$(pwd)/.docker-custom >> $GITHUB_ENV
-      - uses: docker/setup-buildx-action@v3
+      - uses: docker/setup-buildx-action@v2
 
       - uses: docker/login-action@v3
         with:
@@ -791,7 +797,7 @@ jobs:
         run: |
           mkdir -p .docker-custom
           echo DOCKER_CONFIG=$(pwd)/.docker-custom >> $GITHUB_ENV
-      - uses: docker/setup-buildx-action@v3
+      - uses: docker/setup-buildx-action@v2
         with:
           # Disable parallelism for docker buildkit.
           # As we already build everything with `make -j$(nproc)`, running it in additional level of parallelisam blows up the Runner.
@@ -864,7 +870,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.23.2
+      VM_BUILDER_VERSION: v0.28.1
 
     steps:
       - name: Checkout
@@ -1120,18 +1126,34 @@ jobs:
         run: |
           if [[ "$GITHUB_REF_NAME" == "main" ]]; then
             gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f deployPreprodRegion=false
-
-            # TODO: move deployPreprodRegion to release (`"$GITHUB_REF_NAME" == "release"` block), once Staging support different compute tag prefixes for different regions
-            gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f deployPreprodRegion=true
           elif [[ "$GITHUB_REF_NAME" == "release" ]]; then
-            gh workflow --repo neondatabase/aws run deploy-prod.yml --ref main \
+            gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main \
               -f deployPgSniRouter=false \
               -f deployProxy=false \
               -f deployStorage=true \
               -f deployStorageBroker=true \
+              -f deployStorageController=true \
+              -f branch=main \
+              -f dockerTag=${{needs.tag.outputs.build-tag}} \
+              -f deployPreprodRegion=true
+
+            gh workflow --repo neondatabase/aws run deploy-prod.yml --ref main \
+              -f deployStorage=true \
+              -f deployStorageBroker=true \
+              -f deployStorageController=true \
               -f branch=main \
               -f dockerTag=${{needs.tag.outputs.build-tag}}
           elif [[ "$GITHUB_REF_NAME" == "release-proxy" ]]; then
+            gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main \
+              -f deployPgSniRouter=true \
+              -f deployProxy=true \
+              -f deployStorage=false \
+              -f deployStorageBroker=false \
+              -f deployStorageController=false \
+              -f branch=main \
+              -f dockerTag=${{needs.tag.outputs.build-tag}} \
+              -f deployPreprodRegion=true
+
             gh workflow --repo neondatabase/aws run deploy-proxy-prod.yml --ref main \
               -f deployPgSniRouter=true \
               -f deployProxy=true \

.github/workflows/check-build-tools-image.yml

@@ -28,7 +28,9 @@ jobs:
       - name: Get build-tools image tag for the current commit
         id: get-build-tools-tag
         env:
-          COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
+          # Usually, for COMMIT_SHA, we use `github.event.pull_request.head.sha || github.sha`, but here, even for PRs,
+          # we want to use `github.sha` i.e. point to a phantom merge commit to determine the image tag correctly.
+          COMMIT_SHA: ${{ github.sha }}
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
           LAST_BUILD_TOOLS_SHA=$(

.github/workflows/pin-build-tools-image.yml

@@ -20,6 +20,7 @@ defaults:
 
 concurrency:
   group: pin-build-tools-image-${{ inputs.from-tag }}
+  cancel-in-progress: false
 
 permissions: {}
 

.github/workflows/trigger-e2e-tests.yml

@@ -62,14 +62,14 @@ jobs:
 
   trigger-e2e-tests:
     needs: [ tag ]
-    runs-on: [ self-hosted, gen3, small ]
+    runs-on: ubuntu-latest
     env:
       TAG: ${{ needs.tag.outputs.build-tag }}
-    container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
-      options: --init
     steps:
       - name: check if ecr image are present
+        env:
+          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_KEY_DEV }}
         run: |
           for REPO in neon compute-tools compute-node-v14 vm-compute-node-v14 compute-node-v15 vm-compute-node-v15 compute-node-v16 vm-compute-node-v16; do
             OUTPUT=$(aws ecr describe-images --repository-name ${REPO} --region eu-central-1 --query "imageDetails[?imageTags[?contains(@, '${TAG}')]]" --output text)
@@ -79,41 +79,55 @@ jobs:
             fi
           done
 
-      - name: Set PR's status to pending and request a remote CI test
+      - name: Set e2e-platforms
+        id: e2e-platforms
+        env:
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
-          # For pull requests, GH Actions set "github.sha" variable to point at a fake merge commit
-          # but we need to use a real sha of a latest commit in the PR's branch for the e2e job,
-          # to place a job run status update later.
-          COMMIT_SHA=${{ github.event.pull_request.head.sha }}
-          # For non-PR kinds of runs, the above will produce an empty variable, pick the original sha value for those
-          COMMIT_SHA=${COMMIT_SHA:-${{ github.sha }}}
+          # Default set of platforms to run e2e tests on
+          platforms='["docker", "k8s"]'
 
-          REMOTE_REPO="${{ github.repository_owner }}/cloud"
+          # If the PR changes vendor/, pgxn/ or libs/vm_monitor/ directories, or Dockerfile.compute-node, add k8s-neonvm to the list of platforms.
+          # If the workflow run is not a pull request, add k8s-neonvm to the list.
+          if [ "$GITHUB_EVENT_NAME" == "pull_request" ]; then
+            for f in $(gh api "/repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}/files" --paginate --jq '.[].filename'); do
+              case "$f" in
+                vendor/*|pgxn/*|libs/vm_monitor/*|Dockerfile.compute-node)
+                  platforms=$(echo "${platforms}" | jq --compact-output '. += ["k8s-neonvm"] | unique')
+                  ;;
+                *)
+                  # no-op
+                  ;;
+              esac
+            done
+          else
+            platforms=$(echo "${platforms}" | jq --compact-output '. += ["k8s-neonvm"] | unique')
+          fi
 
-          curl -f -X POST \
-          https://api.github.com/repos/${{ github.repository }}/statuses/$COMMIT_SHA \
-          -H "Accept: application/vnd.github.v3+json" \
-          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
-          --data \
-            "{
-              \"state\": \"pending\",
-              \"context\": \"neon-cloud-e2e\",
-              \"description\": \"[$REMOTE_REPO] Remote CI job is about to start\"
-            }"
+          echo "e2e-platforms=${platforms}" | tee -a $GITHUB_OUTPUT
 
-          curl -f -X POST \
-          https://api.github.com/repos/$REMOTE_REPO/actions/workflows/testing.yml/dispatches \
-          -H "Accept: application/vnd.github.v3+json" \
-          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
-          --data \
-            "{
-              \"ref\": \"main\",
-              \"inputs\": {
-                \"ci_job_name\": \"neon-cloud-e2e\",
-                \"commit_hash\": \"$COMMIT_SHA\",
-                \"remote_repo\": \"${{ github.repository }}\",
-                \"storage_image_tag\": \"${TAG}\",
-                \"compute_image_tag\": \"${TAG}\",
-                \"concurrency_group\": \"${{ env.E2E_CONCURRENCY_GROUP }}\"
-              }
-            }"
+      - name: Set PR's status to pending and request a remote CI test
+        env:
+          E2E_PLATFORMS: ${{ steps.e2e-platforms.outputs.e2e-platforms }}
+          COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
+          GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
+        run: |
+          REMOTE_REPO="${GITHUB_REPOSITORY_OWNER}/cloud"
+
+          gh api "/repos/${GITHUB_REPOSITORY}/statuses/${COMMIT_SHA}" \
+            --method POST \
+            --raw-field "state=pending" \
+            --raw-field "description=[$REMOTE_REPO] Remote CI job is about to start" \
+            --raw-field "context=neon-cloud-e2e"
+
+          gh workflow --repo ${REMOTE_REPO} \
+            run testing.yml \
+              --ref "main" \
+              --raw-field "ci_job_name=neon-cloud-e2e" \
+              --raw-field "commit_hash=$COMMIT_SHA" \
+              --raw-field "remote_repo=${GITHUB_REPOSITORY}" \
+              --raw-field "storage_image_tag=${TAG}" \
+              --raw-field "compute_image_tag=${TAG}" \
+              --raw-field "concurrency_group=${E2E_CONCURRENCY_GROUP}" \
+              --raw-field "e2e-platforms=${E2E_PLATFORMS}"

CODEOWNERS

@@ -1,12 +1,13 @@
 /compute_tools/ @neondatabase/control-plane @neondatabase/compute
-/control_plane/attachment_service @neondatabase/storage
+/storage_controller @neondatabase/storage
 /libs/pageserver_api/ @neondatabase/storage
-/libs/postgres_ffi/ @neondatabase/compute
+/libs/postgres_ffi/ @neondatabase/compute @neondatabase/safekeepers
 /libs/remote_storage/ @neondatabase/storage
 /libs/safekeeper_api/ @neondatabase/safekeepers
 /libs/vm_monitor/ @neondatabase/autoscaling
 /pageserver/ @neondatabase/storage
 /pgxn/ @neondatabase/compute
+/pgxn/neon/ @neondatabase/compute @neondatabase/safekeepers
 /proxy/ @neondatabase/proxy
 /safekeeper/ @neondatabase/safekeepers
 /vendor/ @neondatabase/compute

Cargo.toml

@@ -3,7 +3,7 @@ resolver = "2"
 members = [
     "compute_tools",
     "control_plane",
-    "control_plane/attachment_service",
+    "control_plane/storcon_cli",
     "pageserver",
     "pageserver/compaction",
     "pageserver/ctl",
@@ -12,6 +12,7 @@ members = [
     "proxy",
     "safekeeper",
     "storage_broker",
+    "storage_controller",
     "s3_scrubber",
     "workspace_hack",
     "trace",
@@ -43,19 +44,22 @@ license = "Apache-2.0"
 anyhow = { version = "1.0", features = ["backtrace"] }
 arc-swap = "1.6"
 async-compression = { version = "0.4.0", features = ["tokio", "gzip", "zstd"] }
-azure_core = "0.18"
-azure_identity = "0.18"
-azure_storage = "0.18"
-azure_storage_blobs = "0.18"
+atomic-take = "1.1.0"
+azure_core = "0.19"
+azure_identity = "0.19"
+azure_storage = "0.19"
+azure_storage_blobs = "0.19"
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
 aws-config = { version = "1.1.4", default-features = false, features=["rustls"] }
 aws-sdk-s3 = "1.14"
-aws-sdk-secretsmanager = { version = "1.14.0" }
+aws-sdk-iam = "1.15.0"
 aws-smithy-async = { version = "1.1.4", default-features = false, features=["rt-tokio"] }
 aws-smithy-types = "1.1.4"
 aws-credential-types = "1.1.4"
+aws-sigv4 = { version = "1.2.0", features = ["sign-http"] }
+aws-types = "1.1.7"
 axum = { version = "0.6.20", features = ["ws"] }
 base64 = "0.13.0"
 bincode = "1.3"
@@ -76,6 +80,7 @@ either = "1.8"
 enum-map = "2.4.2"
 enumset = "1.0.12"
 fail = "0.5.0"
+fallible-iterator = "0.2"
 fs2 = "0.4.3"
 futures = "0.3"
 futures-core = "0.3"
@@ -88,11 +93,12 @@ hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
 hostname = "0.3.1"
+http = {version = "1.1.0", features = ["std"]}
 http-types = { version = "2", default-features = false }
 humantime = "2.1"
 humantime-serde = "1.1.1"
 hyper = "0.14"
-hyper-tungstenite = "0.11"
+hyper-tungstenite = "0.13.0"
 inotify = "0.10.2"
 ipnet = "2.9.0"
 itertools = "0.10"
@@ -100,8 +106,9 @@ jsonwebtoken = "9"
 lasso = "0.7"
 leaky-bucket = "1.0.1"
 libc = "0.2"
-lz4_flex = "0.11.1"
 md5 = "0.7.0"
+measured = { version = "0.0.21", features=["lasso"] }
+measured-process = { version = "0.0.21" }
 memoffset = "0.8"
 native-tls = "0.2"
 nix = { version = "0.27", features = ["fs", "process", "socket", "signal", "poll"] }
@@ -121,12 +128,12 @@ procfs = "0.14"
 prometheus = {version = "0.13", default_features=false, features = ["process"]} # removes protobuf dependency
 prost = "0.11"
 rand = "0.8"
-redis = { version = "0.24.0", features = ["tokio-rustls-comp", "keep-alive"] }
+redis = { version = "0.25.2", features = ["tokio-rustls-comp", "keep-alive"] }
 regex = "1.10.2"
-reqwest = { version = "0.11", default-features = false, features = ["rustls-tls"] }
-reqwest-tracing = { version = "0.4.7", features = ["opentelemetry_0_20"] }
-reqwest-middleware = "0.2.0"
-reqwest-retry = "0.2.2"
+reqwest = { version = "0.12", default-features = false, features = ["rustls-tls"] }
+reqwest-tracing = { version = "0.5", features = ["opentelemetry_0_20"] }
+reqwest-middleware = "0.3.0"
+reqwest-retry = "0.5"
 routerify = "3"
 rpds = "0.13"
 rustc-hash = "1.1.0"
@@ -136,7 +143,7 @@ rustls-split = "0.3"
 scopeguard = "1.1"
 sysinfo = "0.29.2"
 sd-notify = "0.4.1"
-sentry = { version = "0.31", default-features = false, features = ["backtrace", "contexts", "panic", "rustls", "reqwest" ] }
+sentry = { version = "0.32", default-features = false, features = ["backtrace", "contexts", "panic", "rustls", "reqwest" ] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1"
 serde_path_to_error = "0.1"
@@ -149,11 +156,13 @@ smol_str = { version = "0.2.0", features = ["serde"] }
 socket2 = "0.5"
 strum = "0.24"
 strum_macros = "0.24"
-svg_fmt = "0.4.1"
+"subtle"  = "2.5.0"
+# https://github.com/nical/rust_debug/pull/4
+svg_fmt = { git = "https://github.com/neondatabase/fork--nical--rust_debug", branch = "neon" }
 sync_wrapper = "0.1.2"
 tar = "0.4"
 task-local-extensions = "0.1.4"
-test-context = "0.1"
+test-context = "0.3"
 thiserror = "1.0"
 tikv-jemallocator = "0.5"
 tikv-jemalloc-ctl = "0.5"
@@ -168,10 +177,11 @@ tokio-util = { version = "0.7.10", features = ["io", "rt"] }
 toml = "0.7"
 toml_edit = "0.19"
 tonic = {version = "0.9", features = ["tls", "tls-roots"]}
+tower-service = "0.3.2"
 tracing = "0.1"
 tracing-error = "0.2.0"
-tracing-opentelemetry = "0.20.0"
-tracing-subscriber = { version = "0.3", default_features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json"] }
+tracing-opentelemetry = "0.21.0"
+tracing-subscriber = { version = "0.3", default_features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json", "ansi"] }
 twox-hash = { version = "1.6.3", default-features = false }
 url = "2.2"
 urlencoding = "2.1"

Dockerfile.build-tools

@@ -58,8 +58,14 @@ RUN curl -fsSL "https://github.com/protocolbuffers/protobuf/releases/download/v$
     && mv protoc/include/google /usr/local/include/google \
     && rm -rf protoc.zip protoc
 
+# s5cmd
+ENV S5CMD_VERSION=2.2.2
+RUN curl -sL "https://github.com/peak/s5cmd/releases/download/v${S5CMD_VERSION}/s5cmd_${S5CMD_VERSION}_Linux-$(uname -m | sed 's/x86_64/64bit/g' | sed 's/aarch64/arm64/g').tar.gz" | tar zxvf - s5cmd \
+    && chmod +x s5cmd \
+    && mv s5cmd /usr/local/bin/s5cmd
+
 # LLVM
-ENV LLVM_VERSION=17
+ENV LLVM_VERSION=18
 RUN curl -fsSL 'https://apt.llvm.org/llvm-snapshot.gpg.key' | apt-key add - \
     && echo "deb http://apt.llvm.org/bullseye/ llvm-toolchain-bullseye-${LLVM_VERSION} main" > /etc/apt/sources.list.d/llvm.stable.list \
     && apt update \
@@ -135,7 +141,7 @@ 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=1.76.0
+ENV RUSTC_VERSION=1.78.0
 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 && \
@@ -149,7 +155,7 @@ RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux
     cargo install --git https://github.com/paritytech/cachepot && \
     cargo install rustfilt && \
     cargo install cargo-hakari && \
-    cargo install cargo-deny && \
+    cargo install cargo-deny --locked && \
     cargo install cargo-hack && \
     cargo install cargo-nextest && \
     rm -rf /home/nonroot/.cargo/registry && \

Dockerfile.compute-node

@@ -944,6 +944,9 @@ RUN mkdir /var/db && useradd -m -d /var/db/postgres postgres && \
 COPY --from=postgres-cleanup-layer --chown=postgres /usr/local/pgsql /usr/local
 COPY --from=compute-tools --chown=postgres /home/nonroot/target/release-line-debug-size-lto/compute_ctl /usr/local/bin/compute_ctl
 
+# Create remote extension download directory
+RUN mkdir /usr/local/download_extensions && chown -R postgres:postgres /usr/local/download_extensions
+
 # Install:
 # libreadline8 for psql
 # libicu67, locales for collations (including ICU and plpgsql_check)

Makefile

@@ -25,14 +25,16 @@ ifeq ($(UNAME_S),Linux)
 	# Seccomp BPF is only available for Linux
 	PG_CONFIGURE_OPTS += --with-libseccomp
 else ifeq ($(UNAME_S),Darwin)
-	# macOS with brew-installed openssl requires explicit paths
-	# It can be configured with OPENSSL_PREFIX variable
-	OPENSSL_PREFIX ?= $(shell brew --prefix openssl@3)
-	PG_CONFIGURE_OPTS += --with-includes=$(OPENSSL_PREFIX)/include --with-libraries=$(OPENSSL_PREFIX)/lib
-	PG_CONFIGURE_OPTS += PKG_CONFIG_PATH=$(shell brew --prefix icu4c)/lib/pkgconfig
-	# macOS already has bison and flex in the system, but they are old and result in postgres-v14 target failure
-	# brew formulae are keg-only and not symlinked into HOMEBREW_PREFIX, force their usage
-	EXTRA_PATH_OVERRIDES += $(shell brew --prefix bison)/bin/:$(shell brew --prefix flex)/bin/:
+	ifndef DISABLE_HOMEBREW
+		# macOS with brew-installed openssl requires explicit paths
+		# It can be configured with OPENSSL_PREFIX variable
+		OPENSSL_PREFIX ?= $(shell brew --prefix openssl@3)
+		PG_CONFIGURE_OPTS += --with-includes=$(OPENSSL_PREFIX)/include --with-libraries=$(OPENSSL_PREFIX)/lib
+		PG_CONFIGURE_OPTS += PKG_CONFIG_PATH=$(shell brew --prefix icu4c)/lib/pkgconfig
+		# macOS already has bison and flex in the system, but they are old and result in postgres-v14 target failure
+		# brew formulae are keg-only and not symlinked into HOMEBREW_PREFIX, force their usage
+		EXTRA_PATH_OVERRIDES += $(shell brew --prefix bison)/bin/:$(shell brew --prefix flex)/bin/:
+	endif
 endif
 
 # Use -C option so that when PostgreSQL "make install" installs the

README.md

@@ -238,6 +238,14 @@ If you encounter errors during setting up the initial tenant, it's best to stop
 
 ## Running tests
 
+### Rust unit tests
+
+We are using [`cargo-nextest`](https://nexte.st/) to run the tests in Github Workflows.
+Some crates do not support running plain `cargo test` anymore, prefer `cargo nextest run` instead.
+You can install `cargo-nextest` with `cargo install cargo-nextest`.
+
+### Integration tests
+
 Ensure your dependencies are installed as described [here](https://github.com/neondatabase/neon#dependency-installation-notes).
 
 ```sh

clippy.toml

@@ -2,6 +2,8 @@ disallowed-methods = [
     "tokio::task::block_in_place",
     # Allow this for now, to deny it later once we stop using Handle::block_on completely
     # "tokio::runtime::Handle::block_on",
+    # use tokio_epoll_uring_ext instead
+    "tokio_epoll_uring::thread_local_system",
 ]
 
 disallowed-macros = [

compute_tools/README.md

@@ -32,6 +32,29 @@ compute_ctl -D /var/db/postgres/compute \
             -b /usr/local/bin/postgres
 ```
 
+## State Diagram
+
+Computes can be in various states. Below is a diagram that details how a
+compute moves between states.
+
+```mermaid
+%% https://mermaid.js.org/syntax/stateDiagram.html
+stateDiagram-v2
+  [*] --> Empty : Compute spawned
+  Empty --> ConfigurationPending : Waiting for compute spec
+  ConfigurationPending --> Configuration : Received compute spec
+  Configuration --> Failed : Failed to configure the compute
+  Configuration --> Running : Compute has been configured
+  Empty --> Init : Compute spec is immediately available
+  Empty --> TerminationPending : Requested termination
+  Init --> Failed : Failed to start Postgres
+  Init --> Running : Started Postgres
+  Running --> TerminationPending : Requested termination
+  TerminationPending --> Terminated : Terminated compute
+  Failed --> [*] : Compute exited
+  Terminated --> [*] : Compute exited
+```
+
 ## Tests
 
 Cargo formatter:

compute_tools/src/bin/compute_ctl.rs

@@ -47,7 +47,7 @@ use chrono::Utc;
 use clap::Arg;
 use signal_hook::consts::{SIGQUIT, SIGTERM};
 use signal_hook::{consts::SIGINT, iterator::Signals};
-use tracing::{error, info};
+use tracing::{error, info, warn};
 use url::Url;
 
 use compute_api::responses::ComputeStatus;
@@ -62,6 +62,7 @@ use compute_tools::logger::*;
 use compute_tools::monitor::launch_monitor;
 use compute_tools::params::*;
 use compute_tools::spec::*;
+use compute_tools::swap::resize_swap;
 
 // this is an arbitrary build tag. Fine as a default / for testing purposes
 // in-case of not-set environment var
@@ -110,6 +111,7 @@ fn main() -> Result<()> {
         .expect("Postgres connection string is required");
     let spec_json = matches.get_one::<String>("spec");
     let spec_path = matches.get_one::<String>("spec-path");
+    let resize_swap_on_bind = matches.get_flag("resize-swap-on-bind");
 
     // Extract OpenTelemetry context for the startup actions from the
     // TRACEPARENT and TRACESTATE env variables, and attach it to the current
@@ -226,14 +228,14 @@ fn main() -> Result<()> {
 
     // If this is a pooled VM, prewarm before starting HTTP server and becoming
     // available for binding. Prewarming helps Postgres start quicker later,
-    // because QEMU will already have it's memory allocated from the host, and
+    // because QEMU will already have its memory allocated from the host, and
     // the necessary binaries will already be cached.
     if !spec_set {
         compute.prewarm_postgres()?;
     }
 
-    // Launch http service first, so we were able to serve control-plane
-    // requests, while configuration is still in progress.
+    // Launch http service first, so that we can serve control-plane requests
+    // while configuration is still in progress.
     let _http_handle =
         launch_http_server(http_port, &compute).expect("cannot launch http endpoint thread");
 
@@ -253,21 +255,22 @@ fn main() -> Result<()> {
                 break;
             }
         }
+
+        // Record for how long we slept waiting for the spec.
+        let now = Utc::now();
+        state.metrics.wait_for_spec_ms = now
+            .signed_duration_since(state.start_time)
+            .to_std()
+            .unwrap()
+            .as_millis() as u64;
+
+        // Reset start time, so that the total startup time that is calculated later will
+        // not include the time that we waited for the spec.
+        state.start_time = now;
     }
 
     // We got all we need, update the state.
     let mut state = compute.state.lock().unwrap();
-
-    // Record for how long we slept waiting for the spec.
-    state.metrics.wait_for_spec_ms = Utc::now()
-        .signed_duration_since(state.start_time)
-        .to_std()
-        .unwrap()
-        .as_millis() as u64;
-    // Reset start time to the actual start of the configuration, so that
-    // total startup time was properly measured at the end.
-    state.start_time = Utc::now();
-
     state.status = ComputeStatus::Init;
     compute.state_changed.notify_all();
 
@@ -275,33 +278,72 @@ fn main() -> Result<()> {
         "running compute with features: {:?}",
         state.pspec.as_ref().unwrap().spec.features
     );
+    // before we release the mutex, fetch the swap size (if any) for later.
+    let swap_size_bytes = state.pspec.as_ref().unwrap().spec.swap_size_bytes;
     drop(state);
 
     // Launch remaining service threads
     let _monitor_handle = launch_monitor(&compute);
     let _configurator_handle = launch_configurator(&compute);
 
-    // Start Postgres
+    let mut prestartup_failed = false;
     let mut delay_exit = false;
-    let mut exit_code = None;
-    let pg = match compute.start_compute(extension_server_port) {
-        Ok(pg) => Some(pg),
-        Err(err) => {
-            error!("could not start the compute node: {:#}", err);
-            let mut state = compute.state.lock().unwrap();
-            state.error = Some(format!("{:?}", err));
-            state.status = ComputeStatus::Failed;
-            // Notify others that Postgres failed to start. In case of configuring the
-            // empty compute, it's likely that API handler is still waiting for compute
-            // state change. With this we will notify it that compute is in Failed state,
-            // so control plane will know about it earlier and record proper error instead
-            // of timeout.
-            compute.state_changed.notify_all();
-            drop(state); // unlock
-            delay_exit = true;
-            None
+
+    // Resize swap to the desired size if the compute spec says so
+    if let (Some(size_bytes), true) = (swap_size_bytes, resize_swap_on_bind) {
+        // To avoid 'swapoff' hitting postgres startup, we need to run resize-swap to completion
+        // *before* starting postgres.
+        //
+        // In theory, we could do this asynchronously if SkipSwapon was enabled for VMs, but this
+        // carries a risk of introducing hard-to-debug issues - e.g. if postgres sometimes gets
+        // OOM-killed during startup because swap wasn't available yet.
+        match resize_swap(size_bytes) {
+            Ok(()) => {
+                let size_gib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
+                info!(%size_bytes, %size_gib, "resized swap");
+            }
+            Err(err) => {
+                let err = err.context("failed to resize swap");
+                error!("{err:#}");
+
+                // Mark compute startup as failed; don't try to start postgres, and report this
+                // error to the control plane when it next asks.
+                prestartup_failed = true;
+                let mut state = compute.state.lock().unwrap();
+                state.error = Some(format!("{err:?}"));
+                state.status = ComputeStatus::Failed;
+                compute.state_changed.notify_all();
+                delay_exit = true;
+            }
         }
-    };
+    }
+
+    // Start Postgres
+    let mut pg = None;
+    let mut exit_code = None;
+
+    if !prestartup_failed {
+        pg = match compute.start_compute(extension_server_port) {
+            Ok(pg) => Some(pg),
+            Err(err) => {
+                error!("could not start the compute node: {:#}", err);
+                let mut state = compute.state.lock().unwrap();
+                state.error = Some(format!("{:?}", err));
+                state.status = ComputeStatus::Failed;
+                // Notify others that Postgres failed to start. In case of configuring the
+                // empty compute, it's likely that API handler is still waiting for compute
+                // state change. With this we will notify it that compute is in Failed state,
+                // so control plane will know about it earlier and record proper error instead
+                // of timeout.
+                compute.state_changed.notify_all();
+                drop(state); // unlock
+                delay_exit = true;
+                None
+            }
+        };
+    } else {
+        warn!("skipping postgres startup because pre-startup step failed");
+    }
 
     // Start the vm-monitor if directed to. The vm-monitor only runs on linux
     // because it requires cgroups.
@@ -526,6 +568,11 @@ fn cli() -> clap::Command {
                 )
                 .value_name("FILECACHE_CONNSTR"),
         )
+        .arg(
+            Arg::new("resize-swap-on-bind")
+                .long("resize-swap-on-bind")
+                .action(clap::ArgAction::SetTrue),
+        )
 }
 
 /// When compute_ctl is killed, send also termination signal to sync-safekeepers

compute_tools/src/compute.rs

@@ -818,9 +818,15 @@ impl ComputeNode {
                         Client::connect(zenith_admin_connstr.as_str(), NoTls)
                             .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
                     // Disable forwarding so that users don't get a cloud_admin role
-                    client.simple_query("SET neon.forward_ddl = false")?;
-                    client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
-                    client.simple_query("GRANT zenith_admin TO cloud_admin")?;
+
+                    let mut func = || {
+                        client.simple_query("SET neon.forward_ddl = false")?;
+                        client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
+                        client.simple_query("GRANT zenith_admin TO cloud_admin")?;
+                        Ok::<_, anyhow::Error>(())
+                    };
+                    func().context("apply_config setup cloud_admin")?;
+
                     drop(client);
 
                     // reconnect with connstring with expected name
@@ -832,24 +838,29 @@ impl ComputeNode {
         };
 
         // Disable DDL forwarding because control plane already knows about these roles/databases.
-        client.simple_query("SET neon.forward_ddl = false")?;
+        client
+            .simple_query("SET neon.forward_ddl = false")
+            .context("apply_config SET neon.forward_ddl = false")?;
 
         // Proceed with post-startup configuration. Note, that order of operations is important.
         let spec = &compute_state.pspec.as_ref().expect("spec must be set").spec;
-        create_neon_superuser(spec, &mut client)?;
-        cleanup_instance(&mut client)?;
-        handle_roles(spec, &mut client)?;
-        handle_databases(spec, &mut client)?;
-        handle_role_deletions(spec, connstr.as_str(), &mut client)?;
+        create_neon_superuser(spec, &mut client).context("apply_config create_neon_superuser")?;
+        cleanup_instance(&mut client).context("apply_config cleanup_instance")?;
+        handle_roles(spec, &mut client).context("apply_config handle_roles")?;
+        handle_databases(spec, &mut client).context("apply_config handle_databases")?;
+        handle_role_deletions(spec, connstr.as_str(), &mut client)
+            .context("apply_config handle_role_deletions")?;
         handle_grants(
             spec,
             &mut client,
             connstr.as_str(),
             self.has_feature(ComputeFeature::AnonExtension),
-        )?;
-        handle_extensions(spec, &mut client)?;
-        handle_extension_neon(&mut client)?;
-        create_availability_check_data(&mut client)?;
+        )
+        .context("apply_config handle_grants")?;
+        handle_extensions(spec, &mut client).context("apply_config handle_extensions")?;
+        handle_extension_neon(&mut client).context("apply_config handle_extension_neon")?;
+        create_availability_check_data(&mut client)
+            .context("apply_config create_availability_check_data")?;
 
         // 'Close' connection
         drop(client);
@@ -857,7 +868,7 @@ impl ComputeNode {
         // Run migrations separately to not hold up cold starts
         thread::spawn(move || {
             let mut client = Client::connect(connstr.as_str(), NoTls)?;
-            handle_migrations(&mut client)
+            handle_migrations(&mut client).context("apply_config handle_migrations")
         });
         Ok(())
     }
@@ -1262,10 +1273,12 @@ LIMIT 100",
         .await
         .map_err(DownloadError::Other);
 
-        self.ext_download_progress
-            .write()
-            .expect("bad lock")
-            .insert(ext_archive_name.to_string(), (download_start, true));
+        if download_size.is_ok() {
+            self.ext_download_progress
+                .write()
+                .expect("bad lock")
+                .insert(ext_archive_name.to_string(), (download_start, true));
+        }
 
         download_size
     }

compute_tools/src/config.rs

@@ -6,8 +6,8 @@ use std::path::Path;
 use anyhow::Result;
 
 use crate::pg_helpers::escape_conf_value;
-use crate::pg_helpers::PgOptionsSerialize;
-use compute_api::spec::{ComputeMode, ComputeSpec};
+use crate::pg_helpers::{GenericOptionExt, PgOptionsSerialize};
+use compute_api::spec::{ComputeMode, ComputeSpec, GenericOption};
 
 /// Check that `line` is inside a text file and put it there if it is not.
 /// Create file if it doesn't exist.
@@ -17,6 +17,7 @@ pub fn line_in_file(path: &Path, line: &str) -> Result<bool> {
         .write(true)
         .create(true)
         .append(false)
+        .truncate(false)
         .open(path)?;
     let buf = io::BufReader::new(&file);
     let mut count: usize = 0;
@@ -91,6 +92,27 @@ pub fn write_postgres_conf(
         }
     }
 
+    if cfg!(target_os = "linux") {
+        // Check /proc/sys/vm/overcommit_memory -- if it equals 2 (i.e. linux memory overcommit is
+        // disabled), then the control plane has enabled swap and we should set
+        // dynamic_shared_memory_type = 'mmap'.
+        //
+        // This is (maybe?) temporary - for more, see https://github.com/neondatabase/cloud/issues/12047.
+        let overcommit_memory_contents = std::fs::read_to_string("/proc/sys/vm/overcommit_memory")
+            // ignore any errors - they may be expected to occur under certain situations (e.g. when
+            // not running in Linux).
+            .unwrap_or_else(|_| String::new());
+        if overcommit_memory_contents.trim() == "2" {
+            let opt = GenericOption {
+                name: "dynamic_shared_memory_type".to_owned(),
+                value: Some("mmap".to_owned()),
+                vartype: "enum".to_owned(),
+            };
+
+            write!(file, "{}", opt.to_pg_setting())?;
+        }
+    }
+
     // If there are any extra options in the 'settings' field, append those
     if spec.cluster.settings.is_some() {
         writeln!(file, "# Managed by compute_ctl: begin")?;

compute_tools/src/lib.rs

@@ -14,4 +14,5 @@ pub mod monitor;
 pub mod params;
 pub mod pg_helpers;
 pub mod spec;
+pub mod swap;
 pub mod sync_sk;

compute_tools/src/pg_helpers.rs

@@ -44,7 +44,7 @@ pub fn escape_conf_value(s: &str) -> String {
     format!("'{}'", res)
 }
 
-trait GenericOptionExt {
+pub trait GenericOptionExt {
     fn to_pg_option(&self) -> String;
     fn to_pg_setting(&self) -> String;
 }

compute_tools/src/spec.rs

@@ -2,7 +2,7 @@ use std::fs::File;
 use std::path::Path;
 use std::str::FromStr;
 
-use anyhow::{anyhow, bail, Result};
+use anyhow::{anyhow, bail, Context, Result};
 use postgres::config::Config;
 use postgres::{Client, NoTls};
 use reqwest::StatusCode;
@@ -302,9 +302,9 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
             RoleAction::Create => {
                 // This branch only runs when roles are created through the console, so it is
                 // safe to add more permissions here. BYPASSRLS and REPLICATION are inherited
-                // from neon_superuser. (NOTE: REPLICATION has been removed from here for now).
+                // from neon_superuser.
                 let mut query: String = format!(
-                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS IN ROLE neon_superuser",
+                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
                     name.pg_quote()
                 );
                 info!("running role create query: '{}'", &query);
@@ -490,7 +490,7 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                 "rename_db" => {
                     let new_name = op.new_name.as_ref().unwrap();
 
-                    if existing_dbs.get(&op.name).is_some() {
+                    if existing_dbs.contains_key(&op.name) {
                         let query: String = format!(
                             "ALTER DATABASE {} RENAME TO {}",
                             op.name.pg_quote(),
@@ -698,7 +698,8 @@ pub fn handle_grants(
 
         // it is important to run this after all grants
         if enable_anon_extension {
-            handle_extension_anon(spec, &db.owner, &mut db_client, false)?;
+            handle_extension_anon(spec, &db.owner, &mut db_client, false)
+                .context("handle_grants handle_extension_anon")?;
         }
     }
 
@@ -745,7 +746,12 @@ pub fn handle_extension_neon(client: &mut Client) -> Result<()> {
     // - extension was already installed and is up to date
     let query = "ALTER EXTENSION neon UPDATE";
     info!("update neon extension version with query: {}", query);
-    client.simple_query(query)?;
+    if let Err(e) = client.simple_query(query) {
+        error!(
+            "failed to upgrade neon extension during `handle_extension_neon`: {}",
+            e
+        );
+    }
 
     Ok(())
 }
@@ -804,43 +810,40 @@ $$;"#,
         "",
         "",
         "",
+        "",
         // Add new migrations below.
-        r#"
-DO $$
-DECLARE
-    role_name TEXT;
-BEGIN
-    FOR role_name IN SELECT rolname FROM pg_roles WHERE rolreplication IS TRUE
-    LOOP
-        RAISE NOTICE 'EXECUTING ALTER ROLE % NOREPLICATION', quote_ident(role_name);
-        EXECUTE 'ALTER ROLE ' || quote_ident(role_name) || ' NOREPLICATION';
-    END LOOP;
-END
-$$;"#,
     ];
 
-    let mut query = "CREATE SCHEMA IF NOT EXISTS neon_migration";
-    client.simple_query(query)?;
+    let mut func = || {
+        let query = "CREATE SCHEMA IF NOT EXISTS neon_migration";
+        client.simple_query(query)?;
 
-    query = "CREATE TABLE IF NOT EXISTS neon_migration.migration_id (key INT NOT NULL PRIMARY KEY, id bigint NOT NULL DEFAULT 0)";
-    client.simple_query(query)?;
+        let query = "CREATE TABLE IF NOT EXISTS neon_migration.migration_id (key INT NOT NULL PRIMARY KEY, id bigint NOT NULL DEFAULT 0)";
+        client.simple_query(query)?;
 
-    query = "INSERT INTO neon_migration.migration_id VALUES (0, 0) ON CONFLICT DO NOTHING";
-    client.simple_query(query)?;
+        let query = "INSERT INTO neon_migration.migration_id VALUES (0, 0) ON CONFLICT DO NOTHING";
+        client.simple_query(query)?;
 
-    query = "ALTER SCHEMA neon_migration OWNER TO cloud_admin";
-    client.simple_query(query)?;
+        let query = "ALTER SCHEMA neon_migration OWNER TO cloud_admin";
+        client.simple_query(query)?;
 
-    query = "REVOKE ALL ON SCHEMA neon_migration FROM PUBLIC";
-    client.simple_query(query)?;
+        let query = "REVOKE ALL ON SCHEMA neon_migration FROM PUBLIC";
+        client.simple_query(query)?;
+        Ok::<_, anyhow::Error>(())
+    };
+    func().context("handle_migrations prepare")?;
 
-    query = "SELECT id FROM neon_migration.migration_id";
-    let row = client.query_one(query, &[])?;
+    let query = "SELECT id FROM neon_migration.migration_id";
+    let row = client
+        .query_one(query, &[])
+        .context("handle_migrations get migration_id")?;
     let mut current_migration: usize = row.get::<&str, i64>("id") as usize;
     let starting_migration_id = current_migration;
 
-    query = "BEGIN";
-    client.simple_query(query)?;
+    let query = "BEGIN";
+    client
+        .simple_query(query)
+        .context("handle_migrations begin")?;
 
     while current_migration < migrations.len() {
         let migration = &migrations[current_migration];
@@ -848,7 +851,9 @@ $$;"#,
             info!("Skip migration id={}", current_migration);
         } else {
             info!("Running migration:\n{}\n", migration);
-            client.simple_query(migration)?;
+            client.simple_query(migration).with_context(|| {
+                format!("handle_migrations current_migration={}", current_migration)
+            })?;
         }
         current_migration += 1;
     }
@@ -856,10 +861,14 @@ $$;"#,
         "UPDATE neon_migration.migration_id SET id={}",
         migrations.len()
     );
-    client.simple_query(&setval)?;
+    client
+        .simple_query(&setval)
+        .context("handle_migrations update id")?;
 
-    query = "COMMIT";
-    client.simple_query(query)?;
+    let query = "COMMIT";
+    client
+        .simple_query(query)
+        .context("handle_migrations commit")?;
 
     info!(
         "Ran {} migrations",

compute_tools/src/swap.rs

@@ -0,0 +1,36 @@
+use anyhow::{anyhow, Context};
+use tracing::warn;
+
+pub const RESIZE_SWAP_BIN: &str = "/neonvm/bin/resize-swap";
+
+pub fn resize_swap(size_bytes: u64) -> anyhow::Result<()> {
+    // run `/neonvm/bin/resize-swap --once {size_bytes}`
+    //
+    // Passing '--once' causes resize-swap to delete itself after successful completion, which
+    // means that if compute_ctl restarts later, we won't end up calling 'swapoff' while
+    // postgres is running.
+    //
+    // NOTE: resize-swap is not very clever. If present, --once MUST be the first arg.
+    let child_result = std::process::Command::new("/usr/bin/sudo")
+        .arg(RESIZE_SWAP_BIN)
+        .arg("--once")
+        .arg(size_bytes.to_string())
+        .spawn();
+
+    if matches!(&child_result, Err(e) if e.kind() == std::io::ErrorKind::NotFound) {
+        warn!("ignoring \"not found\" error from resize-swap to avoid swapoff while compute is running");
+        return Ok(());
+    }
+
+    child_result
+        .context("spawn() failed")
+        .and_then(|mut child| child.wait().context("wait() failed"))
+        .and_then(|status| match status.success() {
+            true => Ok(()),
+            false => Err(anyhow!("process exited with {status}")),
+        })
+        // wrap any prior error with the overall context that we couldn't run the command
+        .with_context(|| {
+            format!("could not run `/usr/bin/sudo {RESIZE_SWAP_BIN} --once {size_bytes}`")
+        })
+}

control_plane/Cargo.toml

@@ -12,10 +12,12 @@ clap.workspace = true
 comfy-table.workspace = true
 futures.workspace = true
 git-version.workspace = true
+humantime.workspace = true
 nix.workspace = true
 once_cell.workspace = true
 postgres.workspace = true
 hex.workspace = true
+humantime-serde.workspace = true
 hyper.workspace = true
 regex.workspace = true
 reqwest = { workspace = true, features = ["blocking", "json"] }

control_plane/attachment_service/src/compute_hook.rs

@@ -1,462 +0,0 @@
-use std::{collections::HashMap, time::Duration};
-
-use control_plane::endpoint::{ComputeControlPlane, EndpointStatus};
-use control_plane::local_env::LocalEnv;
-use hyper::{Method, StatusCode};
-use pageserver_api::shard::{ShardCount, ShardNumber, ShardStripeSize, TenantShardId};
-use postgres_connection::parse_host_port;
-use serde::{Deserialize, Serialize};
-use tokio_util::sync::CancellationToken;
-use utils::{
-    backoff::{self},
-    id::{NodeId, TenantId},
-};
-
-use crate::service::Config;
-
-const BUSY_DELAY: Duration = Duration::from_secs(1);
-const SLOWDOWN_DELAY: Duration = Duration::from_secs(5);
-
-pub(crate) const API_CONCURRENCY: usize = 32;
-
-struct ShardedComputeHookTenant {
-    stripe_size: ShardStripeSize,
-    shard_count: ShardCount,
-    shards: Vec<(ShardNumber, NodeId)>,
-}
-
-enum ComputeHookTenant {
-    Unsharded(NodeId),
-    Sharded(ShardedComputeHookTenant),
-}
-
-impl ComputeHookTenant {
-    /// Construct with at least one shard's information
-    fn new(tenant_shard_id: TenantShardId, stripe_size: ShardStripeSize, node_id: NodeId) -> Self {
-        if tenant_shard_id.shard_count.count() > 1 {
-            Self::Sharded(ShardedComputeHookTenant {
-                shards: vec![(tenant_shard_id.shard_number, node_id)],
-                stripe_size,
-                shard_count: tenant_shard_id.shard_count,
-            })
-        } else {
-            Self::Unsharded(node_id)
-        }
-    }
-
-    /// Set one shard's location.  If stripe size or shard count have changed, Self is reset
-    /// and drops existing content.
-    fn update(
-        &mut self,
-        tenant_shard_id: TenantShardId,
-        stripe_size: ShardStripeSize,
-        node_id: NodeId,
-    ) {
-        match self {
-            Self::Unsharded(existing_node_id) if tenant_shard_id.shard_count.count() == 1 => {
-                *existing_node_id = node_id
-            }
-            Self::Sharded(sharded_tenant)
-                if sharded_tenant.stripe_size == stripe_size
-                    && sharded_tenant.shard_count == tenant_shard_id.shard_count =>
-            {
-                if let Some(existing) = sharded_tenant
-                    .shards
-                    .iter()
-                    .position(|s| s.0 == tenant_shard_id.shard_number)
-                {
-                    sharded_tenant.shards.get_mut(existing).unwrap().1 = node_id;
-                } else {
-                    sharded_tenant
-                        .shards
-                        .push((tenant_shard_id.shard_number, node_id));
-                    sharded_tenant.shards.sort_by_key(|s| s.0)
-                }
-            }
-            _ => {
-                // Shard count changed: reset struct.
-                *self = Self::new(tenant_shard_id, stripe_size, node_id);
-            }
-        }
-    }
-}
-
-#[derive(Serialize, Deserialize, Debug)]
-struct ComputeHookNotifyRequestShard {
-    node_id: NodeId,
-    shard_number: ShardNumber,
-}
-
-/// Request body that we send to the control plane to notify it of where a tenant is attached
-#[derive(Serialize, Deserialize, Debug)]
-struct ComputeHookNotifyRequest {
-    tenant_id: TenantId,
-    stripe_size: Option<ShardStripeSize>,
-    shards: Vec<ComputeHookNotifyRequestShard>,
-}
-
-/// Error type for attempts to call into the control plane compute notification hook
-#[derive(thiserror::Error, Debug)]
-pub(crate) enum NotifyError {
-    // Request was not send successfully, e.g. transport error
-    #[error("Sending request: {0}")]
-    Request(#[from] reqwest::Error),
-    // Request could not be serviced right now due to ongoing Operation in control plane, but should be possible soon.
-    #[error("Control plane tenant busy")]
-    Busy,
-    // Explicit 429 response asking us to retry less frequently
-    #[error("Control plane overloaded")]
-    SlowDown,
-    // A 503 response indicates the control plane can't handle the request right now
-    #[error("Control plane unavailable (status {0})")]
-    Unavailable(StatusCode),
-    // API returned unexpected non-success status.  We will retry, but log a warning.
-    #[error("Control plane returned unexpected status {0}")]
-    Unexpected(StatusCode),
-    // We shutdown while sending
-    #[error("Shutting down")]
-    ShuttingDown,
-    // A response indicates we will never succeed, such as 400 or 404
-    #[error("Non-retryable error {0}")]
-    Fatal(StatusCode),
-}
-
-impl ComputeHookTenant {
-    fn maybe_reconfigure(&self, tenant_id: TenantId) -> Option<ComputeHookNotifyRequest> {
-        match self {
-            Self::Unsharded(node_id) => Some(ComputeHookNotifyRequest {
-                tenant_id,
-                shards: vec![ComputeHookNotifyRequestShard {
-                    shard_number: ShardNumber(0),
-                    node_id: *node_id,
-                }],
-                stripe_size: None,
-            }),
-            Self::Sharded(sharded_tenant)
-                if sharded_tenant.shards.len() == sharded_tenant.shard_count.count() as usize =>
-            {
-                Some(ComputeHookNotifyRequest {
-                    tenant_id,
-                    shards: sharded_tenant
-                        .shards
-                        .iter()
-                        .map(|(shard_number, node_id)| ComputeHookNotifyRequestShard {
-                            shard_number: *shard_number,
-                            node_id: *node_id,
-                        })
-                        .collect(),
-                    stripe_size: Some(sharded_tenant.stripe_size),
-                })
-            }
-            Self::Sharded(sharded_tenant) => {
-                // Sharded tenant doesn't yet have information for all its shards
-
-                tracing::info!(
-                    "ComputeHookTenant::maybe_reconfigure: not enough shards ({}/{})",
-                    sharded_tenant.shards.len(),
-                    sharded_tenant.shard_count.count()
-                );
-                None
-            }
-        }
-    }
-}
-
-/// The compute hook is a destination for notifications about changes to tenant:pageserver
-/// mapping.  It aggregates updates for the shards in a tenant, and when appropriate reconfigures
-/// the compute connection string.
-pub(super) struct ComputeHook {
-    config: Config,
-    state: tokio::sync::Mutex<HashMap<TenantId, ComputeHookTenant>>,
-    authorization_header: Option<String>,
-}
-
-impl ComputeHook {
-    pub(super) fn new(config: Config) -> Self {
-        let authorization_header = config
-            .control_plane_jwt_token
-            .clone()
-            .map(|jwt| format!("Bearer {}", jwt));
-
-        Self {
-            state: Default::default(),
-            config,
-            authorization_header,
-        }
-    }
-
-    /// For test environments: use neon_local's LocalEnv to update compute
-    async fn do_notify_local(
-        &self,
-        reconfigure_request: ComputeHookNotifyRequest,
-    ) -> anyhow::Result<()> {
-        let env = match LocalEnv::load_config() {
-            Ok(e) => e,
-            Err(e) => {
-                tracing::warn!("Couldn't load neon_local config, skipping compute update ({e})");
-                return Ok(());
-            }
-        };
-        let cplane =
-            ComputeControlPlane::load(env.clone()).expect("Error loading compute control plane");
-        let ComputeHookNotifyRequest {
-            tenant_id,
-            shards,
-            stripe_size,
-        } = reconfigure_request;
-
-        let compute_pageservers = shards
-            .into_iter()
-            .map(|shard| {
-                let ps_conf = env
-                    .get_pageserver_conf(shard.node_id)
-                    .expect("Unknown pageserver");
-                let (pg_host, pg_port) = parse_host_port(&ps_conf.listen_pg_addr)
-                    .expect("Unable to parse listen_pg_addr");
-                (pg_host, pg_port.unwrap_or(5432))
-            })
-            .collect::<Vec<_>>();
-
-        for (endpoint_name, endpoint) in &cplane.endpoints {
-            if endpoint.tenant_id == tenant_id && endpoint.status() == EndpointStatus::Running {
-                tracing::info!("Reconfiguring endpoint {}", endpoint_name,);
-                endpoint
-                    .reconfigure(compute_pageservers.clone(), stripe_size)
-                    .await?;
-            }
-        }
-
-        Ok(())
-    }
-
-    async fn do_notify_iteration(
-        &self,
-        client: &reqwest::Client,
-        url: &String,
-        reconfigure_request: &ComputeHookNotifyRequest,
-        cancel: &CancellationToken,
-    ) -> Result<(), NotifyError> {
-        let req = client.request(Method::PUT, url);
-        let req = if let Some(value) = &self.authorization_header {
-            req.header(reqwest::header::AUTHORIZATION, value)
-        } else {
-            req
-        };
-
-        tracing::info!(
-            "Sending notify request to {} ({:?})",
-            url,
-            reconfigure_request
-        );
-        let send_result = req.json(&reconfigure_request).send().await;
-        let response = match send_result {
-            Ok(r) => r,
-            Err(e) => return Err(e.into()),
-        };
-
-        // Treat all 2xx responses as success
-        if response.status() >= StatusCode::OK && response.status() < StatusCode::MULTIPLE_CHOICES {
-            if response.status() != StatusCode::OK {
-                // Non-200 2xx response: it doesn't make sense to retry, but this is unexpected, so
-                // log a warning.
-                tracing::warn!(
-                    "Unexpected 2xx response code {} from control plane",
-                    response.status()
-                );
-            }
-
-            return Ok(());
-        }
-
-        // Error response codes
-        match response.status() {
-            StatusCode::TOO_MANY_REQUESTS => {
-                // TODO: 429 handling should be global: set some state visible to other requests
-                // so that they will delay before starting, rather than all notifications trying
-                // once before backing off.
-                tokio::time::timeout(SLOWDOWN_DELAY, cancel.cancelled())
-                    .await
-                    .ok();
-                Err(NotifyError::SlowDown)
-            }
-            StatusCode::LOCKED => {
-                // Delay our retry if busy: the usual fast exponential backoff in backoff::retry
-                // is not appropriate
-                tokio::time::timeout(BUSY_DELAY, cancel.cancelled())
-                    .await
-                    .ok();
-                Err(NotifyError::Busy)
-            }
-            StatusCode::SERVICE_UNAVAILABLE
-            | StatusCode::GATEWAY_TIMEOUT
-            | StatusCode::BAD_GATEWAY => Err(NotifyError::Unavailable(response.status())),
-            StatusCode::BAD_REQUEST | StatusCode::UNAUTHORIZED | StatusCode::FORBIDDEN => {
-                Err(NotifyError::Fatal(response.status()))
-            }
-            _ => Err(NotifyError::Unexpected(response.status())),
-        }
-    }
-
-    async fn do_notify(
-        &self,
-        url: &String,
-        reconfigure_request: ComputeHookNotifyRequest,
-        cancel: &CancellationToken,
-    ) -> Result<(), NotifyError> {
-        let client = reqwest::Client::new();
-        backoff::retry(
-            || self.do_notify_iteration(&client, url, &reconfigure_request, cancel),
-            |e| matches!(e, NotifyError::Fatal(_) | NotifyError::Unexpected(_)),
-            3,
-            10,
-            "Send compute notification",
-            cancel,
-        )
-        .await
-        .ok_or_else(|| NotifyError::ShuttingDown)
-        .and_then(|x| x)
-    }
-
-    /// Call this to notify the compute (postgres) tier of new pageservers to use
-    /// for a tenant.  notify() is called by each shard individually, and this function
-    /// will decide whether an update to the tenant is sent.  An update is sent on the
-    /// condition that:
-    /// - We know a pageserver for every shard.
-    /// - All the shards have the same shard_count (i.e. we are not mid-split)
-    ///
-    /// Cancellation token enables callers to drop out, e.g. if calling from a Reconciler
-    /// that is cancelled.
-    ///
-    /// This function is fallible, including in the case that the control plane is transiently
-    /// unavailable.  A limited number of retries are done internally to efficiently hide short unavailability
-    /// periods, but we don't retry forever.  The **caller** is responsible for handling failures and
-    /// ensuring that they eventually call again to ensure that the compute is eventually notified of
-    /// the proper pageserver nodes for a tenant.
-    #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), node_id))]
-    pub(super) async fn notify(
-        &self,
-        tenant_shard_id: TenantShardId,
-        node_id: NodeId,
-        stripe_size: ShardStripeSize,
-        cancel: &CancellationToken,
-    ) -> Result<(), NotifyError> {
-        let mut locked = self.state.lock().await;
-
-        use std::collections::hash_map::Entry;
-        let tenant = match locked.entry(tenant_shard_id.tenant_id) {
-            Entry::Vacant(e) => e.insert(ComputeHookTenant::new(
-                tenant_shard_id,
-                stripe_size,
-                node_id,
-            )),
-            Entry::Occupied(e) => {
-                let tenant = e.into_mut();
-                tenant.update(tenant_shard_id, stripe_size, node_id);
-                tenant
-            }
-        };
-
-        let reconfigure_request = tenant.maybe_reconfigure(tenant_shard_id.tenant_id);
-        let Some(reconfigure_request) = reconfigure_request else {
-            // The tenant doesn't yet have pageservers for all its shards: we won't notify anything
-            // until it does.
-            tracing::info!("Tenant isn't yet ready to emit a notification");
-            return Ok(());
-        };
-
-        if let Some(notify_url) = &self.config.compute_hook_url {
-            self.do_notify(notify_url, reconfigure_request, cancel)
-                .await
-        } else {
-            self.do_notify_local(reconfigure_request)
-                .await
-                .map_err(|e| {
-                    // This path is for testing only, so munge the error into our prod-style error type.
-                    tracing::error!("Local notification hook failed: {e}");
-                    NotifyError::Fatal(StatusCode::INTERNAL_SERVER_ERROR)
-                })
-        }
-    }
-}
-
-#[cfg(test)]
-pub(crate) mod tests {
-    use pageserver_api::shard::{ShardCount, ShardNumber};
-    use utils::id::TenantId;
-
-    use super::*;
-
-    #[test]
-    fn tenant_updates() -> anyhow::Result<()> {
-        let tenant_id = TenantId::generate();
-        let mut tenant_state = ComputeHookTenant::new(
-            TenantShardId {
-                tenant_id,
-                shard_count: ShardCount::new(0),
-                shard_number: ShardNumber(0),
-            },
-            ShardStripeSize(12345),
-            NodeId(1),
-        );
-
-        // An unsharded tenant is always ready to emit a notification
-        assert!(tenant_state.maybe_reconfigure(tenant_id).is_some());
-        assert_eq!(
-            tenant_state
-                .maybe_reconfigure(tenant_id)
-                .unwrap()
-                .shards
-                .len(),
-            1
-        );
-        assert!(tenant_state
-            .maybe_reconfigure(tenant_id)
-            .unwrap()
-            .stripe_size
-            .is_none());
-
-        // Writing the first shard of a multi-sharded situation (i.e. in a split)
-        // resets the tenant state and puts it in an non-notifying state (need to
-        // see all shards)
-        tenant_state.update(
-            TenantShardId {
-                tenant_id,
-                shard_count: ShardCount::new(2),
-                shard_number: ShardNumber(1),
-            },
-            ShardStripeSize(32768),
-            NodeId(1),
-        );
-        assert!(tenant_state.maybe_reconfigure(tenant_id).is_none());
-
-        // Writing the second shard makes it ready to notify
-        tenant_state.update(
-            TenantShardId {
-                tenant_id,
-                shard_count: ShardCount::new(2),
-                shard_number: ShardNumber(0),
-            },
-            ShardStripeSize(32768),
-            NodeId(1),
-        );
-
-        assert!(tenant_state.maybe_reconfigure(tenant_id).is_some());
-        assert_eq!(
-            tenant_state
-                .maybe_reconfigure(tenant_id)
-                .unwrap()
-                .shards
-                .len(),
-            2
-        );
-        assert_eq!(
-            tenant_state
-                .maybe_reconfigure(tenant_id)
-                .unwrap()
-                .stripe_size,
-            Some(ShardStripeSize(32768))
-        );
-
-        Ok(())
-    }
-}

control_plane/attachment_service/src/metrics.rs

@@ -1,32 +0,0 @@
-use metrics::{register_int_counter, register_int_counter_vec, IntCounter, IntCounterVec};
-use once_cell::sync::Lazy;
-
-pub(crate) struct ReconcilerMetrics {
-    pub(crate) spawned: IntCounter,
-    pub(crate) complete: IntCounterVec,
-}
-
-impl ReconcilerMetrics {
-    // Labels used on [`Self::complete`]
-    pub(crate) const SUCCESS: &'static str = "ok";
-    pub(crate) const ERROR: &'static str = "success";
-    pub(crate) const CANCEL: &'static str = "cancel";
-}
-
-pub(crate) static RECONCILER: Lazy<ReconcilerMetrics> = Lazy::new(|| ReconcilerMetrics {
-    spawned: register_int_counter!(
-        "storage_controller_reconcile_spawn",
-        "Count of how many times we spawn a reconcile task",
-    )
-    .expect("failed to define a metric"),
-    complete: register_int_counter_vec!(
-        "storage_controller_reconcile_complete",
-        "Reconciler tasks completed, broken down by success/failure/cancelled",
-        &["status"],
-    )
-    .expect("failed to define a metric"),
-});
-
-pub fn preinitialize_metrics() {
-    Lazy::force(&RECONCILER);
-}

control_plane/src/background_process.rs

@@ -86,7 +86,10 @@ where
         .stdout(process_log_file)
         .stderr(same_file_for_stderr)
         .args(args);
-    let filled_cmd = fill_remote_storage_secrets_vars(fill_rust_env_vars(background_command));
+
+    let filled_cmd = fill_env_vars_prefixed_neon(fill_remote_storage_secrets_vars(
+        fill_rust_env_vars(background_command),
+    ));
     filled_cmd.envs(envs);
 
     let pid_file_to_check = match &initial_pid_file {
@@ -268,6 +271,15 @@ fn fill_remote_storage_secrets_vars(mut cmd: &mut Command) -> &mut Command {
     cmd
 }
 
+fn fill_env_vars_prefixed_neon(mut cmd: &mut Command) -> &mut Command {
+    for (var, val) in std::env::vars() {
+        if var.starts_with("NEON_PAGESERVER_") {
+            cmd = cmd.env(var, val);
+        }
+    }
+    cmd
+}
+
 /// Add a `pre_exec` to the cmd that, inbetween fork() and exec(),
 /// 1. Claims a pidfile with a fcntl lock on it and
 /// 2. Sets up the pidfile's file descriptor so that it (and the lock)
@@ -294,7 +306,7 @@ where
     //      is in state 'taken' but the thread that would unlock it is
     //      not there.
     //   2. A rust object that represented some external resource in the
-    //      parent now got implicitly copied by the the fork, even though
+    //      parent now got implicitly copied by the fork, even though
     //      the object's type is not `Copy`. The parent program may use
     //      non-copyability as way to enforce unique ownership of an
     //      external resource in the typesystem. The fork breaks that

control_plane/src/bin/neon_local.rs

@@ -14,17 +14,15 @@ use control_plane::pageserver::{PageServerNode, PAGESERVER_REMOTE_STORAGE_DIR};
 use control_plane::safekeeper::SafekeeperNode;
 use control_plane::storage_controller::StorageController;
 use control_plane::{broker, local_env};
-use pageserver_api::controller_api::{
-    NodeAvailability, NodeConfigureRequest, NodeSchedulingPolicy, PlacementPolicy,
+use pageserver_api::config::{
+    DEFAULT_HTTP_LISTEN_PORT as DEFAULT_PAGESERVER_HTTP_PORT,
+    DEFAULT_PG_LISTEN_PORT as DEFAULT_PAGESERVER_PG_PORT,
 };
+use pageserver_api::controller_api::PlacementPolicy;
 use pageserver_api::models::{
     ShardParameters, TenantCreateRequest, TimelineCreateRequest, TimelineInfo,
 };
 use pageserver_api::shard::{ShardCount, ShardStripeSize, TenantShardId};
-use pageserver_api::{
-    DEFAULT_HTTP_LISTEN_PORT as DEFAULT_PAGESERVER_HTTP_PORT,
-    DEFAULT_PG_LISTEN_PORT as DEFAULT_PAGESERVER_PG_PORT,
-};
 use postgres_backend::AuthType;
 use postgres_connection::parse_host_port;
 use safekeeper_api::{
@@ -419,6 +417,54 @@ async fn handle_tenant(
                 println!("{} {:?}", t.id, t.state);
             }
         }
+        Some(("import", import_match)) => {
+            let tenant_id = parse_tenant_id(import_match)?.unwrap_or_else(TenantId::generate);
+
+            let storage_controller = StorageController::from_env(env);
+            let create_response = storage_controller.tenant_import(tenant_id).await?;
+
+            let shard_zero = create_response
+                .shards
+                .first()
+                .expect("Import response omitted shards");
+
+            let attached_pageserver_id = shard_zero.node_id;
+            let pageserver =
+                PageServerNode::from_env(env, env.get_pageserver_conf(attached_pageserver_id)?);
+
+            println!(
+                "Imported tenant {tenant_id}, attached to pageserver {attached_pageserver_id}"
+            );
+
+            let timelines = pageserver
+                .http_client
+                .list_timelines(shard_zero.shard_id)
+                .await?;
+
+            // Pick a 'main' timeline that has no ancestors, the rest will get arbitrary names
+            let main_timeline = timelines
+                .iter()
+                .find(|t| t.ancestor_timeline_id.is_none())
+                .expect("No timelines found")
+                .timeline_id;
+
+            let mut branch_i = 0;
+            for timeline in timelines.iter() {
+                let branch_name = if timeline.timeline_id == main_timeline {
+                    "main".to_string()
+                } else {
+                    branch_i += 1;
+                    format!("branch_{branch_i}")
+                };
+
+                println!(
+                    "Importing timeline {tenant_id}/{} as branch {branch_name}",
+                    timeline.timeline_id
+                );
+
+                env.register_branch_mapping(branch_name, tenant_id, timeline.timeline_id)?;
+            }
+        }
         Some(("create", create_match)) => {
             let tenant_conf: HashMap<_, _> = create_match
                 .get_many::<String>("config")
@@ -437,7 +483,7 @@ async fn handle_tenant(
 
             let placement_policy = match create_match.get_one::<String>("placement-policy") {
                 Some(s) if !s.is_empty() => serde_json::from_str::<PlacementPolicy>(s)?,
-                _ => PlacementPolicy::Single,
+                _ => PlacementPolicy::Attached(0),
             };
 
             let tenant_conf = PageServerNode::parse_config(tenant_conf)?;
@@ -523,88 +569,6 @@ async fn handle_tenant(
                 .with_context(|| format!("Tenant config failed for tenant with id {tenant_id}"))?;
             println!("tenant {tenant_id} successfully configured on the pageserver");
         }
-        Some(("migrate", matches)) => {
-            let tenant_shard_id = get_tenant_shard_id(matches, env)?;
-            let new_pageserver = get_pageserver(env, matches)?;
-            let new_pageserver_id = new_pageserver.conf.id;
-
-            let storage_controller = StorageController::from_env(env);
-            storage_controller
-                .tenant_migrate(tenant_shard_id, new_pageserver_id)
-                .await?;
-
-            println!("tenant {tenant_shard_id} migrated to {}", new_pageserver_id);
-        }
-        Some(("status", matches)) => {
-            let tenant_id = get_tenant_id(matches, env)?;
-
-            let mut shard_table = comfy_table::Table::new();
-            shard_table.set_header(["Shard", "Pageserver", "Physical Size"]);
-
-            let mut tenant_synthetic_size = None;
-
-            let storage_controller = StorageController::from_env(env);
-            for shard in storage_controller.tenant_locate(tenant_id).await?.shards {
-                let pageserver =
-                    PageServerNode::from_env(env, env.get_pageserver_conf(shard.node_id)?);
-
-                let size = pageserver
-                    .http_client
-                    .tenant_details(shard.shard_id)
-                    .await?
-                    .tenant_info
-                    .current_physical_size
-                    .unwrap();
-
-                shard_table.add_row([
-                    format!("{}", shard.shard_id.shard_slug()),
-                    format!("{}", shard.node_id.0),
-                    format!("{} MiB", size / (1024 * 1024)),
-                ]);
-
-                if shard.shard_id.is_zero() {
-                    tenant_synthetic_size =
-                        Some(pageserver.tenant_synthetic_size(shard.shard_id).await?);
-                }
-            }
-
-            let Some(synthetic_size) = tenant_synthetic_size else {
-                bail!("Shard 0 not found")
-            };
-
-            let mut tenant_table = comfy_table::Table::new();
-            tenant_table.add_row(["Tenant ID".to_string(), tenant_id.to_string()]);
-            tenant_table.add_row([
-                "Synthetic size".to_string(),
-                format!("{} MiB", synthetic_size.size.unwrap_or(0) / (1024 * 1024)),
-            ]);
-
-            println!("{tenant_table}");
-            println!("{shard_table}");
-        }
-        Some(("shard-split", matches)) => {
-            let tenant_id = get_tenant_id(matches, env)?;
-            let shard_count: u8 = matches.get_one::<u8>("shard-count").cloned().unwrap_or(0);
-            let shard_stripe_size: Option<ShardStripeSize> = matches
-                .get_one::<Option<ShardStripeSize>>("shard-stripe-size")
-                .cloned()
-                .unwrap();
-
-            let storage_controller = StorageController::from_env(env);
-            let result = storage_controller
-                .tenant_split(tenant_id, shard_count, shard_stripe_size)
-                .await?;
-            println!(
-                "Split tenant {} into shards {}",
-                tenant_id,
-                result
-                    .new_shards
-                    .iter()
-                    .map(|s| format!("{:?}", s))
-                    .collect::<Vec<_>>()
-                    .join(",")
-            );
-        }
 
         Some((sub_name, _)) => bail!("Unexpected tenant subcommand '{}'", sub_name),
         None => bail!("no tenant subcommand provided"),
@@ -873,6 +837,8 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 .copied()
                 .unwrap_or(false);
 
+            let allow_multiple = sub_args.get_flag("allow-multiple");
+
             let mode = match (lsn, hot_standby) {
                 (Some(lsn), false) => ComputeMode::Static(lsn),
                 (None, true) => ComputeMode::Replica,
@@ -890,7 +856,9 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 _ => {}
             }
 
-            cplane.check_conflicting_endpoints(mode, tenant_id, timeline_id)?;
+            if !allow_multiple {
+                cplane.check_conflicting_endpoints(mode, tenant_id, timeline_id)?;
+            }
 
             cplane.new_endpoint(
                 &endpoint_id,
@@ -919,6 +887,8 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
 
             let remote_ext_config = sub_args.get_one::<String>("remote-ext-config");
 
+            let allow_multiple = sub_args.get_flag("allow-multiple");
+
             // If --safekeepers argument is given, use only the listed safekeeper nodes.
             let safekeepers =
                 if let Some(safekeepers_str) = sub_args.get_one::<String>("safekeepers") {
@@ -944,11 +914,13 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 .cloned()
                 .unwrap_or_default();
 
-            cplane.check_conflicting_endpoints(
-                endpoint.mode,
-                endpoint.tenant_id,
-                endpoint.timeline_id,
-            )?;
+            if !allow_multiple {
+                cplane.check_conflicting_endpoints(
+                    endpoint.mode,
+                    endpoint.tenant_id,
+                    endpoint.timeline_id,
+                )?;
+            }
 
             let (pageservers, stripe_size) = if let Some(pageserver_id) = pageserver_id {
                 let conf = env.get_pageserver_conf(pageserver_id).unwrap();
@@ -1142,21 +1114,6 @@ async fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) ->
             }
         }
 
-        Some(("set-state", subcommand_args)) => {
-            let pageserver = get_pageserver(env, subcommand_args)?;
-            let scheduling = subcommand_args.get_one("scheduling");
-            let availability = subcommand_args.get_one("availability");
-
-            let storage_controller = StorageController::from_env(env);
-            storage_controller
-                .node_configure(NodeConfigureRequest {
-                    node_id: pageserver.conf.id,
-                    scheduling: scheduling.cloned(),
-                    availability: availability.cloned(),
-                })
-                .await?;
-        }
-
         Some(("status", subcommand_args)) => {
             match get_pageserver(env, subcommand_args)?.check_status().await {
                 Ok(_) => println!("Page server is up and running"),
@@ -1330,7 +1287,7 @@ async fn try_stop_all(env: &local_env::LocalEnv, immediate: bool) {
     match ComputeControlPlane::load(env.clone()) {
         Ok(cplane) => {
             for (_k, node) in cplane.endpoints {
-                if let Err(e) = node.stop(if immediate { "immediate" } else { "fast " }, false) {
+                if let Err(e) = node.stop(if immediate { "immediate" } else { "fast" }, false) {
                     eprintln!("postgres stop failed: {e:#}");
                 }
             }
@@ -1495,6 +1452,12 @@ fn cli() -> Command {
         .help("If set, will create test user `user` and `neondb` database. Requires `update-catalog = true`")
         .required(false);
 
+    let allow_multiple = Arg::new("allow-multiple")
+        .help("Allow multiple primary endpoints running on the same branch. Shouldn't be used normally, but useful for tests.")
+        .long("allow-multiple")
+        .action(ArgAction::SetTrue)
+        .required(false);
+
     Command::new("Neon CLI")
         .arg_required_else_help(true)
         .version(GIT_VERSION)
@@ -1516,6 +1479,7 @@ fn cli() -> Command {
         .subcommand(
             Command::new("timeline")
             .about("Manage timelines")
+            .arg_required_else_help(true)
             .subcommand(Command::new("list")
                 .about("List all timelines, available to this pageserver")
                 .arg(tenant_id_arg.clone()))
@@ -1578,19 +1542,8 @@ fn cli() -> Command {
             .subcommand(Command::new("config")
                 .arg(tenant_id_arg.clone())
                 .arg(Arg::new("config").short('c').num_args(1).action(ArgAction::Append).required(false)))
-            .subcommand(Command::new("migrate")
-                .about("Migrate a tenant from one pageserver to another")
-                .arg(tenant_id_arg.clone())
-                .arg(pageserver_id_arg.clone()))
-            .subcommand(Command::new("status")
-                .about("Human readable summary of the tenant's shards and attachment locations")
-                .arg(tenant_id_arg.clone()))
-            .subcommand(Command::new("shard-split")
-                .about("Increase the number of shards in the tenant")
-                .arg(tenant_id_arg.clone())
-                .arg(Arg::new("shard-count").value_parser(value_parser!(u8)).long("shard-count").action(ArgAction::Set).help("Number of shards in the new tenant (default 1)"))
-                .arg(Arg::new("shard-stripe-size").value_parser(value_parser!(u32)).long("shard-stripe-size").action(ArgAction::Set).help("Sharding stripe size in pages"))
-                )
+            .subcommand(Command::new("import").arg(tenant_id_arg.clone().required(true))
+                .about("Import a tenant that is present in remote storage, and create branches for its timelines"))
         )
         .subcommand(
             Command::new("pageserver")
@@ -1610,19 +1563,13 @@ fn cli() -> Command {
                     .about("Restart local pageserver")
                     .arg(pageserver_config_args.clone())
                 )
-                .subcommand(Command::new("set-state")
-                    .arg(Arg::new("availability").value_parser(value_parser!(NodeAvailability)).long("availability").action(ArgAction::Set).help("Availability state: offline,active"))
-                    .arg(Arg::new("scheduling").value_parser(value_parser!(NodeSchedulingPolicy)).long("scheduling").action(ArgAction::Set).help("Scheduling state: draining,pause,filling,active"))
-                    .about("Set scheduling or availability state of pageserver node")
-                    .arg(pageserver_config_args.clone())
-                )
         )
         .subcommand(
             Command::new("storage_controller")
                 .arg_required_else_help(true)
                 .about("Manage storage_controller")
-                .subcommand(Command::new("start").about("Start local pageserver").arg(pageserver_config_args.clone()))
-                .subcommand(Command::new("stop").about("Stop local pageserver")
+                .subcommand(Command::new("start").about("Start storage controller"))
+                .subcommand(Command::new("stop").about("Stop storage controller")
                             .arg(stop_mode_arg.clone()))
         )
         .subcommand(
@@ -1668,6 +1615,7 @@ fn cli() -> Command {
                     .arg(pg_version_arg.clone())
                     .arg(hot_standby_arg.clone())
                     .arg(update_catalog)
+                    .arg(allow_multiple.clone())
                 )
                 .subcommand(Command::new("start")
                     .about("Start postgres.\n If the endpoint doesn't exist yet, it is created.")
@@ -1676,6 +1624,7 @@ fn cli() -> Command {
                     .arg(safekeepers_arg)
                     .arg(remote_ext_config_args)
                     .arg(create_test_user)
+                    .arg(allow_multiple.clone())
                 )
                 .subcommand(Command::new("reconfigure")
                             .about("Reconfigure the endpoint")

control_plane/src/endpoint.rs

@@ -12,7 +12,7 @@
 //!
 //! The endpoint is managed by the `compute_ctl` binary. When an endpoint is
 //! started, we launch `compute_ctl` It synchronizes the safekeepers, downloads
-//! the basebackup from the pageserver to initialize the the data directory, and
+//! the basebackup from the pageserver to initialize the data directory, and
 //! finally launches the PostgreSQL process. It watches the PostgreSQL process
 //! until it exits.
 //!
@@ -554,6 +554,7 @@ impl Endpoint {
             format_version: 1.0,
             operation_uuid: None,
             features: self.features.clone(),
+            swap_size_bytes: None,
             cluster: Cluster {
                 cluster_id: None, // project ID: not used
                 name: None,       // project name: not used

control_plane/src/local_env.rs

@@ -17,6 +17,7 @@ use std::net::Ipv4Addr;
 use std::net::SocketAddr;
 use std::path::{Path, PathBuf};
 use std::process::{Command, Stdio};
+use std::time::Duration;
 use utils::{
     auth::{encode_from_key_file, Claims},
     id::{NodeId, TenantId, TenantTimelineId, TimelineId},
@@ -66,6 +67,10 @@ pub struct LocalEnv {
 
     pub broker: NeonBroker,
 
+    // Configuration for the storage controller (1 per neon_local environment)
+    #[serde(default)]
+    pub storage_controller: NeonStorageControllerConf,
+
     /// This Vec must always contain at least one pageserver
     pub pageservers: Vec<PageServerConf>,
 
@@ -98,6 +103,29 @@ pub struct NeonBroker {
     pub listen_addr: SocketAddr,
 }
 
+/// Broker config for cluster internal communication.
+#[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
+#[serde(default)]
+pub struct NeonStorageControllerConf {
+    /// Heartbeat timeout before marking a node offline
+    #[serde(with = "humantime_serde")]
+    pub max_unavailable: Duration,
+}
+
+impl NeonStorageControllerConf {
+    // Use a shorter pageserver unavailability interval than the default to speed up tests.
+    const DEFAULT_MAX_UNAVAILABLE_INTERVAL: std::time::Duration =
+        std::time::Duration::from_secs(10);
+}
+
+impl Default for NeonStorageControllerConf {
+    fn default() -> Self {
+        Self {
+            max_unavailable: Self::DEFAULT_MAX_UNAVAILABLE_INTERVAL,
+        }
+    }
+}
+
 // Dummy Default impl to satisfy Deserialize derive.
 impl Default for NeonBroker {
     fn default() -> Self {
@@ -114,7 +142,7 @@ impl NeonBroker {
 }
 
 #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
-#[serde(default)]
+#[serde(default, deny_unknown_fields)]
 pub struct PageServerConf {
     // node id
     pub id: NodeId,
@@ -126,6 +154,11 @@ pub struct PageServerConf {
     // auth type used for the PG and HTTP ports
     pub pg_auth_type: AuthType,
     pub http_auth_type: AuthType,
+
+    pub(crate) virtual_file_io_engine: Option<String>,
+    pub(crate) get_vectored_impl: Option<String>,
+    pub(crate) get_impl: Option<String>,
+    pub(crate) validate_vectored_get: Option<bool>,
 }
 
 impl Default for PageServerConf {
@@ -136,6 +169,10 @@ impl Default for PageServerConf {
             listen_http_addr: String::new(),
             pg_auth_type: AuthType::Trust,
             http_auth_type: AuthType::Trust,
+            virtual_file_io_engine: None,
+            get_vectored_impl: None,
+            get_impl: None,
+            validate_vectored_get: None,
         }
     }
 }
@@ -151,6 +188,7 @@ pub struct SafekeeperConf {
     pub remote_storage: Option<String>,
     pub backup_threads: Option<u32>,
     pub auth_enabled: bool,
+    pub listen_addr: Option<String>,
 }
 
 impl Default for SafekeeperConf {
@@ -164,6 +202,7 @@ impl Default for SafekeeperConf {
             remote_storage: None,
             backup_threads: None,
             auth_enabled: false,
+            listen_addr: None,
         }
     }
 }
@@ -343,7 +382,10 @@ impl LocalEnv {
 
         // Find neon binaries.
         if env.neon_distrib_dir == Path::new("") {
-            env.neon_distrib_dir = env::current_exe()?.parent().unwrap().to_owned();
+            env::current_exe()?
+                .parent()
+                .unwrap()
+                .clone_into(&mut env.neon_distrib_dir);
         }
 
         if env.pageservers.is_empty() {

control_plane/src/pageserver.rs

@@ -4,7 +4,6 @@
 //!
 //!   .neon/
 //!
-use std::borrow::Cow;
 use std::collections::HashMap;
 
 use std::io;
@@ -78,18 +77,51 @@ impl PageServerNode {
     ///
     /// These all end up on the command line of the `pageserver` binary.
     fn neon_local_overrides(&self, cli_overrides: &[&str]) -> Vec<String> {
-        let id = format!("id={}", self.conf.id);
         // FIXME: the paths should be shell-escaped to handle paths with spaces, quotas etc.
         let pg_distrib_dir_param = format!(
             "pg_distrib_dir='{}'",
             self.env.pg_distrib_dir_raw().display()
         );
 
-        let http_auth_type_param = format!("http_auth_type='{}'", self.conf.http_auth_type);
-        let listen_http_addr_param = format!("listen_http_addr='{}'", self.conf.listen_http_addr);
+        let PageServerConf {
+            id,
+            listen_pg_addr,
+            listen_http_addr,
+            pg_auth_type,
+            http_auth_type,
+            virtual_file_io_engine,
+            get_vectored_impl,
+            get_impl,
+            validate_vectored_get,
+        } = &self.conf;
 
-        let pg_auth_type_param = format!("pg_auth_type='{}'", self.conf.pg_auth_type);
-        let listen_pg_addr_param = format!("listen_pg_addr='{}'", self.conf.listen_pg_addr);
+        let id = format!("id={}", id);
+
+        let http_auth_type_param = format!("http_auth_type='{}'", http_auth_type);
+        let listen_http_addr_param = format!("listen_http_addr='{}'", listen_http_addr);
+
+        let pg_auth_type_param = format!("pg_auth_type='{}'", pg_auth_type);
+        let listen_pg_addr_param = format!("listen_pg_addr='{}'", listen_pg_addr);
+        let virtual_file_io_engine = if let Some(virtual_file_io_engine) = virtual_file_io_engine {
+            format!("virtual_file_io_engine='{virtual_file_io_engine}'")
+        } else {
+            String::new()
+        };
+        let get_vectored_impl = if let Some(get_vectored_impl) = get_vectored_impl {
+            format!("get_vectored_impl='{get_vectored_impl}'")
+        } else {
+            String::new()
+        };
+        let get_impl = if let Some(get_impl) = get_impl {
+            format!("get_impl='{get_impl}'")
+        } else {
+            String::new()
+        };
+        let validate_vectored_get = if let Some(validate_vectored_get) = validate_vectored_get {
+            format!("validate_vectored_get={validate_vectored_get}")
+        } else {
+            String::new()
+        };
 
         let broker_endpoint_param = format!("broker_endpoint='{}'", self.env.broker.client_url());
 
@@ -101,6 +133,10 @@ impl PageServerNode {
             listen_http_addr_param,
             listen_pg_addr_param,
             broker_endpoint_param,
+            virtual_file_io_engine,
+            get_vectored_impl,
+            get_impl,
+            validate_vectored_get,
         ];
 
         if let Some(control_plane_api) = &self.env.control_plane_api {
@@ -111,7 +147,7 @@ impl PageServerNode {
 
             // Storage controller uses the same auth as pageserver: if JWT is enabled
             // for us, we will also need it to talk to them.
-            if matches!(self.conf.http_auth_type, AuthType::NeonJWT) {
+            if matches!(http_auth_type, AuthType::NeonJWT) {
                 let jwt_token = self
                     .env
                     .generate_auth_token(&Claims::new(None, Scope::GenerationsApi))
@@ -129,8 +165,7 @@ impl PageServerNode {
             ));
         }
 
-        if self.conf.http_auth_type != AuthType::Trust || self.conf.pg_auth_type != AuthType::Trust
-        {
+        if *http_auth_type != AuthType::Trust || *pg_auth_type != AuthType::Trust {
             // Keys are generated in the toplevel repo dir, pageservers' workdirs
             // are one level below that, so refer to keys with ../
             overrides.push("auth_validation_public_key_path='../auth_public_key.pem'".to_owned());
@@ -162,7 +197,7 @@ impl PageServerNode {
     }
 
     pub async fn start(&self, config_overrides: &[&str]) -> anyhow::Result<()> {
-        self.start_node(config_overrides, false).await
+        self.start_node(config_overrides).await
     }
 
     fn pageserver_init(&self, config_overrides: &[&str]) -> anyhow::Result<()> {
@@ -183,11 +218,18 @@ impl PageServerNode {
         let datadir_path_str = datadir.to_str().with_context(|| {
             format!("Cannot start pageserver node {node_id} in path that has no string representation: {datadir:?}")
         })?;
-        let mut args = self.pageserver_basic_args(config_overrides, datadir_path_str);
-        args.push(Cow::Borrowed("--init"));
 
+        // `pageserver --init` merges the `--config-override`s into a built-in default config,
+        // then writes out the merged product to `pageserver.toml`.
+        // TODO: just write the full `pageserver.toml` and get rid of `--config-override`.
+        let mut args = vec!["--init", "--workdir", datadir_path_str];
+        let overrides = self.neon_local_overrides(config_overrides);
+        for piece in &overrides {
+            args.push("--config-override");
+            args.push(piece);
+        }
         let init_output = Command::new(self.env.pageserver_bin())
-            .args(args.iter().map(Cow::as_ref))
+            .args(args)
             .envs(self.pageserver_env_variables()?)
             .output()
             .with_context(|| format!("Failed to run pageserver init for node {node_id}"))?;
@@ -212,12 +254,13 @@ impl PageServerNode {
         // situation: the metadata is written by some other script.
         std::fs::write(
             metadata_path,
-            serde_json::to_vec(&serde_json::json!({
-                "host": "localhost",
-                "port": self.pg_connection_config.port(),
-                "http_host": "localhost",
-                "http_port": http_port,
-            }))
+            serde_json::to_vec(&pageserver_api::config::NodeMetadata {
+                postgres_host: "localhost".to_string(),
+                postgres_port: self.pg_connection_config.port(),
+                http_host: "localhost".to_string(),
+                http_port,
+                other: HashMap::new(),
+            })
             .unwrap(),
         )
         .expect("Failed to write metadata file");
@@ -225,11 +268,7 @@ impl PageServerNode {
         Ok(())
     }
 
-    async fn start_node(
-        &self,
-        config_overrides: &[&str],
-        update_config: bool,
-    ) -> anyhow::Result<()> {
+    async fn start_node(&self, config_overrides: &[&str]) -> anyhow::Result<()> {
         // TODO: using a thread here because start_process() is not async but we need to call check_status()
         let datadir = self.repo_path();
         print!(
@@ -246,15 +285,16 @@ impl PageServerNode {
                 self.conf.id, datadir,
             )
         })?;
-        let mut args = self.pageserver_basic_args(config_overrides, datadir_path_str);
-        if update_config {
-            args.push(Cow::Borrowed("--update-config"));
+        let mut args = vec!["-D", datadir_path_str];
+        for config_override in config_overrides {
+            args.push("--config-override");
+            args.push(*config_override);
         }
         background_process::start_process(
             "pageserver",
             &datadir,
             &self.env.pageserver_bin(),
-            args.iter().map(Cow::as_ref),
+            args,
             self.pageserver_env_variables()?,
             background_process::InitialPidFile::Expect(self.pid_file()),
             || async {
@@ -271,22 +311,6 @@ impl PageServerNode {
         Ok(())
     }
 
-    fn pageserver_basic_args<'a>(
-        &self,
-        config_overrides: &'a [&'a str],
-        datadir_path_str: &'a str,
-    ) -> Vec<Cow<'a, str>> {
-        let mut args = vec![Cow::Borrowed("-D"), Cow::Borrowed(datadir_path_str)];
-
-        let overrides = self.neon_local_overrides(config_overrides);
-        for config_override in overrides {
-            args.push(Cow::Borrowed("-c"));
-            args.push(Cow::Owned(config_override));
-        }
-
-        args
-    }
-
     fn pageserver_env_variables(&self) -> anyhow::Result<Vec<(String, String)>> {
         // FIXME: why is this tied to pageserver's auth type? Whether or not the safekeeper
         // needs a token, and how to generate that token, seems independent to whether
@@ -367,6 +391,10 @@ impl PageServerNode {
                 .remove("image_creation_threshold")
                 .map(|x| x.parse::<usize>())
                 .transpose()?,
+            image_layer_creation_check_threshold: settings
+                .remove("image_layer_creation_check_threshold")
+                .map(|x| x.parse::<u8>())
+                .transpose()?,
             pitr_interval: settings.remove("pitr_interval").map(|x| x.to_string()),
             walreceiver_connect_timeout: settings
                 .remove("walreceiver_connect_timeout")
@@ -384,11 +412,6 @@ impl PageServerNode {
                 .map(|x| x.parse::<bool>())
                 .transpose()
                 .context("Failed to parse 'trace_read_requests' as bool")?,
-            image_layer_compression: settings
-                .remove("image_layer_compression")
-                .map(serde_json::from_str)
-                .transpose()
-                .context("Failed to parse 'image_layer_compression' json")?,
             eviction_policy: settings
                 .remove("eviction_policy")
                 .map(serde_json::from_str)
@@ -413,6 +436,11 @@ impl PageServerNode {
                 .map(serde_json::from_str)
                 .transpose()
                 .context("parse `timeline_get_throttle` from json")?,
+            switch_to_aux_file_v2: settings
+                .remove("switch_to_aux_file_v2")
+                .map(|x| x.parse::<bool>())
+                .transpose()
+                .context("Failed to parse 'switch_to_aux_file_v2' as bool")?,
         };
         if !settings.is_empty() {
             bail!("Unrecognized tenant settings: {settings:?}")
@@ -484,6 +512,12 @@ impl PageServerNode {
                     .map(|x| x.parse::<usize>())
                     .transpose()
                     .context("Failed to parse 'image_creation_threshold' as non zero integer")?,
+                image_layer_creation_check_threshold: settings
+                    .remove("image_layer_creation_check_threshold")
+                    .map(|x| x.parse::<u8>())
+                    .transpose()
+                    .context("Failed to parse 'image_creation_check_threshold' as integer")?,
+
                 pitr_interval: settings.remove("pitr_interval").map(|x| x.to_string()),
                 walreceiver_connect_timeout: settings
                     .remove("walreceiver_connect_timeout")
@@ -501,11 +535,6 @@ impl PageServerNode {
                     .map(|x| x.parse::<bool>())
                     .transpose()
                     .context("Failed to parse 'trace_read_requests' as bool")?,
-                image_layer_compression: settings
-                    .remove("image_layer_compression")
-                    .map(serde_json::from_str)
-                    .transpose()
-                    .context("Failed to parse 'image_layer_compression' json")?,
                 eviction_policy: settings
                     .remove("eviction_policy")
                     .map(serde_json::from_str)
@@ -530,6 +559,11 @@ impl PageServerNode {
                     .map(serde_json::from_str)
                     .transpose()
                     .context("parse `timeline_get_throttle` from json")?,
+                switch_to_aux_file_v2: settings
+                    .remove("switch_to_aux_file_v2")
+                    .map(|x| x.parse::<bool>())
+                    .transpose()
+                    .context("Failed to parse 'switch_to_aux_file_v2' as bool")?,
             }
         };
 
@@ -564,13 +598,6 @@ impl PageServerNode {
         Ok(self.http_client.list_timelines(*tenant_shard_id).await?)
     }
 
-    pub async fn tenant_secondary_download(&self, tenant_id: &TenantShardId) -> anyhow::Result<()> {
-        Ok(self
-            .http_client
-            .tenant_secondary_download(*tenant_id)
-            .await?)
-    }
-
     pub async fn timeline_create(
         &self,
         tenant_shard_id: TenantShardId,

control_plane/src/safekeeper.rs

@@ -70,24 +70,31 @@ pub struct SafekeeperNode {
     pub pg_connection_config: PgConnectionConfig,
     pub env: LocalEnv,
     pub http_client: reqwest::Client,
+    pub listen_addr: String,
     pub http_base_url: String,
 }
 
 impl SafekeeperNode {
     pub fn from_env(env: &LocalEnv, conf: &SafekeeperConf) -> SafekeeperNode {
+        let listen_addr = if let Some(ref listen_addr) = conf.listen_addr {
+            listen_addr.clone()
+        } else {
+            "127.0.0.1".to_string()
+        };
         SafekeeperNode {
             id: conf.id,
             conf: conf.clone(),
-            pg_connection_config: Self::safekeeper_connection_config(conf.pg_port),
+            pg_connection_config: Self::safekeeper_connection_config(&listen_addr, conf.pg_port),
             env: env.clone(),
             http_client: reqwest::Client::new(),
-            http_base_url: format!("http://127.0.0.1:{}/v1", conf.http_port),
+            http_base_url: format!("http://{}:{}/v1", listen_addr, conf.http_port),
+            listen_addr,
         }
     }
 
     /// Construct libpq connection string for connecting to this safekeeper.
-    fn safekeeper_connection_config(port: u16) -> PgConnectionConfig {
-        PgConnectionConfig::new_host_port(url::Host::parse("127.0.0.1").unwrap(), port)
+    fn safekeeper_connection_config(addr: &str, port: u16) -> PgConnectionConfig {
+        PgConnectionConfig::new_host_port(url::Host::parse(addr).unwrap(), port)
     }
 
     pub fn datadir_path_by_id(env: &LocalEnv, sk_id: NodeId) -> PathBuf {
@@ -111,8 +118,8 @@ impl SafekeeperNode {
         );
         io::stdout().flush().unwrap();
 
-        let listen_pg = format!("127.0.0.1:{}", self.conf.pg_port);
-        let listen_http = format!("127.0.0.1:{}", self.conf.http_port);
+        let listen_pg = format!("{}:{}", self.listen_addr, self.conf.pg_port);
+        let listen_http = format!("{}:{}", self.listen_addr, self.conf.http_port);
         let id = self.id;
         let datadir = self.datadir_path();
 
@@ -139,7 +146,7 @@ impl SafekeeperNode {
             availability_zone,
         ];
         if let Some(pg_tenant_only_port) = self.conf.pg_tenant_only_port {
-            let listen_pg_tenant_only = format!("127.0.0.1:{}", pg_tenant_only_port);
+            let listen_pg_tenant_only = format!("{}:{}", self.listen_addr, pg_tenant_only_port);
             args.extend(["--listen-pg-tenant-only".to_owned(), listen_pg_tenant_only]);
         }
         if !self.conf.sync {

control_plane/src/storage_controller.rs

@@ -1,6 +1,8 @@
-use crate::{background_process, local_env::LocalEnv};
+use crate::{
+    background_process,
+    local_env::{LocalEnv, NeonStorageControllerConf},
+};
 use camino::{Utf8Path, Utf8PathBuf};
-use hyper::Method;
 use pageserver_api::{
     controller_api::{
         NodeConfigureRequest, NodeRegisterRequest, TenantCreateResponse, TenantLocateResponse,
@@ -14,6 +16,7 @@ use pageserver_api::{
 };
 use pageserver_client::mgmt_api::ResponseErrorMessageExt;
 use postgres_backend::AuthType;
+use reqwest::Method;
 use serde::{de::DeserializeOwned, Deserialize, Serialize};
 use std::{fs, str::FromStr};
 use tokio::process::Command;
@@ -32,6 +35,7 @@ pub struct StorageController {
     public_key: Option<String>,
     postgres_port: u16,
     client: reqwest::Client,
+    config: NeonStorageControllerConf,
 }
 
 const COMMAND: &str = "storage_controller";
@@ -132,6 +136,7 @@ impl StorageController {
             client: reqwest::ClientBuilder::new()
                 .build()
                 .expect("Failed to construct http client"),
+            config: env.storage_controller.clone(),
         }
     }
 
@@ -274,8 +279,11 @@ impl StorageController {
             &self.listen,
             "-p",
             self.path.as_ref(),
+            "--dev",
             "--database-url",
             &database_url,
+            "--max-unavailable-interval",
+            &humantime::Duration::from(self.config.max_unavailable).to_string(),
         ]
         .into_iter()
         .map(|s| s.to_string())
@@ -371,7 +379,7 @@ impl StorageController {
     /// Simple HTTP request wrapper for calling into storage controller
     async fn dispatch<RQ, RS>(
         &self,
-        method: hyper::Method,
+        method: reqwest::Method,
         path: String,
         body: Option<RQ>,
     ) -> anyhow::Result<RS>
@@ -464,11 +472,21 @@ impl StorageController {
             .await
     }
 
+    #[instrument(skip(self))]
+    pub async fn tenant_import(&self, tenant_id: TenantId) -> anyhow::Result<TenantCreateResponse> {
+        self.dispatch::<(), TenantCreateResponse>(
+            Method::POST,
+            format!("debug/v1/tenant/{tenant_id}/import"),
+            None,
+        )
+        .await
+    }
+
     #[instrument(skip(self))]
     pub async fn tenant_locate(&self, tenant_id: TenantId) -> anyhow::Result<TenantLocateResponse> {
         self.dispatch::<(), _>(
             Method::GET,
-            format!("control/v1/tenant/{tenant_id}/locate"),
+            format!("debug/v1/tenant/{tenant_id}/locate"),
             None,
         )
         .await

control_plane/storcon_cli/Cargo.toml

@@ -0,0 +1,23 @@
+[package]
+name = "storcon_cli"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+
+[dependencies]
+anyhow.workspace = true
+clap.workspace = true
+comfy-table.workspace = true
+hyper.workspace = true
+pageserver_api.workspace = true
+pageserver_client.workspace = true
+reqwest.workspace = true
+serde.workspace = true
+serde_json = { workspace = true, features = ["raw_value"] }
+thiserror.workspace = true
+tokio.workspace = true
+tracing.workspace = true
+utils.workspace = true
+workspace_hack.workspace = true
+

control_plane/storcon_cli/src/main.rs

@@ -0,0 +1,680 @@
+use std::{collections::HashMap, str::FromStr, time::Duration};
+
+use clap::{Parser, Subcommand};
+use pageserver_api::{
+    controller_api::{
+        NodeAvailabilityWrapper, NodeDescribeResponse, ShardSchedulingPolicy,
+        TenantDescribeResponse, TenantPolicyRequest,
+    },
+    models::{
+        LocationConfigSecondary, ShardParameters, TenantConfig, TenantConfigRequest,
+        TenantCreateRequest, TenantShardSplitRequest, TenantShardSplitResponse,
+    },
+    shard::{ShardStripeSize, TenantShardId},
+};
+use pageserver_client::mgmt_api::{self, ResponseErrorMessageExt};
+use reqwest::{Method, StatusCode, Url};
+use serde::{de::DeserializeOwned, Serialize};
+use utils::id::{NodeId, TenantId};
+
+use pageserver_api::controller_api::{
+    NodeConfigureRequest, NodeRegisterRequest, NodeSchedulingPolicy, PlacementPolicy,
+    TenantLocateResponse, TenantShardMigrateRequest, TenantShardMigrateResponse,
+};
+
+#[derive(Subcommand, Debug)]
+enum Command {
+    /// Register a pageserver with the storage controller.  This shouldn't usually be necessary,
+    /// since pageservers auto-register when they start up
+    NodeRegister {
+        #[arg(long)]
+        node_id: NodeId,
+
+        #[arg(long)]
+        listen_pg_addr: String,
+        #[arg(long)]
+        listen_pg_port: u16,
+
+        #[arg(long)]
+        listen_http_addr: String,
+        #[arg(long)]
+        listen_http_port: u16,
+    },
+
+    /// Modify a node's configuration in the storage controller
+    NodeConfigure {
+        #[arg(long)]
+        node_id: NodeId,
+
+        /// Availability is usually auto-detected based on heartbeats.  Set 'offline' here to
+        /// manually mark a node offline
+        #[arg(long)]
+        availability: Option<NodeAvailabilityArg>,
+        /// Scheduling policy controls whether tenant shards may be scheduled onto this node.
+        #[arg(long)]
+        scheduling: Option<NodeSchedulingPolicy>,
+    },
+    /// Modify a tenant's policies in the storage controller
+    TenantPolicy {
+        #[arg(long)]
+        tenant_id: TenantId,
+        /// Placement policy controls whether a tenant is `detached`, has only a secondary location (`secondary`),
+        /// or is in the normal attached state with N secondary locations (`attached:N`)
+        #[arg(long)]
+        placement: Option<PlacementPolicyArg>,
+        /// Scheduling policy enables pausing the controller's scheduling activity involving this tenant.  `active` is normal,
+        /// `essential` disables optimization scheduling changes, `pause` disables all scheduling changes, and `stop` prevents
+        /// all reconciliation activity including for scheduling changes already made.  `pause` and `stop` can make a tenant
+        /// unavailable, and are only for use in emergencies.
+        #[arg(long)]
+        scheduling: Option<ShardSchedulingPolicyArg>,
+    },
+    /// List nodes known to the storage controller
+    Nodes {},
+    /// List tenants known to the storage controller
+    Tenants {},
+    /// Create a new tenant in the storage controller, and by extension on pageservers.
+    TenantCreate {
+        #[arg(long)]
+        tenant_id: TenantId,
+    },
+    /// Delete a tenant in the storage controller, and by extension on pageservers.
+    TenantDelete {
+        #[arg(long)]
+        tenant_id: TenantId,
+    },
+    /// Split an existing tenant into a higher number of shards than its current shard count.
+    TenantShardSplit {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        shard_count: u8,
+        /// Optional, in 8kiB pages.  e.g. set 2048 for 16MB stripes.
+        #[arg(long)]
+        stripe_size: Option<u32>,
+    },
+    /// Migrate the attached location for a tenant shard to a specific pageserver.
+    TenantShardMigrate {
+        #[arg(long)]
+        tenant_shard_id: TenantShardId,
+        #[arg(long)]
+        node: NodeId,
+    },
+    /// Modify the pageserver tenant configuration of a tenant: this is the configuration structure
+    /// that is passed through to pageservers, and does not affect storage controller behavior.
+    TenantConfig {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        config: String,
+    },
+    /// Attempt to balance the locations for a tenant across pageservers.  This is a client-side
+    /// alternative to the storage controller's scheduling optimization behavior.
+    TenantScatter {
+        #[arg(long)]
+        tenant_id: TenantId,
+    },
+    /// Print details about a particular tenant, including all its shards' states.
+    TenantDescribe {
+        #[arg(long)]
+        tenant_id: TenantId,
+    },
+    /// For a tenant which hasn't been onboarded to the storage controller yet, add it in secondary
+    /// mode so that it can warm up content on a pageserver.
+    TenantWarmup {
+        #[arg(long)]
+        tenant_id: TenantId,
+    },
+}
+
+#[derive(Parser)]
+#[command(
+    author,
+    version,
+    about,
+    long_about = "CLI for Storage Controller Support/Debug"
+)]
+#[command(arg_required_else_help(true))]
+struct Cli {
+    #[arg(long)]
+    /// URL to storage controller.  e.g. http://127.0.0.1:1234 when using `neon_local`
+    api: Url,
+
+    #[arg(long)]
+    /// JWT token for authenticating with storage controller.  Depending on the API used, this
+    /// should have either `pageserverapi` or `admin` scopes: for convenience, you should mint
+    /// a token with both scopes to use with this tool.
+    jwt: Option<String>,
+
+    #[command(subcommand)]
+    command: Command,
+}
+
+#[derive(Debug, Clone)]
+struct PlacementPolicyArg(PlacementPolicy);
+
+impl FromStr for PlacementPolicyArg {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "detached" => Ok(Self(PlacementPolicy::Detached)),
+            "secondary" => Ok(Self(PlacementPolicy::Secondary)),
+            _ if s.starts_with("attached:") => {
+                let mut splitter = s.split(':');
+                let _prefix = splitter.next().unwrap();
+                match splitter.next().and_then(|s| s.parse::<usize>().ok()) {
+                    Some(n) => Ok(Self(PlacementPolicy::Attached(n))),
+                    None => Err(anyhow::anyhow!(
+                        "Invalid format '{s}', a valid example is 'attached:1'"
+                    )),
+                }
+            }
+            _ => Err(anyhow::anyhow!(
+                "Unknown placement policy '{s}', try detached,secondary,attached:<n>"
+            )),
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+struct ShardSchedulingPolicyArg(ShardSchedulingPolicy);
+
+impl FromStr for ShardSchedulingPolicyArg {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "active" => Ok(Self(ShardSchedulingPolicy::Active)),
+            "essential" => Ok(Self(ShardSchedulingPolicy::Essential)),
+            "pause" => Ok(Self(ShardSchedulingPolicy::Pause)),
+            "stop" => Ok(Self(ShardSchedulingPolicy::Stop)),
+            _ => Err(anyhow::anyhow!(
+                "Unknown scheduling policy '{s}', try active,essential,pause,stop"
+            )),
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+struct NodeAvailabilityArg(NodeAvailabilityWrapper);
+
+impl FromStr for NodeAvailabilityArg {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "active" => Ok(Self(NodeAvailabilityWrapper::Active)),
+            "offline" => Ok(Self(NodeAvailabilityWrapper::Offline)),
+            _ => Err(anyhow::anyhow!("Unknown availability state '{s}'")),
+        }
+    }
+}
+
+struct Client {
+    base_url: Url,
+    jwt_token: Option<String>,
+    client: reqwest::Client,
+}
+
+impl Client {
+    fn new(base_url: Url, jwt_token: Option<String>) -> Self {
+        Self {
+            base_url,
+            jwt_token,
+            client: reqwest::ClientBuilder::new()
+                .build()
+                .expect("Failed to construct http client"),
+        }
+    }
+
+    /// Simple HTTP request wrapper for calling into storage controller
+    async fn dispatch<RQ, RS>(
+        &self,
+        method: Method,
+        path: String,
+        body: Option<RQ>,
+    ) -> mgmt_api::Result<RS>
+    where
+        RQ: Serialize + Sized,
+        RS: DeserializeOwned + Sized,
+    {
+        // The configured URL has the /upcall path prefix for pageservers to use: we will strip that out
+        // for general purpose API access.
+        let url = Url::from_str(&format!(
+            "http://{}:{}/{path}",
+            self.base_url.host_str().unwrap(),
+            self.base_url.port().unwrap()
+        ))
+        .unwrap();
+
+        let mut builder = self.client.request(method, url);
+        if let Some(body) = body {
+            builder = builder.json(&body)
+        }
+        if let Some(jwt_token) = &self.jwt_token {
+            builder = builder.header(
+                reqwest::header::AUTHORIZATION,
+                format!("Bearer {jwt_token}"),
+            );
+        }
+
+        let response = builder.send().await.map_err(mgmt_api::Error::ReceiveBody)?;
+        let response = response.error_from_body().await?;
+
+        response
+            .json()
+            .await
+            .map_err(pageserver_client::mgmt_api::Error::ReceiveBody)
+    }
+}
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    let cli = Cli::parse();
+
+    let storcon_client = Client::new(cli.api.clone(), cli.jwt.clone());
+
+    let mut trimmed = cli.api.to_string();
+    trimmed.pop();
+    let vps_client = mgmt_api::Client::new(trimmed, cli.jwt.as_deref());
+
+    match cli.command {
+        Command::NodeRegister {
+            node_id,
+            listen_pg_addr,
+            listen_pg_port,
+            listen_http_addr,
+            listen_http_port,
+        } => {
+            storcon_client
+                .dispatch::<_, ()>(
+                    Method::POST,
+                    "control/v1/node".to_string(),
+                    Some(NodeRegisterRequest {
+                        node_id,
+                        listen_pg_addr,
+                        listen_pg_port,
+                        listen_http_addr,
+                        listen_http_port,
+                    }),
+                )
+                .await?;
+        }
+        Command::TenantCreate { tenant_id } => {
+            vps_client
+                .tenant_create(&TenantCreateRequest {
+                    new_tenant_id: TenantShardId::unsharded(tenant_id),
+                    generation: None,
+                    shard_parameters: ShardParameters::default(),
+                    placement_policy: Some(PlacementPolicy::Attached(1)),
+                    config: TenantConfig::default(),
+                })
+                .await?;
+        }
+        Command::TenantDelete { tenant_id } => {
+            let status = vps_client
+                .tenant_delete(TenantShardId::unsharded(tenant_id))
+                .await?;
+            tracing::info!("Delete status: {}", status);
+        }
+        Command::Nodes {} => {
+            let resp = storcon_client
+                .dispatch::<(), Vec<NodeDescribeResponse>>(
+                    Method::GET,
+                    "control/v1/node".to_string(),
+                    None,
+                )
+                .await?;
+            let mut table = comfy_table::Table::new();
+            table.set_header(["Id", "Hostname", "Scheduling", "Availability"]);
+            for node in resp {
+                table.add_row([
+                    format!("{}", node.id),
+                    node.listen_http_addr,
+                    format!("{:?}", node.scheduling),
+                    format!("{:?}", node.availability),
+                ]);
+            }
+            println!("{table}");
+        }
+        Command::NodeConfigure {
+            node_id,
+            availability,
+            scheduling,
+        } => {
+            let req = NodeConfigureRequest {
+                node_id,
+                availability: availability.map(|a| a.0),
+                scheduling,
+            };
+            storcon_client
+                .dispatch::<_, ()>(
+                    Method::PUT,
+                    format!("control/v1/node/{node_id}/config"),
+                    Some(req),
+                )
+                .await?;
+        }
+        Command::Tenants {} => {
+            let resp = storcon_client
+                .dispatch::<(), Vec<TenantDescribeResponse>>(
+                    Method::GET,
+                    "control/v1/tenant".to_string(),
+                    None,
+                )
+                .await?;
+            let mut table = comfy_table::Table::new();
+            table.set_header([
+                "TenantId",
+                "ShardCount",
+                "StripeSize",
+                "Placement",
+                "Scheduling",
+            ]);
+            for tenant in resp {
+                let shard_zero = tenant.shards.into_iter().next().unwrap();
+                table.add_row([
+                    format!("{}", tenant.tenant_id),
+                    format!("{}", shard_zero.tenant_shard_id.shard_count.literal()),
+                    format!("{:?}", tenant.stripe_size),
+                    format!("{:?}", tenant.policy),
+                    format!("{:?}", shard_zero.scheduling_policy),
+                ]);
+            }
+
+            println!("{table}");
+        }
+        Command::TenantPolicy {
+            tenant_id,
+            placement,
+            scheduling,
+        } => {
+            let req = TenantPolicyRequest {
+                scheduling: scheduling.map(|s| s.0),
+                placement: placement.map(|p| p.0),
+            };
+            storcon_client
+                .dispatch::<_, ()>(
+                    Method::PUT,
+                    format!("control/v1/tenant/{tenant_id}/policy"),
+                    Some(req),
+                )
+                .await?;
+        }
+        Command::TenantShardSplit {
+            tenant_id,
+            shard_count,
+            stripe_size,
+        } => {
+            let req = TenantShardSplitRequest {
+                new_shard_count: shard_count,
+                new_stripe_size: stripe_size.map(ShardStripeSize),
+            };
+
+            let response = storcon_client
+                .dispatch::<TenantShardSplitRequest, TenantShardSplitResponse>(
+                    Method::PUT,
+                    format!("control/v1/tenant/{tenant_id}/shard_split"),
+                    Some(req),
+                )
+                .await?;
+            println!(
+                "Split tenant {} into {} shards: {}",
+                tenant_id,
+                shard_count,
+                response
+                    .new_shards
+                    .iter()
+                    .map(|s| format!("{:?}", s))
+                    .collect::<Vec<_>>()
+                    .join(",")
+            );
+        }
+        Command::TenantShardMigrate {
+            tenant_shard_id,
+            node,
+        } => {
+            let req = TenantShardMigrateRequest {
+                tenant_shard_id,
+                node_id: node,
+            };
+
+            storcon_client
+                .dispatch::<TenantShardMigrateRequest, TenantShardMigrateResponse>(
+                    Method::PUT,
+                    format!("control/v1/tenant/{tenant_shard_id}/migrate"),
+                    Some(req),
+                )
+                .await?;
+        }
+        Command::TenantConfig { tenant_id, config } => {
+            let tenant_conf = serde_json::from_str(&config)?;
+
+            vps_client
+                .tenant_config(&TenantConfigRequest {
+                    tenant_id,
+                    config: tenant_conf,
+                })
+                .await?;
+        }
+        Command::TenantScatter { tenant_id } => {
+            // Find the shards
+            let locate_response = storcon_client
+                .dispatch::<(), TenantLocateResponse>(
+                    Method::GET,
+                    format!("control/v1/tenant/{tenant_id}/locate"),
+                    None,
+                )
+                .await?;
+            let shards = locate_response.shards;
+
+            let mut node_to_shards: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
+            let shard_count = shards.len();
+            for s in shards {
+                let entry = node_to_shards.entry(s.node_id).or_default();
+                entry.push(s.shard_id);
+            }
+
+            // Load list of available nodes
+            let nodes_resp = storcon_client
+                .dispatch::<(), Vec<NodeDescribeResponse>>(
+                    Method::GET,
+                    "control/v1/node".to_string(),
+                    None,
+                )
+                .await?;
+
+            for node in nodes_resp {
+                if matches!(node.availability, NodeAvailabilityWrapper::Active) {
+                    node_to_shards.entry(node.id).or_default();
+                }
+            }
+
+            let max_shard_per_node = shard_count / node_to_shards.len();
+
+            loop {
+                let mut migrate_shard = None;
+                for shards in node_to_shards.values_mut() {
+                    if shards.len() > max_shard_per_node {
+                        // Pick the emptiest
+                        migrate_shard = Some(shards.pop().unwrap());
+                    }
+                }
+                let Some(migrate_shard) = migrate_shard else {
+                    break;
+                };
+
+                // Pick the emptiest node to migrate to
+                let mut destinations = node_to_shards
+                    .iter()
+                    .map(|(k, v)| (k, v.len()))
+                    .collect::<Vec<_>>();
+                destinations.sort_by_key(|i| i.1);
+                let (destination_node, destination_count) = *destinations.first().unwrap();
+                if destination_count + 1 > max_shard_per_node {
+                    // Even the emptiest destination doesn't have space: we're done
+                    break;
+                }
+                let destination_node = *destination_node;
+
+                node_to_shards
+                    .get_mut(&destination_node)
+                    .unwrap()
+                    .push(migrate_shard);
+
+                println!("Migrate {} -> {} ...", migrate_shard, destination_node);
+
+                storcon_client
+                    .dispatch::<TenantShardMigrateRequest, TenantShardMigrateResponse>(
+                        Method::PUT,
+                        format!("control/v1/tenant/{migrate_shard}/migrate"),
+                        Some(TenantShardMigrateRequest {
+                            tenant_shard_id: migrate_shard,
+                            node_id: destination_node,
+                        }),
+                    )
+                    .await?;
+                println!("Migrate {} -> {} OK", migrate_shard, destination_node);
+            }
+
+            // Spread the shards across the nodes
+        }
+        Command::TenantDescribe { tenant_id } => {
+            let describe_response = storcon_client
+                .dispatch::<(), TenantDescribeResponse>(
+                    Method::GET,
+                    format!("control/v1/tenant/{tenant_id}"),
+                    None,
+                )
+                .await?;
+            let shards = describe_response.shards;
+            let mut table = comfy_table::Table::new();
+            table.set_header(["Shard", "Attached", "Secondary", "Last error", "status"]);
+            for shard in shards {
+                let secondary = shard
+                    .node_secondary
+                    .iter()
+                    .map(|n| format!("{}", n))
+                    .collect::<Vec<_>>()
+                    .join(",");
+
+                let mut status_parts = Vec::new();
+                if shard.is_reconciling {
+                    status_parts.push("reconciling");
+                }
+
+                if shard.is_pending_compute_notification {
+                    status_parts.push("pending_compute");
+                }
+
+                if shard.is_splitting {
+                    status_parts.push("splitting");
+                }
+                let status = status_parts.join(",");
+
+                table.add_row([
+                    format!("{}", shard.tenant_shard_id),
+                    shard
+                        .node_attached
+                        .map(|n| format!("{}", n))
+                        .unwrap_or(String::new()),
+                    secondary,
+                    shard.last_error,
+                    status,
+                ]);
+            }
+            println!("{table}");
+        }
+        Command::TenantWarmup { tenant_id } => {
+            let describe_response = storcon_client
+                .dispatch::<(), TenantDescribeResponse>(
+                    Method::GET,
+                    format!("control/v1/tenant/{tenant_id}"),
+                    None,
+                )
+                .await;
+            match describe_response {
+                Ok(describe) => {
+                    if matches!(describe.policy, PlacementPolicy::Secondary) {
+                        // Fine: it's already known to controller in secondary mode: calling
+                        // again to put it into secondary mode won't cause problems.
+                    } else {
+                        anyhow::bail!("Tenant already present with policy {:?}", describe.policy);
+                    }
+                }
+                Err(mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _)) => {
+                    // Fine: this tenant isn't know to the storage controller yet.
+                }
+                Err(e) => {
+                    // Unexpected API error
+                    return Err(e.into());
+                }
+            }
+
+            vps_client
+                .location_config(
+                    TenantShardId::unsharded(tenant_id),
+                    pageserver_api::models::LocationConfig {
+                        mode: pageserver_api::models::LocationConfigMode::Secondary,
+                        generation: None,
+                        secondary_conf: Some(LocationConfigSecondary { warm: true }),
+                        shard_number: 0,
+                        shard_count: 0,
+                        shard_stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE.0,
+                        tenant_conf: TenantConfig::default(),
+                    },
+                    None,
+                    true,
+                )
+                .await?;
+
+            let describe_response = storcon_client
+                .dispatch::<(), TenantDescribeResponse>(
+                    Method::GET,
+                    format!("control/v1/tenant/{tenant_id}"),
+                    None,
+                )
+                .await?;
+
+            let secondary_ps_id = describe_response
+                .shards
+                .first()
+                .unwrap()
+                .node_secondary
+                .first()
+                .unwrap();
+
+            println!("Tenant {tenant_id} warming up on pageserver {secondary_ps_id}");
+            loop {
+                let (status, progress) = vps_client
+                    .tenant_secondary_download(
+                        TenantShardId::unsharded(tenant_id),
+                        Some(Duration::from_secs(10)),
+                    )
+                    .await?;
+                println!(
+                    "Progress: {}/{} layers, {}/{} bytes",
+                    progress.layers_downloaded,
+                    progress.layers_total,
+                    progress.bytes_downloaded,
+                    progress.bytes_total
+                );
+                match status {
+                    StatusCode::OK => {
+                        println!("Download complete");
+                        break;
+                    }
+                    StatusCode::ACCEPTED => {
+                        // Loop
+                    }
+                    _ => {
+                        anyhow::bail!("Unexpected download status: {status}");
+                    }
+                }
+            }
+        }
+    }
+
+    Ok(())
+}

diesel.toml

@@ -2,8 +2,8 @@
 # see https://diesel.rs/guides/configuring-diesel-cli
 
 [print_schema]
-file = "control_plane/attachment_service/src/schema.rs"
+file = "storage_controller/src/schema.rs"
 custom_type_derives = ["diesel::query_builder::QueryId"]
 
 [migrations_directory]
-dir = "control_plane/attachment_service/migrations"
+dir = "storage_controller/migrations"

docs/rfcs/031-sharding-static.md

@@ -0,0 +1,408 @@
+# Sharding Phase 1: Static Key-space Sharding
+
+## Summary
+
+To enable databases with sizes approaching the capacity of a pageserver's disk,
+it is necessary to break up the storage for the database, or _shard_ it.
+
+Sharding in general is a complex area. This RFC aims to define an initial
+capability that will permit creating large-capacity databases using a static configuration
+defined at time of Tenant creation.
+
+## Motivation
+
+Currently, all data for a Tenant, including all its timelines, is stored on a single
+pageserver. The local storage required may be several times larger than the actual
+database size, due to LSM write inflation.
+
+If a database is larger than what one pageserver can hold, then it becomes impossible
+for the pageserver to hold it in local storage, as it must do to provide service to
+clients.
+
+### Prior art
+
+In Neon:
+
+- Layer File Spreading: https://www.notion.so/neondatabase/One-Pager-Layer-File-Spreading-Konstantin-21fd9b11b618475da5f39c61dd8ab7a4
+- Layer File SPreading: https://www.notion.so/neondatabase/One-Pager-Layer-File-Spreading-Christian-eb6b64182a214e11b3fceceee688d843
+- Key Space partitioning: https://www.notion.so/neondatabase/One-Pager-Key-Space-Partitioning-Stas-8e3a28a600a04a25a68523f42a170677
+
+Prior art in other distributed systems is too broad to capture here: pretty much
+any scale out storage system does something like this.
+
+## Requirements
+
+- Enable creating a large (for example, 16TiB) database without requiring dedicated
+  pageserver nodes.
+- Share read/write bandwidth costs for large databases across pageservers, as well
+  as storage capacity, in order to avoid large capacity databases acting as I/O hotspots
+  that disrupt service to other tenants.
+- Our data distribution scheme should handle sparse/nonuniform keys well, since postgres
+  does not write out a single contiguous ranges of page numbers.
+
+_Note: the definition of 'large database' is arbitrary, but the lower bound is to ensure that a database
+that a user might create on a current-gen enterprise SSD should also work well on
+Neon. The upper bound is whatever postgres can handle: i.e. we must make sure that the
+pageserver backend is not the limiting factor in the database size_.
+
+## Non Goals
+
+- Independently distributing timelines within the same tenant. If a tenant has many
+  timelines, then sharding may be a less efficient mechanism for distributing load than
+  sharing out timelines between pageservers.
+- Distributing work in the LSN dimension: this RFC focuses on the Key dimension only,
+  based on the idea that separate mechanisms will make sense for each dimension.
+
+## Impacted Components
+
+pageserver, control plane, postgres/smgr
+
+## Terminology
+
+**Key**: a postgres page number, qualified by relation. In the sense that the pageserver is a versioned key-value store,
+the page number is the key in that store. `Key` is a literal data type in existing code.
+
+**LSN dimension**: this just means the range of LSNs (history), when talking about the range
+of keys and LSNs as a two dimensional space.
+
+## Implementation
+
+### Key sharding vs. LSN sharding
+
+When we think of sharding across the two dimensional key/lsn space, this is an
+opportunity to think about how the two dimensions differ:
+
+- Sharding the key space distributes the _write_ workload of ingesting data
+  and compacting. This work must be carefully managed so that exactly one
+  node owns a given key.
+- Sharding the LSN space distributes the _historical read_ workload. This work
+  can be done by anyone without any special coordination, as long as they can
+  see the remote index and layers.
+
+The key sharding is the harder part, and also the more urgent one, to support larger
+capacity databases. Because distributing historical LSN read work is a relatively
+simpler problem that most users don't have, we defer it to future work. It is anticipated
+that some quite simple P2P offload model will enable distributing work for historical
+reads: a node which is low on space can call out to peer to ask it to download and
+serve reads from a historical layer.
+
+### Key mapping scheme
+
+Having decided to focus on key sharding, we must next decide how we will map
+keys to shards. It is proposed to use a "wide striping" approach, to obtain a good compromise
+between data locality and avoiding entire large relations mapping to the same shard.
+
+We will define two spaces:
+
+- Key space: unsigned integer
+- Shard space: integer from 0 to N-1, where we have N shards.
+
+### Key -> Shard mapping
+
+Keys are currently defined in the pageserver's getpage@lsn interface as follows:
+
+```
+pub struct Key {
+    pub field1: u8,
+    pub field2: u32,
+    pub field3: u32,
+    pub field4: u32,
+    pub field5: u8,
+    pub field6: u32,
+}
+
+
+fn rel_block_to_key(rel: RelTag, blknum: BlockNumber) -> Key {
+    Key {
+        field1: 0x00,
+        field2: rel.spcnode,
+        field3: rel.dbnode,
+        field4: rel.relnode,
+        field5: rel.forknum,
+        field6: blknum,
+    }
+}
+```
+
+_Note: keys for relation metadata are ignored here, as this data will be mirrored to all
+shards. For distribution purposes, we only care about user data keys_
+
+The properties we want from our Key->Shard mapping are:
+
+- Locality in `blknum`, such that adjacent `blknum` will usually map to
+  the same stripe and consequently land on the same shard, even though the overall
+  collection of blocks in a relation will be spread over many stripes and therefore
+  many shards.
+- Avoid the same blknum on different relations landing on the same stripe, so that
+  with many small relations we do not end up aliasing data to the same stripe/shard.
+- Avoid vulnerability to aliasing in the values of relation identity fields, such that
+  if there are patterns in the value of `relnode`, these do not manifest as patterns
+  in data placement.
+
+To accomplish this, the blknum is used to select a stripe, and stripes are
+assigned to shards in a pseudorandom order via a hash. The motivation for
+pseudo-random distribution (rather than sequential mapping of stripe to shard)
+is to avoid I/O hotspots when sequentially reading multiple relations: we don't want
+all relations' stripes to touch pageservers in the same order.
+
+To map a `Key` to a shard:
+
+- Hash the `Key` field 4 (relNode).
+- Divide field 6 (`blknum`) field by the stripe size in pages, and combine the
+  hash of this with the hash from the previous step.
+- The total hash modulo the shard count gives the shard holding this key.
+
+Why don't we use the other fields in the Key?
+
+- We ignore `forknum` for key mapping, because it distinguishes different classes of data
+  in the same relation, and we would like to keep the data in a relation together.
+- We would like to use spcNode and dbNode, but cannot. Postgres database creation operations can refer to an existing database as a template, such that the created
+  database's blocks differ only by spcNode and dbNode from the original. To enable running
+  this type of creation without cross-pageserver communication, we must ensure that these
+  blocks map to the same shard -- we do this by excluding spcNode and dbNode from the hash.
+
+### Data placement examples
+
+For example, consider the extreme large databases cases of postgres data layout in a system with 8 shards
+and a stripe size of 32k pages:
+
+- A single large relation: `blknum` division will break the data up into 4096
+  stripes, which will be scattered across the shards.
+- 4096 relations of of 32k pages each: each relation will map to exactly one stripe,
+  and that stripe will be placed according to the hash of the key fields 4. The
+  data placement will be statistically uniform across shards.
+
+Data placement will be more uneven on smaller databases:
+
+- A tenant with 2 shards and 2 relations of one stripe size each: there is a 50% chance
+  that both relations land on the same shard and no data lands on the other shard.
+- A tenant with 8 shards and one relation of size 12 stripes: 4 shards will have double
+  the data of the other four shards.
+
+These uneven cases for small amounts of data do not matter, as long as the stripe size
+is an order of magnitude smaller than the amount of data we are comfortable holding
+in a single shard: if our system handles shard sizes up to 10-100GB, then it is not an issue if
+a tenant has some shards with 256MB size and some shards with 512MB size, even though
+the standard deviation of shard size within the tenant is very high. Our key mapping
+scheme provides a statistical guarantee that as the tenant's overall data size increases,
+uniformity of placement will improve.
+
+### Important Types
+
+#### `ShardIdentity`
+
+Provides the information needed to know whether a particular key belongs
+to a particular shard:
+
+- Layout version
+- Stripe size
+- Shard count
+- Shard index
+
+This structure's size is constant. Note that if we had used a differnet key
+mapping scheme such as consistent hashing with explicit hash ranges assigned
+to each shard, then the ShardIdentity's size would grow with the shard count: the simpler
+key mapping scheme used here enables a small fixed size ShardIdentity.
+
+### Pageserver changes
+
+#### Structural
+
+Everywhere the Pageserver currently deals with Tenants, it will move to dealing with
+`TenantShard`s, which are just a `Tenant` plus a `ShardIdentity` telling it which part
+of the keyspace it owns. An un-sharded tenant is just a `TenantShard` whose `ShardIdentity`
+covers the whole keyspace.
+
+When the pageserver writes layers and index_part.json to remote storage, it must
+include the shard index & count in the name, to avoid collisions (the count is
+necessary for future-proofing: the count will vary in time). These keys
+will also include a generation number: the [generation numbers](025-generation-numbers.md) system will work
+exactly the same for TenantShards as it does for Tenants today: each shard will have
+its own generation number.
+
+#### Storage Format: Keys
+
+For tenants with >1 shard, layer files implicitly become sparse: within the key
+range described in the layer name, the layer file for a shard will only hold the
+content relevant to stripes assigned to the shard.
+
+For this reason, the LayerFileName within a tenant is no longer unique: different shards
+may use the same LayerFileName to refer to different data. We may solve this simply
+by including the shard number in the keys used for layers.
+
+The shard number will be included as a prefix (as part of tenant ID), like this:
+
+`pageserver/v1/tenants/<tenant_id>-<shard_number><shard_count>/timelines/<timeline id>/<layer file name>-<generation>`
+
+`pageserver/v1/tenants/<tenant_id>-<shard_number><shard_count>/timelines/<timeline id>/index_part.json-<generation>`
+
+Reasons for this particular format:
+
+- Use of a prefix is convenient for implementation (no need to carry the shard ID everywhere
+  we construct a layer file name), and enables efficient listing of index_parts within
+  a particular shard-timeline prefix.
+- Including the shard _count_ as well as shard number means that in future when we implement
+  shard splitting, it will be possible for a parent shard and one of its children to write
+  the same layer file without a name collision. For example, a parent shard 0_1 might split
+  into two (0_2, 1_2), and in the process of splitting shard 0_2 could write a layer or index_part
+  that is distinct from what shard 0_1 would have written at the same place.
+
+In practice, we expect shard counts to be relatively small, so a `u8` will be sufficient,
+and therefore the shard part of the path can be a fixed-length hex string like `{:02X}{:02X}`,
+for example a single-shard tenant's prefix will be `0001`.
+
+For backward compatibility, we may define a special `ShardIdentity` that has shard_count==0,
+and use this as a cue to construct paths with no prefix at all.
+
+#### Storage Format: Indices
+
+In the phase 1 described in this RFC, shards only reference layers they write themselves. However,
+when we implement shard splitting in future, it will be useful to enable shards to reference layers
+written by other shards (specifically the parent shard during a split), so that shards don't
+have to exhaustively copy all data into their own shard-prefixed keys.
+
+To enable this, the `IndexPart` structure will be extended to store the (shard number, shard count)
+tuple on each layer, such that it can construct paths for layers written by other shards. This
+naturally raises the question of who "owns" such layers written by ancestral shards: this problem
+will be addressed in phase 2.
+
+For backward compatibility, any index entry without shard information will be assumed to be
+in the legacy shardidentity.
+
+#### WAL Ingest
+
+In Phase 1, all shards will subscribe to the safekeeper to download WAL content. They will filter
+it down to the pages relevant to their shard:
+
+- For ordinary user data writes, only retain a write if it matches the ShardIdentity
+- For metadata describing relations etc, all shards retain these writes.
+
+The pageservers must somehow give the safekeeper correct feedback on remote_consistent_lsn:
+one solution here is for the 0th shard to periodically peek at the IndexParts for all the other shards,
+and have only the 0th shard populate remote_consistent_lsn. However, this is relatively
+expensive: if the safekeeper can be made shard-aware then it could be taught to use
+the max() of all shards' remote_consistent_lsns to decide when to trim the WAL.
+
+#### Compaction/GC
+
+No changes needed.
+
+The pageserver doesn't have to do anything special during compaction
+or GC. It is implicitly operating on the subset of keys that map to its ShardIdentity.
+This will result in sparse layer files, containing keys only in the stripes that this
+shard owns. Where optimizations currently exist in compaction for spotting "gaps" in
+the key range, these should be updated to ignore gaps that are due to sharding, to
+avoid spuriously splitting up layers ito stripe-sized pieces.
+
+### Compute Endpoints
+
+Compute endpoints will need to:
+
+- Accept a vector of connection strings as part of their configuration from the control plane
+- Route pageserver requests according to mapping the hash of key to the correct
+  entry in the vector of connection strings.
+
+Doing this in compute rather than routing requests via a single pageserver is
+necessary to enable sharding tenants without adding latency from extra hops.
+
+### Control Plane
+
+Tenants, or _Projects_ in the control plane, will each own a set of TenantShards (this will
+be 1 for small tenants). Logic for placement of tenant shards is just the same as the current logic for placing
+tenants.
+
+Tenant lifecycle operations like deletion will require fanning-out to all the shards
+in the tenant. The same goes for timeline creation and deletion: a timeline should
+not be considered created until it has been created in all shards.
+
+#### Selectively enabling sharding for large tenants
+
+Initially, we will explicitly enable sharding for large tenants only.
+
+In future, this hint mechanism will become optional when we implement automatic
+re-sharding of tenants.
+
+## Future Phases
+
+This section exists to indicate what will likely come next after this phase.
+
+Phases 2a and 2b are amenable to execution in parallel.
+
+### Phase 2a: WAL fan-out
+
+**Problem**: when all shards consume the whole WAL, the network bandwidth used
+for transmitting the WAL from safekeeper to pageservers is multiplied by a factor
+of the shard count.
+
+Network bandwidth is not our most pressing bottleneck, but it is likely to become
+a problem if we set a modest shard count (~8) on a significant number of tenants,
+especially as those larger tenants which we shard are also likely to have higher
+write bandwidth than average.
+
+### Phase 2b: Shard Splitting
+
+**Problem**: the number of shards in a tenant is defined at creation time and cannot
+be changed. This causes excessive sharding for most small tenants, and an upper
+bound on scale for very large tenants.
+
+To address this, a _splitting_ feature will later be added. One shard can split its
+data into a number of children by doing a special compaction operation to generate
+image layers broken up child-shard-wise, and then writing out an `index_part.json` for
+each child. This will then require external coordination (by the control plane) to
+safely attach these new child shards and then move them around to distribute work.
+The opposite _merging_ operation can also be imagined, but is unlikely to be implemented:
+once a Tenant has been sharded, the marginal efficiency benefit of merging is unlikely to justify
+the risk/complexity of implementing such a rarely-encountered scenario.
+
+### Phase N (future): distributed historical reads
+
+**Problem**: while sharding based on key is good for handling changes in overall
+database size, it is less suitable for spiky/unpredictable changes in the read
+workload to historical layers. Sudden increases in historical reads could result
+in sudden increases in local disk capacity required for a TenantShard.
+
+Example: the extreme case of this would be to run a tenant for a year, then create branches
+with ancestors at monthly intervals. This could lead to a sudden 12x inflation in
+the on-disk capacity footprint of a TenantShard, since it would be serving reads
+from all those disparate historical layers.
+
+If we can respond fast enough, then key-sharding a tenant more finely can help with
+this, but splitting may be a relatively expensive operation and the increased historical
+read load may be transient.
+
+A separate mechanism for handling heavy historical reads could be something like
+a gossip mechanism for pageservers to communicate
+about their workload, and then a getpageatlsn offload mechanism where one pageserver can
+ask another to go read the necessary layers from remote storage to serve the read. This
+requires relativly little coordination because it is read-only: any node can service any
+read. All reads to a particular shard would still flow through one node, but the
+disk capactity & I/O impact of servicing the read would be distributed.
+
+## FAQ/Alternatives
+
+### Why stripe the data, rather than using contiguous ranges of keyspace for each shard?
+
+When a database is growing under a write workload, writes may predominantly hit the
+end of the keyspace, creating a bandwidth hotspot on that shard. Similarly, if the user
+is intensively re-writing a particular relation, if that relation lived in a particular
+shard then it would not achieve our goal of distributing the write work across shards.
+
+### Why not proxy read requests through one pageserver, so that endpoints don't have to change?
+
+1. This would not achieve scale-out of network bandwidth: a busy tenant with a large
+   database would still cause a load hotspot on the pageserver routing its read requests.
+2. The additional hop through the "proxy" pageserver would add latency and overall
+   resource cost (CPU, network bandwidth)
+
+### Layer File Spreading: use one pageserver as the owner of a tenant, and have it spread out work on a per-layer basis to peers
+
+In this model, there would be no explicit sharding of work, but the pageserver to which
+a tenant is attached would not hold all layers on its disk: instead, it would call out
+to peers to have them store some layers, and call out to those peers to request reads
+in those layers.
+
+This mechanism will work well for distributing work in the LSN dimension, but in the key
+space dimension it has the major limitation of requiring one node to handle all
+incoming writes, and compactions. Even if the write workload for a large database
+fits in one pageserver, it will still be a hotspot and such tenants may still
+de-facto require their own pageserver.

docs/rfcs/032-shard-splitting.md

@@ -0,0 +1,479 @@
+# Shard splitting
+
+## Summary
+
+This RFC describes a new pageserver API for splitting an existing tenant shard into
+multiple shards, and describes how to use this API to safely increase the total
+shard count of a tenant.
+
+## Motivation
+
+In the [sharding RFC](031-sharding-static.md), a mechanism was introduced to scale
+tenants beyond the capacity of a single pageserver by breaking up the key space
+into stripes, and distributing these stripes across many pageservers. However,
+the shard count was defined once at tenant creation time and not varied thereafter.
+
+In practice, the expected size of a database is rarely known at creation time, and
+it is inefficient to enable sharding for very small tenants: we need to be
+able to create a tenant with a small number of shards (such as 1), and later expand
+when it becomes clear that the tenant has grown in size to a point where sharding
+is beneficial.
+
+### Prior art
+
+Many distributed systems have the problem of choosing how many shards to create for
+tenants that do not specify an expected size up-front. There are a couple of general
+approaches:
+
+- Write to a key space in order, and start a new shard when the highest key advances
+  past some point. This doesn't work well for Neon, because we write to our key space
+  in many different contiguous ranges (per relation), rather than in one contiguous
+  range. To adapt to this kind of model, we would need a sharding scheme where each
+  relation had its own range of shards, which would be inefficient for the common
+  case of databases with many small relations.
+- Monitor the system, and automatically re-shard at some size threshold. For
+  example in Ceph, the [pg_autoscaler](https://github.com/ceph/ceph/blob/49c27499af4ee9a90f69fcc6bf3597999d6efc7b/src/pybind/mgr/pg_autoscaler/module.py)
+  component monitors the size of each RADOS Pool, and adjusts the number of Placement
+  Groups (Ceph's shard equivalent).
+
+## Requirements
+
+- A configurable capacity limit per-shard is enforced.
+- Changes in shard count do not interrupt service beyond requiring postgres
+  to reconnect (i.e. milliseconds).
+- Human being does not have to choose shard count
+
+## Non Goals
+
+- Shard splitting is always a tenant-global operation: we will not enable splitting
+  one shard while leaving others intact.
+- The inverse operation (shard merging) is not described in this RFC. This is a lower
+  priority than splitting, because databases grow more often than they shrink, and
+  a database with many shards will still work properly if the stored data shrinks, just
+  with slightly more overhead (e.g. redundant WAL replication)
+- Shard splitting is only initiated based on capacity bounds, not load. Splitting
+  a tenant based on load will make sense for some medium-capacity, high-load workloads,
+  but is more complex to reason about and likely is not desirable until we have
+  shard merging to reduce the shard count again if the database becomes less busy.
+
+## Impacted Components
+
+pageserver, storage controller
+
+(the _storage controller_ is the evolution of what was called `attachment_service` in our test environment)
+
+## Terminology
+
+**Parent** shards are the shards that exist before a split. **Child** shards are
+the new shards created during a split.
+
+**Shard** is synonymous with _tenant shard_.
+
+**Shard Index** is the 2-tuple of shard number and shard count, written in
+paths as {:02x}{:02x}, e.g. `0001`.
+
+## Background
+
+In the implementation section, a couple of existing aspects of sharding are important
+to remember:
+
+- Shard identifiers contain the shard number and count, so that "shard 0 of 1" (`0001`) is
+  a distinct shard from "shard 0 of 2" (`0002`). This is the case in key paths, local
+  storage paths, and remote index metadata.
+- Remote layer file paths contain the shard index of the shard that created them, and
+  remote indices contain the same index to enable building the layer file path. A shard's
+  index may reference layers that were created by another shard.
+- Local tenant shard directories include the shard index. All layers downloaded by
+  a tenant shard are stored in this shard-prefixed path, even if those layers were
+  initially created by another shard: tenant shards do not read and write one anothers'
+  paths.
+- The `Tenant` pageserver type represents one tenant _shard_, not the whole tenant.
+  This is for historical reasons and will be cleaned up in future, but the existing
+  name is used here to help comprehension when reading code.
+
+## Implementation
+
+Note: this section focuses on the correctness of the core split process. This will
+be fairly inefficient in a naive implementation, and several important optimizations
+are described in a later section.
+
+There are broadly two parts to the implementation:
+
+1. The pageserver split API, which splits one shard on one pageserver
+2. The overall tenant split proccess which is coordinated by the storage controller,
+   and calls into the pageserver split API as needed.
+
+### Pageserver Split API
+
+The pageserver will expose a new API endpoint at `/v1/tenant/:tenant_shard_id/shard_split`
+that takes the new total shard count in the body.
+
+The pageserver split API operates on one tenant shard, on one pageserver. External
+coordination is required to use it safely, this is described in the later
+'Split procedure' section.
+
+#### Preparation
+
+First identify the shard indices for the new child shards. These are deterministic,
+calculated from the parent shard's index, and the number of children being created (this
+is an input to the API, and validated to be a power of two). In a trivial example, splitting
+0001 in two always results in 0002 and 0102.
+
+Child shard indices are chosen such that the childrens' parts of the keyspace will
+be subsets of the parent's parts of the keyspace.
+
+#### Step 1: write new remote indices
+
+In remote storage, splitting is very simple: we may just write new index_part.json
+objects for each child shard, containing exactly the same layers as the parent shard.
+
+The children will have more data than they need, but this avoids any exhausive
+re-writing or copying of layer files.
+
+The index key path includes a generation number: the parent shard's current
+attached generation number will also be used for the child shards' indices. This
+makes the operation safely retryable: if everything crashes and restarts, we may
+call the split API again on the parent shard, and the result will be some new remote
+indices for the child shards, under a higher generation number.
+
+#### Step 2: start new `Tenant` objects
+
+A new `Tenant` object may be instantiated for each child shard, while the parent
+shard still exists. When calling the tenant_spawn function for this object,
+the remote index from step 1 will be read, and the child shard will start
+to ingest WAL to catch up from whatever was in the remote storage at step 1.
+
+We now wait for child shards' WAL ingestion to catch up with the parent shard,
+so that we can safely tear down the parent shard without risking an availability
+gap to clients reading recent LSNs.
+
+#### Step 3: tear down parent `Tenant` object
+
+Once child shards are running and have caught up with WAL ingest, we no longer
+need the parent shard. Note that clients may still be using it -- when we
+shut it down, any page_service handlers will also shut down, causing clients
+to disconnect. When the client reconnects, it will re-lookup the tenant,
+and hit the child shard instead of the parent (shard lookup from page_service
+should bias toward higher ShardCount shards).
+
+Note that at this stage the page service client has not yet been notified of
+any split. In the trivial single split example:
+
+- Shard 0001 is gone: Tenant object torn down
+- Shards 0002 and 0102 are running on the same pageserver where Shard 0001 used to live.
+- Clients will continue to connect to that server thinking that shard 0001 is there,
+  and all requests will work, because any key that was in shard 0001 is definitely
+  available in either shard 0002 or shard 0102.
+- Eventually, the storage controller (not the pageserver) will decide to migrate
+  some child shards away: at that point it will do a live migration, ensuring
+  that the client has an updated configuration before it detaches anything
+  from the original server.
+
+#### Complete
+
+When we send a 200 response to the split request, we are promising the caller:
+
+- That the child shards are persistent in remote storage
+- That the parent shard has been shut down
+
+This enables the caller to proceed with the overall shard split operation, which
+may involve other shards on other pageservers.
+
+### Storage Controller Split procedure
+
+Splitting a tenant requires calling the pageserver split API, and tracking
+enough state to ensure recovery + completion in the event of any component (pageserver
+or storage controller) crashing (or request timing out) during the split.
+
+1. call the split API on all existing shards. Ensure that the resulting
+   child shards are pinned to their pageservers until _all_ the split calls are done.
+   This pinning may be implemented as a "split bit" on the tenant shards, that
+   blocks any migrations, and also acts as a sign that if we restart, we must go
+   through some recovery steps to resume the split.
+2. Once all the split calls are done, we may unpin the child shards (clear
+   the split bit). The split is now complete: subsequent steps are just migrations,
+   not strictly part of the split.
+3. Try to schedule new pageserver locations for the child shards, using
+   a soft anti-affinity constraint to place shards from the same tenant onto different
+   pageservers.
+
+Updating computes about the new shard count is not necessary until we migrate
+any of the child shards away from the parent's location.
+
+### Recovering from failures
+
+#### Rolling back an incomplete split
+
+An incomplete shard split may be rolled back quite simply, by attaching the parent shards to pageservers,
+and detaching child shards. This will lose any WAL ingested into the children after the parents
+were detached earlier, but the parents will catch up.
+
+No special pageserver API is needed for this. From the storage controllers point of view, the
+procedure is:
+
+1. For all parent shards in the tenant, ensure they are attached
+2. For all child shards, ensure they are not attached
+3. Drop child shards from the storage controller's database, and clear the split bit on the parent shards.
+
+Any remote storage content for child shards is left behind. This is similar to other cases where
+we may leave garbage objects in S3 (e.g. when we upload a layer but crash before uploading an
+index that references it). Future online scrub/cleanup functionality can remove these objects, or
+they will be removed when the tenant is deleted, as tenant deletion lists all objects in the prefix,
+which would include any child shards that were rolled back.
+
+If any timelines had been created on child shards, they will be lost when rolling back. To mitigate
+this, we will **block timeline creation during splitting**, so that we can safely roll back until
+the split is complete, without risking losing timelines.
+
+Rolling back an incomplete split will happen automatically if a split fails due to some fatal
+reason, and will not be accessible via an API:
+
+- A pageserver fails to complete its split API request after too many retries
+- A pageserver returns a fatal unexpected error such as 400 or 500
+- The storage controller database returns a non-retryable error
+- Some internal invariant is violated in the storage controller split code
+
+#### Rolling back a complete split
+
+A complete shard split may be rolled back similarly to an incomplete split, with the following
+modifications:
+
+- The parent shards will no longer exist in the storage controller database, so these must
+  be re-synthesized somehow: the hard part of this is figuring the parent shards' generations. This
+  may be accomplished either by probing in S3, or by retaining some tombstone state for deleted
+  shards in the storage controller database.
+- Any timelines that were created after the split complete will disappear when rolling back
+  to the tenant shards. For this reason, rolling back after a complete split should only
+  be done due to serious issues where loss of recently created timelines is acceptable, or
+  in cases where we have confirmed that no timelines were created in the intervening period.
+- Parent shards' layers must not have been deleted: this property will come "for free" when
+  we first roll out sharding, by simply not implementing deletion of parent layers after
+  a split. When we do implement such deletion (see "Cleaning up parent-shard layers" in the
+  Optimizations section), it should apply a TTL to layers such that we have a
+  defined walltime window in which rollback will be possible.
+
+The storage controller will expose an API for rolling back a complete split, for use
+in the field if we encounter some critical bug with a post-split tenant.
+
+#### Retrying API calls during Pageserver Restart
+
+When a pageserver restarts during a split API call, it may witness on-disk content for both parent and
+child shards from an ongoing split. This does not intrinsically break anything, and the
+pageserver may include all these shards in its `/re-attach` request to the storage controller.
+
+In order to support such restarts, it is important that the storage controller stores
+persistent records of each child shard before it calls into a pageserver, as these child shards
+may require generation increments via a `/re-attach` request.
+
+The pageserver restart will also result in a failed API call from the storage controller's point
+of view. Recall that if _any_ pageserver fails to split, the overall split operation may not
+complete, and all shards must remain pinned to their current pageserver locations until the
+split is done.
+
+The pageserver API calls during splitting will retry on transient errors, so that
+short availability gaps do not result in a failure of the overall operation. The
+split in progress will be automatically rolled back if the threshold for API
+retries is reached (e.g. if a pageserver stays offline for longer than a typical
+restart).
+
+#### Rollback on Storage Controller Restart
+
+On startup, the storage controller will inspect the split bit for tenant shards that
+it loads from the database. If any splits are in progress:
+
+- Database content will be reverted to the parent shards
+- Child shards will be dropped from memory
+- The parent and child shards will be included in the general startup reconciliation that
+  the storage controller does: any child shards will be detached from pageservers because
+  they don't exist in the storage controller's expected set of shards, and parent shards
+  will be attached if they aren't already.
+
+#### Storage controller API request failures/retries
+
+The split request handler will implement idempotency: if the [`Tenant`] requested to split
+doesn't exist, we will check for the would-be child shards, and if they already exist,
+we consider the request complete.
+
+If a request is retried while the original request is still underway, then the split
+request handler will notice an InProgress marker in TenantManager, and return 503
+to encourage the client to backoff/retry. This is the same as the general pageserver
+API handling for calls that try to act on an InProgress shard.
+
+#### Compute start/restart during a split
+
+If a compute starts up during split, it will be configured with the old sharding
+configuration. This will work for reads irrespective of the progress of the split
+as long as no child hards have been migrated away from their original location, and
+this is guaranteed in the split procedure (see earlier section).
+
+#### Pageserver fails permanently during a split
+
+If a pageserver permanently fails (i.e. the storage controller availability state for it
+goes to Offline) while a split is in progress, the splitting operation will roll back, and
+during the roll back it will skip any API calls to the offline pageserver. If the offline
+pageserver becomes available again, any stale locations will be cleaned up via the normal reconciliation process (the `/re-attach` API).
+
+### Handling secondary locations
+
+For correctness, it is not necessary to split secondary locations. We can simply detach
+the secondary locations for parent shards, and then attach new secondary locations
+for child shards.
+
+Clearly this is not optimal, as it will result in re-downloads of layer files that
+were already present on disk. See "Splitting secondary locations"
+
+### Conditions to trigger a split
+
+The pageserver will expose a new API for reporting on shards that are candidates
+for split: this will return a top-N report of the largest tenant shards by
+physical size (remote size). This should exclude any tenants that are already
+at the maximum configured shard count.
+
+The API would look something like:
+`/v1/top_n_tenant?shard_count_lt=8&sort_by=resident_size`
+
+The storage controller will poll that API across all pageservers it manages at some appropriate interval (e.g. 60 seconds).
+
+A split operation will be started when the tenant exceeds some threshold. This threshold
+should be _less than_ how large we actually want shards to be, perhaps much less. That's to
+minimize the amount of work involved in splitting -- if we want 100GiB shards, we shouldn't
+wait for a tenant to exceed 100GiB before we split anything. Some data analysis of existing
+tenant size distribution may be useful here: if we can make a statement like "usually, if
+a tenant has exceeded 20GiB they're probably going to exceed 100GiB later", then we might
+make our policy to split a tenant at 20GiB.
+
+The finest split we can do is by factors of two, but we can do higher-cardinality splits
+too, and this will help to reduce the overhead of repeatedly re-splitting a tenant
+as it grows. An example of a very simple heuristic for early deployment of the splitting
+feature would be: "Split tenants into 8 shards when their physical size exceeds 64GiB": that
+would give us two kinds of tenant (1 shard and 8 shards), and the confidence that once we had
+split a tenant, it will not need re-splitting soon after.
+
+## Optimizations
+
+### Flush parent shard to remote storage during split
+
+Any data that is in WAL but not remote storage at time of split will need
+to be replayed by child shards when they start for the first time. To minimize
+this work, we may flush the parent shard to remote storage before writing the
+remote indices for child shards.
+
+It is important that this flush is subject to some time bounds: we may be splitting
+in response to a surge of write ingest, so it may be time-critical to split. A
+few seconds to flush latest data should be sufficient to optimize common cases without
+running the risk of holding up a split for a harmful length of time when a parent
+shard is being written heavily. If the flush doesn't complete in time, we may proceed
+to shut down the parent shard and carry on with the split.
+
+### Hard linking parent layers into child shard directories
+
+Before we start the Tenant objects for child shards, we may pre-populate their
+local storage directories with hard links to the layer files already present
+in the parent shard's local directory. When the child shard starts and downloads
+its remote index, it will find all those layer files already present on local disk.
+
+This avoids wasting download capacity and makes splitting faster, but more importantly
+it avoids taking up a factor of N more disk space when splitting 1 shard into N.
+
+This mechanism will work well in typical flows where shards are migrated away
+promptly after a split, but for the general case including what happens when
+layers are evicted and re-downloaded after a split, see the 'Proactive compaction'
+section below.
+
+### Filtering during compaction
+
+Compaction, especially image layer generation, should skip any keys that are
+present in a shard's layer files, but do not match the shard's ShardIdentity's
+is_key_local() check. This avoids carrying around data for longer than necessary
+in post-split compactions.
+
+This was already implemented in https://github.com/neondatabase/neon/pull/6246
+
+### Proactive compaction
+
+In remote storage, there is little reason to rewrite any data on a shard split:
+all the children can reference parent layers via the very cheap write of the child
+index_part.json.
+
+In local storage, things are more nuanced. During the initial split there is no
+capacity cost to duplicating parent layers, if we implement the hard linking
+optimization described above. However, as soon as any layers are evicted from
+local disk and re-downloaded, the downloaded layers will not be hard-links any more:
+they'll have real capacity footprint. That isn't a problem if we migrate child shards
+away from the parent node swiftly, but it risks a significant over-use of local disk
+space if we do not.
+
+For example, if we did an 8-way split of a shard, and then _didn't_ migrate 7 of
+the shards elsewhere, then churned all the layers in all the shards via eviction,
+then we would blow up the storage capacity used on the node by 8x. If we're splitting
+a 100GB shard, that could take the pageserver to the point of exhausting disk space.
+
+To avoid this scenario, we could implement a special compaction mode where we just
+read historic layers, drop unwanted keys, and write back the layer file. This
+is pretty expensive, but useful if we have split a large shard and are not going to
+migrate the child shards away.
+
+The heuristic conditions for triggering such a compaction are:
+
+- A) eviction plus time: if a child shard
+  has existed for more than a time threshold, and has been requested to perform at least one eviction, then it becomes urgent for this child shard to execute a proactive compaction to reduce its storage footprint, at the cost of I/O load.
+- B) resident size plus time: we may inspect the resident layers and calculate how
+  many of them include the overhead of storing pre-split keys. After some time
+  threshold (different to the one in case A) we still have such layers occupying
+  local disk space, then we should proactively compact them.
+
+### Cleaning up parent-shard layers
+
+It is functionally harmless to leave parent shard layers in remote storage indefinitely.
+They would be cleaned up in the event of the tenant's deletion.
+
+As an optimization to avoid leaking remote storage capacity (which costs money), we may
+lazily clean up parent shard layers once no child shards reference them.
+
+This may be done _very_ lazily: e.g. check every PITR interval. The cleanup procedure is:
+
+- list all the key prefixes beginning with the tenant ID, and select those shard prefixes
+  which do not belong to the most-recently-split set of shards (_ancestral shards_, i.e. `shard*count < max(shard_count) over all shards)`, and those shard prefixes which do have the latest shard count (_current shards_)
+- If there are no _ancestral shard_ prefixes found, we have nothing to clean up and
+  may drop out now.
+- find the latest-generation index for each _current shard_, read all and accumulate the set of layers belonging to ancestral shards referenced by these indices.
+- for all ancestral shards, list objects in the prefix and delete any layer which was not
+  referenced by a current shard.
+
+If this cleanup is scheduled for 1-2 PITR periods after the split, there is a good chance that child shards will have written their own image layers covering the whole keyspace, such that all parent shard layers will be deletable.
+
+The cleanup may be done by the scrubber (external process), or we may choose to have
+the zeroth shard in the latest generation do the work -- there is no obstacle to one shard
+reading the other shard's indices at runtime, and we do not require visibility of the
+latest index writes.
+
+Cleanup should be artificially delayed by some period (for example 24 hours) to ensure
+that we retain the option to roll back a split in case of bugs.
+
+### Splitting secondary locations
+
+We may implement a pageserver API similar to the main splitting API, which does a simpler
+operation for secondary locations: it would not write anything to S3, instead it would simply
+create the child shard directory on local disk, hard link in directories from the parent,
+and set up the in memory (TenantSlot) state for the children.
+
+Similar to attached locations, a subset of secondary locations will probably need re-locating
+after the split is complete, to avoid leaving multiple child shards on the same pageservers,
+where they may use excessive space for the tenant.
+
+## FAQ/Alternatives
+
+### What should the thresholds be set to?
+
+Shard size limit: the pre-sharding default capacity quota for databases was 200GiB, so this could be a starting point for the per-shard size limit.
+
+Max shard count:
+
+- The safekeeper overhead to sharding is currently O(N) network bandwidth because
+  the un-filtered WAL is sent to all shards. To avoid this growing out of control,
+  a limit of 8 shards should be temporarily imposed until WAL filtering is implemented
+  on the safekeeper.
+- there is also little benefit to increasing the shard count beyond the number
+  of pageservers in a region.
+
+### Is it worth just rewriting all the data during a split to simplify reasoning about space?

docs/sourcetree.md

@@ -7,6 +7,11 @@ Below you will find a brief overview of each subdir in the source tree in alphab
 Neon storage broker, providing messaging between safekeepers and pageservers.
 [storage_broker.md](./storage_broker.md)
 
+`storage_controller`:
+
+Neon storage controller, manages a cluster of pageservers and exposes an API that enables
+managing a many-sharded tenant as a single entity.
+
 `/control_plane`:
 
 Local control plane.

docs/storage_controller.md

@@ -0,0 +1,150 @@
+# Storage Controller
+
+## Concepts
+
+The storage controller sits between administrative API clients and pageservers, and handles the details of mapping tenants to pageserver tenant shards. For example, creating a tenant is one API call to the storage controller,
+which is mapped into many API calls to many pageservers (for multiple shards, and for secondary locations).
+
+It implements a pageserver-compatible API that may be used for CRUD operations on tenants and timelines, translating these requests into appropriate operations on the shards within a tenant, which may be on many different pageservers. Using this API, the storage controller may be used in the same way as the pageserver's administrative HTTP API, hiding
+the underlying details of how data is spread across multiple nodes.
+
+The storage controller also manages generations, high availability (via secondary locations) and live migrations for tenants under its management. This is done with a reconciliation loop pattern, where tenants have an “intent” state and a “reconcile” task that tries to make the outside world match the intent.
+
+## APIs
+
+The storage controller’s HTTP server implements four logically separate APIs:
+
+- `/v1/...` path is the pageserver-compatible API. This has to be at the path root because that’s where clients expect to find it on a pageserver.
+- `/control/v1/...` path is the storage controller’s API, which enables operations such as registering and management pageservers, or executing shard splits.
+- `/debug/v1/...` path contains endpoints which are either exclusively used in tests, or are for use by engineers when supporting a deployed system.
+- `/upcall/v1/...` path contains endpoints that are called by pageservers. This includes the `/re-attach` and `/validate` APIs used by pageservers
+  to ensure data safety with generation numbers.
+
+The API is authenticated with a JWT token, and tokens must have scope `pageserverapi` (i.e. the same scope as pageservers’ APIs).
+
+See the `http.rs` file in the source for where the HTTP APIs are implemented.
+
+## Database
+
+The storage controller uses a postgres database to persist a subset of its state. Note that the storage controller does _not_ keep all its state in the database: this is a design choice to enable most operations to be done efficiently in memory, rather than having to read from the database. See `persistence.rs` for a more comprehensive comment explaining what we do and do not persist: a useful metaphor is that we persist objects like tenants and nodes, but we do not
+persist the _relationships_ between them: the attachment state of a tenant's shards to nodes is kept in memory and
+rebuilt on startup.
+
+The file `persistence.rs` contains all the code for accessing the database, and has a large doc comment that goes into more detail about exactly what we persist and why.
+
+The `diesel` crate is used for defining models & migrations.
+
+Running a local cluster with `cargo neon` automatically starts a vanilla postgress process to host the storage controller’s database.
+
+### Diesel tip: migrations
+
+If you need to modify the database schema, here’s how to create a migration:
+
+- Install the diesel CLI with `cargo install diesel_cli`
+- Use `diesel migration generate <name>` to create a new migration
+- Populate the SQL files in the `migrations/` subdirectory
+- Use `DATABASE_URL=... diesel migration run` to apply the migration you just wrote: this will update the `[schema.rs](http://schema.rs)` file automatically.
+  - This requires a running database: the easiest way to do that is to just run `cargo neon init ; cargo neon start`, which will leave a database available at `postgresql://localhost:1235/attachment_service`
+- Commit the migration files and the changes to schema.rs
+- If you need to iterate, you can rewind migrations with `diesel migration revert -a` and then `diesel migration run` again.
+- The migrations are build into the storage controller binary, and automatically run at startup after it is deployed, so once you’ve committed a migration no further steps are needed.
+
+## storcon_cli
+
+The `storcon_cli` tool enables interactive management of the storage controller. This is usually
+only necessary for debug, but may also be used to manage nodes (e.g. marking a node as offline).
+
+`storcon_cli --help` includes details on commands.
+
+# Deploying
+
+This section is aimed at engineers deploying the storage controller outside of Neon's cloud platform, as
+part of a self-hosted system.
+
+_General note: since the default `neon_local` environment includes a storage controller, this is a useful
+reference when figuring out deployment._
+
+## Database
+
+It is **essential** that the database used by the storage controller is durable (**do not store it on ephemeral
+local disk**). This database contains pageserver generation numbers, which are essential to data safety on the pageserver.
+
+The resource requirements for the database are very low: a single CPU core and 1GiB of memory should work well for most deployments. The physical size of the database is typically under a gigabyte.
+
+Set the URL to the database using the `--database-url` CLI option.
+
+There is no need to run migrations manually: the storage controller automatically applies migrations
+when it starts up.
+
+## Configure pageservers to use the storage controller
+
+1. The pageserver `control_plane_api` and `control_plane_api_token` should be set in the `pageserver.toml` file. The API setting should
+   point to the "upcall" prefix, for example `http://127.0.0.1:1234/upcall/v1/` is used in neon_local clusters.
+2. Create a `metadata.json` file in the same directory as `pageserver.toml`: this enables the pageserver to automatically register itself
+   with the storage controller when it starts up. See the example below for the format of this file.
+
+### Example `metadata.json`
+
+```
+{"host":"acmehost.localdomain","http_host":"acmehost.localdomain","http_port":9898,"port":64000}
+```
+
+- `port` and `host` refer to the _postgres_ port and host, and these must be accessible from wherever
+  postgres runs.
+- `http_port` and `http_host` refer to the pageserver's HTTP api, this must be accessible from where
+  the storage controller runs.
+
+## Handle compute notifications.
+
+The storage controller independently moves tenant attachments between pageservers in response to
+changes such as a pageserver node becoming unavailable, or the tenant's shard count changing. To enable
+postgres clients to handle such changes, the storage controller calls an API hook when a tenant's pageserver
+location changes.
+
+The hook is configured using the storage controller's `--compute-hook-url` CLI option. If the hook requires
+JWT auth, the token may be provided with `--control-plane-jwt-token`. The hook will be invoked with a `PUT` request.
+
+In the Neon cloud service, this hook is implemented by Neon's internal cloud control plane. In `neon_local` systems
+the storage controller integrates directly with neon_local to reconfigure local postgres processes instead of calling
+the compute hook.
+
+When implementing an on-premise Neon deployment, you must implement a service that handles the compute hook. This is not complicated:
+the request body has format of the `ComputeHookNotifyRequest` structure, provided below for convenience.
+
+```
+struct ComputeHookNotifyRequestShard {
+    node_id: NodeId,
+    shard_number: ShardNumber,
+}
+
+struct ComputeHookNotifyRequest {
+    tenant_id: TenantId,
+    stripe_size: Option<ShardStripeSize>,
+    shards: Vec<ComputeHookNotifyRequestShard>,
+}
+```
+
+When a notification is received:
+
+1. Modify postgres configuration for this tenant:
+
+   - set `neon.pageserver_connstr` to a comma-separated list of postgres connection strings to pageservers according to the `shards` list. The
+     shards identified by `NodeId` must be converted to the address+port of the node.
+   - if stripe_size is not None, set `neon.stripe_size` to this value
+
+2. Send SIGHUP to postgres to reload configuration
+3. Respond with 200 to the notification request. Do not return success if postgres was not updated: if an error is returned, the controller
+   will retry the notification until it succeeds..
+
+### Example notification body
+
+```
+{
+  "tenant_id": "1f359dd625e519a1a4e8d7509690f6fc",
+  "stripe_size": 32768,
+  "shards": [
+      {"node_id": 344, "shard_number": 0},
+      {"node_id": 722, "shard_number": 1},
+  ],
+}
+```

libs/compute_api/src/spec.rs

@@ -33,6 +33,23 @@ pub struct ComputeSpec {
     #[serde(default)]
     pub features: Vec<ComputeFeature>,
 
+    /// If compute_ctl was passed `--resize-swap-on-bind`, a value of `Some(_)` instructs
+    /// compute_ctl to `/neonvm/bin/resize-swap` with the given size, when the spec is first
+    /// received.
+    ///
+    /// Both this field and `--resize-swap-on-bind` are required, so that the control plane's
+    /// spec generation doesn't need to be aware of the actual compute it's running on, while
+    /// guaranteeing gradual rollout of swap. Otherwise, without `--resize-swap-on-bind`, we could
+    /// end up trying to resize swap in VMs without it -- or end up *not* resizing swap, thus
+    /// giving every VM much more swap than it should have (32GiB).
+    ///
+    /// Eventually we may remove `--resize-swap-on-bind` and exclusively use `swap_size_bytes` for
+    /// enabling the swap resizing behavior once rollout is complete.
+    ///
+    /// See neondatabase/cloud#12047 for more.
+    #[serde(default)]
+    pub swap_size_bytes: Option<u64>,
+
     /// Expected cluster state at the end of transition process.
     pub cluster: Cluster,
     pub delta_operations: Option<Vec<DeltaOp>>,

libs/metrics/Cargo.toml

@@ -10,11 +10,13 @@ libc.workspace = true
 once_cell.workspace = true
 chrono.workspace = true
 twox-hash.workspace = true
+measured.workspace = true
 
 workspace_hack.workspace = true
 
 [target.'cfg(target_os = "linux")'.dependencies]
 procfs.workspace = true
+measured-process.workspace = true
 
 [dev-dependencies]
 rand = "0.8"

libs/metrics/src/hll.rs

@@ -7,14 +7,19 @@
 //! use significantly less memory than this, but can only approximate the cardinality.
 
 use std::{
-    collections::HashMap,
-    hash::{BuildHasher, BuildHasherDefault, Hash, Hasher},
-    sync::{atomic::AtomicU8, Arc, RwLock},
+    hash::{BuildHasher, BuildHasherDefault, Hash},
+    sync::atomic::AtomicU8,
 };
 
-use prometheus::{
-    core::{self, Describer},
-    proto, Opts,
+use measured::{
+    label::{LabelGroupVisitor, LabelName, LabelValue, LabelVisitor},
+    metric::{
+        group::{Encoding, MetricValue},
+        name::MetricNameEncoder,
+        Metric, MetricType, MetricVec,
+    },
+    text::TextEncoder,
+    LabelGroup,
 };
 use twox_hash::xxh3;
 
@@ -40,7 +45,7 @@ macro_rules! register_hll {
     }};
 
     ($N:literal, $NAME:expr, $HELP:expr $(,)?) => {{
-        $crate::register_hll!($N, $crate::opts!($NAME, $HELP), $LABELS_NAMES)
+        $crate::register_hll!($N, $crate::opts!($NAME, $HELP))
     }};
 }
 
@@ -93,203 +98,25 @@ macro_rules! register_hll {
 /// ```
 ///
 /// See <https://en.wikipedia.org/wiki/HyperLogLog#Practical_considerations> for estimates on alpha
-#[derive(Clone)]
-pub struct HyperLogLogVec<const N: usize> {
-    core: Arc<HyperLogLogVecCore<N>>,
+pub type HyperLogLogVec<L, const N: usize> = MetricVec<HyperLogLogState<N>, L>;
+pub type HyperLogLog<const N: usize> = Metric<HyperLogLogState<N>>;
+
+pub struct HyperLogLogState<const N: usize> {
+    shards: [AtomicU8; N],
 }
-
-struct HyperLogLogVecCore<const N: usize> {
-    pub children: RwLock<HashMap<u64, HyperLogLog<N>, BuildHasherDefault<xxh3::Hash64>>>,
-    pub desc: core::Desc,
-    pub opts: Opts,
-}
-
-impl<const N: usize> core::Collector for HyperLogLogVec<N> {
-    fn desc(&self) -> Vec<&core::Desc> {
-        vec![&self.core.desc]
-    }
-
-    fn collect(&self) -> Vec<proto::MetricFamily> {
-        let mut m = proto::MetricFamily::default();
-        m.set_name(self.core.desc.fq_name.clone());
-        m.set_help(self.core.desc.help.clone());
-        m.set_field_type(proto::MetricType::GAUGE);
-
-        let mut metrics = Vec::new();
-        for child in self.core.children.read().unwrap().values() {
-            child.core.collect_into(&mut metrics);
-        }
-        m.set_metric(metrics);
-
-        vec![m]
+impl<const N: usize> Default for HyperLogLogState<N> {
+    fn default() -> Self {
+        #[allow(clippy::declare_interior_mutable_const)]
+        const ZERO: AtomicU8 = AtomicU8::new(0);
+        Self { shards: [ZERO; N] }
     }
 }
 
-impl<const N: usize> HyperLogLogVec<N> {
-    /// Create a new [`HyperLogLogVec`] based on the provided
-    /// [`Opts`] and partitioned by the given label names. At least one label name must be
-    /// provided.
-    pub fn new(opts: Opts, label_names: &[&str]) -> prometheus::Result<Self> {
-        assert!(N.is_power_of_two());
-        let variable_names = label_names.iter().map(|s| (*s).to_owned()).collect();
-        let opts = opts.variable_labels(variable_names);
-
-        let desc = opts.describe()?;
-        let v = HyperLogLogVecCore {
-            children: RwLock::new(HashMap::default()),
-            desc,
-            opts,
-        };
-
-        Ok(Self { core: Arc::new(v) })
-    }
-
-    /// `get_metric_with_label_values` returns the [`HyperLogLog<P>`] for the given slice
-    /// of label values (same order as the VariableLabels in Desc). If that combination of
-    /// label values is accessed for the first time, a new [`HyperLogLog<P>`] is created.
-    ///
-    /// An error is returned if the number of label values is not the same as the
-    /// number of VariableLabels in Desc.
-    pub fn get_metric_with_label_values(
-        &self,
-        vals: &[&str],
-    ) -> prometheus::Result<HyperLogLog<N>> {
-        self.core.get_metric_with_label_values(vals)
-    }
-
-    /// `with_label_values` works as `get_metric_with_label_values`, but panics if an error
-    /// occurs.
-    pub fn with_label_values(&self, vals: &[&str]) -> HyperLogLog<N> {
-        self.get_metric_with_label_values(vals).unwrap()
-    }
+impl<const N: usize> MetricType for HyperLogLogState<N> {
+    type Metadata = ();
 }
 
-impl<const N: usize> HyperLogLogVecCore<N> {
-    pub fn get_metric_with_label_values(
-        &self,
-        vals: &[&str],
-    ) -> prometheus::Result<HyperLogLog<N>> {
-        let h = self.hash_label_values(vals)?;
-
-        if let Some(metric) = self.children.read().unwrap().get(&h).cloned() {
-            return Ok(metric);
-        }
-
-        self.get_or_create_metric(h, vals)
-    }
-
-    pub(crate) fn hash_label_values(&self, vals: &[&str]) -> prometheus::Result<u64> {
-        if vals.len() != self.desc.variable_labels.len() {
-            return Err(prometheus::Error::InconsistentCardinality {
-                expect: self.desc.variable_labels.len(),
-                got: vals.len(),
-            });
-        }
-
-        let mut h = xxh3::Hash64::default();
-        for val in vals {
-            h.write(val.as_bytes());
-        }
-
-        Ok(h.finish())
-    }
-
-    fn get_or_create_metric(
-        &self,
-        hash: u64,
-        label_values: &[&str],
-    ) -> prometheus::Result<HyperLogLog<N>> {
-        let mut children = self.children.write().unwrap();
-        // Check exist first.
-        if let Some(metric) = children.get(&hash).cloned() {
-            return Ok(metric);
-        }
-
-        let metric = HyperLogLog::with_opts_and_label_values(&self.opts, label_values)?;
-        children.insert(hash, metric.clone());
-        Ok(metric)
-    }
-}
-
-/// HLL is a probabilistic cardinality measure.
-///
-/// How to use this time-series for a metric name `my_metrics_total_hll`:
-///
-/// ```promql
-/// # harmonic mean
-/// 1 / (
-///     sum (
-///         2 ^ -(
-///             # HLL merge operation
-///             max (my_metrics_total_hll{}) by (hll_shard, other_labels...)
-///         )
-///     ) without (hll_shard)
-/// )
-/// * alpha
-/// * shards_count
-/// * shards_count
-/// ```
-///
-/// If you want an estimate over time, you can use the following query:
-///
-/// ```promql
-/// # harmonic mean
-/// 1 / (
-///     sum (
-///         2 ^ -(
-///             # HLL merge operation
-///             max (
-///                 max_over_time(my_metrics_total_hll{}[$__rate_interval])
-///             ) by (hll_shard, other_labels...)
-///         )
-///     ) without (hll_shard)
-/// )
-/// * alpha
-/// * shards_count
-/// * shards_count
-/// ```
-///
-/// In the case of low cardinality, you might want to use the linear counting approximation:
-///
-/// ```promql
-/// # LinearCounting(m, V) = m log (m / V)
-/// shards_count * ln(shards_count /
-///     # calculate V = how many shards contain a 0
-///     count(max (proxy_connecting_endpoints{}) by (hll_shard, protocol) == 0) without (hll_shard)
-/// )
-/// ```
-///
-/// See <https://en.wikipedia.org/wiki/HyperLogLog#Practical_considerations> for estimates on alpha
-#[derive(Clone)]
-pub struct HyperLogLog<const N: usize> {
-    core: Arc<HyperLogLogCore<N>>,
-}
-
-impl<const N: usize> HyperLogLog<N> {
-    /// Create a [`HyperLogLog`] with the `name` and `help` arguments.
-    pub fn new<S1: Into<String>, S2: Into<String>>(name: S1, help: S2) -> prometheus::Result<Self> {
-        assert!(N.is_power_of_two());
-        let opts = Opts::new(name, help);
-        Self::with_opts(opts)
-    }
-
-    /// Create a [`HyperLogLog`] with the `opts` options.
-    pub fn with_opts(opts: Opts) -> prometheus::Result<Self> {
-        Self::with_opts_and_label_values(&opts, &[])
-    }
-
-    fn with_opts_and_label_values(opts: &Opts, label_values: &[&str]) -> prometheus::Result<Self> {
-        let desc = opts.describe()?;
-        let labels = make_label_pairs(&desc, label_values)?;
-
-        let v = HyperLogLogCore {
-            shards: [0; N].map(AtomicU8::new),
-            desc,
-            labels,
-        };
-        Ok(Self { core: Arc::new(v) })
-    }
-
+impl<const N: usize> HyperLogLogState<N> {
     pub fn measure(&self, item: &impl Hash) {
         // changing the hasher will break compatibility with previous measurements.
         self.record(BuildHasherDefault::<xxh3::Hash64>::default().hash_one(item));
@@ -299,42 +126,11 @@ impl<const N: usize> HyperLogLog<N> {
         let p = N.ilog2() as u8;
         let j = hash & (N as u64 - 1);
         let rho = (hash >> p).leading_zeros() as u8 + 1 - p;
-        self.core.shards[j as usize].fetch_max(rho, std::sync::atomic::Ordering::Relaxed);
-    }
-}
-
-struct HyperLogLogCore<const N: usize> {
-    shards: [AtomicU8; N],
-    desc: core::Desc,
-    labels: Vec<proto::LabelPair>,
-}
-
-impl<const N: usize> core::Collector for HyperLogLog<N> {
-    fn desc(&self) -> Vec<&core::Desc> {
-        vec![&self.core.desc]
+        self.shards[j as usize].fetch_max(rho, std::sync::atomic::Ordering::Relaxed);
     }
 
-    fn collect(&self) -> Vec<proto::MetricFamily> {
-        let mut m = proto::MetricFamily::default();
-        m.set_name(self.core.desc.fq_name.clone());
-        m.set_help(self.core.desc.help.clone());
-        m.set_field_type(proto::MetricType::GAUGE);
-
-        let mut metrics = Vec::new();
-        self.core.collect_into(&mut metrics);
-        m.set_metric(metrics);
-
-        vec![m]
-    }
-}
-
-impl<const N: usize> HyperLogLogCore<N> {
-    fn collect_into(&self, metrics: &mut Vec<proto::Metric>) {
-        self.shards.iter().enumerate().for_each(|(i, x)| {
-            let mut shard_label = proto::LabelPair::default();
-            shard_label.set_name("hll_shard".to_owned());
-            shard_label.set_value(format!("{i}"));
-
+    fn take_sample(&self) -> [u8; N] {
+        self.shards.each_ref().map(|x| {
             // We reset the counter to 0 so we can perform a cardinality measure over any time slice in prometheus.
 
             // This seems like it would be a race condition,
@@ -344,85 +140,90 @@ impl<const N: usize> HyperLogLogCore<N> {
 
             // TODO: maybe we shouldn't reset this on every collect, instead, only after a time window.
             // this would mean that a dev port-forwarding the metrics url won't break the sampling.
-            let v = x.swap(0, std::sync::atomic::Ordering::Relaxed);
-
-            let mut m = proto::Metric::default();
-            let mut c = proto::Gauge::default();
-            c.set_value(v as f64);
-            m.set_gauge(c);
-
-            let mut labels = Vec::with_capacity(self.labels.len() + 1);
-            labels.extend_from_slice(&self.labels);
-            labels.push(shard_label);
-
-            m.set_label(labels);
-            metrics.push(m);
+            x.swap(0, std::sync::atomic::Ordering::Relaxed)
         })
     }
 }
-
-fn make_label_pairs(
-    desc: &core::Desc,
-    label_values: &[&str],
-) -> prometheus::Result<Vec<proto::LabelPair>> {
-    if desc.variable_labels.len() != label_values.len() {
-        return Err(prometheus::Error::InconsistentCardinality {
-            expect: desc.variable_labels.len(),
-            got: label_values.len(),
-        });
+impl<W: std::io::Write, const N: usize> measured::metric::MetricEncoding<TextEncoder<W>>
+    for HyperLogLogState<N>
+{
+    fn write_type(
+        name: impl MetricNameEncoder,
+        enc: &mut TextEncoder<W>,
+    ) -> Result<(), std::io::Error> {
+        enc.write_type(&name, measured::text::MetricType::Gauge)
     }
+    fn collect_into(
+        &self,
+        _: &(),
+        labels: impl LabelGroup,
+        name: impl MetricNameEncoder,
+        enc: &mut TextEncoder<W>,
+    ) -> Result<(), std::io::Error> {
+        struct I64(i64);
+        impl LabelValue for I64 {
+            fn visit<V: LabelVisitor>(&self, v: V) -> V::Output {
+                v.write_int(self.0)
+            }
+        }
 
-    let total_len = desc.variable_labels.len() + desc.const_label_pairs.len();
-    if total_len == 0 {
-        return Ok(vec![]);
-    }
+        struct HllShardLabel {
+            hll_shard: i64,
+        }
 
-    if desc.variable_labels.is_empty() {
-        return Ok(desc.const_label_pairs.clone());
-    }
+        impl LabelGroup for HllShardLabel {
+            fn visit_values(&self, v: &mut impl LabelGroupVisitor) {
+                const LE: &LabelName = LabelName::from_str("hll_shard");
+                v.write_value(LE, &I64(self.hll_shard));
+            }
+        }
 
-    let mut label_pairs = Vec::with_capacity(total_len);
-    for (i, n) in desc.variable_labels.iter().enumerate() {
-        let mut label_pair = proto::LabelPair::default();
-        label_pair.set_name(n.clone());
-        label_pair.set_value(label_values[i].to_owned());
-        label_pairs.push(label_pair);
+        self.take_sample()
+            .into_iter()
+            .enumerate()
+            .try_for_each(|(hll_shard, val)| {
+                enc.write_metric_value(
+                    name.by_ref(),
+                    labels.by_ref().compose_with(HllShardLabel {
+                        hll_shard: hll_shard as i64,
+                    }),
+                    MetricValue::Int(val as i64),
+                )
+            })
     }
-
-    for label_pair in &desc.const_label_pairs {
-        label_pairs.push(label_pair.clone());
-    }
-    label_pairs.sort();
-    Ok(label_pairs)
 }
 
 #[cfg(test)]
 mod tests {
     use std::collections::HashSet;
 
-    use prometheus::{proto, Opts};
+    use measured::{label::StaticLabelSet, FixedCardinalityLabel};
     use rand::{rngs::StdRng, Rng, SeedableRng};
     use rand_distr::{Distribution, Zipf};
 
     use crate::HyperLogLogVec;
 
-    fn collect(hll: &HyperLogLogVec<32>) -> Vec<proto::Metric> {
-        let mut metrics = vec![];
-        hll.core
-            .children
-            .read()
-            .unwrap()
-            .values()
-            .for_each(|c| c.core.collect_into(&mut metrics));
-        metrics
+    #[derive(FixedCardinalityLabel, Clone, Copy)]
+    #[label(singleton = "x")]
+    enum Label {
+        A,
+        B,
     }
-    fn get_cardinality(metrics: &[proto::Metric], filter: impl Fn(&proto::Metric) -> bool) -> f64 {
+
+    fn collect(hll: &HyperLogLogVec<StaticLabelSet<Label>, 32>) -> ([u8; 32], [u8; 32]) {
+        // cannot go through the `hll.collect_family_into` interface yet...
+        // need to see if I can fix the conflicting impls problem in measured.
+        (
+            hll.get_metric(hll.with_labels(Label::A)).take_sample(),
+            hll.get_metric(hll.with_labels(Label::B)).take_sample(),
+        )
+    }
+
+    fn get_cardinality(samples: &[[u8; 32]]) -> f64 {
         let mut buckets = [0.0; 32];
-        for metric in metrics.chunks_exact(32) {
-            if filter(&metric[0]) {
-                for (i, m) in metric.iter().enumerate() {
-                    buckets[i] = f64::max(buckets[i], m.get_gauge().get_value());
-                }
+        for &sample in samples {
+            for (i, m) in sample.into_iter().enumerate() {
+                buckets[i] = f64::max(buckets[i], m as f64);
             }
         }
 
@@ -437,7 +238,7 @@ mod tests {
     }
 
     fn test_cardinality(n: usize, dist: impl Distribution<f64>) -> ([usize; 3], [f64; 3]) {
-        let hll = HyperLogLogVec::<32>::new(Opts::new("foo", "bar"), &["x"]).unwrap();
+        let hll = HyperLogLogVec::<StaticLabelSet<Label>, 32>::new();
 
         let mut iter = StdRng::seed_from_u64(0x2024_0112).sample_iter(dist);
         let mut set_a = HashSet::new();
@@ -445,18 +246,20 @@ mod tests {
 
         for x in iter.by_ref().take(n) {
             set_a.insert(x.to_bits());
-            hll.with_label_values(&["a"]).measure(&x.to_bits());
+            hll.get_metric(hll.with_labels(Label::A))
+                .measure(&x.to_bits());
         }
         for x in iter.by_ref().take(n) {
             set_b.insert(x.to_bits());
-            hll.with_label_values(&["b"]).measure(&x.to_bits());
+            hll.get_metric(hll.with_labels(Label::B))
+                .measure(&x.to_bits());
         }
         let merge = &set_a | &set_b;
 
-        let metrics = collect(&hll);
-        let len = get_cardinality(&metrics, |_| true);
-        let len_a = get_cardinality(&metrics, |l| l.get_label()[0].get_value() == "a");
-        let len_b = get_cardinality(&metrics, |l| l.get_label()[0].get_value() == "b");
+        let (a, b) = collect(&hll);
+        let len = get_cardinality(&[a, b]);
+        let len_a = get_cardinality(&[a]);
+        let len_b = get_cardinality(&[b]);
 
         ([merge.len(), set_a.len(), set_b.len()], [len, len_a, len_b])
     }

libs/metrics/src/lib.rs

@@ -4,6 +4,17 @@
 //! a default registry.
 #![deny(clippy::undocumented_unsafe_blocks)]
 
+use measured::{
+    label::{LabelGroupSet, LabelGroupVisitor, LabelName, NoLabels},
+    metric::{
+        counter::CounterState,
+        gauge::GaugeState,
+        group::{Encoding, MetricValue},
+        name::{MetricName, MetricNameEncoder},
+        MetricEncoding, MetricFamilyEncoding,
+    },
+    FixedCardinalityLabel, LabelGroup, MetricGroup,
+};
 use once_cell::sync::Lazy;
 use prometheus::core::{
     Atomic, AtomicU64, Collector, GenericCounter, GenericCounterVec, GenericGauge, GenericGaugeVec,
@@ -11,6 +22,7 @@ use prometheus::core::{
 pub use prometheus::opts;
 pub use prometheus::register;
 pub use prometheus::Error;
+use prometheus::Registry;
 pub use prometheus::{core, default_registry, proto};
 pub use prometheus::{exponential_buckets, linear_buckets};
 pub use prometheus::{register_counter_vec, Counter, CounterVec};
@@ -23,13 +35,12 @@ pub use prometheus::{register_int_counter_vec, IntCounterVec};
 pub use prometheus::{register_int_gauge, IntGauge};
 pub use prometheus::{register_int_gauge_vec, IntGaugeVec};
 pub use prometheus::{Encoder, TextEncoder};
-use prometheus::{Registry, Result};
 
 pub mod launch_timestamp;
 mod wrappers;
 pub use wrappers::{CountedReader, CountedWriter};
 mod hll;
-pub use hll::{HyperLogLog, HyperLogLogVec};
+pub use hll::{HyperLogLog, HyperLogLogState, HyperLogLogVec};
 #[cfg(target_os = "linux")]
 pub mod more_process_metrics;
 
@@ -59,7 +70,7 @@ static INTERNAL_REGISTRY: Lazy<Registry> = Lazy::new(Registry::new);
 /// Register a collector in the internal registry. MUST be called before the first call to `gather()`.
 /// Otherwise, we can have a deadlock in the `gather()` call, trying to register a new collector
 /// while holding the lock.
-pub fn register_internal(c: Box<dyn Collector>) -> Result<()> {
+pub fn register_internal(c: Box<dyn Collector>) -> prometheus::Result<()> {
     INTERNAL_REGISTRY.register(c)
 }
 
@@ -96,6 +107,127 @@ pub const DISK_WRITE_SECONDS_BUCKETS: &[f64] = &[
     0.000_050, 0.000_100, 0.000_500, 0.001, 0.003, 0.005, 0.01, 0.05, 0.1, 0.3, 0.5,
 ];
 
+pub struct BuildInfo {
+    pub revision: &'static str,
+    pub build_tag: &'static str,
+}
+
+// todo: allow label group without the set
+impl LabelGroup for BuildInfo {
+    fn visit_values(&self, v: &mut impl LabelGroupVisitor) {
+        const REVISION: &LabelName = LabelName::from_str("revision");
+        v.write_value(REVISION, &self.revision);
+        const BUILD_TAG: &LabelName = LabelName::from_str("build_tag");
+        v.write_value(BUILD_TAG, &self.build_tag);
+    }
+}
+
+impl<T: Encoding> MetricFamilyEncoding<T> for BuildInfo
+where
+    GaugeState: MetricEncoding<T>,
+{
+    fn collect_family_into(
+        &self,
+        name: impl measured::metric::name::MetricNameEncoder,
+        enc: &mut T,
+    ) -> Result<(), T::Err> {
+        enc.write_help(&name, "Build/version information")?;
+        GaugeState::write_type(&name, enc)?;
+        GaugeState {
+            count: std::sync::atomic::AtomicI64::new(1),
+        }
+        .collect_into(&(), self, name, enc)
+    }
+}
+
+#[derive(MetricGroup)]
+#[metric(new(build_info: BuildInfo))]
+pub struct NeonMetrics {
+    #[cfg(target_os = "linux")]
+    #[metric(namespace = "process")]
+    #[metric(init = measured_process::ProcessCollector::for_self())]
+    process: measured_process::ProcessCollector,
+
+    #[metric(namespace = "libmetrics")]
+    #[metric(init = LibMetrics::new(build_info))]
+    libmetrics: LibMetrics,
+}
+
+#[derive(MetricGroup)]
+#[metric(new(build_info: BuildInfo))]
+pub struct LibMetrics {
+    #[metric(init = build_info)]
+    build_info: BuildInfo,
+
+    #[metric(flatten)]
+    rusage: Rusage,
+
+    serve_count: CollectionCounter,
+}
+
+fn write_gauge<Enc: Encoding>(
+    x: i64,
+    labels: impl LabelGroup,
+    name: impl MetricNameEncoder,
+    enc: &mut Enc,
+) -> Result<(), Enc::Err> {
+    enc.write_metric_value(name, labels, MetricValue::Int(x))
+}
+
+#[derive(Default)]
+struct Rusage;
+
+#[derive(FixedCardinalityLabel, Clone, Copy)]
+#[label(singleton = "io_operation")]
+enum IoOp {
+    Read,
+    Write,
+}
+
+impl<T: Encoding> MetricGroup<T> for Rusage
+where
+    GaugeState: MetricEncoding<T>,
+{
+    fn collect_group_into(&self, enc: &mut T) -> Result<(), T::Err> {
+        const DISK_IO: &MetricName = MetricName::from_str("disk_io_bytes_total");
+        const MAXRSS: &MetricName = MetricName::from_str("maxrss_kb");
+
+        let ru = get_rusage_stats();
+
+        enc.write_help(
+            DISK_IO,
+            "Bytes written and read from disk, grouped by the operation (read|write)",
+        )?;
+        GaugeState::write_type(DISK_IO, enc)?;
+        write_gauge(ru.ru_inblock * BYTES_IN_BLOCK, IoOp::Read, DISK_IO, enc)?;
+        write_gauge(ru.ru_oublock * BYTES_IN_BLOCK, IoOp::Write, DISK_IO, enc)?;
+
+        enc.write_help(MAXRSS, "Memory usage (Maximum Resident Set Size)")?;
+        GaugeState::write_type(MAXRSS, enc)?;
+        write_gauge(ru.ru_maxrss, IoOp::Read, MAXRSS, enc)?;
+
+        Ok(())
+    }
+}
+
+#[derive(Default)]
+struct CollectionCounter(CounterState);
+
+impl<T: Encoding> MetricFamilyEncoding<T> for CollectionCounter
+where
+    CounterState: MetricEncoding<T>,
+{
+    fn collect_family_into(
+        &self,
+        name: impl measured::metric::name::MetricNameEncoder,
+        enc: &mut T,
+    ) -> Result<(), T::Err> {
+        self.0.inc();
+        enc.write_help(&name, "Number of metric requests made")?;
+        self.0.collect_into(&(), NoLabels, name, enc)
+    }
+}
+
 pub fn set_build_info_metric(revision: &str, build_tag: &str) {
     let metric = register_int_gauge_vec!(
         "libmetrics_build_info",
@@ -105,6 +237,7 @@ pub fn set_build_info_metric(revision: &str, build_tag: &str) {
     .expect("Failed to register build info metric");
     metric.with_label_values(&[revision, build_tag]).set(1);
 }
+const BYTES_IN_BLOCK: i64 = 512;
 
 // Records I/O stats in a "cross-platform" way.
 // Compiles both on macOS and Linux, but current macOS implementation always returns 0 as values for I/O stats.
@@ -117,14 +250,22 @@ pub fn set_build_info_metric(revision: &str, build_tag: &str) {
 fn update_rusage_metrics() {
     let rusage_stats = get_rusage_stats();
 
-    const BYTES_IN_BLOCK: i64 = 512;
     DISK_IO_BYTES
         .with_label_values(&["read"])
         .set(rusage_stats.ru_inblock * BYTES_IN_BLOCK);
     DISK_IO_BYTES
         .with_label_values(&["write"])
         .set(rusage_stats.ru_oublock * BYTES_IN_BLOCK);
-    MAXRSS_KB.set(rusage_stats.ru_maxrss);
+
+    // On macOS, the unit of maxrss is bytes; on Linux, it's kilobytes. https://stackoverflow.com/a/59915669
+    #[cfg(target_os = "macos")]
+    {
+        MAXRSS_KB.set(rusage_stats.ru_maxrss / 1024);
+    }
+    #[cfg(not(target_os = "macos"))]
+    {
+        MAXRSS_KB.set(rusage_stats.ru_maxrss);
+    }
 }
 
 fn get_rusage_stats() -> libc::rusage {
@@ -151,6 +292,7 @@ macro_rules! register_int_counter_pair_vec {
         }
     }};
 }
+
 /// Create an [`IntCounterPair`] and registers to default registry.
 #[macro_export(local_inner_macros)]
 macro_rules! register_int_counter_pair {
@@ -188,7 +330,10 @@ impl<P: Atomic> GenericCounterPairVec<P> {
     ///
     /// An error is returned if the number of label values is not the same as the
     /// number of VariableLabels in Desc.
-    pub fn get_metric_with_label_values(&self, vals: &[&str]) -> Result<GenericCounterPair<P>> {
+    pub fn get_metric_with_label_values(
+        &self,
+        vals: &[&str],
+    ) -> prometheus::Result<GenericCounterPair<P>> {
         Ok(GenericCounterPair {
             inc: self.inc.get_metric_with_label_values(vals)?,
             dec: self.dec.get_metric_with_label_values(vals)?,
@@ -201,7 +346,7 @@ impl<P: Atomic> GenericCounterPairVec<P> {
         self.get_metric_with_label_values(vals).unwrap()
     }
 
-    pub fn remove_label_values(&self, res: &mut [Result<()>; 2], vals: &[&str]) {
+    pub fn remove_label_values(&self, res: &mut [prometheus::Result<()>; 2], vals: &[&str]) {
         res[0] = self.inc.remove_label_values(vals);
         res[1] = self.dec.remove_label_values(vals);
     }
@@ -285,3 +430,171 @@ pub type IntCounterPair = GenericCounterPair<AtomicU64>;
 
 /// A guard for [`IntCounterPair`] that will decrement the gauge on drop
 pub type IntCounterPairGuard = GenericCounterPairGuard<AtomicU64>;
+
+pub trait CounterPairAssoc {
+    const INC_NAME: &'static MetricName;
+    const DEC_NAME: &'static MetricName;
+
+    const INC_HELP: &'static str;
+    const DEC_HELP: &'static str;
+
+    type LabelGroupSet: LabelGroupSet;
+}
+
+pub struct CounterPairVec<A: CounterPairAssoc> {
+    vec: measured::metric::MetricVec<MeasuredCounterPairState, A::LabelGroupSet>,
+}
+
+impl<A: CounterPairAssoc> Default for CounterPairVec<A>
+where
+    A::LabelGroupSet: Default,
+{
+    fn default() -> Self {
+        Self {
+            vec: Default::default(),
+        }
+    }
+}
+
+impl<A: CounterPairAssoc> CounterPairVec<A> {
+    pub fn guard(
+        &self,
+        labels: <A::LabelGroupSet as LabelGroupSet>::Group<'_>,
+    ) -> MeasuredCounterPairGuard<'_, A> {
+        let id = self.vec.with_labels(labels);
+        self.vec.get_metric(id).inc.inc();
+        MeasuredCounterPairGuard { vec: &self.vec, id }
+    }
+    pub fn inc(&self, labels: <A::LabelGroupSet as LabelGroupSet>::Group<'_>) {
+        let id = self.vec.with_labels(labels);
+        self.vec.get_metric(id).inc.inc();
+    }
+    pub fn dec(&self, labels: <A::LabelGroupSet as LabelGroupSet>::Group<'_>) {
+        let id = self.vec.with_labels(labels);
+        self.vec.get_metric(id).dec.inc();
+    }
+    pub fn remove_metric(
+        &self,
+        labels: <A::LabelGroupSet as LabelGroupSet>::Group<'_>,
+    ) -> Option<MeasuredCounterPairState> {
+        let id = self.vec.with_labels(labels);
+        self.vec.remove_metric(id)
+    }
+}
+
+impl<T, A> ::measured::metric::group::MetricGroup<T> for CounterPairVec<A>
+where
+    T: ::measured::metric::group::Encoding,
+    A: CounterPairAssoc,
+    ::measured::metric::counter::CounterState: ::measured::metric::MetricEncoding<T>,
+{
+    fn collect_group_into(&self, enc: &mut T) -> Result<(), T::Err> {
+        // write decrement first to avoid a race condition where inc - dec < 0
+        T::write_help(enc, A::DEC_NAME, A::DEC_HELP)?;
+        self.vec
+            .collect_family_into(A::DEC_NAME, &mut Dec(&mut *enc))?;
+
+        T::write_help(enc, A::INC_NAME, A::INC_HELP)?;
+        self.vec
+            .collect_family_into(A::INC_NAME, &mut Inc(&mut *enc))?;
+
+        Ok(())
+    }
+}
+
+#[derive(MetricGroup, Default)]
+pub struct MeasuredCounterPairState {
+    pub inc: CounterState,
+    pub dec: CounterState,
+}
+
+impl measured::metric::MetricType for MeasuredCounterPairState {
+    type Metadata = ();
+}
+
+pub struct MeasuredCounterPairGuard<'a, A: CounterPairAssoc> {
+    vec: &'a measured::metric::MetricVec<MeasuredCounterPairState, A::LabelGroupSet>,
+    id: measured::metric::LabelId<A::LabelGroupSet>,
+}
+
+impl<A: CounterPairAssoc> Drop for MeasuredCounterPairGuard<'_, A> {
+    fn drop(&mut self) {
+        self.vec.get_metric(self.id).dec.inc();
+    }
+}
+
+/// [`MetricEncoding`] for [`MeasuredCounterPairState`] that only writes the inc counter to the inner encoder.
+struct Inc<T>(T);
+/// [`MetricEncoding`] for [`MeasuredCounterPairState`] that only writes the dec counter to the inner encoder.
+struct Dec<T>(T);
+
+impl<T: Encoding> Encoding for Inc<T> {
+    type Err = T::Err;
+
+    fn write_help(&mut self, name: impl MetricNameEncoder, help: &str) -> Result<(), Self::Err> {
+        self.0.write_help(name, help)
+    }
+
+    fn write_metric_value(
+        &mut self,
+        name: impl MetricNameEncoder,
+        labels: impl LabelGroup,
+        value: MetricValue,
+    ) -> Result<(), Self::Err> {
+        self.0.write_metric_value(name, labels, value)
+    }
+}
+
+impl<T: Encoding> MetricEncoding<Inc<T>> for MeasuredCounterPairState
+where
+    CounterState: MetricEncoding<T>,
+{
+    fn write_type(name: impl MetricNameEncoder, enc: &mut Inc<T>) -> Result<(), T::Err> {
+        CounterState::write_type(name, &mut enc.0)
+    }
+    fn collect_into(
+        &self,
+        metadata: &(),
+        labels: impl LabelGroup,
+        name: impl MetricNameEncoder,
+        enc: &mut Inc<T>,
+    ) -> Result<(), T::Err> {
+        self.inc.collect_into(metadata, labels, name, &mut enc.0)
+    }
+}
+
+impl<T: Encoding> Encoding for Dec<T> {
+    type Err = T::Err;
+
+    fn write_help(&mut self, name: impl MetricNameEncoder, help: &str) -> Result<(), Self::Err> {
+        self.0.write_help(name, help)
+    }
+
+    fn write_metric_value(
+        &mut self,
+        name: impl MetricNameEncoder,
+        labels: impl LabelGroup,
+        value: MetricValue,
+    ) -> Result<(), Self::Err> {
+        self.0.write_metric_value(name, labels, value)
+    }
+}
+
+/// Write the dec counter to the encoder
+impl<T: Encoding> MetricEncoding<Dec<T>> for MeasuredCounterPairState
+where
+    CounterState: MetricEncoding<T>,
+{
+    fn write_type(name: impl MetricNameEncoder, enc: &mut Dec<T>) -> Result<(), T::Err> {
+        CounterState::write_type(name, &mut enc.0)
+    }
+    fn collect_into(
+        &self,
+        metadata: &(),
+        labels: impl LabelGroup,
+        name: impl MetricNameEncoder,
+        enc: &mut Dec<T>,
+    ) -> Result<(), T::Err> {
+        self.dec.collect_into(metadata, labels, name, &mut enc.0)
+    }
+}

libs/pageserver_api/src/config.rs

@@ -0,0 +1,31 @@
+use std::collections::HashMap;
+
+use const_format::formatcp;
+
+#[cfg(test)]
+mod tests;
+
+pub const DEFAULT_PG_LISTEN_PORT: u16 = 64000;
+pub const DEFAULT_PG_LISTEN_ADDR: &str = formatcp!("127.0.0.1:{DEFAULT_PG_LISTEN_PORT}");
+pub const DEFAULT_HTTP_LISTEN_PORT: u16 = 9898;
+pub const DEFAULT_HTTP_LISTEN_ADDR: &str = formatcp!("127.0.0.1:{DEFAULT_HTTP_LISTEN_PORT}");
+
+// Certain metadata (e.g. externally-addressable name, AZ) is delivered
+// as a separate structure.  This information is not neeed by the pageserver
+// itself, it is only used for registering the pageserver with the control
+// plane and/or storage controller.
+//
+#[derive(PartialEq, Eq, Debug, serde::Serialize, serde::Deserialize)]
+pub struct NodeMetadata {
+    #[serde(rename = "host")]
+    pub postgres_host: String,
+    #[serde(rename = "port")]
+    pub postgres_port: u16,
+    pub http_host: String,
+    pub http_port: u16,
+
+    // Deployment tools may write fields to the metadata file beyond what we
+    // use in this type: this type intentionally only names fields that require.
+    #[serde(flatten)]
+    pub other: HashMap<String, serde_json::Value>,
+}

libs/pageserver_api/src/config/tests.rs

@@ -0,0 +1,22 @@
+use super::*;
+
+#[test]
+fn test_node_metadata_v1_backward_compatibilty() {
+    let v1 = serde_json::to_vec(&serde_json::json!({
+        "host": "localhost",
+        "port": 23,
+        "http_host": "localhost",
+        "http_port": 42,
+    }));
+
+    assert_eq!(
+        serde_json::from_slice::<NodeMetadata>(&v1.unwrap()).unwrap(),
+        NodeMetadata {
+            postgres_host: "localhost".to_string(),
+            postgres_port: 23,
+            http_host: "localhost".to_string(),
+            http_port: 42,
+            other: HashMap::new(),
+        }
+    )
+}

libs/pageserver_api/src/controller_api.rs

@@ -2,11 +2,14 @@ use std::str::FromStr;
 
 /// Request/response types for the storage controller
 /// API (`/control/v1` prefix).  Implemented by the server
-/// in [`attachment_service::http`]
+/// in [`storage_controller::http`]
 use serde::{Deserialize, Serialize};
-use utils::id::NodeId;
+use utils::id::{NodeId, TenantId};
 
-use crate::{models::ShardParameters, shard::TenantShardId};
+use crate::{
+    models::{ShardParameters, TenantConfig},
+    shard::{ShardStripeSize, TenantShardId},
+};
 
 #[derive(Serialize, Deserialize)]
 pub struct TenantCreateResponseShard {
@@ -35,10 +38,16 @@ pub struct NodeRegisterRequest {
 pub struct NodeConfigureRequest {
     pub node_id: NodeId,
 
-    pub availability: Option<NodeAvailability>,
+    pub availability: Option<NodeAvailabilityWrapper>,
     pub scheduling: Option<NodeSchedulingPolicy>,
 }
 
+#[derive(Serialize, Deserialize)]
+pub struct TenantPolicyRequest {
+    pub placement: Option<PlacementPolicy>,
+    pub scheduling: Option<ShardSchedulingPolicy>,
+}
+
 #[derive(Serialize, Deserialize, Debug)]
 pub struct TenantLocateResponseShard {
     pub shard_id: TenantShardId,
@@ -57,6 +66,48 @@ pub struct TenantLocateResponse {
     pub shard_params: ShardParameters,
 }
 
+#[derive(Serialize, Deserialize)]
+pub struct TenantDescribeResponse {
+    pub tenant_id: TenantId,
+    pub shards: Vec<TenantDescribeResponseShard>,
+    pub stripe_size: ShardStripeSize,
+    pub policy: PlacementPolicy,
+    pub config: TenantConfig,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct NodeDescribeResponse {
+    pub id: NodeId,
+
+    pub availability: NodeAvailabilityWrapper,
+    pub scheduling: NodeSchedulingPolicy,
+
+    pub listen_http_addr: String,
+    pub listen_http_port: u16,
+
+    pub listen_pg_addr: String,
+    pub listen_pg_port: u16,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct TenantDescribeResponseShard {
+    pub tenant_shard_id: TenantShardId,
+
+    pub node_attached: Option<NodeId>,
+    pub node_secondary: Vec<NodeId>,
+
+    pub last_error: String,
+
+    /// A task is currently running to reconcile this tenant's intent state with the state on pageservers
+    pub is_reconciling: bool,
+    /// This shard failed in sending a compute notification to the cloud control plane, and a retry is pending.
+    pub is_pending_compute_notification: bool,
+    /// A shard split is currently underway
+    pub is_splitting: bool,
+
+    pub scheduling_policy: ShardSchedulingPolicy,
+}
+
 /// Explicitly migrating a particular shard is a low level operation
 /// TODO: higher level "Reschedule tenant" operation where the request
 /// specifies some constraints, e.g. asking it to get off particular node(s)
@@ -66,29 +117,94 @@ pub struct TenantShardMigrateRequest {
     pub node_id: NodeId,
 }
 
-#[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq)]
+/// Utilisation score indicating how good a candidate a pageserver
+/// is for scheduling the next tenant. See [`crate::models::PageserverUtilization`].
+/// Lower values are better.
+#[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, PartialOrd, Ord, Debug)]
+pub struct UtilizationScore(pub u64);
+
+impl UtilizationScore {
+    pub fn worst() -> Self {
+        UtilizationScore(u64::MAX)
+    }
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy, Debug)]
+#[serde(into = "NodeAvailabilityWrapper")]
 pub enum NodeAvailability {
     // Normal, happy state
-    Active,
+    Active(UtilizationScore),
     // Offline: Tenants shouldn't try to attach here, but they may assume that their
     // secondary locations on this node still exist.  Newly added nodes are in this
     // state until we successfully contact them.
     Offline,
 }
 
-impl FromStr for NodeAvailability {
-    type Err = anyhow::Error;
+impl PartialEq for NodeAvailability {
+    fn eq(&self, other: &Self) -> bool {
+        use NodeAvailability::*;
+        matches!((self, other), (Active(_), Active(_)) | (Offline, Offline))
+    }
+}
 
-    fn from_str(s: &str) -> Result<Self, Self::Err> {
-        match s {
-            "active" => Ok(Self::Active),
-            "offline" => Ok(Self::Offline),
-            _ => Err(anyhow::anyhow!("Unknown availability state '{s}'")),
+impl Eq for NodeAvailability {}
+
+// This wrapper provides serde functionality and it should only be used to
+// communicate with external callers which don't know or care about the
+// utilisation score of the pageserver it is targeting.
+#[derive(Serialize, Deserialize, Clone, Copy, Debug)]
+pub enum NodeAvailabilityWrapper {
+    Active,
+    Offline,
+}
+
+impl From<NodeAvailabilityWrapper> for NodeAvailability {
+    fn from(val: NodeAvailabilityWrapper) -> Self {
+        match val {
+            // Assume the worst utilisation score to begin with. It will later be updated by
+            // the heartbeats.
+            NodeAvailabilityWrapper::Active => NodeAvailability::Active(UtilizationScore::worst()),
+            NodeAvailabilityWrapper::Offline => NodeAvailability::Offline,
         }
     }
 }
 
-#[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq)]
+impl From<NodeAvailability> for NodeAvailabilityWrapper {
+    fn from(val: NodeAvailability) -> Self {
+        match val {
+            NodeAvailability::Active(_) => NodeAvailabilityWrapper::Active,
+            NodeAvailability::Offline => NodeAvailabilityWrapper::Offline,
+        }
+    }
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, Debug)]
+pub enum ShardSchedulingPolicy {
+    // Normal mode: the tenant's scheduled locations may be updated at will, including
+    // for non-essential optimization.
+    Active,
+
+    // Disable optimizations, but permit scheduling when necessary to fulfil the PlacementPolicy.
+    // For example, this still permits a node's attachment location to change to a secondary in
+    // response to a node failure, or to assign a new secondary if a node was removed.
+    Essential,
+
+    // No scheduling: leave the shard running wherever it currently is.  Even if the shard is
+    // unavailable, it will not be rescheduled to another node.
+    Pause,
+
+    // No reconciling: we will make no location_conf API calls to pageservers at all.  If the
+    // shard is unavailable, it stays that way.  If a node fails, this shard doesn't get failed over.
+    Stop,
+}
+
+impl Default for ShardSchedulingPolicy {
+    fn default() -> Self {
+        Self::Active
+    }
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, Debug)]
 pub enum NodeSchedulingPolicy {
     Active,
     Filling,
@@ -127,11 +243,8 @@ impl From<NodeSchedulingPolicy> for String {
 /// to create secondary locations.
 #[derive(Clone, Serialize, Deserialize, Debug, PartialEq, Eq)]
 pub enum PlacementPolicy {
-    /// Cheapest way to attach a tenant: just one pageserver, no secondary
-    Single,
-    /// Production-ready way to attach a tenant: one attached pageserver and
-    /// some number of secondaries.
-    Double(usize),
+    /// Normal live state: one attached pageserver and zero or more secondaries.
+    Attached(usize),
     /// Create one secondary mode locations. This is useful when onboarding
     /// a tenant, or for an idle tenant that we might want to bring online quickly.
     Secondary,
@@ -153,14 +266,14 @@ mod test {
     /// Check stability of PlacementPolicy's serialization
     #[test]
     fn placement_policy_encoding() -> anyhow::Result<()> {
-        let v = PlacementPolicy::Double(1);
+        let v = PlacementPolicy::Attached(1);
         let encoded = serde_json::to_string(&v)?;
-        assert_eq!(encoded, "{\"Double\":1}");
+        assert_eq!(encoded, "{\"Attached\":1}");
         assert_eq!(serde_json::from_str::<PlacementPolicy>(&encoded)?, v);
 
-        let v = PlacementPolicy::Single;
+        let v = PlacementPolicy::Detached;
         let encoded = serde_json::to_string(&v)?;
-        assert_eq!(encoded, "\"Single\"");
+        assert_eq!(encoded, "\"Detached\"");
         assert_eq!(serde_json::from_str::<PlacementPolicy>(&encoded)?, v);
         Ok(())
     }

libs/pageserver_api/src/key.rs

@@ -1,5 +1,6 @@
 use anyhow::{bail, Result};
 use byteorder::{ByteOrder, BE};
+use bytes::BufMut;
 use postgres_ffi::relfile_utils::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
 use postgres_ffi::{Oid, TransactionId};
 use serde::{Deserialize, Serialize};
@@ -21,15 +22,107 @@ pub struct Key {
     pub field6: u32,
 }
 
+/// The storage key size.
 pub const KEY_SIZE: usize = 18;
 
+/// The metadata key size. 2B fewer than the storage key size because field2 is not fully utilized.
+/// See [`Key::to_i128`] for more information on the encoding.
+pub const METADATA_KEY_SIZE: usize = 16;
+
+/// The key prefix start range for the metadata keys. All keys with the first byte >= 0x40 is a metadata key.
+pub const METADATA_KEY_BEGIN_PREFIX: u8 = 0x60;
+pub const METADATA_KEY_END_PREFIX: u8 = 0x7F;
+
+/// The (reserved) key prefix of relation sizes.
+pub const RELATION_SIZE_PREFIX: u8 = 0x61;
+
+/// The key prefix of AUX file keys.
+pub const AUX_KEY_PREFIX: u8 = 0x62;
+
+/// Check if the key falls in the range of metadata keys.
+pub const fn is_metadata_key_slice(key: &[u8]) -> bool {
+    key[0] >= METADATA_KEY_BEGIN_PREFIX && key[0] < METADATA_KEY_END_PREFIX
+}
+
 impl Key {
+    /// Check if the key falls in the range of metadata keys.
+    pub const fn is_metadata_key(&self) -> bool {
+        self.field1 >= METADATA_KEY_BEGIN_PREFIX && self.field1 < METADATA_KEY_END_PREFIX
+    }
+
+    /// Encode a metadata key to a storage key.
+    pub fn from_metadata_key_fixed_size(key: &[u8; METADATA_KEY_SIZE]) -> Self {
+        assert!(is_metadata_key_slice(key), "key not in metadata key range");
+        Key {
+            field1: key[0],
+            field2: u16::from_be_bytes(key[1..3].try_into().unwrap()) as u32,
+            field3: u32::from_be_bytes(key[3..7].try_into().unwrap()),
+            field4: u32::from_be_bytes(key[7..11].try_into().unwrap()),
+            field5: key[11],
+            field6: u32::from_be_bytes(key[12..16].try_into().unwrap()),
+        }
+    }
+
+    /// Encode a metadata key to a storage key.
+    pub fn from_metadata_key(key: &[u8]) -> Self {
+        Self::from_metadata_key_fixed_size(key.try_into().expect("expect 16 byte metadata key"))
+    }
+
+    /// Extract a metadata key to a writer. The result should always be 16 bytes.
+    pub fn extract_metadata_key_to_writer(&self, mut writer: impl BufMut) {
+        writer.put_u8(self.field1);
+        assert!(self.field2 <= 0xFFFF);
+        writer.put_u16(self.field2 as u16);
+        writer.put_u32(self.field3);
+        writer.put_u32(self.field4);
+        writer.put_u8(self.field5);
+        writer.put_u32(self.field6);
+    }
+
+    /// Get the range of metadata keys.
+    pub const fn metadata_key_range() -> Range<Self> {
+        Key {
+            field1: METADATA_KEY_BEGIN_PREFIX,
+            field2: 0,
+            field3: 0,
+            field4: 0,
+            field5: 0,
+            field6: 0,
+        }..Key {
+            field1: METADATA_KEY_END_PREFIX,
+            field2: 0,
+            field3: 0,
+            field4: 0,
+            field5: 0,
+            field6: 0,
+        }
+    }
+
+    /// Get the range of aux keys.
+    pub fn metadata_aux_key_range() -> Range<Self> {
+        Key {
+            field1: AUX_KEY_PREFIX,
+            field2: 0,
+            field3: 0,
+            field4: 0,
+            field5: 0,
+            field6: 0,
+        }..Key {
+            field1: AUX_KEY_PREFIX + 1,
+            field2: 0,
+            field3: 0,
+            field4: 0,
+            field5: 0,
+            field6: 0,
+        }
+    }
+
     /// 'field2' is used to store tablespaceid for relations and small enum numbers for other relish.
     /// As long as Neon does not support tablespace (because of lack of access to local file system),
     /// we can assume that only some predefined namespace OIDs are used which can fit in u16
     pub fn to_i128(&self) -> i128 {
         assert!(self.field2 < 0xFFFF || self.field2 == 0xFFFFFFFF || self.field2 == 0x22222222);
-        (((self.field1 & 0xf) as i128) << 120)
+        (((self.field1 & 0x7F) as i128) << 120)
             | (((self.field2 & 0xFFFF) as i128) << 104)
             | ((self.field3 as i128) << 72)
             | ((self.field4 as i128) << 40)
@@ -39,7 +132,7 @@ impl Key {
 
     pub const fn from_i128(x: i128) -> Self {
         Key {
-            field1: ((x >> 120) & 0xf) as u8,
+            field1: ((x >> 120) & 0x7F) as u8,
             field2: ((x >> 104) & 0xFFFF) as u32,
             field3: (x >> 72) as u32,
             field4: (x >> 40) as u32,
@@ -48,11 +141,11 @@ impl Key {
         }
     }
 
-    pub fn next(&self) -> Key {
+    pub const fn next(&self) -> Key {
         self.add(1)
     }
 
-    pub fn add(&self, x: u32) -> Key {
+    pub const fn add(&self, x: u32) -> Key {
         let mut key = *self;
 
         let r = key.field6.overflowing_add(x);
@@ -81,6 +174,8 @@ impl Key {
         key
     }
 
+    /// Convert a 18B slice to a key. This function should not be used for metadata keys because field2 is handled differently.
+    /// Use [`Key::from_metadata_key`] instead.
     pub fn from_slice(b: &[u8]) -> Self {
         Key {
             field1: b[0],
@@ -92,6 +187,8 @@ impl Key {
         }
     }
 
+    /// Convert a key to a 18B slice. This function should not be used for metadata keys because field2 is handled differently.
+    /// Use [`Key::extract_metadata_key_to_writer`] instead.
     pub fn write_to_byte_slice(&self, buf: &mut [u8]) {
         buf[0] = self.field1;
         BE::write_u32(&mut buf[1..5], self.field2);
@@ -475,12 +572,17 @@ pub const AUX_FILES_KEY: Key = Key {
 // Reverse mappings for a few Keys.
 // These are needed by WAL redo manager.
 
+/// Non inherited range for vectored get.
+pub const NON_INHERITED_RANGE: Range<Key> = AUX_FILES_KEY..AUX_FILES_KEY.next();
+/// Sparse keyspace range for vectored get. Missing key error will be ignored for this range.
+pub const NON_INHERITED_SPARSE_RANGE: Range<Key> = Key::metadata_key_range();
+
 // AUX_FILES currently stores only data for logical replication (slots etc), and
 // we don't preserve these on a branch because safekeepers can't follow timeline
 // switch (and generally it likely should be optional), so ignore these.
 #[inline(always)]
 pub fn is_inherited_key(key: Key) -> bool {
-    key != AUX_FILES_KEY
+    !NON_INHERITED_RANGE.contains(&key) && !NON_INHERITED_SPARSE_RANGE.contains(&key)
 }
 
 #[inline(always)]
@@ -556,11 +658,14 @@ impl std::str::FromStr for Key {
 mod tests {
     use std::str::FromStr;
 
+    use crate::key::is_metadata_key_slice;
     use crate::key::Key;
 
     use rand::Rng;
     use rand::SeedableRng;
 
+    use super::AUX_KEY_PREFIX;
+
     #[test]
     fn display_fromstr_bijection() {
         let mut rng = rand::rngs::StdRng::seed_from_u64(42);
@@ -576,4 +681,16 @@ mod tests {
 
         assert_eq!(key, Key::from_str(&format!("{key}")).unwrap());
     }
+
+    #[test]
+    fn test_metadata_keys() {
+        let mut metadata_key = vec![AUX_KEY_PREFIX];
+        metadata_key.extend_from_slice(&[0xFF; 15]);
+        let encoded_key = Key::from_metadata_key(&metadata_key);
+        let mut output_key = Vec::new();
+        encoded_key.extract_metadata_key_to_writer(&mut output_key);
+        assert_eq!(metadata_key, output_key);
+        assert!(encoded_key.is_metadata_key());
+        assert!(is_metadata_key_slice(&metadata_key));
+    }
 }

libs/pageserver_api/src/keyspace.rs

@@ -1,7 +1,10 @@
 use postgres_ffi::BLCKSZ;
 use std::ops::Range;
 
-use crate::key::Key;
+use crate::{
+    key::Key,
+    shard::{ShardCount, ShardIdentity},
+};
 use itertools::Itertools;
 
 ///
@@ -14,44 +17,279 @@ pub struct KeySpace {
     pub ranges: Vec<Range<Key>>,
 }
 
-impl KeySpace {
+/// A wrapper type for sparse keyspaces.
+#[derive(Clone, Debug, Default, PartialEq, Eq)]
+pub struct SparseKeySpace(pub KeySpace);
+
+/// Represents a contiguous half-open range of the keyspace, masked according to a particular
+/// ShardNumber's stripes: within this range of keys, only some "belong" to the current
+/// shard.
+///
+/// When we iterate over keys within this object, we will skip any keys that don't belong
+/// to this shard.
+///
+/// The start + end keys may not belong to the shard: these specify where layer files should
+/// start  + end, but we will never actually read/write those keys.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct ShardedRange<'a> {
+    pub shard_identity: &'a ShardIdentity,
+    pub range: Range<Key>,
+}
+
+// Calculate the size of a range within the blocks of the same relation, or spanning only the
+// top page in the previous relation's space.
+fn contiguous_range_len(range: &Range<Key>) -> u32 {
+    debug_assert!(is_contiguous_range(range));
+    if range.start.field6 == 0xffffffff {
+        range.end.field6 + 1
+    } else {
+        range.end.field6 - range.start.field6
+    }
+}
+
+/// Return true if this key range includes only keys in the same relation's data blocks, or
+/// just spanning one relation and the logical size (0xffffffff) block of the relation before it.
+///
+/// Contiguous in this context means we know the keys are in use _somewhere_, but it might not
+/// be on our shard.  Later in ShardedRange we do the extra work to figure out how much
+/// of a given contiguous range is present on one shard.
+///
+/// This matters, because:
+/// - Within such ranges, keys are used contiguously.  Outside such ranges it is sparse.
+/// - Within such ranges, we may calculate distances using simple subtraction of field6.
+fn is_contiguous_range(range: &Range<Key>) -> bool {
+    range.start.field1 == range.end.field1
+        && range.start.field2 == range.end.field2
+        && range.start.field3 == range.end.field3
+        && range.start.field4 == range.end.field4
+        && (range.start.field5 == range.end.field5
+            || (range.start.field6 == 0xffffffff && range.start.field5 + 1 == range.end.field5))
+}
+
+impl<'a> ShardedRange<'a> {
+    pub fn new(range: Range<Key>, shard_identity: &'a ShardIdentity) -> Self {
+        Self {
+            shard_identity,
+            range,
+        }
+    }
+
+    /// Break up this range into chunks, each of which has at least one local key in it if the
+    /// total range has at least one local key.
+    pub fn fragment(self, target_nblocks: u32) -> Vec<(u32, Range<Key>)> {
+        // Optimization for single-key case (e.g. logical size keys)
+        if self.range.end == self.range.start.add(1) {
+            return vec![(
+                if self.shard_identity.is_key_disposable(&self.range.start) {
+                    0
+                } else {
+                    1
+                },
+                self.range,
+            )];
+        }
+
+        if !is_contiguous_range(&self.range) {
+            // Ranges that span relations are not fragmented.  We only get these ranges as a result
+            // of operations that act on existing layers, so we trust that the existing range is
+            // reasonably small.
+            return vec![(u32::MAX, self.range)];
+        }
+
+        let mut fragments: Vec<(u32, Range<Key>)> = Vec::new();
+
+        let mut cursor = self.range.start;
+        while cursor < self.range.end {
+            let advance_by = self.distance_to_next_boundary(cursor);
+            let is_fragment_disposable = self.shard_identity.is_key_disposable(&cursor);
+
+            // If the previous fragment is undersized, then we seek to consume enough
+            // blocks to complete it.
+            let (want_blocks, merge_last_fragment) = match fragments.last_mut() {
+                Some(frag) if frag.0 < target_nblocks => (target_nblocks - frag.0, Some(frag)),
+                Some(frag) => {
+                    // Prev block is complete, want the full number.
+                    (
+                        target_nblocks,
+                        if is_fragment_disposable {
+                            // If this current range will be empty (not shard-local data), we will merge into previous
+                            Some(frag)
+                        } else {
+                            None
+                        },
+                    )
+                }
+                None => {
+                    // First iteration, want the full number
+                    (target_nblocks, None)
+                }
+            };
+
+            let advance_by = if is_fragment_disposable {
+                advance_by
+            } else {
+                std::cmp::min(advance_by, want_blocks)
+            };
+
+            let next_cursor = cursor.add(advance_by);
+
+            let this_frag = (
+                if is_fragment_disposable {
+                    0
+                } else {
+                    advance_by
+                },
+                cursor..next_cursor,
+            );
+            cursor = next_cursor;
+
+            if let Some(last_fragment) = merge_last_fragment {
+                // Previous fragment was short or this one is empty, merge into it
+                last_fragment.0 += this_frag.0;
+                last_fragment.1.end = this_frag.1.end;
+            } else {
+                fragments.push(this_frag);
+            }
+        }
+
+        fragments
+    }
+
+    /// Estimate the physical pages that are within this range, on this shard.  This returns
+    /// u32::MAX if the range spans relations: this return value should be interpreted as "large".
+    pub fn page_count(&self) -> u32 {
+        // Special cases for single keys like logical sizes
+        if self.range.end == self.range.start.add(1) {
+            return if self.shard_identity.is_key_disposable(&self.range.start) {
+                0
+            } else {
+                1
+            };
+        }
+
+        // We can only do an authentic calculation of contiguous key ranges
+        if !is_contiguous_range(&self.range) {
+            return u32::MAX;
+        }
+
+        // Special case for single sharded tenants: our logical and physical sizes are the same
+        if self.shard_identity.count < ShardCount::new(2) {
+            return contiguous_range_len(&self.range);
+        }
+
+        // Normal path: step through stripes and part-stripes in the range, evaluate whether each one belongs
+        // to Self, and add the stripe's block count to our total if so.
+        let mut result: u64 = 0;
+        let mut cursor = self.range.start;
+        while cursor < self.range.end {
+            // Count up to the next stripe_size boundary or end of range
+            let advance_by = self.distance_to_next_boundary(cursor);
+
+            // If this blocks in this stripe belong to us, add them to our count
+            if !self.shard_identity.is_key_disposable(&cursor) {
+                result += advance_by as u64;
+            }
+
+            cursor = cursor.add(advance_by);
+        }
+
+        if result > u32::MAX as u64 {
+            u32::MAX
+        } else {
+            result as u32
+        }
+    }
+
+    /// Advance the cursor to the next potential fragment boundary: this is either
+    /// a stripe boundary, or the end of the range.
+    fn distance_to_next_boundary(&self, cursor: Key) -> u32 {
+        let distance_to_range_end = contiguous_range_len(&(cursor..self.range.end));
+
+        if self.shard_identity.count < ShardCount::new(2) {
+            // Optimization: don't bother stepping through stripes if the tenant isn't sharded.
+            return distance_to_range_end;
+        }
+
+        if cursor.field6 == 0xffffffff {
+            // We are wrapping from one relation's logical size to the next relation's first data block
+            return 1;
+        }
+
+        let stripe_index = cursor.field6 / self.shard_identity.stripe_size.0;
+        let stripe_remainder = self.shard_identity.stripe_size.0
+            - (cursor.field6 - stripe_index * self.shard_identity.stripe_size.0);
+
+        if cfg!(debug_assertions) {
+            // We should never overflow field5 and field6 -- our callers check this earlier
+            // and would have returned their u32::MAX cases if the input range violated this.
+            let next_cursor = cursor.add(stripe_remainder);
+            debug_assert!(
+                next_cursor.field1 == cursor.field1
+                    && next_cursor.field2 == cursor.field2
+                    && next_cursor.field3 == cursor.field3
+                    && next_cursor.field4 == cursor.field4
+                    && next_cursor.field5 == cursor.field5
+            )
+        }
+
+        std::cmp::min(stripe_remainder, distance_to_range_end)
+    }
+
+    /// Whereas `page_count` estimates the number of pages physically in this range on this shard,
+    /// this function simply calculates the number of pages in the space, without accounting for those
+    /// pages that would not actually be stored on this node.
     ///
+    /// Don't use this function in code that works with physical entities like layer files.
+    fn raw_size(range: &Range<Key>) -> u32 {
+        if is_contiguous_range(range) {
+            contiguous_range_len(range)
+        } else {
+            u32::MAX
+        }
+    }
+}
+
+impl KeySpace {
+    /// Create a key space with a single range.
+    pub fn single(key_range: Range<Key>) -> Self {
+        Self {
+            ranges: vec![key_range],
+        }
+    }
+
     /// Partition a key space into roughly chunks of roughly 'target_size' bytes
     /// in each partition.
     ///
-    pub fn partition(&self, target_size: u64) -> KeyPartitioning {
+    pub fn partition(&self, shard_identity: &ShardIdentity, target_size: u64) -> KeyPartitioning {
         // Assume that each value is 8k in size.
-        let target_nblocks = (target_size / BLCKSZ as u64) as usize;
+        let target_nblocks = (target_size / BLCKSZ as u64) as u32;
 
         let mut parts = Vec::new();
         let mut current_part = Vec::new();
         let mut current_part_size: usize = 0;
         for range in &self.ranges {
-            // If appending the next contiguous range in the keyspace to the current
-            // partition would cause it to be too large, start a new partition.
-            let this_size = key_range_size(range) as usize;
-            if current_part_size + this_size > target_nblocks && !current_part.is_empty() {
-                parts.push(KeySpace {
-                    ranges: current_part,
-                });
-                current_part = Vec::new();
-                current_part_size = 0;
-            }
+            // While doing partitioning, wrap the range in ShardedRange so that our size calculations
+            // will respect shard striping rather than assuming all keys within a range are present.
+            let range = ShardedRange::new(range.clone(), shard_identity);
 
-            // If the next range is larger than 'target_size', split it into
-            // 'target_size' chunks.
-            let mut remain_size = this_size;
-            let mut start = range.start;
-            while remain_size > target_nblocks {
-                let next = start.add(target_nblocks as u32);
-                parts.push(KeySpace {
-                    ranges: vec![start..next],
-                });
-                start = next;
-                remain_size -= target_nblocks
+            // Chunk up the range into parts that each contain up to target_size local blocks
+            for (frag_on_shard_size, frag_range) in range.fragment(target_nblocks) {
+                // If appending the next contiguous range in the keyspace to the current
+                // partition would cause it to be too large, and our current partition
+                // covers at least one block that is physically present in this shard,
+                // then start a new partition
+                if current_part_size + frag_on_shard_size as usize > target_nblocks as usize
+                    && current_part_size > 0
+                {
+                    parts.push(KeySpace {
+                        ranges: current_part,
+                    });
+                    current_part = Vec::new();
+                    current_part_size = 0;
+                }
+                current_part.push(frag_range.start..frag_range.end);
+                current_part_size += frag_on_shard_size as usize;
             }
-            current_part.push(start..range.end);
-            current_part_size += remain_size;
         }
 
         // add last partition that wasn't full yet.
@@ -64,6 +302,10 @@ impl KeySpace {
         KeyPartitioning { parts }
     }
 
+    pub fn is_empty(&self) -> bool {
+        self.total_raw_size() == 0
+    }
+
     /// Merge another keyspace into the current one.
     /// Note: the keyspaces must not ovelap (enforced via assertions)
     pub fn merge(&mut self, other: &KeySpace) {
@@ -94,12 +336,13 @@ impl KeySpace {
 
     /// Remove all keys in `other` from `self`.
     /// This can involve splitting or removing of existing ranges.
-    pub fn remove_overlapping_with(&mut self, other: &KeySpace) {
+    /// Returns the removed keyspace
+    pub fn remove_overlapping_with(&mut self, other: &KeySpace) -> KeySpace {
         let (self_start, self_end) = match (self.start(), self.end()) {
             (Some(start), Some(end)) => (start, end),
             _ => {
                 // self is empty
-                return;
+                return KeySpace::default();
             }
         };
 
@@ -112,30 +355,37 @@ impl KeySpace {
             .skip_while(|range| self_start >= range.end)
             .take_while(|range| self_end > range.start);
 
+        let mut removed_accum = KeySpaceRandomAccum::new();
         for range in other_ranges {
             while let Some(overlap_at) = self.overlaps_at(range) {
                 let overlapped = self.ranges[overlap_at].clone();
 
                 if overlapped.start < range.start && overlapped.end <= range.end {
                     // Higher part of the range is completely overlapped.
+                    removed_accum.add_range(range.start..self.ranges[overlap_at].end);
                     self.ranges[overlap_at].end = range.start;
                 }
                 if overlapped.start >= range.start && overlapped.end > range.end {
                     // Lower part of the range is completely overlapped.
+                    removed_accum.add_range(self.ranges[overlap_at].start..range.end);
                     self.ranges[overlap_at].start = range.end;
                 }
                 if overlapped.start < range.start && overlapped.end > range.end {
                     // Middle part of the range is overlapped.
+                    removed_accum.add_range(range.clone());
                     self.ranges[overlap_at].end = range.start;
                     self.ranges
                         .insert(overlap_at + 1, range.end..overlapped.end);
                 }
                 if overlapped.start >= range.start && overlapped.end <= range.end {
                     // Whole range is overlapped
+                    removed_accum.add_range(self.ranges[overlap_at].clone());
                     self.ranges.remove(overlap_at);
                 }
             }
         }
+
+        removed_accum.to_keyspace()
     }
 
     pub fn start(&self) -> Option<Key> {
@@ -146,11 +396,11 @@ impl KeySpace {
         self.ranges.last().map(|range| range.end)
     }
 
-    #[allow(unused)]
-    pub fn total_size(&self) -> usize {
+    /// The size of the keyspace in pages, before accounting for sharding
+    pub fn total_raw_size(&self) -> usize {
         self.ranges
             .iter()
-            .map(|range| key_range_size(range) as usize)
+            .map(|range| ShardedRange::raw_size(range) as usize)
             .sum()
     }
 
@@ -170,6 +420,11 @@ impl KeySpace {
     pub fn overlaps(&self, range: &Range<Key>) -> bool {
         self.overlaps_at(range).is_some()
     }
+
+    /// Check if the keyspace contains a key
+    pub fn contains(&self, key: &Key) -> bool {
+        self.overlaps(&(*key..key.next()))
+    }
 }
 
 ///
@@ -184,10 +439,33 @@ pub struct KeyPartitioning {
     pub parts: Vec<KeySpace>,
 }
 
+/// Represents a partitioning of the sparse key space.
+#[derive(Clone, Debug, Default)]
+pub struct SparseKeyPartitioning {
+    pub parts: Vec<SparseKeySpace>,
+}
+
 impl KeyPartitioning {
     pub fn new() -> Self {
         KeyPartitioning { parts: Vec::new() }
     }
+
+    /// Convert a key partitioning to a sparse partition.
+    pub fn into_sparse(self) -> SparseKeyPartitioning {
+        SparseKeyPartitioning {
+            parts: self.parts.into_iter().map(SparseKeySpace).collect(),
+        }
+    }
+}
+
+impl SparseKeyPartitioning {
+    /// Note: use this function with caution. Attempt to handle a sparse keyspace in the same way as a dense keyspace will
+    /// cause long/dead loops.
+    pub fn into_dense(self) -> KeyPartitioning {
+        KeyPartitioning {
+            parts: self.parts.into_iter().map(|x| x.0).collect(),
+        }
+    }
 }
 
 ///
@@ -219,7 +497,7 @@ impl KeySpaceAccum {
 
     #[inline(always)]
     pub fn add_range(&mut self, range: Range<Key>) {
-        self.size += key_range_size(&range) as u64;
+        self.size += ShardedRange::raw_size(&range) as u64;
 
         match self.accum.as_mut() {
             Some(accum) => {
@@ -251,7 +529,9 @@ impl KeySpaceAccum {
         std::mem::take(self).to_keyspace()
     }
 
-    pub fn size(&self) -> u64 {
+    // The total number of keys in this object, ignoring any sharding effects that might cause some of
+    // the keys to be omitted in storage on this shard.
+    pub fn raw_size(&self) -> u64 {
         self.size
     }
 }
@@ -307,36 +587,19 @@ impl KeySpaceRandomAccum {
     }
 }
 
-#[inline(always)]
-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 rand::{RngCore, SeedableRng};
+
+    use crate::{
+        models::ShardParameters,
+        shard::{ShardCount, ShardNumber},
+    };
+
     use super::*;
     use std::fmt::Write;
 
@@ -379,14 +642,17 @@ mod tests {
             accum.add_range(range.clone());
         }
 
-        let expected_size: u64 = ranges.iter().map(|r| key_range_size(r) as u64).sum();
-        assert_eq!(accum.size(), expected_size);
+        let expected_size: u64 = ranges
+            .iter()
+            .map(|r| ShardedRange::raw_size(r) as u64)
+            .sum();
+        assert_eq!(accum.raw_size(), expected_size);
 
         assert_ks_eq(&accum.consume_keyspace(), ranges.clone());
-        assert_eq!(accum.size(), 0);
+        assert_eq!(accum.raw_size(), 0);
 
         assert_ks_eq(&accum.consume_keyspace(), vec![]);
-        assert_eq!(accum.size(), 0);
+        assert_eq!(accum.raw_size(), 0);
 
         for range in &ranges {
             accum.add_range(range.clone());
@@ -553,7 +819,16 @@ mod tests {
                 Key::from_i128(11)..Key::from_i128(13),
             ],
         };
-        key_space1.remove_overlapping_with(&key_space2);
+        let removed = key_space1.remove_overlapping_with(&key_space2);
+        let removed_expected = KeySpace {
+            ranges: vec![
+                Key::from_i128(2)..Key::from_i128(3),
+                Key::from_i128(6)..Key::from_i128(7),
+                Key::from_i128(11)..Key::from_i128(12),
+            ],
+        };
+        assert_eq!(removed, removed_expected);
+
         assert_eq!(
             key_space1.ranges,
             vec![
@@ -583,7 +858,17 @@ mod tests {
                 Key::from_i128(14)..Key::from_i128(17),
             ],
         };
-        key_space1.remove_overlapping_with(&key_space2);
+
+        let removed = key_space1.remove_overlapping_with(&key_space2);
+        let removed_expected = KeySpace {
+            ranges: vec![
+                Key::from_i128(3)..Key::from_i128(5),
+                Key::from_i128(8)..Key::from_i128(10),
+                Key::from_i128(14)..Key::from_i128(15),
+            ],
+        };
+        assert_eq!(removed, removed_expected);
+
         assert_eq!(
             key_space1.ranges,
             vec![
@@ -610,7 +895,11 @@ mod tests {
                 Key::from_i128(15)..Key::from_i128(17),
             ],
         };
-        key_space1.remove_overlapping_with(&key_space2);
+
+        let removed = key_space1.remove_overlapping_with(&key_space2);
+        let removed_expected = KeySpace::default();
+        assert_eq!(removed, removed_expected);
+
         assert_eq!(
             key_space1.ranges,
             vec![
@@ -637,7 +926,17 @@ mod tests {
         let key_space2 = KeySpace {
             ranges: vec![Key::from_i128(9)..Key::from_i128(19)],
         };
-        key_space1.remove_overlapping_with(&key_space2);
+
+        let removed = key_space1.remove_overlapping_with(&key_space2);
+        let removed_expected = KeySpace {
+            ranges: vec![
+                Key::from_i128(9)..Key::from_i128(10),
+                Key::from_i128(12)..Key::from_i128(15),
+                Key::from_i128(17)..Key::from_i128(19),
+            ],
+        };
+        assert_eq!(removed, removed_expected);
+
         assert_eq!(
             key_space1.ranges,
             vec![
@@ -650,4 +949,412 @@ mod tests {
             ]
         );
     }
+    #[test]
+    fn sharded_range_relation_gap() {
+        let shard_identity = ShardIdentity::new(
+            ShardNumber(0),
+            ShardCount::new(4),
+            ShardParameters::DEFAULT_STRIPE_SIZE,
+        )
+        .unwrap();
+
+        let range = ShardedRange::new(
+            Range {
+                start: Key::from_hex("000000067F00000005000040100300000000").unwrap(),
+                end: Key::from_hex("000000067F00000005000040130000004000").unwrap(),
+            },
+            &shard_identity,
+        );
+
+        // Key range spans relations, expect MAX
+        assert_eq!(range.page_count(), u32::MAX);
+    }
+
+    #[test]
+    fn shard_identity_keyspaces_single_key() {
+        let shard_identity = ShardIdentity::new(
+            ShardNumber(1),
+            ShardCount::new(4),
+            ShardParameters::DEFAULT_STRIPE_SIZE,
+        )
+        .unwrap();
+
+        let range = ShardedRange::new(
+            Range {
+                start: Key::from_hex("000000067f000000010000007000ffffffff").unwrap(),
+                end: Key::from_hex("000000067f00000001000000700100000000").unwrap(),
+            },
+            &shard_identity,
+        );
+        // Single-key range on logical size key
+        assert_eq!(range.page_count(), 1);
+    }
+
+    /// Test the helper that we use to identify ranges which go outside the data blocks of a single relation
+    #[test]
+    fn contiguous_range_check() {
+        assert!(!is_contiguous_range(
+            &(Key::from_hex("000000067f00000001000004df00fffffffe").unwrap()
+                ..Key::from_hex("000000067f00000001000004df0100000003").unwrap())
+        ),);
+
+        // The ranges goes all the way up to the 0xffffffff, including it: this is
+        // not considered a rel block range because 0xffffffff stores logical sizes,
+        // not blocks.
+        assert!(!is_contiguous_range(
+            &(Key::from_hex("000000067f00000001000004df00fffffffe").unwrap()
+                ..Key::from_hex("000000067f00000001000004df0100000000").unwrap())
+        ),);
+
+        // Keys within the normal data region of a relation
+        assert!(is_contiguous_range(
+            &(Key::from_hex("000000067f00000001000004df0000000000").unwrap()
+                ..Key::from_hex("000000067f00000001000004df0000000080").unwrap())
+        ),);
+
+        // The logical size key of one forkno, then some blocks in the next
+        assert!(is_contiguous_range(
+            &(Key::from_hex("000000067f00000001000004df00ffffffff").unwrap()
+                ..Key::from_hex("000000067f00000001000004df0100000080").unwrap())
+        ),);
+    }
+
+    #[test]
+    fn shard_identity_keyspaces_forkno_gap() {
+        let shard_identity = ShardIdentity::new(
+            ShardNumber(1),
+            ShardCount::new(4),
+            ShardParameters::DEFAULT_STRIPE_SIZE,
+        )
+        .unwrap();
+
+        let range = ShardedRange::new(
+            Range {
+                start: Key::from_hex("000000067f00000001000004df00fffffffe").unwrap(),
+                end: Key::from_hex("000000067f00000001000004df0100000003").unwrap(),
+            },
+            &shard_identity,
+        );
+
+        // Range spanning the end of one forkno and the start of the next: we do not attempt to
+        // calculate a valid size, because we have no way to know if they keys between start
+        // and end are actually in use.
+        assert_eq!(range.page_count(), u32::MAX);
+    }
+
+    #[test]
+    fn shard_identity_keyspaces_one_relation() {
+        for shard_number in 0..4 {
+            let shard_identity = ShardIdentity::new(
+                ShardNumber(shard_number),
+                ShardCount::new(4),
+                ShardParameters::DEFAULT_STRIPE_SIZE,
+            )
+            .unwrap();
+
+            let range = ShardedRange::new(
+                Range {
+                    start: Key::from_hex("000000067f00000001000000ae0000000000").unwrap(),
+                    end: Key::from_hex("000000067f00000001000000ae0000000001").unwrap(),
+                },
+                &shard_identity,
+            );
+
+            // Very simple case: range covering block zero of one relation, where that block maps to shard zero
+            if shard_number == 0 {
+                assert_eq!(range.page_count(), 1);
+            } else {
+                // Other shards should perceive the range's size as zero
+                assert_eq!(range.page_count(), 0);
+            }
+        }
+    }
+
+    /// Test helper: construct a ShardedRange and call fragment() on it, returning
+    /// the total page count in the range and the fragments.
+    fn do_fragment(
+        range_start: Key,
+        range_end: Key,
+        shard_identity: &ShardIdentity,
+        target_nblocks: u32,
+    ) -> (u32, Vec<(u32, Range<Key>)>) {
+        let range = ShardedRange::new(
+            Range {
+                start: range_start,
+                end: range_end,
+            },
+            shard_identity,
+        );
+
+        let page_count = range.page_count();
+        let fragments = range.fragment(target_nblocks);
+
+        // Invariant: we always get at least one fragment
+        assert!(!fragments.is_empty());
+
+        // Invariant: the first/last fragment start/end should equal the input start/end
+        assert_eq!(fragments.first().unwrap().1.start, range_start);
+        assert_eq!(fragments.last().unwrap().1.end, range_end);
+
+        if page_count > 0 {
+            // Invariant: every fragment must contain at least one shard-local page, if the
+            // total range contains at least one shard-local page
+            let all_nonzero = fragments.iter().all(|f| f.0 > 0);
+            if !all_nonzero {
+                eprintln!("Found a zero-length fragment: {:?}", fragments);
+            }
+            assert!(all_nonzero);
+        } else {
+            // A range with no shard-local pages should always be returned as a single fragment
+            assert_eq!(fragments, vec![(0, range_start..range_end)]);
+        }
+
+        // Invariant: fragments must be ordered and non-overlapping
+        let mut last: Option<Range<Key>> = None;
+        for frag in &fragments {
+            if let Some(last) = last {
+                assert!(frag.1.start >= last.end);
+                assert!(frag.1.start > last.start);
+            }
+            last = Some(frag.1.clone())
+        }
+
+        // Invariant: fragments respect target_nblocks
+        for frag in &fragments {
+            assert!(frag.0 == u32::MAX || frag.0 <= target_nblocks);
+        }
+
+        (page_count, fragments)
+    }
+
+    /// Really simple tests for fragment(), on a range that just contains a single stripe
+    /// for a single tenant.
+    #[test]
+    fn sharded_range_fragment_simple() {
+        let shard_identity = ShardIdentity::new(
+            ShardNumber(0),
+            ShardCount::new(4),
+            ShardParameters::DEFAULT_STRIPE_SIZE,
+        )
+        .unwrap();
+
+        // A range which we happen to know covers exactly one stripe which belongs to this shard
+        let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
+        let input_end = Key::from_hex("000000067f00000001000000ae0000008000").unwrap();
+
+        // Ask for stripe_size blocks, we get the whole stripe
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 32768),
+            (32768, vec![(32768, input_start..input_end)])
+        );
+
+        // Ask for more, we still get the whole stripe
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 10000000),
+            (32768, vec![(32768, input_start..input_end)])
+        );
+
+        // Ask for target_nblocks of half the stripe size, we get two halves
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 16384),
+            (
+                32768,
+                vec![
+                    (16384, input_start..input_start.add(16384)),
+                    (16384, input_start.add(16384)..input_end)
+                ]
+            )
+        );
+    }
+
+    #[test]
+    fn sharded_range_fragment_multi_stripe() {
+        let shard_identity = ShardIdentity::new(
+            ShardNumber(0),
+            ShardCount::new(4),
+            ShardParameters::DEFAULT_STRIPE_SIZE,
+        )
+        .unwrap();
+
+        // A range which covers multiple stripes, exactly one of which belongs to the current shard.
+        let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
+        let input_end = Key::from_hex("000000067f00000001000000ae0000020000").unwrap();
+        // Ask for all the blocks, get a fragment that covers the whole range but reports
+        // its size to be just the blocks belonging to our shard.
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 131072),
+            (32768, vec![(32768, input_start..input_end)])
+        );
+
+        // Ask for a sub-stripe quantity
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 16000),
+            (
+                32768,
+                vec![
+                    (16000, input_start..input_start.add(16000)),
+                    (16000, input_start.add(16000)..input_start.add(32000)),
+                    (768, input_start.add(32000)..input_end),
+                ]
+            )
+        );
+
+        // Try on a range that starts slightly after our owned stripe
+        assert_eq!(
+            do_fragment(input_start.add(1), input_end, &shard_identity, 131072),
+            (32767, vec![(32767, input_start.add(1)..input_end)])
+        );
+    }
+
+    /// Test our calculations work correctly when we start a range from the logical size key of
+    /// a previous relation.
+    #[test]
+    fn sharded_range_fragment_starting_from_logical_size() {
+        let input_start = Key::from_hex("000000067f00000001000000ae00ffffffff").unwrap();
+        let input_end = Key::from_hex("000000067f00000001000000ae0100008000").unwrap();
+
+        // Shard 0 owns the first stripe in the relation, and the preceding logical size is shard local too
+        let shard_identity = ShardIdentity::new(
+            ShardNumber(0),
+            ShardCount::new(4),
+            ShardParameters::DEFAULT_STRIPE_SIZE,
+        )
+        .unwrap();
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 0x10000),
+            (0x8001, vec![(0x8001, input_start..input_end)])
+        );
+
+        // Shard 1 does not own the first stripe in the relation, but it does own the logical size (all shards
+        // store all logical sizes)
+        let shard_identity = ShardIdentity::new(
+            ShardNumber(1),
+            ShardCount::new(4),
+            ShardParameters::DEFAULT_STRIPE_SIZE,
+        )
+        .unwrap();
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 0x10000),
+            (0x1, vec![(0x1, input_start..input_end)])
+        );
+    }
+
+    /// Test that ShardedRange behaves properly when used on un-sharded data
+    #[test]
+    fn sharded_range_fragment_unsharded() {
+        let shard_identity = ShardIdentity::unsharded();
+
+        let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
+        let input_end = Key::from_hex("000000067f00000001000000ae0000010000").unwrap();
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 0x8000),
+            (
+                0x10000,
+                vec![
+                    (0x8000, input_start..input_start.add(0x8000)),
+                    (0x8000, input_start.add(0x8000)..input_start.add(0x10000))
+                ]
+            )
+        );
+    }
+
+    #[test]
+    fn sharded_range_fragment_cross_relation() {
+        let shard_identity = ShardIdentity::unsharded();
+
+        // A range that spans relations: expect fragmentation to give up and return a u32::MAX size
+        let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
+        let input_end = Key::from_hex("000000068f00000001000000ae0000010000").unwrap();
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 0x8000),
+            (u32::MAX, vec![(u32::MAX, input_start..input_end),])
+        );
+
+        // Same, but using a sharded identity
+        let shard_identity = ShardIdentity::new(
+            ShardNumber(0),
+            ShardCount::new(4),
+            ShardParameters::DEFAULT_STRIPE_SIZE,
+        )
+        .unwrap();
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 0x8000),
+            (u32::MAX, vec![(u32::MAX, input_start..input_end),])
+        );
+    }
+
+    #[test]
+    fn sharded_range_fragment_tiny_nblocks() {
+        let shard_identity = ShardIdentity::unsharded();
+
+        // A range that spans relations: expect fragmentation to give up and return a u32::MAX size
+        let input_start = Key::from_hex("000000067F00000001000004E10000000000").unwrap();
+        let input_end = Key::from_hex("000000067F00000001000004E10000000038").unwrap();
+        assert_eq!(
+            do_fragment(input_start, input_end, &shard_identity, 16),
+            (
+                0x38,
+                vec![
+                    (16, input_start..input_start.add(16)),
+                    (16, input_start.add(16)..input_start.add(32)),
+                    (16, input_start.add(32)..input_start.add(48)),
+                    (8, input_start.add(48)..input_end),
+                ]
+            )
+        );
+    }
+
+    #[test]
+    fn sharded_range_fragment_fuzz() {
+        // Use a fixed seed: we don't want to explicitly pick values, but we do want
+        // the test to be reproducible.
+        let mut prng = rand::rngs::StdRng::seed_from_u64(0xdeadbeef);
+
+        for _i in 0..1000 {
+            let shard_identity = if prng.next_u32() % 2 == 0 {
+                ShardIdentity::unsharded()
+            } else {
+                let shard_count = prng.next_u32() % 127 + 1;
+                ShardIdentity::new(
+                    ShardNumber((prng.next_u32() % shard_count) as u8),
+                    ShardCount::new(shard_count as u8),
+                    ShardParameters::DEFAULT_STRIPE_SIZE,
+                )
+                .unwrap()
+            };
+
+            let target_nblocks = prng.next_u32() % 65536 + 1;
+
+            let start_offset = prng.next_u32() % 16384;
+
+            // Try ranges up to 4GiB in size, that are always at least 1
+            let range_size = prng.next_u32() % 8192 + 1;
+
+            // A range that spans relations: expect fragmentation to give up and return a u32::MAX size
+            let input_start = Key::from_hex("000000067F00000001000004E10000000000")
+                .unwrap()
+                .add(start_offset);
+            let input_end = input_start.add(range_size);
+
+            // This test's main success conditions are the invariants baked into do_fragment
+            let (_total_size, fragments) =
+                do_fragment(input_start, input_end, &shard_identity, target_nblocks);
+
+            // Pick a random key within the range and check it appears in the output
+            let example_key = input_start.add(prng.next_u32() % range_size);
+
+            // Panic on unwrap if it isn't found
+            let example_key_frag = fragments
+                .iter()
+                .find(|f| f.1.contains(&example_key))
+                .unwrap();
+
+            // Check that the fragment containing our random key has a nonzero size if
+            // that key is shard-local
+            let example_key_local = !shard_identity.is_key_disposable(&example_key);
+            if example_key_local {
+                assert!(example_key_frag.0 > 0);
+            }
+        }
+    }
 }

libs/pageserver_api/src/lib.rs

@@ -1,6 +1,5 @@
 #![deny(unsafe_code)]
 #![deny(clippy::undocumented_unsafe_blocks)]
-use const_format::formatcp;
 
 pub mod controller_api;
 pub mod key;
@@ -11,7 +10,4 @@ pub mod shard;
 /// Public API types
 pub mod upcall_api;
 
-pub const DEFAULT_PG_LISTEN_PORT: u16 = 64000;
-pub const DEFAULT_PG_LISTEN_ADDR: &str = formatcp!("127.0.0.1:{DEFAULT_PG_LISTEN_PORT}");
-pub const DEFAULT_HTTP_LISTEN_PORT: u16 = 9898;
-pub const DEFAULT_HTTP_LISTEN_ADDR: &str = formatcp!("127.0.0.1:{DEFAULT_HTTP_LISTEN_PORT}");
+pub mod config;

libs/pageserver_api/src/models.rs

@@ -4,6 +4,7 @@ pub mod utilization;
 pub use utilization::PageserverUtilization;
 
 use std::{
+    borrow::Cow,
     collections::HashMap,
     io::{BufRead, Read},
     num::{NonZeroU64, NonZeroUsize},
@@ -19,6 +20,7 @@ use utils::{
     history_buffer::HistoryBufferWithDropCounter,
     id::{NodeId, TenantId, TimelineId},
     lsn::Lsn,
+    serde_system_time,
 };
 
 use crate::controller_api::PlacementPolicy;
@@ -294,13 +296,14 @@ pub struct TenantConfig {
     pub lagging_wal_timeout: Option<String>,
     pub max_lsn_wal_lag: Option<NonZeroU64>,
     pub trace_read_requests: Option<bool>,
-    pub image_layer_compression: Option<CompressionAlgorithm>,
     pub eviction_policy: Option<EvictionPolicy>,
     pub min_resident_size_override: Option<u64>,
     pub evictions_low_residence_duration_metric_threshold: Option<String>,
     pub heatmap_period: Option<String>,
     pub lazy_slru_download: Option<bool>,
     pub timeline_get_throttle: Option<ThrottleConfig>,
+    pub image_layer_creation_check_threshold: Option<u8>,
+    pub switch_to_aux_file_v2: Option<bool>,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
@@ -328,23 +331,6 @@ pub enum CompactionAlgorithm {
     Tiered,
 }
 
-#[derive(
-    Debug,
-    Clone,
-    Copy,
-    PartialEq,
-    Eq,
-    Serialize,
-    Deserialize,
-    strum_macros::FromRepr,
-    enum_map::Enum,
-)]
-#[repr(u8)]
-pub enum CompressionAlgorithm {
-    NoCompression,
-    LZ4,
-}
-
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 pub struct EvictionPolicyLayerAccessThreshold {
     #[serde(with = "humantime_serde")]
@@ -444,7 +430,6 @@ pub struct StatusResponse {
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct TenantLocationConfigRequest {
-    pub tenant_id: Option<TenantShardId>,
     #[serde(flatten)]
     pub config: LocationConfig, // as we have a flattened field, we should reject all unknown fields in it
 }
@@ -595,7 +580,7 @@ pub struct TimelineInfo {
     pub walreceiver_status: String,
 }
 
-#[derive(Debug, Clone, Serialize)]
+#[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct LayerMapInfo {
     pub in_memory_layers: Vec<InMemoryLayerInfo>,
     pub historic_layers: Vec<HistoricLayerInfo>,
@@ -613,7 +598,7 @@ pub enum LayerAccessKind {
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct LayerAccessStatFullDetails {
     pub when_millis_since_epoch: u64,
-    pub task_kind: &'static str,
+    pub task_kind: Cow<'static, str>,
     pub access_kind: LayerAccessKind,
 }
 
@@ -672,23 +657,23 @@ impl LayerResidenceEvent {
     }
 }
 
-#[derive(Debug, Clone, Serialize)]
+#[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct LayerAccessStats {
     pub access_count_by_access_kind: HashMap<LayerAccessKind, u64>,
-    pub task_kind_access_flag: Vec<&'static str>,
+    pub task_kind_access_flag: Vec<Cow<'static, str>>,
     pub first: Option<LayerAccessStatFullDetails>,
     pub accesses_history: HistoryBufferWithDropCounter<LayerAccessStatFullDetails, 16>,
     pub residence_events_history: HistoryBufferWithDropCounter<LayerResidenceEvent, 16>,
 }
 
-#[derive(Debug, Clone, Serialize)]
+#[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(tag = "kind")]
 pub enum InMemoryLayerInfo {
     Open { lsn_start: Lsn },
     Frozen { lsn_start: Lsn, lsn_end: Lsn },
 }
 
-#[derive(Debug, Clone, Serialize)]
+#[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(tag = "kind")]
 pub enum HistoricLayerInfo {
     Delta {
@@ -710,6 +695,32 @@ pub enum HistoricLayerInfo {
     },
 }
 
+impl HistoricLayerInfo {
+    pub fn layer_file_name(&self) -> &str {
+        match self {
+            HistoricLayerInfo::Delta {
+                layer_file_name, ..
+            } => layer_file_name,
+            HistoricLayerInfo::Image {
+                layer_file_name, ..
+            } => layer_file_name,
+        }
+    }
+    pub fn is_remote(&self) -> bool {
+        match self {
+            HistoricLayerInfo::Delta { remote, .. } => *remote,
+            HistoricLayerInfo::Image { remote, .. } => *remote,
+        }
+    }
+    pub fn set_remote(&mut self, value: bool) {
+        let field = match self {
+            HistoricLayerInfo::Delta { remote, .. } => remote,
+            HistoricLayerInfo::Image { remote, .. } => remote,
+        };
+        *field = value;
+    }
+}
+
 #[derive(Debug, Serialize, Deserialize)]
 pub struct DownloadRemoteLayersTaskSpawnRequest {
     pub max_concurrent_downloads: NonZeroUsize,
@@ -736,10 +747,48 @@ pub struct TimelineGcRequest {
     pub gc_horizon: Option<u64>,
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct WalRedoManagerProcessStatus {
+    pub pid: u32,
+    /// The strum-generated `into::<&'static str>()` for `pageserver::walredo::ProcessKind`.
+    /// `ProcessKind` are a transitory thing, so, they have no enum representation in `pageserver_api`.
+    pub kind: Cow<'static, str>,
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct WalRedoManagerStatus {
     pub last_redo_at: Option<chrono::DateTime<chrono::Utc>>,
-    pub pid: Option<u32>,
+    pub process: Option<WalRedoManagerProcessStatus>,
+}
+
+/// The progress of a secondary tenant is mostly useful when doing a long running download: e.g. initiating
+/// a download job, timing out while waiting for it to run, and then inspecting this status to understand
+/// what's happening.
+#[derive(Default, Debug, Serialize, Deserialize, Clone)]
+pub struct SecondaryProgress {
+    /// The remote storage LastModified time of the heatmap object we last downloaded.
+    pub heatmap_mtime: Option<serde_system_time::SystemTime>,
+
+    /// The number of layers currently on-disk
+    pub layers_downloaded: usize,
+    /// The number of layers in the most recently seen heatmap
+    pub layers_total: usize,
+
+    /// The number of layer bytes currently on-disk
+    pub bytes_downloaded: u64,
+    /// The number of layer bytes in the most recently seen heatmap
+    pub bytes_total: u64,
+}
+
+#[derive(Serialize, Deserialize, Debug)]
+pub struct TenantScanRemoteStorageShard {
+    pub tenant_shard_id: TenantShardId,
+    pub generation: Option<u32>,
+}
+
+#[derive(Serialize, Deserialize, Debug, Default)]
+pub struct TenantScanRemoteStorageResponse {
+    pub shards: Vec<TenantScanRemoteStorageShard>,
 }
 
 pub mod virtual_file {
@@ -809,39 +858,72 @@ impl TryFrom<u8> for PagestreamBeMessageTag {
     }
 }
 
+// In the V2 protocol version, a GetPage request contains two LSN values:
+//
+// request_lsn: Get the page version at this point in time.  Lsn::Max is a special value that means
+// "get the latest version present". It's used by the primary server, which knows that no one else
+// is writing WAL. 'not_modified_since' must be set to a proper value even if request_lsn is
+// Lsn::Max. Standby servers use the current replay LSN as the request LSN.
+//
+// not_modified_since: Hint to the pageserver that the client knows that the page has not been
+// modified between 'not_modified_since' and the request LSN. It's always correct to set
+// 'not_modified_since equal' to 'request_lsn' (unless Lsn::Max is used as the 'request_lsn'), but
+// passing an earlier LSN can speed up the request, by allowing the pageserver to process the
+// request without waiting for 'request_lsn' to arrive.
+//
+// The legacy V1 interface contained only one LSN, and a boolean 'latest' flag. The V1 interface was
+// sufficient for the primary; the 'lsn' was equivalent to the 'not_modified_since' value, and
+// 'latest' was set to true. The V2 interface was added because there was no correct way for a
+// standby to request a page at a particular non-latest LSN, and also include the
+// 'not_modified_since' hint. That led to an awkward choice of either using an old LSN in the
+// request, if the standby knows that the page hasn't been modified since, and risk getting an error
+// if that LSN has fallen behind the GC horizon, or requesting the current replay LSN, which could
+// require the pageserver unnecessarily to wait for the WAL to arrive up to that point. The new V2
+// interface allows sending both LSNs, and let the pageserver do the right thing. There is no
+// difference in the responses between V1 and V2.
+//
+// The Request structs below reflect the V2 interface. If V1 is used, the parse function
+// maps the old format requests to the new format.
+//
+#[derive(Clone, Copy)]
+pub enum PagestreamProtocolVersion {
+    V1,
+    V2,
+}
+
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamExistsRequest {
-    pub latest: bool,
-    pub lsn: Lsn,
+    pub request_lsn: Lsn,
+    pub not_modified_since: Lsn,
     pub rel: RelTag,
 }
 
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamNblocksRequest {
-    pub latest: bool,
-    pub lsn: Lsn,
+    pub request_lsn: Lsn,
+    pub not_modified_since: Lsn,
     pub rel: RelTag,
 }
 
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamGetPageRequest {
-    pub latest: bool,
-    pub lsn: Lsn,
+    pub request_lsn: Lsn,
+    pub not_modified_since: Lsn,
     pub rel: RelTag,
     pub blkno: u32,
 }
 
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamDbSizeRequest {
-    pub latest: bool,
-    pub lsn: Lsn,
+    pub request_lsn: Lsn,
+    pub not_modified_since: Lsn,
     pub dbnode: u32,
 }
 
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamGetSlruSegmentRequest {
-    pub latest: bool,
-    pub lsn: Lsn,
+    pub request_lsn: Lsn,
+    pub not_modified_since: Lsn,
     pub kind: u8,
     pub segno: u32,
 }
@@ -888,14 +970,16 @@ pub struct TenantHistorySize {
 }
 
 impl PagestreamFeMessage {
+    /// Serialize a compute -> pageserver message. This is currently only used in testing
+    /// tools. Always uses protocol version 2.
     pub fn serialize(&self) -> Bytes {
         let mut bytes = BytesMut::new();
 
         match self {
             Self::Exists(req) => {
                 bytes.put_u8(0);
-                bytes.put_u8(u8::from(req.latest));
-                bytes.put_u64(req.lsn.0);
+                bytes.put_u64(req.request_lsn.0);
+                bytes.put_u64(req.not_modified_since.0);
                 bytes.put_u32(req.rel.spcnode);
                 bytes.put_u32(req.rel.dbnode);
                 bytes.put_u32(req.rel.relnode);
@@ -904,8 +988,8 @@ impl PagestreamFeMessage {
 
             Self::Nblocks(req) => {
                 bytes.put_u8(1);
-                bytes.put_u8(u8::from(req.latest));
-                bytes.put_u64(req.lsn.0);
+                bytes.put_u64(req.request_lsn.0);
+                bytes.put_u64(req.not_modified_since.0);
                 bytes.put_u32(req.rel.spcnode);
                 bytes.put_u32(req.rel.dbnode);
                 bytes.put_u32(req.rel.relnode);
@@ -914,8 +998,8 @@ impl PagestreamFeMessage {
 
             Self::GetPage(req) => {
                 bytes.put_u8(2);
-                bytes.put_u8(u8::from(req.latest));
-                bytes.put_u64(req.lsn.0);
+                bytes.put_u64(req.request_lsn.0);
+                bytes.put_u64(req.not_modified_since.0);
                 bytes.put_u32(req.rel.spcnode);
                 bytes.put_u32(req.rel.dbnode);
                 bytes.put_u32(req.rel.relnode);
@@ -925,15 +1009,15 @@ impl PagestreamFeMessage {
 
             Self::DbSize(req) => {
                 bytes.put_u8(3);
-                bytes.put_u8(u8::from(req.latest));
-                bytes.put_u64(req.lsn.0);
+                bytes.put_u64(req.request_lsn.0);
+                bytes.put_u64(req.not_modified_since.0);
                 bytes.put_u32(req.dbnode);
             }
 
             Self::GetSlruSegment(req) => {
                 bytes.put_u8(4);
-                bytes.put_u8(u8::from(req.latest));
-                bytes.put_u64(req.lsn.0);
+                bytes.put_u64(req.request_lsn.0);
+                bytes.put_u64(req.not_modified_since.0);
                 bytes.put_u8(req.kind);
                 bytes.put_u32(req.segno);
             }
@@ -942,18 +1026,40 @@ impl PagestreamFeMessage {
         bytes.into()
     }
 
-    pub fn parse<R: std::io::Read>(body: &mut R) -> anyhow::Result<PagestreamFeMessage> {
-        // TODO these gets can fail
-
+    pub fn parse<R: std::io::Read>(
+        body: &mut R,
+        protocol_version: PagestreamProtocolVersion,
+    ) -> anyhow::Result<PagestreamFeMessage> {
         // these correspond to the NeonMessageTag enum in pagestore_client.h
         //
         // TODO: consider using protobuf or serde bincode for less error prone
         // serialization.
         let msg_tag = body.read_u8()?;
+
+        let (request_lsn, not_modified_since) = match protocol_version {
+            PagestreamProtocolVersion::V2 => (
+                Lsn::from(body.read_u64::<BigEndian>()?),
+                Lsn::from(body.read_u64::<BigEndian>()?),
+            ),
+            PagestreamProtocolVersion::V1 => {
+                // In the old protocol, each message starts with a boolean 'latest' flag,
+                // followed by 'lsn'. Convert that to the two LSNs, 'request_lsn' and
+                // 'not_modified_since', used in the new protocol version.
+                let latest = body.read_u8()? != 0;
+                let request_lsn = Lsn::from(body.read_u64::<BigEndian>()?);
+                if latest {
+                    (Lsn::MAX, request_lsn) // get latest version
+                } else {
+                    (request_lsn, request_lsn) // get version at specified LSN
+                }
+            }
+        };
+
+        // The rest of the messages are the same between V1 and V2
         match msg_tag {
             0 => Ok(PagestreamFeMessage::Exists(PagestreamExistsRequest {
-                latest: body.read_u8()? != 0,
-                lsn: Lsn::from(body.read_u64::<BigEndian>()?),
+                request_lsn,
+                not_modified_since,
                 rel: RelTag {
                     spcnode: body.read_u32::<BigEndian>()?,
                     dbnode: body.read_u32::<BigEndian>()?,
@@ -962,8 +1068,8 @@ impl PagestreamFeMessage {
                 },
             })),
             1 => Ok(PagestreamFeMessage::Nblocks(PagestreamNblocksRequest {
-                latest: body.read_u8()? != 0,
-                lsn: Lsn::from(body.read_u64::<BigEndian>()?),
+                request_lsn,
+                not_modified_since,
                 rel: RelTag {
                     spcnode: body.read_u32::<BigEndian>()?,
                     dbnode: body.read_u32::<BigEndian>()?,
@@ -972,8 +1078,8 @@ impl PagestreamFeMessage {
                 },
             })),
             2 => Ok(PagestreamFeMessage::GetPage(PagestreamGetPageRequest {
-                latest: body.read_u8()? != 0,
-                lsn: Lsn::from(body.read_u64::<BigEndian>()?),
+                request_lsn,
+                not_modified_since,
                 rel: RelTag {
                     spcnode: body.read_u32::<BigEndian>()?,
                     dbnode: body.read_u32::<BigEndian>()?,
@@ -983,14 +1089,14 @@ impl PagestreamFeMessage {
                 blkno: body.read_u32::<BigEndian>()?,
             })),
             3 => Ok(PagestreamFeMessage::DbSize(PagestreamDbSizeRequest {
-                latest: body.read_u8()? != 0,
-                lsn: Lsn::from(body.read_u64::<BigEndian>()?),
+                request_lsn,
+                not_modified_since,
                 dbnode: body.read_u32::<BigEndian>()?,
             })),
             4 => Ok(PagestreamFeMessage::GetSlruSegment(
                 PagestreamGetSlruSegmentRequest {
-                    latest: body.read_u8()? != 0,
-                    lsn: Lsn::from(body.read_u64::<BigEndian>()?),
+                    request_lsn,
+                    not_modified_since,
                     kind: body.read_u8()?,
                     segno: body.read_u32::<BigEndian>()?,
                 },
@@ -1118,8 +1224,8 @@ mod tests {
         // Test serialization/deserialization of PagestreamFeMessage
         let messages = vec![
             PagestreamFeMessage::Exists(PagestreamExistsRequest {
-                latest: true,
-                lsn: Lsn(4),
+                request_lsn: Lsn(4),
+                not_modified_since: Lsn(3),
                 rel: RelTag {
                     forknum: 1,
                     spcnode: 2,
@@ -1128,8 +1234,8 @@ mod tests {
                 },
             }),
             PagestreamFeMessage::Nblocks(PagestreamNblocksRequest {
-                latest: false,
-                lsn: Lsn(4),
+                request_lsn: Lsn(4),
+                not_modified_since: Lsn(4),
                 rel: RelTag {
                     forknum: 1,
                     spcnode: 2,
@@ -1138,8 +1244,8 @@ mod tests {
                 },
             }),
             PagestreamFeMessage::GetPage(PagestreamGetPageRequest {
-                latest: true,
-                lsn: Lsn(4),
+                request_lsn: Lsn(4),
+                not_modified_since: Lsn(3),
                 rel: RelTag {
                     forknum: 1,
                     spcnode: 2,
@@ -1149,14 +1255,16 @@ mod tests {
                 blkno: 7,
             }),
             PagestreamFeMessage::DbSize(PagestreamDbSizeRequest {
-                latest: true,
-                lsn: Lsn(4),
+                request_lsn: Lsn(4),
+                not_modified_since: Lsn(3),
                 dbnode: 7,
             }),
         ];
         for msg in messages {
             let bytes = msg.serialize();
-            let reconstructed = PagestreamFeMessage::parse(&mut bytes.reader()).unwrap();
+            let reconstructed =
+                PagestreamFeMessage::parse(&mut bytes.reader(), PagestreamProtocolVersion::V2)
+                    .unwrap();
             assert!(msg == reconstructed);
         }
     }

libs/pageserver_api/src/models/partitioning.rs

@@ -1,9 +1,11 @@
 use utils::lsn::Lsn;
 
+use crate::keyspace::SparseKeySpace;
+
 #[derive(Debug, PartialEq, Eq)]
 pub struct Partitioning {
     pub keys: crate::keyspace::KeySpace,
-
+    pub sparse_keys: crate::keyspace::SparseKeySpace,
     pub at_lsn: Lsn,
 }
 
@@ -32,6 +34,8 @@ impl serde::Serialize for Partitioning {
         let mut map = serializer.serialize_map(Some(2))?;
         map.serialize_key("keys")?;
         map.serialize_value(&KeySpace(&self.keys))?;
+        map.serialize_key("sparse_keys")?;
+        map.serialize_value(&KeySpace(&self.sparse_keys.0))?;
         map.serialize_key("at_lsn")?;
         map.serialize_value(&WithDisplay(&self.at_lsn))?;
         map.end()
@@ -99,6 +103,7 @@ impl<'a> serde::Deserialize<'a> for Partitioning {
         #[derive(serde::Deserialize)]
         struct De {
             keys: KeySpace,
+            sparse_keys: KeySpace,
             #[serde_as(as = "serde_with::DisplayFromStr")]
             at_lsn: Lsn,
         }
@@ -107,6 +112,7 @@ impl<'a> serde::Deserialize<'a> for Partitioning {
         Ok(Self {
             at_lsn: de.at_lsn,
             keys: de.keys.0,
+            sparse_keys: SparseKeySpace(de.sparse_keys.0),
         })
     }
 }
@@ -133,6 +139,12 @@ mod tests {
                 "030000000000000000000000000000000003"
               ]
             ],
+            "sparse_keys": [
+              [
+                "620000000000000000000000000000000000",
+                "620000000000000000000000000000000003"
+              ]
+            ],
             "at_lsn": "0/2240160"
         }
         "#;

libs/pageserver_api/src/models/utilization.rs

@@ -1,4 +1,4 @@
-use std::time::SystemTime;
+use utils::serde_system_time::SystemTime;
 
 /// Pageserver current utilization and scoring for how good candidate the pageserver would be for
 /// the next tenant.
@@ -7,7 +7,7 @@ use std::time::SystemTime;
 ///
 /// `format: int64` fields must use `ser_saturating_u63` because openapi generated clients might
 /// not handle full u64 values properly.
-#[derive(serde::Serialize, Debug)]
+#[derive(serde::Serialize, serde::Deserialize, Debug, Clone)]
 pub struct PageserverUtilization {
     /// Used disk space
     #[serde(serialize_with = "ser_saturating_u63")]
@@ -21,17 +21,9 @@ pub struct PageserverUtilization {
     /// When was this snapshot captured, pageserver local time.
     ///
     /// Use millis to give confidence that the value is regenerated often enough.
-    #[serde(serialize_with = "ser_rfc3339_millis")]
     pub captured_at: SystemTime,
 }
 
-fn ser_rfc3339_millis<S: serde::Serializer>(
-    ts: &SystemTime,
-    serializer: S,
-) -> Result<S::Ok, S::Error> {
-    serializer.collect_str(&humantime::format_rfc3339_millis(*ts))
-}
-
 /// openapi knows only `format: int64`, so avoid outputting a non-parseable value by generated clients.
 ///
 /// Instead of newtype, use this because a newtype would get require handling deserializing values
@@ -58,7 +50,9 @@ mod tests {
             disk_usage_bytes: u64::MAX,
             free_space_bytes: 0,
             utilization_score: u64::MAX,
-            captured_at: SystemTime::UNIX_EPOCH + Duration::from_secs(1708509779),
+            captured_at: SystemTime(
+                std::time::SystemTime::UNIX_EPOCH + Duration::from_secs(1708509779),
+            ),
         };
 
         let s = serde_json::to_string(&doc).unwrap();

libs/pageserver_api/src/shard.rs

@@ -5,21 +5,99 @@ use crate::{
     models::ShardParameters,
 };
 use hex::FromHex;
+use postgres_ffi::relfile_utils::INIT_FORKNUM;
 use serde::{Deserialize, Serialize};
 use utils::id::TenantId;
 
+/// See docs/rfcs/031-sharding-static.md for an overview of sharding.
+///
+/// This module contains a variety of types used to represent the concept of sharding
+/// a Neon tenant across multiple physical shards.  Since there are quite a few of these,
+/// we provide an summary here.
+///
+/// Types used to describe shards:
+/// - [`ShardCount`] describes how many shards make up a tenant, plus the magic `unsharded` value
+///   which identifies a tenant which is not shard-aware.  This means its storage paths do not include
+///   a shard suffix.
+/// - [`ShardNumber`] is simply the zero-based index of a shard within a tenant.
+/// - [`ShardIndex`] is the 2-tuple of `ShardCount` and `ShardNumber`, it's just like a `TenantShardId`
+///   without the tenant ID.  This is useful for things that are implicitly scoped to a particular
+///   tenant, such as layer files.
+/// - [`ShardIdentity`]` is the full description of a particular shard's parameters, in sufficient
+///   detail to convert a [`Key`] to a [`ShardNumber`] when deciding where to write/read.
+/// - The [`ShardSlug`] is a terse formatter for ShardCount and ShardNumber, written as
+///   four hex digits.  An unsharded tenant is `0000`.
+/// - [`TenantShardId`] is the unique ID of a particular shard within a particular tenant
+///
+/// Types used to describe the parameters for data distribution in a sharded tenant:
+/// - [`ShardStripeSize`] controls how long contiguous runs of [`Key`]s (stripes) are when distributed across
+///   multiple shards.  Its value is given in 8kiB pages.
+/// - [`ShardLayout`] describes the data distribution scheme, and at time of writing is
+///   always zero: this is provided for future upgrades that might introduce different
+///   data distribution schemes.
+///
+/// Examples:
+/// - A legacy unsharded tenant has one shard with ShardCount(0), ShardNumber(0), and its slug is 0000
+/// - A single sharded tenant has one shard with ShardCount(1), ShardNumber(0), and its slug is 0001
+/// - In a tenant with 4 shards, each shard has ShardCount(N), ShardNumber(i) where i in 0..N-1 (inclusive),
+///   and their slugs are 0004, 0104, 0204, and 0304.
+
 #[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug, Hash)]
 pub struct ShardNumber(pub u8);
 
 #[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug, Hash)]
 pub struct ShardCount(u8);
 
+/// Combination of ShardNumber and ShardCount.  For use within the context of a particular tenant,
+/// when we need to know which 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, Hash)]
+pub struct ShardIndex {
+    pub shard_number: ShardNumber,
+    pub shard_count: ShardCount,
+}
+
+/// The ShardIdentity contains enough information to map a [`Key`] to a [`ShardNumber`],
+/// and to check whether that [`ShardNumber`] is the same as the current shard.
+#[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Debug)]
+pub struct ShardIdentity {
+    pub number: ShardNumber,
+    pub count: ShardCount,
+    pub stripe_size: ShardStripeSize,
+    layout: ShardLayout,
+}
+
+/// Formatting helper, for generating the `shard_id` label in traces.
+struct ShardSlug<'a>(&'a TenantShardId);
+
+/// TenantShardId globally identifies a particular shard in a particular tenant.
+///
+/// These are written as `<TenantId>-<ShardSlug>`, for example:
+///   # The second shard in a two-shard tenant
+///   072f1291a5310026820b2fe4b2968934-0102
+///
+/// If the `ShardCount` is _unsharded_, the `TenantShardId` is written without
+/// a shard suffix and is equivalent to the encoding of a `TenantId`: this enables
+/// an unsharded [`TenantShardId`] to be used interchangably with a [`TenantId`].
+///
+/// The human-readable encoding of an unsharded TenantShardId, such as used in API URLs,
+/// is both forward and backward compatible with TenantId: a legacy TenantId can be
+/// decoded as a TenantShardId, and when re-encoded it will be parseable
+/// as a TenantId.
+#[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy, Hash)]
+pub struct TenantShardId {
+    pub tenant_id: TenantId,
+    pub shard_number: ShardNumber,
+    pub shard_count: ShardCount,
+}
+
 impl ShardCount {
     pub const MAX: Self = Self(u8::MAX);
 
     /// The internal value of a ShardCount may be zero, which means "1 shard, but use
     /// legacy format for TenantShardId that excludes the shard suffix", also known
-    /// as `TenantShardId::unsharded`.
+    /// as [`TenantShardId::unsharded`].
     ///
     /// This method returns the actual number of shards, i.e. if our internal value is
     /// zero, we return 1 (unsharded tenants have 1 shard).
@@ -38,6 +116,9 @@ impl ShardCount {
         self.0
     }
 
+    /// Whether the `ShardCount` is for an unsharded tenant, so uses one shard but
+    /// uses the legacy format for `TenantShardId`. See also the documentation for
+    /// [`Self::count`].
     pub fn is_unsharded(&self) -> bool {
         self.0 == 0
     }
@@ -53,33 +134,6 @@ impl ShardNumber {
     pub const MAX: Self = Self(u8::MAX);
 }
 
-/// TenantShardId identify the units of work for the Pageserver.
-///
-/// These are written as `<tenant_id>-<shard number><shard-count>`, for example:
-///
-///   # The second shard in a two-shard tenant
-///   072f1291a5310026820b2fe4b2968934-0102
-///
-/// Historically, tenants could not have multiple shards, and were identified
-/// by TenantId.  To support this, TenantShardId has a special legacy
-/// mode where `shard_count` is equal to zero: this represents a single-sharded
-/// tenant which should be written as a TenantId with no suffix.
-///
-/// The human-readable encoding of TenantShardId, such as used in API URLs,
-/// is both forward and backward compatible: a legacy TenantId can be
-/// decoded as a TenantShardId, and when re-encoded it will be parseable
-/// as a TenantId.
-///
-/// Note that the binary encoding is _not_ backward compatible, because
-/// at the time sharding is introduced, there are no existing binary structures
-/// containing TenantId that we need to handle.
-#[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy, Hash)]
-pub struct TenantShardId {
-    pub tenant_id: TenantId,
-    pub shard_number: ShardNumber,
-    pub shard_count: ShardCount,
-}
-
 impl TenantShardId {
     pub fn unsharded(tenant_id: TenantId) -> Self {
         Self {
@@ -111,10 +165,13 @@ impl TenantShardId {
     }
 
     /// Convenience for code that has special behavior on the 0th shard.
-    pub fn is_zero(&self) -> bool {
+    pub fn is_shard_zero(&self) -> bool {
         self.shard_number == ShardNumber(0)
     }
 
+    /// The "unsharded" value is distinct from simply having a single shard: it represents
+    /// a tenant which is not shard-aware at all, and whose storage paths will not include
+    /// a shard suffix.
     pub fn is_unsharded(&self) -> bool {
         self.shard_number == ShardNumber(0) && self.shard_count.is_unsharded()
     }
@@ -150,9 +207,6 @@ impl TenantShardId {
     }
 }
 
-/// Formatting helper
-struct ShardSlug<'a>(&'a TenantShardId);
-
 impl<'a> std::fmt::Display for ShardSlug<'a> {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         write!(
@@ -222,16 +276,6 @@ impl From<[u8; 18]> for TenantShardId {
     }
 }
 
-/// For use within the context of a particular tenant, when we need to know which
-/// 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, Hash)]
-pub struct ShardIndex {
-    pub shard_number: ShardNumber,
-    pub shard_count: ShardCount,
-}
-
 impl ShardIndex {
     pub fn new(number: ShardNumber, count: ShardCount) -> Self {
         Self {
@@ -246,6 +290,9 @@ impl ShardIndex {
         }
     }
 
+    /// The "unsharded" value is distinct from simply having a single shard: it represents
+    /// a tenant which is not shard-aware at all, and whose storage paths will not include
+    /// a shard suffix.
     pub fn is_unsharded(&self) -> bool {
         self.shard_number == ShardNumber(0) && self.shard_count == ShardCount(0)
     }
@@ -313,6 +360,8 @@ impl Serialize for TenantShardId {
         if serializer.is_human_readable() {
             serializer.collect_str(self)
         } else {
+            // Note: while human encoding of [`TenantShardId`] is backward and forward
+            // compatible, this binary encoding is not.
             let mut packed: [u8; 18] = [0; 18];
             packed[0..16].clone_from_slice(&self.tenant_id.as_arr());
             packed[16] = self.shard_number.0;
@@ -390,16 +439,6 @@ const LAYOUT_BROKEN: ShardLayout = ShardLayout(255);
 /// Default stripe size in pages: 256MiB divided by 8kiB page size.
 const DEFAULT_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(256 * 1024 / 8);
 
-/// The ShardIdentity contains the information needed for one member of map
-/// to resolve a key to a shard, and then check whether that shard is ==self.
-#[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Debug)]
-pub struct ShardIdentity {
-    pub number: ShardNumber,
-    pub count: ShardCount,
-    pub stripe_size: ShardStripeSize,
-    layout: ShardLayout,
-}
-
 #[derive(thiserror::Error, Debug, PartialEq, Eq)]
 pub enum ShardConfigError {
     #[error("Invalid shard count")]
@@ -414,7 +453,7 @@ impl ShardIdentity {
     /// An identity with number=0 count=0 is a "none" identity, which represents legacy
     /// tenants.  Modern single-shard tenants should not use this: they should
     /// have number=0 count=1.
-    pub fn unsharded() -> Self {
+    pub const fn unsharded() -> Self {
         Self {
             number: ShardNumber(0),
             count: ShardCount(0),
@@ -439,6 +478,9 @@ impl ShardIdentity {
         }
     }
 
+    /// The "unsharded" value is distinct from simply having a single shard: it represents
+    /// a tenant which is not shard-aware at all, and whose storage paths will not include
+    /// a shard suffix.
     pub fn is_unsharded(&self) -> bool {
         self.number == ShardNumber(0) && self.count == ShardCount(0)
     }
@@ -487,6 +529,8 @@ impl ShardIdentity {
     }
 
     /// Return true if the key should be ingested by this shard
+    ///
+    /// Shards must ingest _at least_ keys which return true from this check.
     pub fn is_key_local(&self, key: &Key) -> bool {
         assert!(!self.is_broken());
         if self.count < ShardCount(2) || (key_is_shard0(key) && self.number == ShardNumber(0)) {
@@ -497,7 +541,9 @@ impl ShardIdentity {
     }
 
     /// Return true if the key should be discarded if found in this shard's
-    /// data store, e.g. during compaction after a split
+    /// data store, e.g. during compaction after a split.
+    ///
+    /// Shards _may_ drop keys which return false here, but are not obliged to.
     pub fn is_key_disposable(&self, key: &Key) -> bool {
         if key_is_shard0(key) {
             // Q: Why can't we dispose of shard0 content if we're not shard 0?
@@ -523,7 +569,7 @@ impl ShardIdentity {
 
     /// Convenience for checking if this identity is the 0th shard in a tenant,
     /// for special cases on shard 0 such as ingesting relation sizes.
-    pub fn is_zero(&self) -> bool {
+    pub fn is_shard_zero(&self) -> bool {
         self.number == ShardNumber(0)
     }
 }
@@ -606,7 +652,13 @@ fn key_is_shard0(key: &Key) -> bool {
     // relation pages are distributed to shards other than shard zero. Everything else gets
     // stored on shard 0.  This guarantees that shard 0 can independently serve basebackup
     // requests, and any request other than those for particular blocks in relations.
-    !is_rel_block_key(key)
+    //
+    // The only exception to this rule is "initfork" data -- this relates to postgres's UNLOGGED table
+    // type. These are special relations, usually with only 0 or 1 blocks, and we store them on shard 0
+    // because they must be included in basebackups.
+    let is_initfork = key.field5 == INIT_FORKNUM;
+
+    !is_rel_block_key(key) || is_initfork
 }
 
 /// Provide the same result as the function in postgres `hashfn.h` with the same name

libs/pageserver_api/src/upcall_api.rs

@@ -6,7 +6,9 @@
 use serde::{Deserialize, Serialize};
 use utils::id::NodeId;
 
-use crate::{controller_api::NodeRegisterRequest, shard::TenantShardId};
+use crate::{
+    controller_api::NodeRegisterRequest, models::LocationConfigMode, shard::TenantShardId,
+};
 
 /// Upcall message sent by the pageserver to the configured `control_plane_api` on
 /// startup.
@@ -20,12 +22,20 @@ pub struct ReAttachRequest {
     pub register: Option<NodeRegisterRequest>,
 }
 
-#[derive(Serialize, Deserialize)]
-pub struct ReAttachResponseTenant {
-    pub id: TenantShardId,
-    pub gen: u32,
+fn default_mode() -> LocationConfigMode {
+    LocationConfigMode::AttachedSingle
 }
 
+#[derive(Serialize, Deserialize, Debug)]
+pub struct ReAttachResponseTenant {
+    pub id: TenantShardId,
+    /// Mandatory if LocationConfigMode is None or set to an Attached* mode
+    pub gen: Option<u32>,
+
+    /// Default value only for backward compat: this field should be set
+    #[serde(default = "default_mode")]
+    pub mode: LocationConfigMode,
+}
 #[derive(Serialize, Deserialize)]
 pub struct ReAttachResponse {
     pub tenants: Vec<ReAttachResponseTenant>,

libs/postgres_ffi/src/lib.rs

@@ -118,7 +118,9 @@ pub use v14::bindings::{TimeLineID, TimestampTz, XLogRecPtr, XLogSegNo};
 // Likewise for these, although the assumption that these don't change is a little more iffy.
 pub use v14::bindings::{MultiXactOffset, MultiXactStatus};
 pub use v14::bindings::{PageHeaderData, XLogRecord};
-pub use v14::xlog_utils::{XLOG_SIZE_OF_XLOG_RECORD, XLOG_SIZE_OF_XLOG_SHORT_PHD};
+pub use v14::xlog_utils::{
+    XLOG_SIZE_OF_XLOG_LONG_PHD, XLOG_SIZE_OF_XLOG_RECORD, XLOG_SIZE_OF_XLOG_SHORT_PHD,
+};
 
 pub use v14::bindings::{CheckPoint, ControlFileData};
 

libs/postgres_ffi/src/xlog_utils.rs

@@ -331,7 +331,10 @@ impl CheckPoint {
     /// Returns 'true' if the XID was updated.
     pub fn update_next_xid(&mut self, xid: u32) -> bool {
         // nextXid should be greater than any XID in WAL, so increment provided XID and check for wraparround.
-        let mut new_xid = std::cmp::max(xid.wrapping_add(1), pg_constants::FIRST_NORMAL_TRANSACTION_ID);
+        let mut new_xid = std::cmp::max(
+            xid.wrapping_add(1),
+            pg_constants::FIRST_NORMAL_TRANSACTION_ID,
+        );
         // To reduce number of metadata checkpoints, we forward align XID on XID_CHECKPOINT_INTERVAL.
         // XID_CHECKPOINT_INTERVAL should not be larger than BLCKSZ*CLOG_XACTS_PER_BYTE
         new_xid =
@@ -367,8 +370,16 @@ pub fn generate_wal_segment(segno: u64, system_id: u64, lsn: Lsn) -> Result<Byte
     let seg_off = lsn.segment_offset(WAL_SEGMENT_SIZE);
 
     let first_page_only = seg_off < XLOG_BLCKSZ;
-    let (shdr_rem_len, infoflags) = if first_page_only {
-        (seg_off, pg_constants::XLP_FIRST_IS_CONTRECORD)
+    // If first records starts in the middle of the page, pretend in page header
+    // there is a fake record which ends where first real record starts. This
+    // makes pg_waldump etc happy.
+    let (shdr_rem_len, infoflags) = if first_page_only && seg_off > 0 {
+        assert!(seg_off >= XLOG_SIZE_OF_XLOG_LONG_PHD);
+        // xlp_rem_len doesn't include page header, hence the subtraction.
+        (
+            seg_off - XLOG_SIZE_OF_XLOG_LONG_PHD,
+            pg_constants::XLP_FIRST_IS_CONTRECORD,
+        )
     } else {
         (0, 0)
     };
@@ -397,20 +408,22 @@ pub fn generate_wal_segment(segno: u64, system_id: u64, lsn: Lsn) -> Result<Byte
 
     if !first_page_only {
         let block_offset = lsn.page_offset_in_segment(WAL_SEGMENT_SIZE) as usize;
+        // see comments above about XLP_FIRST_IS_CONTRECORD and xlp_rem_len.
+        let (xlp_rem_len, xlp_info) = if page_off > 0 {
+            assert!(page_off >= XLOG_SIZE_OF_XLOG_SHORT_PHD as u64);
+            (
+                (page_off - XLOG_SIZE_OF_XLOG_SHORT_PHD as u64) as u32,
+                pg_constants::XLP_FIRST_IS_CONTRECORD,
+            )
+        } else {
+            (0, 0)
+        };
         let header = XLogPageHeaderData {
             xlp_magic: XLOG_PAGE_MAGIC as u16,
-            xlp_info: if page_off >= pg_constants::SIZE_OF_PAGE_HEADER as u64 {
-                pg_constants::XLP_FIRST_IS_CONTRECORD
-            } else {
-                0
-            },
+            xlp_info,
             xlp_tli: PG_TLI,
             xlp_pageaddr: lsn.page_lsn().0,
-            xlp_rem_len: if page_off >= pg_constants::SIZE_OF_PAGE_HEADER as u64 {
-                page_off as u32
-            } else {
-                0u32
-            },
+            xlp_rem_len,
             ..Default::default() // Put 0 in padding fields.
         };
         let hdr_bytes = header.encode()?;

libs/postgres_ffi/wal_craft/src/bin/wal_craft.rs

@@ -1,5 +1,6 @@
 use anyhow::*;
 use clap::{value_parser, Arg, ArgMatches, Command};
+use postgres::Client;
 use std::{path::PathBuf, str::FromStr};
 use wal_craft::*;
 
@@ -8,8 +9,8 @@ fn main() -> Result<()> {
         .init();
     let arg_matches = cli().get_matches();
 
-    let wal_craft = |arg_matches: &ArgMatches, client| {
-        let (intermediate_lsns, end_of_wal_lsn) = match arg_matches
+    let wal_craft = |arg_matches: &ArgMatches, client: &mut Client| {
+        let intermediate_lsns = match arg_matches
             .get_one::<String>("type")
             .map(|s| s.as_str())
             .context("'type' is required")?
@@ -25,6 +26,7 @@ fn main() -> Result<()> {
             LastWalRecordCrossingSegment::NAME => LastWalRecordCrossingSegment::craft(client)?,
             a => panic!("Unknown --type argument: {a}"),
         };
+        let end_of_wal_lsn = client.pg_current_wal_insert_lsn()?;
         for lsn in intermediate_lsns {
             println!("intermediate_lsn = {lsn}");
         }

libs/postgres_ffi/wal_craft/src/lib.rs

@@ -4,8 +4,9 @@ use log::*;
 use postgres::types::PgLsn;
 use postgres::Client;
 use postgres_ffi::{WAL_SEGMENT_SIZE, XLOG_BLCKSZ};
-use postgres_ffi::{XLOG_SIZE_OF_XLOG_RECORD, XLOG_SIZE_OF_XLOG_SHORT_PHD};
-use std::cmp::Ordering;
+use postgres_ffi::{
+    XLOG_SIZE_OF_XLOG_LONG_PHD, XLOG_SIZE_OF_XLOG_RECORD, XLOG_SIZE_OF_XLOG_SHORT_PHD,
+};
 use std::path::{Path, PathBuf};
 use std::process::Command;
 use std::time::{Duration, Instant};
@@ -232,59 +233,62 @@ pub fn ensure_server_config(client: &mut impl postgres::GenericClient) -> anyhow
 pub trait Crafter {
     const NAME: &'static str;
 
-    /// Generates WAL using the client `client`. Returns a pair of:
-    /// * A vector of some valid "interesting" intermediate LSNs which one may start reading from.
-    ///   May include or exclude Lsn(0) and the end-of-wal.
-    /// * The expected end-of-wal LSN.
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)>;
+    /// Generates WAL using the client `client`. Returns a vector of some valid
+    /// "interesting" intermediate LSNs which one may start reading from.
+    /// test_end_of_wal uses this to check various starting points.
+    ///
+    /// Note that postgres is generally keen about writing some WAL. While we
+    /// try to disable it (autovacuum, big wal_writer_delay, etc) it is always
+    /// possible, e.g. xl_running_xacts are dumped each 15s. So checks about
+    /// stable WAL end would be flaky unless postgres is shut down. For this
+    /// reason returning potential end of WAL here is pointless. Most of the
+    /// time this doesn't happen though, so it is reasonable to create needed
+    /// WAL structure and immediately kill postgres like test_end_of_wal does.
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>>;
 }
 
+/// Wraps some WAL craft function, providing current LSN to it before the
+/// insertion and flushing WAL afterwards. Also pushes initial LSN to the
+/// result.
 fn craft_internal<C: postgres::GenericClient>(
     client: &mut C,
-    f: impl Fn(&mut C, PgLsn) -> anyhow::Result<(Vec<PgLsn>, Option<PgLsn>)>,
-) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
+    f: impl Fn(&mut C, PgLsn) -> anyhow::Result<Vec<PgLsn>>,
+) -> anyhow::Result<Vec<PgLsn>> {
     ensure_server_config(client)?;
 
     let initial_lsn = client.pg_current_wal_insert_lsn()?;
     info!("LSN initial = {}", initial_lsn);
 
-    let (mut intermediate_lsns, last_lsn) = f(client, initial_lsn)?;
-    let last_lsn = match last_lsn {
-        None => client.pg_current_wal_insert_lsn()?,
-        Some(last_lsn) => {
-            let insert_lsn = client.pg_current_wal_insert_lsn()?;
-            match last_lsn.cmp(&insert_lsn) {
-                Ordering::Less => bail!(
-                    "Some records were inserted after the crafted WAL: {} vs {}",
-                    last_lsn,
-                    insert_lsn
-                ),
-                Ordering::Equal => last_lsn,
-                Ordering::Greater => bail!("Reported LSN is greater than insert_lsn"),
-            }
-        }
-    };
+    let mut intermediate_lsns = f(client, initial_lsn)?;
     if !intermediate_lsns.starts_with(&[initial_lsn]) {
         intermediate_lsns.insert(0, initial_lsn);
     }
 
-    // Some records may be not flushed, e.g. non-transactional logical messages.
-    client.execute("select neon_xlogflush(pg_current_wal_insert_lsn())", &[])?;
-    match last_lsn.cmp(&client.pg_current_wal_flush_lsn()?) {
-        Ordering::Less => bail!("Some records were flushed after the crafted WAL"),
-        Ordering::Equal => {}
-        Ordering::Greater => bail!("Reported LSN is greater than flush_lsn"),
+    // Some records may be not flushed, e.g. non-transactional logical messages. Flush now.
+    //
+    // If the previous WAL record ended exactly at page boundary, pg_current_wal_insert_lsn
+    // returns the position just after the page header on the next page. That's where the next
+    // record will be inserted. But the page header hasn't actually been written to the WAL
+    // yet, and if you try to flush it, you get a "request to flush past end of generated WAL"
+    // error. Because of that, if the insert location is just after a page header, back off to
+    // previous page boundary.
+    let mut lsn = u64::from(client.pg_current_wal_insert_lsn()?);
+    if lsn % WAL_SEGMENT_SIZE as u64 == XLOG_SIZE_OF_XLOG_LONG_PHD as u64 {
+        lsn -= XLOG_SIZE_OF_XLOG_LONG_PHD as u64;
+    } else if lsn % XLOG_BLCKSZ as u64 == XLOG_SIZE_OF_XLOG_SHORT_PHD as u64 {
+        lsn -= XLOG_SIZE_OF_XLOG_SHORT_PHD as u64;
     }
-    Ok((intermediate_lsns, last_lsn))
+    client.execute("select neon_xlogflush($1)", &[&PgLsn::from(lsn)])?;
+    Ok(intermediate_lsns)
 }
 
 pub struct Simple;
 impl Crafter for Simple {
     const NAME: &'static str = "simple";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
         craft_internal(client, |client, _| {
             client.execute("CREATE table t(x int)", &[])?;
-            Ok((Vec::new(), None))
+            Ok(Vec::new())
         })
     }
 }
@@ -292,67 +296,85 @@ impl Crafter for Simple {
 pub struct LastWalRecordXlogSwitch;
 impl Crafter for LastWalRecordXlogSwitch {
     const NAME: &'static str = "last_wal_record_xlog_switch";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
-        // Do not use generate_internal because here we end up with flush_lsn exactly on
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
+        // Do not use craft_internal because here we end up with flush_lsn exactly on
         // the segment boundary and insert_lsn after the initial page header, which is unusual.
         ensure_server_config(client)?;
 
         client.execute("CREATE table t(x int)", &[])?;
         let before_xlog_switch = client.pg_current_wal_insert_lsn()?;
-        let after_xlog_switch: PgLsn = client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
-        let next_segment = PgLsn::from(0x0200_0000);
+        // pg_switch_wal returns end of last record of the switched segment,
+        // i.e. end of SWITCH itself.
+        let xlog_switch_record_end: PgLsn = client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
+        let before_xlog_switch_u64 = u64::from(before_xlog_switch);
+        let next_segment = PgLsn::from(
+            before_xlog_switch_u64 - (before_xlog_switch_u64 % WAL_SEGMENT_SIZE as u64)
+                + WAL_SEGMENT_SIZE as u64,
+        );
         ensure!(
-            after_xlog_switch <= next_segment,
-            "XLOG_SWITCH message ended after the expected segment boundary: {} > {}",
-            after_xlog_switch,
+            xlog_switch_record_end <= next_segment,
+            "XLOG_SWITCH record ended after the expected segment boundary: {} > {}",
+            xlog_switch_record_end,
             next_segment
         );
-        Ok((vec![before_xlog_switch, after_xlog_switch], next_segment))
+        Ok(vec![before_xlog_switch, xlog_switch_record_end])
     }
 }
 
 pub struct LastWalRecordXlogSwitchEndsOnPageBoundary;
+/// Craft xlog SWITCH record ending at page boundary.
 impl Crafter for LastWalRecordXlogSwitchEndsOnPageBoundary {
     const NAME: &'static str = "last_wal_record_xlog_switch_ends_on_page_boundary";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
         // Do not use generate_internal because here we end up with flush_lsn exactly on
         // the segment boundary and insert_lsn after the initial page header, which is unusual.
         ensure_server_config(client)?;
 
         client.execute("CREATE table t(x int)", &[])?;
 
-        // Add padding so the XLOG_SWITCH record ends exactly on XLOG_BLCKSZ boundary.
-        // We will use logical message as the padding. We start with detecting how much WAL
-        // it takes for one logical message, considering all alignments and headers.
-        let base_wal_advance = {
+        // Add padding so the XLOG_SWITCH record ends exactly on XLOG_BLCKSZ boundary.  We
+        // will use carefully-sized logical messages to advance WAL insert location such
+        // that there is just enough space on the page for the XLOG_SWITCH record.
+        loop {
+            // We start with measuring how much WAL it takes for one logical message,
+            // considering all alignments and headers.
             let before_lsn = client.pg_current_wal_insert_lsn()?;
-            // Small non-empty message bigger than few bytes is more likely than an empty
-            // message to have the same format as the big padding message.
             client.execute(
                 "SELECT pg_logical_emit_message(false, 'swch', REPEAT('a', 10))",
                 &[],
             )?;
-            // The XLOG_SWITCH record has no data => its size is exactly XLOG_SIZE_OF_XLOG_RECORD.
-            (u64::from(client.pg_current_wal_insert_lsn()?) - u64::from(before_lsn)) as usize
-                + XLOG_SIZE_OF_XLOG_RECORD
-        };
-        let mut remaining_lsn =
-            XLOG_BLCKSZ - u64::from(client.pg_current_wal_insert_lsn()?) as usize % XLOG_BLCKSZ;
-        if remaining_lsn < base_wal_advance {
-            remaining_lsn += XLOG_BLCKSZ;
+            let after_lsn = client.pg_current_wal_insert_lsn()?;
+
+            // Did the record cross a page boundary? If it did, start over. Crossing a
+            // page boundary adds to the apparent size of the record because of the page
+            // header, which throws off the calculation.
+            if u64::from(before_lsn) / XLOG_BLCKSZ as u64
+                != u64::from(after_lsn) / XLOG_BLCKSZ as u64
+            {
+                continue;
+            }
+            // base_size is the size of a logical message without the payload
+            let base_size = u64::from(after_lsn) - u64::from(before_lsn) - 10;
+
+            // Is there enough space on the page for another logical message and an
+            // XLOG_SWITCH? If not, start over.
+            let page_remain = XLOG_BLCKSZ as u64 - u64::from(after_lsn) % XLOG_BLCKSZ as u64;
+            if page_remain < base_size - XLOG_SIZE_OF_XLOG_RECORD as u64 {
+                continue;
+            }
+
+            // We will write another logical message, such that after the logical message
+            // record, there will be space for exactly one XLOG_SWITCH. How large should
+            // the logical message's payload be? An XLOG_SWITCH record has no data => its
+            // size is exactly XLOG_SIZE_OF_XLOG_RECORD.
+            let repeats = page_remain - base_size - XLOG_SIZE_OF_XLOG_RECORD as u64;
+
+            client.execute(
+                "SELECT pg_logical_emit_message(false, 'swch', REPEAT('a', $1))",
+                &[&(repeats as i32)],
+            )?;
+            break;
         }
-        let repeats = 10 + remaining_lsn - base_wal_advance;
-        info!(
-            "current_wal_insert_lsn={}, remaining_lsn={}, base_wal_advance={}, repeats={}",
-            client.pg_current_wal_insert_lsn()?,
-            remaining_lsn,
-            base_wal_advance,
-            repeats
-        );
-        client.execute(
-            "SELECT pg_logical_emit_message(false, 'swch', REPEAT('a', $1))",
-            &[&(repeats as i32)],
-        )?;
         info!(
             "current_wal_insert_lsn={}, XLOG_SIZE_OF_XLOG_RECORD={}",
             client.pg_current_wal_insert_lsn()?,
@@ -361,28 +383,29 @@ impl Crafter for LastWalRecordXlogSwitchEndsOnPageBoundary {
 
         // Emit the XLOG_SWITCH
         let before_xlog_switch = client.pg_current_wal_insert_lsn()?;
-        let after_xlog_switch: PgLsn = client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
+        let xlog_switch_record_end: PgLsn = client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
         let next_segment = PgLsn::from(0x0200_0000);
         ensure!(
-            after_xlog_switch < next_segment,
-            "XLOG_SWITCH message ended on or after the expected segment boundary: {} > {}",
-            after_xlog_switch,
+            xlog_switch_record_end < next_segment,
+            "XLOG_SWITCH record ended on or after the expected segment boundary: {} > {}",
+            xlog_switch_record_end,
             next_segment
         );
         ensure!(
-            u64::from(after_xlog_switch) as usize % XLOG_BLCKSZ == XLOG_SIZE_OF_XLOG_SHORT_PHD,
+            u64::from(xlog_switch_record_end) as usize % XLOG_BLCKSZ == XLOG_SIZE_OF_XLOG_SHORT_PHD,
             "XLOG_SWITCH message ended not on page boundary: {}, offset = {}",
-            after_xlog_switch,
-            u64::from(after_xlog_switch) as usize % XLOG_BLCKSZ
+            xlog_switch_record_end,
+            u64::from(xlog_switch_record_end) as usize % XLOG_BLCKSZ
         );
-        Ok((vec![before_xlog_switch, after_xlog_switch], next_segment))
+        Ok(vec![before_xlog_switch, xlog_switch_record_end])
     }
 }
 
-fn craft_single_logical_message(
+/// Write ~16MB logical message; it should cross WAL segment.
+fn craft_seg_size_logical_message(
     client: &mut impl postgres::GenericClient,
     transactional: bool,
-) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
+) -> anyhow::Result<Vec<PgLsn>> {
     craft_internal(client, |client, initial_lsn| {
         ensure!(
             initial_lsn < PgLsn::from(0x0200_0000 - 1024 * 1024),
@@ -405,34 +428,24 @@ fn craft_single_logical_message(
             "Logical message crossed two segments"
         );
 
-        if transactional {
-            // Transactional logical messages are part of a transaction, so the one above is
-            // followed by a small COMMIT record.
-
-            let after_message_lsn = client.pg_current_wal_insert_lsn()?;
-            ensure!(
-                message_lsn < after_message_lsn,
-                "No record found after the emitted message"
-            );
-            Ok((vec![message_lsn], Some(after_message_lsn)))
-        } else {
-            Ok((Vec::new(), Some(message_lsn)))
-        }
+        Ok(vec![message_lsn])
     })
 }
 
 pub struct WalRecordCrossingSegmentFollowedBySmallOne;
 impl Crafter for WalRecordCrossingSegmentFollowedBySmallOne {
     const NAME: &'static str = "wal_record_crossing_segment_followed_by_small_one";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
-        craft_single_logical_message(client, true)
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
+        // Transactional message crossing WAL segment will be followed by small
+        // commit record.
+        craft_seg_size_logical_message(client, true)
     }
 }
 
 pub struct LastWalRecordCrossingSegment;
 impl Crafter for LastWalRecordCrossingSegment {
     const NAME: &'static str = "last_wal_record_crossing_segment";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
-        craft_single_logical_message(client, false)
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
+        craft_seg_size_logical_message(client, false)
     }
 }

libs/postgres_ffi/wal_craft/src/xlog_utils_test.rs

@@ -11,13 +11,15 @@ use utils::const_assert;
 use utils::lsn::Lsn;
 
 fn init_logging() {
-    let _ = env_logger::Builder::from_env(env_logger::Env::default().default_filter_or(
-        format!("crate=info,postgres_ffi::{PG_MAJORVERSION}::xlog_utils=trace"),
-    ))
+    let _ = env_logger::Builder::from_env(env_logger::Env::default().default_filter_or(format!(
+        "crate=info,postgres_ffi::{PG_MAJORVERSION}::xlog_utils=trace"
+    )))
     .is_test(true)
     .try_init();
 }
 
+/// Test that find_end_of_wal returns the same results as pg_dump on various
+/// WALs created by Crafter.
 fn test_end_of_wal<C: crate::Crafter>(test_name: &str) {
     use crate::*;
 
@@ -38,13 +40,13 @@ fn test_end_of_wal<C: crate::Crafter>(test_name: &str) {
     }
     cfg.initdb().unwrap();
     let srv = cfg.start_server().unwrap();
-    let (intermediate_lsns, expected_end_of_wal_partial) =
-        C::craft(&mut srv.connect_with_timeout().unwrap()).unwrap();
+    let intermediate_lsns = C::craft(&mut srv.connect_with_timeout().unwrap()).unwrap();
     let intermediate_lsns: Vec<Lsn> = intermediate_lsns
         .iter()
         .map(|&lsn| u64::from(lsn).into())
         .collect();
-    let expected_end_of_wal: Lsn = u64::from(expected_end_of_wal_partial).into();
+    // Kill postgres. Note that it might have inserted to WAL something after
+    // 'craft' did its job.
     srv.kill();
 
     // Check find_end_of_wal on the initial WAL
@@ -56,7 +58,7 @@ fn test_end_of_wal<C: crate::Crafter>(test_name: &str) {
         .filter(|fname| IsXLogFileName(fname))
         .max()
         .unwrap();
-    check_pg_waldump_end_of_wal(&cfg, &last_segment, expected_end_of_wal);
+    let expected_end_of_wal = find_pg_waldump_end_of_wal(&cfg, &last_segment);
     for start_lsn in intermediate_lsns
         .iter()
         .chain(std::iter::once(&expected_end_of_wal))
@@ -91,11 +93,7 @@ fn test_end_of_wal<C: crate::Crafter>(test_name: &str) {
     }
 }
 
-fn check_pg_waldump_end_of_wal(
-    cfg: &crate::Conf,
-    last_segment: &str,
-    expected_end_of_wal: Lsn,
-) {
+fn find_pg_waldump_end_of_wal(cfg: &crate::Conf, last_segment: &str) -> Lsn {
     // Get the actual end of WAL by pg_waldump
     let waldump_output = cfg
         .pg_waldump("000000010000000000000001", last_segment)
@@ -113,11 +111,8 @@ fn check_pg_waldump_end_of_wal(
         }
     };
     let waldump_wal_end = Lsn::from_str(caps.get(1).unwrap().as_str()).unwrap();
-    info!(
-        "waldump erred on {}, expected wal end at {}",
-        waldump_wal_end, expected_end_of_wal
-    );
-    assert_eq!(waldump_wal_end, expected_end_of_wal);
+    info!("waldump erred on {}", waldump_wal_end);
+    waldump_wal_end
 }
 
 fn check_end_of_wal(
@@ -210,9 +205,9 @@ pub fn test_update_next_xid() {
 #[test]
 pub fn test_encode_logical_message() {
     let expected = [
-        64, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 21, 0, 0, 170, 34, 166, 227, 255,
-        38, 0, 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 112, 114,
-        101, 102, 105, 120, 0, 109, 101, 115, 115, 97, 103, 101,
+        64, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 21, 0, 0, 170, 34, 166, 227, 255, 38,
+        0, 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 112, 114, 101, 102,
+        105, 120, 0, 109, 101, 115, 115, 97, 103, 101,
     ];
     let actual = encode_logical_message("prefix", "message");
     assert_eq!(expected, actual[..]);

libs/remote_storage/Cargo.toml

@@ -18,6 +18,7 @@ camino.workspace = true
 humantime.workspace = true
 hyper = { workspace = true, features = ["stream"] }
 futures.workspace = true
+rand.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 tokio = { workspace = true, features = ["sync", "fs", "io-util"] }
@@ -37,6 +38,7 @@ azure_storage_blobs.workspace = true
 futures-util.workspace = true
 http-types.workspace = true
 itertools.workspace = true
+sync_wrapper = { workspace = true, features = ["futures"] }
 
 [dev-dependencies]
 camino-tempfile.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -3,6 +3,7 @@
 use std::borrow::Cow;
 use std::collections::HashMap;
 use std::env;
+use std::io;
 use std::num::NonZeroU32;
 use std::pin::Pin;
 use std::str::FromStr;
@@ -20,6 +21,7 @@ use azure_storage_blobs::blob::CopyStatus;
 use azure_storage_blobs::prelude::ClientBuilder;
 use azure_storage_blobs::{blob::operations::GetBlobBuilder, prelude::ContainerClient};
 use bytes::Bytes;
+use futures::future::Either;
 use futures::stream::Stream;
 use futures_util::StreamExt;
 use futures_util::TryStreamExt;
@@ -128,12 +130,12 @@ impl AzureBlobStorage {
         let kind = RequestKind::Get;
 
         let _permit = self.permit(kind, cancel).await?;
+        let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
+        let cancel_or_timeout_ = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
 
         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
 
         let download = async {
             let response = builder
@@ -152,30 +154,46 @@ impl AzureBlobStorage {
                 Err(_elapsed) => Err(DownloadError::Timeout),
             });
 
-            let mut response = std::pin::pin!(response);
+            let mut response = Box::pin(response);
 
-            let mut bufs = Vec::new();
-            while let Some(part) = response.next().await {
-                let part = part?;
-                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())));
-                }
-                let data = part
-                    .data
-                    .collect()
-                    .await
-                    .map_err(|e| DownloadError::Other(e.into()))?;
-                bufs.push(data);
+            let Some(part) = response.next().await else {
+                return Err(DownloadError::Other(anyhow::anyhow!(
+                    "Azure GET response contained no response body"
+                )));
+            };
+            let part = part?;
+            if etag.is_none() {
+                etag = Some(part.blob.properties.etag);
             }
+            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())));
+            }
+
+            // unwrap safety: if these were None, bufs would be empty and we would have returned an error already
+            let etag = etag.unwrap();
+            let last_modified = last_modified.unwrap();
+
+            let tail_stream = response
+                .map(|part| match part {
+                    Ok(part) => Either::Left(part.data.map(|r| r.map_err(io::Error::other))),
+                    Err(e) => {
+                        Either::Right(futures::stream::once(async { Err(io::Error::other(e)) }))
+                    }
+                })
+                .flatten();
+            let stream = part
+                .data
+                .map(|r| r.map_err(io::Error::other))
+                .chain(sync_wrapper::SyncStream::new(tail_stream));
+            //.chain(SyncStream::from_pin(Box::pin(tail_stream)));
+
+            let download_stream = crate::support::DownloadStream::new(cancel_or_timeout_, stream);
+
             Ok(Download {
-                download_stream: Box::pin(futures::stream::iter(bufs.into_iter().map(Ok))),
+                download_stream: Box::pin(download_stream),
                 etag,
                 last_modified,
                 metadata: Some(StorageMetadata(metadata)),
@@ -184,7 +202,10 @@ impl AzureBlobStorage {
 
         tokio::select! {
             bufs = download => bufs,
-            _ = cancel.cancelled() => Err(DownloadError::Cancelled),
+            cancel_or_timeout = cancel_or_timeout => match cancel_or_timeout {
+                TimeoutOrCancel::Timeout => Err(DownloadError::Timeout),
+                TimeoutOrCancel::Cancel => Err(DownloadError::Cancelled),
+            },
         }
     }
 

libs/remote_storage/src/lib.rs

@@ -21,11 +21,13 @@ use std::{
     fmt::Debug,
     num::{NonZeroU32, NonZeroUsize},
     pin::Pin,
+    str::FromStr,
     sync::Arc,
     time::{Duration, SystemTime},
 };
 
 use anyhow::{bail, Context};
+use aws_sdk_s3::types::StorageClass;
 use camino::{Utf8Path, Utf8PathBuf};
 
 use bytes::Bytes;
@@ -42,6 +44,9 @@ pub use self::{
 };
 use s3_bucket::RequestKind;
 
+/// Azure SDK's ETag type is a simple String wrapper: we use this internally instead of repeating it here.
+pub use azure_core::Etag;
+
 pub use error::{DownloadError, TimeTravelError, TimeoutOrCancel};
 
 /// Currently, sync happens with AWS S3, that has two limits on requests per second:
@@ -50,11 +55,11 @@ pub use error::{DownloadError, TimeTravelError, TimeoutOrCancel};
 /// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
 /// <https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/>
 pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
-/// We set this a little bit low as we currently buffer the entire file into RAM
+/// Set this limit analogously to the S3 limit
 ///
 /// Here, a limit of max 20k concurrent connections was noted.
 /// <https://learn.microsoft.com/en-us/answers/questions/1301863/is-there-any-limitation-to-concurrent-connections>
-pub const DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT: usize = 30;
+pub const DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT: usize = 100;
 /// No limits on the client side, which currenltly means 1000 for AWS S3.
 /// <https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax>
 pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
@@ -131,6 +136,11 @@ impl RemotePath {
     pub fn strip_prefix(&self, p: &RemotePath) -> Result<&Utf8Path, std::path::StripPrefixError> {
         self.0.strip_prefix(&p.0)
     }
+
+    pub fn add_trailing_slash(&self) -> Self {
+        // Unwrap safety inputs are guararnteed to be valid UTF-8
+        Self(format!("{}/", self.0).try_into().unwrap())
+    }
 }
 
 /// We don't need callers to be able to pass arbitrary delimiters: just control
@@ -154,47 +164,21 @@ pub struct Listing {
 /// providing basic CRUD operations for storage files.
 #[allow(async_fn_in_trait)]
 pub trait RemoteStorage: Send + Sync + 'static {
-    /// Lists all top level subdirectories for a given prefix
-    /// Note: here we assume that if the prefix is passed it was obtained via remote_object_id
-    /// which already takes into account any kind of global prefix (prefix_in_bucket for S3 or storage_root for LocalFS)
-    /// so this method doesnt need to.
-    async fn list_prefixes(
-        &self,
-        prefix: Option<&RemotePath>,
-        cancel: &CancellationToken,
-    ) -> Result<Vec<RemotePath>, DownloadError> {
-        let result = self
-            .list(prefix, ListingMode::WithDelimiter, None, cancel)
-            .await?
-            .prefixes;
-        Ok(result)
-    }
-    /// Lists all files in directory "recursively"
-    /// (not really recursively, because AWS has a flat namespace)
-    /// Note: This is subtely different than list_prefixes,
-    /// because it is for listing files instead of listing
-    /// names sharing common prefixes.
-    /// For example,
-    /// list_files("foo/bar") = ["foo/bar/cat123.txt",
-    /// "foo/bar/cat567.txt", "foo/bar/dog123.txt", "foo/bar/dog456.txt"]
-    /// whereas,
-    /// list_prefixes("foo/bar/") = ["cat", "dog"]
-    /// See `test_real_s3.rs` for more details.
+    /// List objects in remote storage, with semantics matching AWS S3's ListObjectsV2.
+    /// (see `<https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html>`)
+    ///
+    /// Note that the prefix is relative to any `prefix_in_bucket` configured for the client, not
+    /// from the absolute root of the bucket.
+    ///
+    /// `mode` configures whether to use a delimiter.  Without a delimiter all keys
+    /// within the prefix are listed in the `keys` of the result.  With a delimiter, any "directories" at the top level of
+    /// the prefix are returned in the `prefixes` of the result, and keys in the top level of the prefix are
+    /// returned in `keys` ().
+    ///
+    /// `max_keys` controls the maximum number of keys that will be returned.  If this is None, this function
+    /// will iteratively call listobjects until it runs out of keys.  Note that this is not safe to use on
+    /// unlimted size buckets, as the full list of objects is allocated into a monolithic data structure.
     ///
-    /// max_keys limits max number of keys returned; None means unlimited.
-    async fn list_files(
-        &self,
-        prefix: Option<&RemotePath>,
-        max_keys: Option<NonZeroU32>,
-        cancel: &CancellationToken,
-    ) -> Result<Vec<RemotePath>, DownloadError> {
-        let result = self
-            .list(prefix, ListingMode::NoDelimiter, max_keys, cancel)
-            .await?
-            .keys;
-        Ok(result)
-    }
-
     async fn list(
         &self,
         prefix: Option<&RemotePath>,
@@ -291,9 +275,9 @@ pub type DownloadStream =
 pub struct Download {
     pub download_stream: DownloadStream,
     /// The last time the file was modified (`last-modified` HTTP header)
-    pub last_modified: Option<SystemTime>,
+    pub last_modified: SystemTime,
     /// A way to identify this specific version of the resource (`etag` HTTP header)
-    pub etag: Option<String>,
+    pub etag: Etag,
     /// Extra key-value data, associated with the current remote file.
     pub metadata: Option<StorageMetadata>,
 }
@@ -333,41 +317,6 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
         }
     }
 
-    // A function for listing all the files in a "directory"
-    // Example:
-    // list_files("foo/bar") = ["foo/bar/a.txt", "foo/bar/b.txt"]
-    //
-    // max_keys limits max number of keys returned; None means unlimited.
-    pub async fn list_files(
-        &self,
-        folder: Option<&RemotePath>,
-        max_keys: Option<NonZeroU32>,
-        cancel: &CancellationToken,
-    ) -> Result<Vec<RemotePath>, DownloadError> {
-        match self {
-            Self::LocalFs(s) => s.list_files(folder, max_keys, cancel).await,
-            Self::AwsS3(s) => s.list_files(folder, max_keys, cancel).await,
-            Self::AzureBlob(s) => s.list_files(folder, max_keys, cancel).await,
-            Self::Unreliable(s) => s.list_files(folder, max_keys, cancel).await,
-        }
-    }
-
-    // lists common *prefixes*, if any of files
-    // Example:
-    // list_prefixes("foo123","foo567","bar123","bar432") = ["foo", "bar"]
-    pub async fn list_prefixes(
-        &self,
-        prefix: Option<&RemotePath>,
-        cancel: &CancellationToken,
-    ) -> Result<Vec<RemotePath>, DownloadError> {
-        match self {
-            Self::LocalFs(s) => s.list_prefixes(prefix, cancel).await,
-            Self::AwsS3(s) => s.list_prefixes(prefix, cancel).await,
-            Self::AzureBlob(s) => s.list_prefixes(prefix, cancel).await,
-            Self::Unreliable(s) => s.list_prefixes(prefix, cancel).await,
-        }
-    }
-
     /// See [`RemoteStorage::upload`]
     pub async fn upload(
         &self,
@@ -562,6 +511,16 @@ impl GenericRemoteStorage {
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct StorageMetadata(HashMap<String, String>);
 
+impl<const N: usize> From<[(&str, &str); N]> for StorageMetadata {
+    fn from(arr: [(&str, &str); N]) -> Self {
+        let map: HashMap<String, String> = arr
+            .iter()
+            .map(|(k, v)| (k.to_string(), v.to_string()))
+            .collect();
+        Self(map)
+    }
+}
+
 /// External backup storage configuration, enough for creating a client for that storage.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct RemoteStorageConfig {
@@ -606,6 +565,7 @@ pub struct S3Config {
     /// See [`DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT`] for more details.
     pub concurrency_limit: NonZeroUsize,
     pub max_keys_per_list_response: Option<i32>,
+    pub upload_storage_class: Option<StorageClass>,
 }
 
 impl Debug for S3Config {
@@ -734,6 +694,18 @@ impl RemoteStorageConfig {
                     endpoint,
                     concurrency_limit,
                     max_keys_per_list_response,
+                    upload_storage_class: toml
+                        .get("upload_storage_class")
+                        .map(|prefix_in_bucket| -> anyhow::Result<_> {
+                            let s = parse_toml_string("upload_storage_class", prefix_in_bucket)?;
+                            let storage_class = StorageClass::from_str(&s).expect("infallible");
+                            #[allow(deprecated)]
+                            if matches!(storage_class, StorageClass::Unknown(_)) {
+                                bail!("Specified storage class unknown to SDK: '{s}'. Allowed values: {:?}", StorageClass::values());
+                            }
+                            Ok(storage_class)
+                        })
+                        .transpose()?,
                 })
             }
             (_, _, _, Some(_), None) => {

libs/remote_storage/src/local_fs.rs

@@ -5,12 +5,10 @@
 //! volume is mounted to the local FS.
 
 use std::{
-    borrow::Cow,
-    future::Future,
+    collections::HashSet,
     io::ErrorKind,
     num::NonZeroU32,
-    pin::Pin,
-    time::{Duration, SystemTime},
+    time::{Duration, SystemTime, UNIX_EPOCH},
 };
 
 use anyhow::{bail, ensure, Context};
@@ -22,14 +20,15 @@ use tokio::{
     io::{self, AsyncReadExt, AsyncSeekExt, AsyncWriteExt},
 };
 use tokio_util::{io::ReaderStream, sync::CancellationToken};
-use tracing::*;
-use utils::{crashsafe::path_with_suffix_extension, fs_ext::is_directory_empty};
+use utils::crashsafe::path_with_suffix_extension;
 
 use crate::{
     Download, DownloadError, Listing, ListingMode, RemotePath, TimeTravelError, TimeoutOrCancel,
+    REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
 use super::{RemoteStorage, StorageMetadata};
+use crate::Etag;
 
 const LOCAL_FS_TEMP_FILE_SUFFIX: &str = "___temp";
 
@@ -92,7 +91,47 @@ impl LocalFs {
 
     #[cfg(test)]
     async fn list_all(&self) -> anyhow::Result<Vec<RemotePath>> {
-        Ok(get_all_files(&self.storage_root, true)
+        use std::{future::Future, pin::Pin};
+        fn get_all_files<'a, P>(
+            directory_path: P,
+        ) -> Pin<Box<dyn Future<Output = anyhow::Result<Vec<Utf8PathBuf>>> + Send + Sync + 'a>>
+        where
+            P: AsRef<Utf8Path> + Send + Sync + 'a,
+        {
+            Box::pin(async move {
+                let directory_path = directory_path.as_ref();
+                if directory_path.exists() {
+                    if directory_path.is_dir() {
+                        let mut paths = Vec::new();
+                        let mut dir_contents = fs::read_dir(directory_path).await?;
+                        while let Some(dir_entry) = dir_contents.next_entry().await? {
+                            let file_type = dir_entry.file_type().await?;
+                            let entry_path =
+                                Utf8PathBuf::from_path_buf(dir_entry.path()).map_err(|pb| {
+                                    anyhow::Error::msg(format!(
+                                        "non-Unicode path: {}",
+                                        pb.to_string_lossy()
+                                    ))
+                                })?;
+                            if file_type.is_symlink() {
+                                tracing::debug!("{entry_path:?} is a symlink, skipping")
+                            } else if file_type.is_dir() {
+                                paths.extend(get_all_files(&entry_path).await?.into_iter())
+                            } else {
+                                paths.push(entry_path);
+                            }
+                        }
+                        Ok(paths)
+                    } else {
+                        bail!("Path {directory_path:?} is not a directory")
+                    }
+                } else {
+                    Ok(Vec::new())
+                }
+            })
+        }
+
+        Ok(get_all_files(&self.storage_root)
             .await?
             .into_iter()
             .map(|path| {
@@ -119,6 +158,14 @@ impl LocalFs {
         // S3 object list prefixes can be arbitrary strings, but when reading
         // the local filesystem we need a directory to start calling read_dir on.
         let mut initial_dir = full_path.clone();
+
+        // If there's no trailing slash, we have to start looking from one above: even if
+        // `initial_dir` is a directory, we should still list any prefixes in the parent
+        // that start with the same string.
+        if !full_path.to_string().ends_with('/') {
+            initial_dir.pop();
+        }
+
         loop {
             // Did we make it to the root?
             if initial_dir.parent().is_none() {
@@ -197,6 +244,7 @@ impl LocalFs {
             fs::OpenOptions::new()
                 .write(true)
                 .create(true)
+                .truncate(true)
                 .open(&temp_file_path)
                 .await
                 .with_context(|| {
@@ -293,61 +341,66 @@ impl RemoteStorage for LocalFs {
         let op = async {
             let mut result = Listing::default();
 
-            if let ListingMode::NoDelimiter = mode {
-                let keys = self
-                    .list_recursive(prefix)
-                    .await
-                    .map_err(DownloadError::Other)?;
-
-                result.keys = keys
-                    .into_iter()
-                    .filter(|k| {
-                        let path = k.with_base(&self.storage_root);
-                        !path.is_dir()
-                    })
-                    .collect();
-
-                if let Some(max_keys) = max_keys {
-                    result.keys.truncate(max_keys.get() as usize);
-                }
-
-                return Ok(result);
-            }
-
-            let path = match prefix {
-                Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
-                None => Cow::Borrowed(&self.storage_root),
-            };
-
-            let prefixes_to_filter = get_all_files(path.as_ref(), false)
+            // Filter out directories: in S3 directories don't exist, only the keys within them do.
+            let keys = self
+                .list_recursive(prefix)
                 .await
                 .map_err(DownloadError::Other)?;
+            let keys = keys
+                .into_iter()
+                .filter(|k| {
+                    let path = k.with_base(&self.storage_root);
+                    !path.is_dir()
+                })
+                .collect();
 
-            // filter out empty directories to mirror s3 behavior.
-            for prefix in prefixes_to_filter {
-                if prefix.is_dir()
-                    && is_directory_empty(&prefix)
-                        .await
-                        .map_err(DownloadError::Other)?
-                {
-                    continue;
-                }
-
-                let stripped = prefix
-                    .strip_prefix(&self.storage_root)
-                    .context("Failed to strip prefix")
-                    .and_then(RemotePath::new)
-                    .expect(
-                        "We list files for storage root, hence should be able to remote the prefix",
-                    );
-
-                if prefix.is_dir() {
-                    result.prefixes.push(stripped);
-                } else {
-                    result.keys.push(stripped);
+            if let ListingMode::NoDelimiter = mode {
+                result.keys = keys;
+            } else {
+                let mut prefixes = HashSet::new();
+                for key in keys {
+                    // If the part after the prefix includes a "/", take only the first part and put it in `prefixes`.
+                    let relative_key = if let Some(prefix) = prefix {
+                        let mut prefix = prefix.clone();
+                        // We only strip the dirname of the prefix, so that when we strip it from the start of keys we
+                        // end up with full file/dir names.
+                        let prefix_full_local_path = prefix.with_base(&self.storage_root);
+                        let has_slash = prefix.0.to_string().ends_with('/');
+                        let strip_prefix = if prefix_full_local_path.is_dir() && has_slash {
+                            prefix
+                        } else {
+                            prefix.0.pop();
+                            prefix
+                        };
+
+                        RemotePath::new(key.strip_prefix(&strip_prefix).unwrap()).unwrap()
+                    } else {
+                        key
+                    };
+
+                    let relative_key = format!("{}", relative_key);
+                    if relative_key.contains(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                        let first_part = relative_key
+                            .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
+                            .next()
+                            .unwrap()
+                            .to_owned();
+                        prefixes.insert(first_part);
+                    } else {
+                        result
+                            .keys
+                            .push(RemotePath::from_string(&relative_key).unwrap());
+                    }
                 }
+                result.prefixes = prefixes
+                    .into_iter()
+                    .map(|s| RemotePath::from_string(&s).unwrap())
+                    .collect();
             }
 
+            if let Some(max_keys) = max_keys {
+                result.keys.truncate(max_keys.get() as usize);
+            }
             Ok(result)
         };
 
@@ -406,35 +459,37 @@ impl RemoteStorage for LocalFs {
         cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         let target_path = from.with_base(&self.storage_root);
-        if file_exists(&target_path).map_err(DownloadError::BadInput)? {
-            let source = ReaderStream::new(
-                fs::OpenOptions::new()
-                    .read(true)
-                    .open(&target_path)
-                    .await
-                    .with_context(|| {
-                        format!("Failed to open source file {target_path:?} to use in the download")
-                    })
-                    .map_err(DownloadError::Other)?,
-            );
 
-            let metadata = self
-                .read_storage_metadata(&target_path)
+        let file_metadata = file_metadata(&target_path).await?;
+
+        let source = ReaderStream::new(
+            fs::OpenOptions::new()
+                .read(true)
+                .open(&target_path)
                 .await
-                .map_err(DownloadError::Other)?;
+                .with_context(|| {
+                    format!("Failed to open source file {target_path:?} to use in the download")
+                })
+                .map_err(DownloadError::Other)?,
+        );
 
-            let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
-            let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
+        let metadata = self
+            .read_storage_metadata(&target_path)
+            .await
+            .map_err(DownloadError::Other)?;
 
-            Ok(Download {
-                metadata,
-                last_modified: None,
-                etag: None,
-                download_stream: Box::pin(source),
-            })
-        } else {
-            Err(DownloadError::NotFound)
-        }
+        let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
+        let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
+
+        let etag = mock_etag(&file_metadata);
+        Ok(Download {
+            metadata,
+            last_modified: file_metadata
+                .modified()
+                .map_err(|e| DownloadError::Other(anyhow::anyhow!(e).context("Reading mtime")))?,
+            etag,
+            download_stream: Box::pin(source),
+        })
     }
 
     async fn download_byte_range(
@@ -452,50 +507,51 @@ impl RemoteStorage for LocalFs {
                 return Err(DownloadError::Other(anyhow::anyhow!("Invalid range, start ({start_inclusive}) and end_exclusive ({end_exclusive:?}) difference is zero bytes")));
             }
         }
+
         let target_path = from.with_base(&self.storage_root);
-        if file_exists(&target_path).map_err(DownloadError::BadInput)? {
-            let mut source = tokio::fs::OpenOptions::new()
-                .read(true)
-                .open(&target_path)
-                .await
-                .with_context(|| {
-                    format!("Failed to open source file {target_path:?} to use in the download")
-                })
-                .map_err(DownloadError::Other)?;
-
-            let len = source
-                .metadata()
-                .await
-                .context("query file length")
-                .map_err(DownloadError::Other)?
-                .len();
-
-            source
-                .seek(io::SeekFrom::Start(start_inclusive))
-                .await
-                .context("Failed to seek to the range start in a local storage file")
-                .map_err(DownloadError::Other)?;
-
-            let metadata = self
-                .read_storage_metadata(&target_path)
-                .await
-                .map_err(DownloadError::Other)?;
-
-            let source = source.take(end_exclusive.unwrap_or(len) - start_inclusive);
-            let source = ReaderStream::new(source);
-
-            let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
-            let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
-
-            Ok(Download {
-                metadata,
-                last_modified: None,
-                etag: None,
-                download_stream: Box::pin(source),
+        let file_metadata = file_metadata(&target_path).await?;
+        let mut source = tokio::fs::OpenOptions::new()
+            .read(true)
+            .open(&target_path)
+            .await
+            .with_context(|| {
+                format!("Failed to open source file {target_path:?} to use in the download")
             })
-        } else {
-            Err(DownloadError::NotFound)
-        }
+            .map_err(DownloadError::Other)?;
+
+        let len = source
+            .metadata()
+            .await
+            .context("query file length")
+            .map_err(DownloadError::Other)?
+            .len();
+
+        source
+            .seek(io::SeekFrom::Start(start_inclusive))
+            .await
+            .context("Failed to seek to the range start in a local storage file")
+            .map_err(DownloadError::Other)?;
+
+        let metadata = self
+            .read_storage_metadata(&target_path)
+            .await
+            .map_err(DownloadError::Other)?;
+
+        let source = source.take(end_exclusive.unwrap_or(len) - start_inclusive);
+        let source = ReaderStream::new(source);
+
+        let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
+        let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
+
+        let etag = mock_etag(&file_metadata);
+        Ok(Download {
+            metadata,
+            last_modified: file_metadata
+                .modified()
+                .map_err(|e| DownloadError::Other(anyhow::anyhow!(e).context("Reading mtime")))?,
+            etag,
+            download_stream: Box::pin(source),
+        })
     }
 
     async fn delete(&self, path: &RemotePath, _cancel: &CancellationToken) -> anyhow::Result<()> {
@@ -555,50 +611,6 @@ fn storage_metadata_path(original_path: &Utf8Path) -> Utf8PathBuf {
     path_with_suffix_extension(original_path, "metadata")
 }
 
-fn get_all_files<'a, P>(
-    directory_path: P,
-    recursive: bool,
-) -> Pin<Box<dyn Future<Output = anyhow::Result<Vec<Utf8PathBuf>>> + Send + Sync + 'a>>
-where
-    P: AsRef<Utf8Path> + Send + Sync + 'a,
-{
-    Box::pin(async move {
-        let directory_path = directory_path.as_ref();
-        if directory_path.exists() {
-            if directory_path.is_dir() {
-                let mut paths = Vec::new();
-                let mut dir_contents = fs::read_dir(directory_path).await?;
-                while let Some(dir_entry) = dir_contents.next_entry().await? {
-                    let file_type = dir_entry.file_type().await?;
-                    let entry_path =
-                        Utf8PathBuf::from_path_buf(dir_entry.path()).map_err(|pb| {
-                            anyhow::Error::msg(format!(
-                                "non-Unicode path: {}",
-                                pb.to_string_lossy()
-                            ))
-                        })?;
-                    if file_type.is_symlink() {
-                        debug!("{entry_path:?} is a symlink, skipping")
-                    } else if file_type.is_dir() {
-                        if recursive {
-                            paths.extend(get_all_files(&entry_path, true).await?.into_iter())
-                        } else {
-                            paths.push(entry_path)
-                        }
-                    } else {
-                        paths.push(entry_path);
-                    }
-                }
-                Ok(paths)
-            } else {
-                bail!("Path {directory_path:?} is not a directory")
-            }
-        } else {
-            Ok(Vec::new())
-        }
-    })
-}
-
 async fn create_target_directory(target_file_path: &Utf8Path) -> anyhow::Result<()> {
     let target_dir = match target_file_path.parent() {
         Some(parent_dir) => parent_dir,
@@ -610,13 +622,22 @@ async fn create_target_directory(target_file_path: &Utf8Path) -> anyhow::Result<
     Ok(())
 }
 
-fn file_exists(file_path: &Utf8Path) -> anyhow::Result<bool> {
-    if file_path.exists() {
-        ensure!(file_path.is_file(), "file path '{file_path}' is not a file");
-        Ok(true)
-    } else {
-        Ok(false)
-    }
+async fn file_metadata(file_path: &Utf8Path) -> Result<std::fs::Metadata, DownloadError> {
+    tokio::fs::metadata(&file_path).await.map_err(|e| {
+        if e.kind() == ErrorKind::NotFound {
+            DownloadError::NotFound
+        } else {
+            DownloadError::BadInput(e.into())
+        }
+    })
+}
+
+// Use mtime as stand-in for ETag.  We could calculate a meaningful one by md5'ing the contents of files we
+// read, but that's expensive and the local_fs test helper's whole reason for existence is to run small tests
+// quickly, with less overhead than using a mock S3 server.
+fn mock_etag(meta: &std::fs::Metadata) -> Etag {
+    let mtime = meta.modified().expect("Filesystem mtime missing");
+    format!("{}", mtime.duration_since(UNIX_EPOCH).unwrap().as_millis()).into()
 }
 
 #[cfg(test)]
@@ -909,13 +930,18 @@ mod fs_tests {
         // No delimiter: should recursively list everything
         let (storage, cancel) = create_storage()?;
         let child = upload_dummy_file(&storage, "grandparent/parent/child", None, &cancel).await?;
+        let child_sibling =
+            upload_dummy_file(&storage, "grandparent/parent/child_sibling", None, &cancel).await?;
         let uncle = upload_dummy_file(&storage, "grandparent/uncle", None, &cancel).await?;
 
         let listing = storage
             .list(None, ListingMode::NoDelimiter, None, &cancel)
             .await?;
         assert!(listing.prefixes.is_empty());
-        assert_eq!(listing.keys, [uncle.clone(), child.clone()].to_vec());
+        assert_eq!(
+            listing.keys.into_iter().collect::<HashSet<_>>(),
+            HashSet::from([uncle.clone(), child.clone(), child_sibling.clone()])
+        );
 
         // Delimiter: should only go one deep
         let listing = storage
@@ -928,7 +954,25 @@ mod fs_tests {
         );
         assert!(listing.keys.is_empty());
 
-        // Delimiter & prefix
+        // Delimiter & prefix with a trailing slash
+        let listing = storage
+            .list(
+                Some(&RemotePath::from_string("timelines/some_timeline/grandparent/").unwrap()),
+                ListingMode::WithDelimiter,
+                None,
+                &cancel,
+            )
+            .await?;
+        assert_eq!(
+            listing.keys,
+            [RemotePath::from_string("uncle").unwrap()].to_vec()
+        );
+        assert_eq!(
+            listing.prefixes,
+            [RemotePath::from_string("parent").unwrap()].to_vec()
+        );
+
+        // Delimiter and prefix without a trailing slash
         let listing = storage
             .list(
                 Some(&RemotePath::from_string("timelines/some_timeline/grandparent").unwrap()),
@@ -937,12 +981,66 @@ mod fs_tests {
                 &cancel,
             )
             .await?;
+        assert_eq!(listing.keys, [].to_vec());
         assert_eq!(
             listing.prefixes,
-            [RemotePath::from_string("timelines/some_timeline/grandparent/parent").unwrap()]
-                .to_vec()
+            [RemotePath::from_string("grandparent").unwrap()].to_vec()
+        );
+
+        // Delimiter and prefix that's partway through a path component
+        let listing = storage
+            .list(
+                Some(&RemotePath::from_string("timelines/some_timeline/grandp").unwrap()),
+                ListingMode::WithDelimiter,
+                None,
+                &cancel,
+            )
+            .await?;
+        assert_eq!(listing.keys, [].to_vec());
+        assert_eq!(
+            listing.prefixes,
+            [RemotePath::from_string("grandparent").unwrap()].to_vec()
+        );
+
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn list_part_component() -> anyhow::Result<()> {
+        // No delimiter: should recursively list everything
+        let (storage, cancel) = create_storage()?;
+
+        // Imitates what happens in a tenant path when we have an unsharded path and a sharded path, and do a listing
+        // of the unsharded path: although there is a "directory" at the unsharded path, it should be handled as
+        // a freeform prefix.
+        let _child_a =
+            upload_dummy_file(&storage, "grandparent/tenant-01/child", None, &cancel).await?;
+        let _child_b =
+            upload_dummy_file(&storage, "grandparent/tenant/child", None, &cancel).await?;
+
+        // Delimiter and prefix that's partway through a path component
+        let listing = storage
+            .list(
+                Some(
+                    &RemotePath::from_string("timelines/some_timeline/grandparent/tenant").unwrap(),
+                ),
+                ListingMode::WithDelimiter,
+                None,
+                &cancel,
+            )
+            .await?;
+        assert_eq!(listing.keys, [].to_vec());
+
+        let mut found_prefixes = listing.prefixes.clone();
+        found_prefixes.sort();
+        assert_eq!(
+            found_prefixes,
+            [
+                RemotePath::from_string("tenant").unwrap(),
+                RemotePath::from_string("tenant-01").unwrap(),
+            ]
+            .to_vec()
         );
-        assert_eq!(listing.keys, [uncle.clone()].to_vec());
 
         Ok(())
     }

libs/remote_storage/src/s3_bucket.rs

@@ -30,13 +30,13 @@ use aws_sdk_s3::{
     config::{AsyncSleep, Builder, IdentityCache, Region, SharedAsyncSleep},
     error::SdkError,
     operation::get_object::GetObjectError,
-    types::{Delete, DeleteMarkerEntry, ObjectIdentifier, ObjectVersion},
+    types::{Delete, DeleteMarkerEntry, ObjectIdentifier, ObjectVersion, StorageClass},
     Client,
 };
 use aws_smithy_async::rt::sleep::TokioSleep;
 
-use aws_smithy_types::byte_stream::ByteStream;
 use aws_smithy_types::{body::SdkBody, DateTime};
+use aws_smithy_types::{byte_stream::ByteStream, date_time::ConversionError};
 use bytes::Bytes;
 use futures::stream::Stream;
 use hyper::Body;
@@ -62,6 +62,7 @@ pub struct S3Bucket {
     bucket_name: String,
     prefix_in_bucket: Option<String>,
     max_keys_per_list_response: Option<i32>,
+    upload_storage_class: Option<StorageClass>,
     concurrency_limiter: ConcurrencyLimiter,
     // Per-request timeout. Accessible for tests.
     pub timeout: Duration,
@@ -154,6 +155,7 @@ impl S3Bucket {
             max_keys_per_list_response: aws_config.max_keys_per_list_response,
             prefix_in_bucket,
             concurrency_limiter: ConcurrencyLimiter::new(aws_config.concurrency_limit.get()),
+            upload_storage_class: aws_config.upload_storage_class.clone(),
             timeout,
         })
     }
@@ -178,10 +180,7 @@ impl S3Bucket {
 
     pub fn relative_path_to_s3_object(&self, path: &RemotePath) -> String {
         assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
-        let path_string = path
-            .get_path()
-            .as_str()
-            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR);
+        let path_string = path.get_path().as_str();
         match &self.prefix_in_bucket {
             Some(prefix) => prefix.clone() + "/" + path_string,
             None => path_string.to_string(),
@@ -287,8 +286,17 @@ impl S3Bucket {
         let remaining = self.timeout.saturating_sub(started_at.elapsed());
 
         let metadata = object_output.metadata().cloned().map(StorageMetadata);
-        let etag = object_output.e_tag;
-        let last_modified = object_output.last_modified.and_then(|t| t.try_into().ok());
+        let etag = object_output
+            .e_tag
+            .ok_or(DownloadError::Other(anyhow::anyhow!("Missing ETag header")))?
+            .into();
+        let last_modified = object_output
+            .last_modified
+            .ok_or(DownloadError::Other(anyhow::anyhow!(
+                "Missing LastModified header"
+            )))?
+            .try_into()
+            .map_err(|e: ConversionError| DownloadError::Other(e.into()))?;
 
         let body = object_output.body;
         let body = ByteStreamAsStream::from(body);
@@ -462,16 +470,11 @@ impl RemoteStorage for S3Bucket {
         // get the passed prefix or if it is not set use prefix_in_bucket value
         let list_prefix = prefix
             .map(|p| self.relative_path_to_s3_object(p))
-            .or_else(|| self.prefix_in_bucket.clone())
-            .map(|mut p| {
-                // required to end with a separator
-                // otherwise request will return only the entry of a prefix
-                if matches!(mode, ListingMode::WithDelimiter)
-                    && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
-                {
-                    p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
-                }
-                p
+            .or_else(|| {
+                self.prefix_in_bucket.clone().map(|mut s| {
+                    s.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
+                    s
+                })
             });
 
         let _permit = self.permit(kind, cancel).await?;
@@ -540,11 +543,15 @@ impl RemoteStorage for S3Bucket {
                 }
             }
 
-            result.prefixes.extend(
-                prefixes
-                    .iter()
-                    .filter_map(|o| Some(self.s3_object_to_relative_path(o.prefix()?))),
-            );
+            // S3 gives us prefixes like "foo/", we return them like "foo"
+            result.prefixes.extend(prefixes.iter().filter_map(|o| {
+                Some(
+                    self.s3_object_to_relative_path(
+                        o.prefix()?
+                            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR),
+                    ),
+                )
+            }));
 
             continuation_token = match response.next_continuation_token {
                 Some(new_token) => Some(new_token),
@@ -577,6 +584,7 @@ impl RemoteStorage for S3Bucket {
             .bucket(self.bucket_name.clone())
             .key(self.relative_path_to_s3_object(to))
             .set_metadata(metadata.map(|m| m.0))
+            .set_storage_class(self.upload_storage_class.clone())
             .content_length(from_size_bytes.try_into()?)
             .body(bytes_stream)
             .send();
@@ -628,6 +636,7 @@ impl RemoteStorage for S3Bucket {
             .copy_object()
             .bucket(self.bucket_name.clone())
             .key(self.relative_path_to_s3_object(to))
+            .set_storage_class(self.upload_storage_class.clone())
             .copy_source(copy_source)
             .send();
 
@@ -885,6 +894,7 @@ impl RemoteStorage for S3Bucket {
                                     .copy_object()
                                     .bucket(self.bucket_name.clone())
                                     .key(key)
+                                    .set_storage_class(self.upload_storage_class.clone())
                                     .copy_source(&source_id)
                                     .send();
 
@@ -1041,22 +1051,22 @@ mod tests {
             Some("/test/prefix/"),
         ];
         let expected_outputs = [
-            vec!["", "some/path", "some/path"],
-            vec!["/", "/some/path", "/some/path"],
+            vec!["", "some/path", "some/path/"],
+            vec!["/", "/some/path", "/some/path/"],
             vec![
                 "test/prefix/",
                 "test/prefix/some/path",
-                "test/prefix/some/path",
+                "test/prefix/some/path/",
             ],
             vec![
                 "test/prefix/",
                 "test/prefix/some/path",
-                "test/prefix/some/path",
+                "test/prefix/some/path/",
             ],
             vec![
                 "test/prefix/",
                 "test/prefix/some/path",
-                "test/prefix/some/path",
+                "test/prefix/some/path/",
             ],
         ];
 
@@ -1068,6 +1078,7 @@ mod tests {
                 endpoint: None,
                 concurrency_limit: NonZeroUsize::new(100).unwrap(),
                 max_keys_per_list_response: Some(5),
+                upload_storage_class: None,
             };
             let storage =
                 S3Bucket::new(&config, std::time::Duration::ZERO).expect("remote storage init");

libs/remote_storage/src/simulate_failures.rs

@@ -107,27 +107,6 @@ impl UnreliableWrapper {
 type VoidStorage = crate::LocalFs;
 
 impl RemoteStorage for UnreliableWrapper {
-    async fn list_prefixes(
-        &self,
-        prefix: Option<&RemotePath>,
-        cancel: &CancellationToken,
-    ) -> Result<Vec<RemotePath>, DownloadError> {
-        self.attempt(RemoteOp::ListPrefixes(prefix.cloned()))
-            .map_err(DownloadError::Other)?;
-        self.inner.list_prefixes(prefix, cancel).await
-    }
-
-    async fn list_files(
-        &self,
-        folder: Option<&RemotePath>,
-        max_keys: Option<NonZeroU32>,
-        cancel: &CancellationToken,
-    ) -> Result<Vec<RemotePath>, DownloadError> {
-        self.attempt(RemoteOp::ListPrefixes(folder.cloned()))
-            .map_err(DownloadError::Other)?;
-        self.inner.list_files(folder, max_keys, cancel).await
-    }
-
     async fn list(
         &self,
         prefix: Option<&RemotePath>,

libs/remote_storage/tests/common/tests.rs

@@ -1,5 +1,6 @@
 use anyhow::Context;
 use camino::Utf8Path;
+use remote_storage::ListingMode;
 use remote_storage::RemotePath;
 use std::sync::Arc;
 use std::{collections::HashSet, num::NonZeroU32};
@@ -54,9 +55,9 @@ async fn pagination_should_work(ctx: &mut MaybeEnabledStorageWithTestBlobs) -> a
     let base_prefix = RemotePath::new(Utf8Path::new(ctx.enabled.base_prefix))
         .context("common_prefix construction")?;
     let root_remote_prefixes = test_client
-        .list_prefixes(None, &cancel)
-        .await
-        .context("client list root prefixes failure")?
+        .list(None, ListingMode::WithDelimiter, None, &cancel)
+        .await?
+        .prefixes
         .into_iter()
         .collect::<HashSet<_>>();
     assert_eq!(
@@ -65,9 +66,14 @@ async fn pagination_should_work(ctx: &mut MaybeEnabledStorageWithTestBlobs) -> a
     );
 
     let nested_remote_prefixes = test_client
-        .list_prefixes(Some(&base_prefix), &cancel)
-        .await
-        .context("client list nested prefixes failure")?
+        .list(
+            Some(&base_prefix.add_trailing_slash()),
+            ListingMode::WithDelimiter,
+            None,
+            &cancel,
+        )
+        .await?
+        .prefixes
         .into_iter()
         .collect::<HashSet<_>>();
     let remote_only_prefixes = nested_remote_prefixes
@@ -90,11 +96,13 @@ async fn pagination_should_work(ctx: &mut MaybeEnabledStorageWithTestBlobs) -> a
 ///
 /// 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`
+///    1. `list(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
+///    2. `list("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
 #[test_context(MaybeEnabledStorageWithSimpleTestBlobs)]
 #[tokio::test]
-async fn list_files_works(ctx: &mut MaybeEnabledStorageWithSimpleTestBlobs) -> anyhow::Result<()> {
+async fn list_no_delimiter_works(
+    ctx: &mut MaybeEnabledStorageWithSimpleTestBlobs,
+) -> anyhow::Result<()> {
     let ctx = match ctx {
         MaybeEnabledStorageWithSimpleTestBlobs::Enabled(ctx) => ctx,
         MaybeEnabledStorageWithSimpleTestBlobs::Disabled => return Ok(()),
@@ -107,29 +115,36 @@ async fn list_files_works(ctx: &mut MaybeEnabledStorageWithSimpleTestBlobs) -> a
     let base_prefix =
         RemotePath::new(Utf8Path::new("folder1")).context("common_prefix construction")?;
     let root_files = test_client
-        .list_files(None, None, &cancel)
+        .list(None, ListingMode::NoDelimiter, None, &cancel)
         .await
         .context("client list root files failure")?
+        .keys
         .into_iter()
         .collect::<HashSet<_>>();
     assert_eq!(
         root_files,
         ctx.remote_blobs.clone(),
-        "remote storage list_files on root mismatches with the uploads."
+        "remote storage list on root mismatches with the uploads."
     );
 
     // Test that max_keys limit works. In total there are about 21 files (see
     // upload_simple_remote_data call in test_real_s3.rs).
     let limited_root_files = test_client
-        .list_files(None, Some(NonZeroU32::new(2).unwrap()), &cancel)
+        .list(
+            None,
+            ListingMode::NoDelimiter,
+            Some(NonZeroU32::new(2).unwrap()),
+            &cancel,
+        )
         .await
         .context("client list root files failure")?;
-    assert_eq!(limited_root_files.len(), 2);
+    assert_eq!(limited_root_files.keys.len(), 2);
 
     let nested_remote_files = test_client
-        .list_files(Some(&base_prefix), None, &cancel)
+        .list(Some(&base_prefix), ListingMode::NoDelimiter, None, &cancel)
         .await
         .context("client list nested files failure")?
+        .keys
         .into_iter()
         .collect::<HashSet<_>>();
     let trim_remote_blobs: HashSet<_> = ctx
@@ -141,7 +156,7 @@ async fn list_files_works(ctx: &mut MaybeEnabledStorageWithSimpleTestBlobs) -> a
         .collect();
     assert_eq!(
         nested_remote_files, trim_remote_blobs,
-        "remote storage list_files on subdirrectory mismatches with the uploads."
+        "remote storage list on subdirrectory mismatches with the uploads."
     );
     Ok(())
 }
@@ -199,7 +214,11 @@ async fn delete_objects_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<(
 
     ctx.client.delete_objects(&[path1, path2], &cancel).await?;
 
-    let prefixes = ctx.client.list_prefixes(None, &cancel).await?;
+    let prefixes = ctx
+        .client
+        .list(None, ListingMode::WithDelimiter, None, &cancel)
+        .await?
+        .prefixes;
 
     assert_eq!(prefixes.len(), 1);
 

libs/remote_storage/tests/test_real_azure.rs

@@ -57,7 +57,6 @@ enum MaybeEnabledStorage {
     Disabled,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorage {
     async fn setup() -> Self {
         ensure_logging_ready();
@@ -86,7 +85,6 @@ struct AzureWithTestBlobs {
     remote_blobs: HashSet<RemotePath>,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorageWithTestBlobs {
     async fn setup() -> Self {
         ensure_logging_ready();
@@ -134,10 +132,6 @@ impl AsyncTestContext for MaybeEnabledStorageWithTestBlobs {
     }
 }
 
-// NOTE: the setups for the list_prefixes test and the list_files test are very similar
-// However, they are not idential. The list_prefixes function is concerned with listing prefixes,
-// whereas the list_files function is concerned with listing files.
-// See `RemoteStorage::list_files` documentation for more details
 enum MaybeEnabledStorageWithSimpleTestBlobs {
     Enabled(AzureWithSimpleTestBlobs),
     Disabled,
@@ -148,7 +142,6 @@ struct AzureWithSimpleTestBlobs {
     remote_blobs: HashSet<RemotePath>,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorageWithSimpleTestBlobs {
     async fn setup() -> Self {
         ensure_logging_ready();

libs/remote_storage/tests/test_real_s3.rs

@@ -12,8 +12,8 @@ use anyhow::Context;
 use camino::Utf8Path;
 use futures_util::StreamExt;
 use remote_storage::{
-    DownloadError, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
-    S3Config,
+    DownloadError, GenericRemoteStorage, ListingMode, RemotePath, RemoteStorageConfig,
+    RemoteStorageKind, S3Config,
 };
 use test_context::test_context;
 use test_context::AsyncTestContext;
@@ -75,11 +75,14 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
         client: &Arc<GenericRemoteStorage>,
         cancel: &CancellationToken,
     ) -> anyhow::Result<HashSet<RemotePath>> {
-        Ok(retry(|| client.list_files(None, None, cancel))
-            .await
-            .context("list root files failure")?
-            .into_iter()
-            .collect::<HashSet<_>>())
+        Ok(
+            retry(|| client.list(None, ListingMode::NoDelimiter, None, cancel))
+                .await
+                .context("list root files failure")?
+                .keys
+                .into_iter()
+                .collect::<HashSet<_>>(),
+        )
     }
 
     let cancel = CancellationToken::new();
@@ -118,7 +121,7 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     // A little check to ensure that our clock is not too far off from the S3 clock
     {
         let dl = retry(|| ctx.client.download(&path2, &cancel)).await?;
-        let last_modified = dl.last_modified.unwrap();
+        let last_modified = dl.last_modified;
         let half_wt = WAIT_TIME.mul_f32(0.5);
         let t0_hwt = t0 + half_wt;
         let t1_hwt = t1 - half_wt;
@@ -219,7 +222,6 @@ enum MaybeEnabledStorage {
     Disabled,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorage {
     async fn setup() -> Self {
         ensure_logging_ready();
@@ -248,7 +250,6 @@ struct S3WithTestBlobs {
     remote_blobs: HashSet<RemotePath>,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorageWithTestBlobs {
     async fn setup() -> Self {
         ensure_logging_ready();
@@ -296,10 +297,6 @@ impl AsyncTestContext for MaybeEnabledStorageWithTestBlobs {
     }
 }
 
-// NOTE: the setups for the list_prefixes test and the list_files test are very similar
-// However, they are not idential. The list_prefixes function is concerned with listing prefixes,
-// whereas the list_files function is concerned with listing files.
-// See `RemoteStorage::list_files` documentation for more details
 enum MaybeEnabledStorageWithSimpleTestBlobs {
     Enabled(S3WithSimpleTestBlobs),
     Disabled,
@@ -310,7 +307,6 @@ struct S3WithSimpleTestBlobs {
     remote_blobs: HashSet<RemotePath>,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorageWithSimpleTestBlobs {
     async fn setup() -> Self {
         ensure_logging_ready();
@@ -384,6 +380,7 @@ fn create_s3_client(
             endpoint: None,
             concurrency_limit: NonZeroUsize::new(100).unwrap(),
             max_keys_per_list_response,
+            upload_storage_class: None,
         }),
         timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
     };

libs/tenant_size_model/tests/tests.rs

@@ -247,7 +247,7 @@ fn scenario_4() {
     //
     // This is in total 5000 + 1000 + 5000 + 1000 = 12000
     //
-    // (If we used the the method from the previous scenario, and
+    // (If we used the method from the previous scenario, and
     // kept only snapshot at the branch point, we'd need to keep
     // all the WAL between 10000-18000 on the main branch, so
     // the total size would be 5000 + 1000 + 8000 = 14000. The

libs/utils/Cargo.toml

@@ -13,6 +13,7 @@ testing = ["fail/failpoints"]
 [dependencies]
 arc-swap.workspace = true
 sentry.workspace = true
+async-compression.workspace = true
 async-trait.workspace = true
 anyhow.workspace = true
 bincode.workspace = true
@@ -21,6 +22,7 @@ camino.workspace = true
 chrono.workspace = true
 heapless.workspace = true
 hex = { workspace = true, features = ["serde"] }
+humantime.workspace = true
 hyper = { workspace = true, features = ["full"] }
 fail.workspace = true
 futures = { workspace = true}
@@ -36,6 +38,7 @@ serde_json.workspace = true
 signal-hook.workspace = true
 thiserror.workspace = true
 tokio.workspace = true
+tokio-tar.workspace = true
 tokio-util.workspace = true
 tracing.workspace = true
 tracing-error.workspace = true
@@ -46,6 +49,7 @@ strum.workspace = true
 strum_macros.workspace = true
 url.workspace = true
 uuid.workspace = true
+walkdir.workspace = true
 
 pq_proto.workspace = true
 postgres_connection.workspace = true

libs/utils/src/env.rs

@@ -0,0 +1,21 @@
+//! Wrapper around `std::env::var` for parsing environment variables.
+
+use std::{fmt::Display, str::FromStr};
+
+pub fn var<V, E>(varname: &str) -> Option<V>
+where
+    V: FromStr<Err = E>,
+    E: Display,
+{
+    match std::env::var(varname) {
+        Ok(s) => Some(
+            s.parse()
+                .map_err(|e| format!("failed to parse env var {varname}: {e:#}"))
+                .unwrap(),
+        ),
+        Err(std::env::VarError::NotPresent) => None,
+        Err(std::env::VarError::NotUnicode(_)) => {
+            panic!("env var {varname} is not unicode")
+        }
+    }
+}

libs/utils/src/generation.rs

@@ -34,6 +34,8 @@ pub enum Generation {
 /// scenarios where pageservers might otherwise issue conflicting writes to
 /// remote storage
 impl Generation {
+    pub const MAX: Self = Self::Valid(u32::MAX);
+
     /// Create a new Generation that represents a legacy key format with
     /// no generation suffix
     pub fn none() -> Self {

libs/utils/src/history_buffer.rs

@@ -47,9 +47,10 @@ impl<T, const L: usize> ops::Deref for HistoryBufferWithDropCounter<T, L> {
     }
 }
 
-#[derive(serde::Serialize)]
+#[derive(serde::Serialize, serde::Deserialize)]
 struct SerdeRepr<T> {
     buffer: Vec<T>,
+    buffer_size: usize,
     drop_count: u64,
 }
 
@@ -61,6 +62,7 @@ where
         let HistoryBufferWithDropCounter { buffer, drop_count } = value;
         SerdeRepr {
             buffer: buffer.iter().cloned().collect(),
+            buffer_size: L,
             drop_count: *drop_count,
         }
     }
@@ -78,19 +80,52 @@ where
     }
 }
 
+impl<'de, T, const L: usize> serde::de::Deserialize<'de> for HistoryBufferWithDropCounter<T, L>
+where
+    T: Clone + serde::Deserialize<'de>,
+{
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        let SerdeRepr {
+            buffer: des_buffer,
+            drop_count,
+            buffer_size,
+        } = SerdeRepr::<T>::deserialize(deserializer)?;
+        if buffer_size != L {
+            use serde::de::Error;
+            return Err(D::Error::custom(format!(
+                "invalid buffer_size, expecting {L} got {buffer_size}"
+            )));
+        }
+        let mut buffer = HistoryBuffer::new();
+        buffer.extend(des_buffer);
+        Ok(HistoryBufferWithDropCounter { buffer, drop_count })
+    }
+}
+
 #[cfg(test)]
 mod test {
     use super::HistoryBufferWithDropCounter;
 
     #[test]
     fn test_basics() {
-        let mut b = HistoryBufferWithDropCounter::<_, 2>::default();
+        let mut b = HistoryBufferWithDropCounter::<usize, 2>::default();
         b.write(1);
         b.write(2);
         b.write(3);
         assert!(b.iter().any(|e| *e == 2));
         assert!(b.iter().any(|e| *e == 3));
         assert!(!b.iter().any(|e| *e == 1));
+
+        // round-trip serde
+        let round_tripped: HistoryBufferWithDropCounter<usize, 2> =
+            serde_json::from_str(&serde_json::to_string(&b).unwrap()).unwrap();
+        assert_eq!(
+            round_tripped.iter().cloned().collect::<Vec<_>>(),
+            b.iter().cloned().collect::<Vec<_>>()
+        );
     }
 
     #[test]

libs/utils/src/http/endpoint.rs

@@ -245,7 +245,7 @@ impl std::io::Write for ChannelWriter {
     }
 }
 
-async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
     SERVE_METRICS_COUNT.inc();
 
     let started_at = std::time::Instant::now();
@@ -367,7 +367,6 @@ pub fn make_router() -> RouterBuilder<hyper::Body, ApiError> {
         .middleware(Middleware::post_with_info(
             add_request_id_header_to_response,
         ))
-        .get("/metrics", |r| request_span(r, prometheus_metrics_handler))
         .err_handler(route_error_handler)
 }
 

libs/utils/src/lib.rs

@@ -63,6 +63,7 @@ pub mod measured_stream;
 
 pub mod serde_percent;
 pub mod serde_regex;
+pub mod serde_system_time;
 
 pub mod pageserver_feedback;
 
@@ -87,6 +88,12 @@ pub mod failpoint_support;
 
 pub mod yielding_loop;
 
+pub mod zstd;
+
+pub mod env;
+
+pub mod poison;
+
 /// This is a shortcut to embed git sha into binaries and avoid copying the same build script to all packages
 ///
 /// we have several cases:

libs/utils/src/lock_file.rs

@@ -63,6 +63,7 @@ impl UnwrittenLockFile {
 pub fn create_exclusive(lock_file_path: &Utf8Path) -> anyhow::Result<UnwrittenLockFile> {
     let lock_file = fs::OpenOptions::new()
         .create(true) // O_CREAT
+        .truncate(true)
         .write(true)
         .open(lock_file_path)
         .context("open lock file")?;

libs/utils/src/pageserver_feedback.rs

@@ -29,12 +29,10 @@ pub struct PageserverFeedback {
     // Serialize with RFC3339 format.
     #[serde(with = "serde_systemtime")]
     pub replytime: SystemTime,
+    /// Used to track feedbacks from different shards. Always zero for unsharded tenants.
+    pub shard_number: u32,
 }
 
-// NOTE: Do not forget to increment this number when adding new fields to PageserverFeedback.
-// Do not remove previously available fields because this might be backwards incompatible.
-pub const PAGESERVER_FEEDBACK_FIELDS_NUMBER: u8 = 5;
-
 impl PageserverFeedback {
     pub fn empty() -> PageserverFeedback {
         PageserverFeedback {
@@ -43,6 +41,7 @@ impl PageserverFeedback {
             remote_consistent_lsn: Lsn::INVALID,
             disk_consistent_lsn: Lsn::INVALID,
             replytime: *PG_EPOCH,
+            shard_number: 0,
         }
     }
 
@@ -59,17 +58,26 @@ impl PageserverFeedback {
     //
     // TODO: change serialized fields names once all computes migrate to rename.
     pub fn serialize(&self, buf: &mut BytesMut) {
-        buf.put_u8(PAGESERVER_FEEDBACK_FIELDS_NUMBER); // # of keys
+        let buf_ptr = buf.len();
+        buf.put_u8(0); // # of keys, will be filled later
+        let mut nkeys = 0;
+
+        nkeys += 1;
         buf.put_slice(b"current_timeline_size\0");
         buf.put_i32(8);
         buf.put_u64(self.current_timeline_size);
 
+        nkeys += 1;
         buf.put_slice(b"ps_writelsn\0");
         buf.put_i32(8);
         buf.put_u64(self.last_received_lsn.0);
+
+        nkeys += 1;
         buf.put_slice(b"ps_flushlsn\0");
         buf.put_i32(8);
         buf.put_u64(self.disk_consistent_lsn.0);
+
+        nkeys += 1;
         buf.put_slice(b"ps_applylsn\0");
         buf.put_i32(8);
         buf.put_u64(self.remote_consistent_lsn.0);
@@ -80,9 +88,19 @@ impl PageserverFeedback {
             .expect("failed to serialize pg_replytime earlier than PG_EPOCH")
             .as_micros() as i64;
 
+        nkeys += 1;
         buf.put_slice(b"ps_replytime\0");
         buf.put_i32(8);
         buf.put_i64(timestamp);
+
+        if self.shard_number > 0 {
+            nkeys += 1;
+            buf.put_slice(b"shard_number\0");
+            buf.put_i32(4);
+            buf.put_u32(self.shard_number);
+        }
+
+        buf[buf_ptr] = nkeys;
     }
 
     // Deserialize PageserverFeedback message
@@ -125,9 +143,8 @@ impl PageserverFeedback {
                 }
                 b"shard_number" => {
                     let len = buf.get_i32();
-                    // TODO: this will be implemented in the next update,
-                    //  for now, we just skip the value.
-                    buf.advance(len as usize);
+                    assert_eq!(len, 4);
+                    rf.shard_number = buf.get_u32();
                 }
                 _ => {
                     let len = buf.get_i32();
@@ -200,10 +217,7 @@ mod tests {
         rf.serialize(&mut data);
 
         // Add an extra field to the buffer and adjust number of keys
-        if let Some(first) = data.first_mut() {
-            *first = PAGESERVER_FEEDBACK_FIELDS_NUMBER + 1;
-        }
-
+        data[0] += 1;
         data.put_slice(b"new_field_one\0");
         data.put_i32(8);
         data.put_u64(42);

libs/utils/src/poison.rs

@@ -0,0 +1,121 @@
+//!  Protect a piece of state from reuse after it is left in an inconsistent state.
+//!
+//!  # Example
+//!
+//!  ```
+//!  # tokio_test::block_on(async {
+//!  use utils::poison::Poison;
+//!  use std::time::Duration;
+//!
+//!  struct State {
+//!    clean: bool,
+//!  }
+//!  let state = tokio::sync::Mutex::new(Poison::new("mystate", State { clean: true }));
+//!
+//!  let mut mutex_guard = state.lock().await;
+//!  let mut poison_guard = mutex_guard.check_and_arm()?;
+//!  let state = poison_guard.data_mut();
+//!  state.clean = false;
+//!  // If we get cancelled at this await point, subsequent check_and_arm() calls will fail.
+//!  tokio::time::sleep(Duration::from_secs(10)).await;
+//!  state.clean = true;
+//!  poison_guard.disarm();
+//!  # Ok::<(), utils::poison::Error>(())
+//!  # });
+//!  ```
+
+use tracing::warn;
+
+pub struct Poison<T> {
+    what: &'static str,
+    state: State,
+    data: T,
+}
+
+#[derive(Clone, Copy)]
+enum State {
+    Clean,
+    Armed,
+    Poisoned { at: chrono::DateTime<chrono::Utc> },
+}
+
+impl<T> Poison<T> {
+    /// We log `what` `warning!` level if the [`Guard`] gets dropped without being [`Guard::disarm`]ed.
+    pub fn new(what: &'static str, data: T) -> Self {
+        Self {
+            what,
+            state: State::Clean,
+            data,
+        }
+    }
+
+    /// Check for poisoning and return a [`Guard`] that provides access to the wrapped state.
+    pub fn check_and_arm(&mut self) -> Result<Guard<T>, Error> {
+        match self.state {
+            State::Clean => {
+                self.state = State::Armed;
+                Ok(Guard(self))
+            }
+            State::Armed => unreachable!("transient state"),
+            State::Poisoned { at } => Err(Error::Poisoned {
+                what: self.what,
+                at,
+            }),
+        }
+    }
+}
+
+/// Use [`Self::data`] and [`Self::data_mut`] to access the wrapped state.
+/// Once modifications are done, use [`Self::disarm`].
+/// If [`Guard`] gets dropped instead of calling [`Self::disarm`], the state is poisoned
+/// and subsequent calls to [`Poison::check_and_arm`] will fail with an error.
+pub struct Guard<'a, T>(&'a mut Poison<T>);
+
+impl<'a, T> Guard<'a, T> {
+    pub fn data(&self) -> &T {
+        &self.0.data
+    }
+    pub fn data_mut(&mut self) -> &mut T {
+        &mut self.0.data
+    }
+
+    pub fn disarm(self) {
+        match self.0.state {
+            State::Clean => unreachable!("we set it to Armed in check_and_arm()"),
+            State::Armed => {
+                self.0.state = State::Clean;
+            }
+            State::Poisoned { at } => {
+                unreachable!("we fail check_and_arm() if it's in that state: {at}")
+            }
+        }
+    }
+}
+
+impl<'a, T> Drop for Guard<'a, T> {
+    fn drop(&mut self) {
+        match self.0.state {
+            State::Clean => {
+                // set by disarm()
+            }
+            State::Armed => {
+                // still armed => poison it
+                let at = chrono::Utc::now();
+                self.0.state = State::Poisoned { at };
+                warn!(at=?at, "poisoning {}", self.0.what);
+            }
+            State::Poisoned { at } => {
+                unreachable!("we fail check_and_arm() if it's in that state: {at}")
+            }
+        }
+    }
+}
+
+#[derive(thiserror::Error, Debug)]
+pub enum Error {
+    #[error("poisoned at {at}: {what}")]
+    Poisoned {
+        what: &'static str,
+        at: chrono::DateTime<chrono::Utc>,
+    },
+}

libs/utils/src/seqwait.rs

@@ -2,11 +2,10 @@
 
 use std::cmp::{Eq, Ordering};
 use std::collections::BinaryHeap;
-use std::fmt::Debug;
 use std::mem;
 use std::sync::Mutex;
 use std::time::Duration;
-use tokio::sync::watch::{channel, Receiver, Sender};
+use tokio::sync::watch::{self, channel};
 use tokio::time::timeout;
 
 /// An error happened while waiting for a number
@@ -35,23 +34,73 @@ pub trait MonotonicCounter<V> {
     fn cnt_value(&self) -> V;
 }
 
-/// Internal components of a `SeqWait`
-struct SeqWaitInt<S, V>
+/// Heap of waiters, lowest numbers pop first.
+struct Waiters<V>
 where
-    S: MonotonicCounter<V>,
     V: Ord,
 {
-    waiters: BinaryHeap<Waiter<V>>,
-    current: S,
-    shutdown: bool,
+    heap: BinaryHeap<Waiter<V>>,
+    /// Number of the first waiter in the heap, or None if there are no waiters.
+    status_channel: watch::Sender<Option<V>>,
+}
+
+impl<V> Waiters<V>
+where
+    V: Ord + Copy,
+{
+    fn new() -> Self {
+        Waiters {
+            heap: BinaryHeap::new(),
+            status_channel: channel(None).0,
+        }
+    }
+
+    /// `status_channel` contains the number of the first waiter in the heap.
+    /// This function should be called whenever waiters heap changes.
+    fn update_status(&self) {
+        let first_waiter = self.heap.peek().map(|w| w.wake_num);
+        let _ = self.status_channel.send_replace(first_waiter);
+    }
+
+    /// Add new waiter to the heap, return a channel that will be notified when the number arrives.
+    fn add(&mut self, num: V) -> watch::Receiver<()> {
+        let (tx, rx) = channel(());
+        self.heap.push(Waiter {
+            wake_num: num,
+            wake_channel: tx,
+        });
+        self.update_status();
+        rx
+    }
+
+    /// Pop all waiters <= num from the heap. Collect channels in a vector,
+    /// so that caller can wake them up.
+    fn pop_leq(&mut self, num: V) -> Vec<watch::Sender<()>> {
+        let mut wake_these = Vec::new();
+        while let Some(n) = self.heap.peek() {
+            if n.wake_num > num {
+                break;
+            }
+            wake_these.push(self.heap.pop().unwrap().wake_channel);
+        }
+        self.update_status();
+        wake_these
+    }
+
+    /// Used on shutdown to efficiently drop all waiters.
+    fn take_all(&mut self) -> BinaryHeap<Waiter<V>> {
+        let heap = mem::take(&mut self.heap);
+        self.update_status();
+        heap
+    }
 }
 
 struct Waiter<T>
 where
     T: Ord,
 {
-    wake_num: T,              // wake me when this number arrives ...
-    wake_channel: Sender<()>, // ... by sending a message to this channel
+    wake_num: T,                     // wake me when this number arrives ...
+    wake_channel: watch::Sender<()>, // ... by sending a message to this channel
 }
 
 // BinaryHeap is a max-heap, and we want a min-heap. Reverse the ordering here
@@ -76,6 +125,17 @@ impl<T: Ord> PartialEq for Waiter<T> {
 
 impl<T: Ord> Eq for Waiter<T> {}
 
+/// Internal components of a `SeqWait`
+struct SeqWaitInt<S, V>
+where
+    S: MonotonicCounter<V>,
+    V: Ord,
+{
+    waiters: Waiters<V>,
+    current: S,
+    shutdown: bool,
+}
+
 /// A tool for waiting on a sequence number
 ///
 /// This provides a way to wait the arrival of a number.
@@ -108,7 +168,7 @@ where
     /// Create a new `SeqWait`, initialized to a particular number
     pub fn new(starting_num: S) -> Self {
         let internal = SeqWaitInt {
-            waiters: BinaryHeap::new(),
+            waiters: Waiters::new(),
             current: starting_num,
             shutdown: false,
         };
@@ -128,9 +188,8 @@ where
             // Block any future waiters from starting
             internal.shutdown = true;
 
-            // This will steal the entire waiters map.
-            // When we drop it all waiters will be woken.
-            mem::take(&mut internal.waiters)
+            // Take all waiters to drop them later.
+            internal.waiters.take_all()
 
             // Drop the lock as we exit this scope.
         };
@@ -182,9 +241,21 @@ where
         }
     }
 
+    /// Check if [`Self::wait_for`] or [`Self::wait_for_timeout`] would wait if called with `num`.
+    pub fn would_wait_for(&self, num: V) -> Result<(), V> {
+        let internal = self.internal.lock().unwrap();
+        let cnt = internal.current.cnt_value();
+        drop(internal);
+        if cnt >= num {
+            Ok(())
+        } else {
+            Err(cnt)
+        }
+    }
+
     /// Register and return a channel that will be notified when a number arrives,
     /// or None, if it has already arrived.
-    fn queue_for_wait(&self, num: V) -> Result<Option<Receiver<()>>, SeqWaitError> {
+    fn queue_for_wait(&self, num: V) -> Result<Option<watch::Receiver<()>>, SeqWaitError> {
         let mut internal = self.internal.lock().unwrap();
         if internal.current.cnt_value() >= num {
             return Ok(None);
@@ -193,12 +264,8 @@ where
             return Err(SeqWaitError::Shutdown);
         }
 
-        // Create a new channel.
-        let (tx, rx) = channel(());
-        internal.waiters.push(Waiter {
-            wake_num: num,
-            wake_channel: tx,
-        });
+        // Add waiter channel to the queue.
+        let rx = internal.waiters.add(num);
         // Drop the lock as we exit this scope.
         Ok(Some(rx))
     }
@@ -219,16 +286,8 @@ where
             }
             internal.current.cnt_advance(num);
 
-            // Pop all waiters <= num from the heap. Collect them in a vector, and
-            // wake them up after releasing the lock.
-            let mut wake_these = Vec::new();
-            while let Some(n) = internal.waiters.peek() {
-                if n.wake_num > num {
-                    break;
-                }
-                wake_these.push(internal.waiters.pop().unwrap().wake_channel);
-            }
-            wake_these
+            // Pop all waiters <= num from the heap.
+            internal.waiters.pop_leq(num)
         };
 
         for tx in wake_these {
@@ -243,6 +302,23 @@ where
     pub fn load(&self) -> S {
         self.internal.lock().unwrap().current
     }
+
+    /// Get a Receiver for the current status.
+    ///
+    /// The current status is the number of the first waiter in the queue,
+    /// or None if there are no waiters.
+    ///
+    /// This receiver will be notified whenever the status changes.
+    /// It is useful for receiving notifications when the first waiter
+    /// starts waiting for a number, or when there are no more waiters left.
+    pub fn status_receiver(&self) -> watch::Receiver<Option<V>> {
+        self.internal
+            .lock()
+            .unwrap()
+            .waiters
+            .status_channel
+            .subscribe()
+    }
 }
 
 #[cfg(test)]

libs/utils/src/serde_system_time.rs

@@ -0,0 +1,55 @@
+//! A `serde::{Deserialize,Serialize}` type for SystemTime with RFC3339 format and millisecond precision.
+
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, serde::Serialize, serde::Deserialize)]
+#[serde(transparent)]
+pub struct SystemTime(
+    #[serde(
+        deserialize_with = "deser_rfc3339_millis",
+        serialize_with = "ser_rfc3339_millis"
+    )]
+    pub std::time::SystemTime,
+);
+
+fn ser_rfc3339_millis<S: serde::ser::Serializer>(
+    ts: &std::time::SystemTime,
+    serializer: S,
+) -> Result<S::Ok, S::Error> {
+    serializer.collect_str(&humantime::format_rfc3339_millis(*ts))
+}
+
+fn deser_rfc3339_millis<'de, D>(deserializer: D) -> Result<std::time::SystemTime, D::Error>
+where
+    D: serde::de::Deserializer<'de>,
+{
+    let s: String = serde::de::Deserialize::deserialize(deserializer)?;
+    humantime::parse_rfc3339(&s).map_err(serde::de::Error::custom)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Helper function to make a SystemTime have millisecond precision by truncating additional nanoseconds.
+    fn to_millisecond_precision(time: SystemTime) -> SystemTime {
+        match time.0.duration_since(std::time::SystemTime::UNIX_EPOCH) {
+            Ok(duration) => {
+                let total_millis = duration.as_secs() * 1_000 + u64::from(duration.subsec_millis());
+                SystemTime(
+                    std::time::SystemTime::UNIX_EPOCH
+                        + std::time::Duration::from_millis(total_millis),
+                )
+            }
+            Err(_) => time,
+        }
+    }
+
+    #[test]
+    fn test_serialize_deserialize() {
+        let input = SystemTime(std::time::SystemTime::now());
+        let expected_serialized = format!("\"{}\"", humantime::format_rfc3339_millis(input.0));
+        let serialized = serde_json::to_string(&input).unwrap();
+        assert_eq!(expected_serialized, serialized);
+        let deserialized: SystemTime = serde_json::from_str(&expected_serialized).unwrap();
+        assert_eq!(to_millisecond_precision(input), deserialized);
+    }
+}

libs/utils/src/sync/heavier_once_cell.rs

@@ -110,6 +110,49 @@ impl<T> OnceCell<T> {
         }
     }
 
+    /// Returns a guard to an existing initialized value, or returns an unique initialization
+    /// permit which can be used to initialize this `OnceCell` using `OnceCell::set`.
+    pub async fn get_or_init_detached(&self) -> Result<Guard<'_, T>, InitPermit> {
+        // It looks like OnceCell::get_or_init could be implemented using this method instead of
+        // duplication. However, that makes the future be !Send due to possibly holding on to the
+        // MutexGuard over an await point.
+        loop {
+            let sem = {
+                let guard = self.inner.lock().unwrap();
+                if guard.value.is_some() {
+                    return Ok(Guard(guard));
+                }
+                guard.init_semaphore.clone()
+            };
+
+            {
+                let permit = {
+                    // increment the count for the duration of queued
+                    let _guard = CountWaitingInitializers::start(self);
+                    sem.acquire().await
+                };
+
+                let Ok(permit) = permit else {
+                    let guard = self.inner.lock().unwrap();
+                    if !Arc::ptr_eq(&sem, &guard.init_semaphore) {
+                        // there was a take_and_deinit in between
+                        continue;
+                    }
+                    assert!(
+                        guard.value.is_some(),
+                        "semaphore got closed, must be initialized"
+                    );
+                    return Ok(Guard(guard));
+                };
+
+                permit.forget();
+            }
+
+            let permit = InitPermit(sem);
+            return Err(permit);
+        }
+    }
+
     /// Assuming a permit is held after previous call to [`Guard::take_and_deinit`], it can be used
     /// to complete initializing the inner value.
     ///
@@ -149,6 +192,14 @@ impl<T> OnceCell<T> {
         }
     }
 
+    /// Like [`Guard::take_and_deinit`], but will return `None` if this OnceCell was never
+    /// initialized.
+    pub fn take_and_deinit(&mut self) -> Option<(T, InitPermit)> {
+        let inner = self.inner.get_mut().unwrap();
+
+        inner.take_and_deinit()
+    }
+
     /// Return the number of [`Self::get_or_init`] calls waiting for initialization to complete.
     pub fn initializer_count(&self) -> usize {
         self.initializers.load(Ordering::Relaxed)
@@ -202,16 +253,24 @@ impl<'a, T> Guard<'a, T> {
     ///
     /// The permit will be on a semaphore part of the new internal value, and any following
     /// [`OnceCell::get_or_init`] will wait on it to complete.
-    pub fn take_and_deinit(&mut self) -> (T, InitPermit) {
+    pub fn take_and_deinit(mut self) -> (T, InitPermit) {
+        self.0
+            .take_and_deinit()
+            .expect("guard is not created unless value has been initialized")
+    }
+}
+
+impl<T> Inner<T> {
+    pub fn take_and_deinit(&mut self) -> Option<(T, InitPermit)> {
+        let value = self.value.take()?;
+
         let mut swapped = Inner::default();
         let sem = swapped.init_semaphore.clone();
         // acquire and forget right away, moving the control over to InitPermit
         sem.try_acquire().expect("we just created this").forget();
-        std::mem::swap(&mut *self.0, &mut swapped);
-        swapped
-            .value
-            .map(|v| (v, InitPermit(sem)))
-            .expect("guard is not created unless value has been initialized")
+        let permit = InitPermit(sem);
+        std::mem::swap(self, &mut swapped);
+        Some((value, permit))
     }
 }
 
@@ -220,6 +279,13 @@ impl<'a, T> Guard<'a, T> {
 /// On drop, this type will return the permit.
 pub struct InitPermit(Arc<tokio::sync::Semaphore>);
 
+impl std::fmt::Debug for InitPermit {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let ptr = Arc::as_ptr(&self.0) as *const ();
+        f.debug_tuple("InitPermit").field(&ptr).finish()
+    }
+}
+
 impl Drop for InitPermit {
     fn drop(&mut self) {
         assert_eq!(
@@ -481,4 +547,57 @@ mod tests {
 
         assert_eq!("t1", *cell.get().unwrap());
     }
+
+    #[tokio::test(start_paused = true)]
+    async fn detached_init_smoke() {
+        let target = OnceCell::default();
+
+        let Err(permit) = target.get_or_init_detached().await else {
+            unreachable!("it is not initialized")
+        };
+
+        tokio::time::timeout(
+            std::time::Duration::from_secs(3600 * 24 * 7 * 365),
+            target.get_or_init(|permit2| async { Ok::<_, Infallible>((11, permit2)) }),
+        )
+        .await
+        .expect_err("should timeout since we are already holding the permit");
+
+        target.set(42, permit);
+
+        let (_answer, permit) = {
+            let guard = target
+                .get_or_init(|permit| async { Ok::<_, Infallible>((11, permit)) })
+                .await
+                .unwrap();
+
+            assert_eq!(*guard, 42);
+
+            guard.take_and_deinit()
+        };
+
+        assert!(target.get().is_none());
+
+        target.set(11, permit);
+
+        assert_eq!(*target.get().unwrap(), 11);
+    }
+
+    #[tokio::test]
+    async fn take_and_deinit_on_mut() {
+        use std::convert::Infallible;
+
+        let mut target = OnceCell::<u32>::default();
+        assert!(target.take_and_deinit().is_none());
+
+        target
+            .get_or_init(|permit| async move { Ok::<_, Infallible>((42, permit)) })
+            .await
+            .unwrap();
+
+        let again = target.take_and_deinit();
+        assert!(matches!(again, Some((42, _))), "{again:?}");
+
+        assert!(target.take_and_deinit().is_none());
+    }
 }

libs/utils/src/vec_map.rs

@@ -1,27 +1,60 @@
 use std::{alloc::Layout, cmp::Ordering, ops::RangeBounds};
 
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum VecMapOrdering {
+    Greater,
+    GreaterOrEqual,
+}
+
 /// Ordered map datastructure implemented in a Vec.
 /// Append only - can only add keys that are larger than the
 /// current max key.
+/// Ordering can be adjusted using [`VecMapOrdering`]
+/// during `VecMap` construction.
 #[derive(Clone, Debug)]
-pub struct VecMap<K, V>(Vec<(K, V)>);
+pub struct VecMap<K, V> {
+    data: Vec<(K, V)>,
+    ordering: VecMapOrdering,
+}
 
 impl<K, V> Default for VecMap<K, V> {
     fn default() -> Self {
-        VecMap(Default::default())
+        VecMap {
+            data: Default::default(),
+            ordering: VecMapOrdering::Greater,
+        }
     }
 }
 
-#[derive(Debug)]
-pub struct InvalidKey;
+#[derive(thiserror::Error, Debug)]
+pub enum VecMapError {
+    #[error("Key violates ordering constraint")]
+    InvalidKey,
+    #[error("Mismatched ordering constraints")]
+    ExtendOrderingError,
+}
 
 impl<K: Ord, V> VecMap<K, V> {
+    pub fn new(ordering: VecMapOrdering) -> Self {
+        Self {
+            data: Vec::new(),
+            ordering,
+        }
+    }
+
+    pub fn with_capacity(capacity: usize, ordering: VecMapOrdering) -> Self {
+        Self {
+            data: Vec::with_capacity(capacity),
+            ordering,
+        }
+    }
+
     pub fn is_empty(&self) -> bool {
-        self.0.is_empty()
+        self.data.is_empty()
     }
 
     pub fn as_slice(&self) -> &[(K, V)] {
-        self.0.as_slice()
+        self.data.as_slice()
     }
 
     /// This function may panic if given a range where the lower bound is
@@ -29,7 +62,7 @@ impl<K: Ord, V> VecMap<K, V> {
     pub fn slice_range<R: RangeBounds<K>>(&self, range: R) -> &[(K, V)] {
         use std::ops::Bound::*;
 
-        let binary_search = |k: &K| self.0.binary_search_by_key(&k, extract_key);
+        let binary_search = |k: &K| self.data.binary_search_by_key(&k, extract_key);
 
         let start_idx = match range.start_bound() {
             Unbounded => 0,
@@ -41,7 +74,7 @@ impl<K: Ord, V> VecMap<K, V> {
         };
 
         let end_idx = match range.end_bound() {
-            Unbounded => self.0.len(),
+            Unbounded => self.data.len(),
             Included(k) => match binary_search(k) {
                 Ok(idx) => idx + 1,
                 Err(idx) => idx,
@@ -49,34 +82,30 @@ impl<K: Ord, V> VecMap<K, V> {
             Excluded(k) => binary_search(k).unwrap_or_else(std::convert::identity),
         };
 
-        &self.0[start_idx..end_idx]
+        &self.data[start_idx..end_idx]
     }
 
     /// Add a key value pair to the map.
-    /// If `key` is less than or equal to the current maximum key
-    /// the pair will not be added and InvalidKey error will be returned.
-    pub fn append(&mut self, key: K, value: V) -> Result<usize, InvalidKey> {
-        if let Some((last_key, _last_value)) = self.0.last() {
-            if &key <= last_key {
-                return Err(InvalidKey);
-            }
-        }
+    /// If `key` is not respective of the `self` ordering the
+    /// pair will not be added and `InvalidKey` error will be returned.
+    pub fn append(&mut self, key: K, value: V) -> Result<usize, VecMapError> {
+        self.validate_key_order(&key)?;
 
         let delta_size = self.instrument_vec_op(|vec| vec.push((key, value)));
         Ok(delta_size)
     }
 
     /// Update the maximum key value pair or add a new key value pair to the map.
-    /// If `key` is less than the current maximum key no updates or additions
-    /// will occur and InvalidKey error will be returned.
+    /// If `key` is not respective of the `self` ordering no updates or additions
+    /// will occur and `InvalidKey` error will be returned.
     pub fn append_or_update_last(
         &mut self,
         key: K,
         mut value: V,
-    ) -> Result<(Option<V>, usize), InvalidKey> {
-        if let Some((last_key, last_value)) = self.0.last_mut() {
+    ) -> Result<(Option<V>, usize), VecMapError> {
+        if let Some((last_key, last_value)) = self.data.last_mut() {
             match key.cmp(last_key) {
-                Ordering::Less => return Err(InvalidKey),
+                Ordering::Less => return Err(VecMapError::InvalidKey),
                 Ordering::Equal => {
                     std::mem::swap(last_value, &mut value);
                     const DELTA_SIZE: usize = 0;
@@ -100,40 +129,67 @@ impl<K: Ord, V> VecMap<K, V> {
         V: Clone,
     {
         let split_idx = self
-            .0
+            .data
             .binary_search_by_key(&cutoff, extract_key)
             .unwrap_or_else(std::convert::identity);
 
         (
-            VecMap(self.0[..split_idx].to_vec()),
-            VecMap(self.0[split_idx..].to_vec()),
+            VecMap {
+                data: self.data[..split_idx].to_vec(),
+                ordering: self.ordering,
+            },
+            VecMap {
+                data: self.data[split_idx..].to_vec(),
+                ordering: self.ordering,
+            },
         )
     }
 
     /// Move items from `other` to the end of `self`, leaving `other` empty.
-    /// If any keys in `other` is less than or equal to any key in `self`,
-    /// `InvalidKey` error will be returned and no mutation will occur.
-    pub fn extend(&mut self, other: &mut Self) -> Result<usize, InvalidKey> {
-        let self_last_opt = self.0.last().map(extract_key);
-        let other_first_opt = other.0.last().map(extract_key);
+    /// If the `other` ordering is different from `self` ordering
+    /// `ExtendOrderingError` error will be returned.
+    /// If any keys in `other` is not respective of the ordering defined in
+    /// `self`, `InvalidKey` error will be returned and no mutation will occur.
+    pub fn extend(&mut self, other: &mut Self) -> Result<usize, VecMapError> {
+        if self.ordering != other.ordering {
+            return Err(VecMapError::ExtendOrderingError);
+        }
 
-        if let (Some(self_last), Some(other_first)) = (self_last_opt, other_first_opt) {
-            if self_last >= other_first {
-                return Err(InvalidKey);
+        let other_first_opt = other.data.last().map(extract_key);
+        if let Some(other_first) = other_first_opt {
+            self.validate_key_order(other_first)?;
+        }
+
+        let delta_size = self.instrument_vec_op(|vec| vec.append(&mut other.data));
+        Ok(delta_size)
+    }
+
+    /// Validate the current last key in `self` and key being
+    /// inserted against the order defined in `self`.
+    fn validate_key_order(&self, key: &K) -> Result<(), VecMapError> {
+        if let Some(last_key) = self.data.last().map(extract_key) {
+            match (&self.ordering, &key.cmp(last_key)) {
+                (VecMapOrdering::Greater, Ordering::Less | Ordering::Equal) => {
+                    return Err(VecMapError::InvalidKey);
+                }
+                (VecMapOrdering::Greater, Ordering::Greater) => {}
+                (VecMapOrdering::GreaterOrEqual, Ordering::Less) => {
+                    return Err(VecMapError::InvalidKey);
+                }
+                (VecMapOrdering::GreaterOrEqual, Ordering::Equal | Ordering::Greater) => {}
             }
         }
 
-        let delta_size = self.instrument_vec_op(|vec| vec.append(&mut other.0));
-        Ok(delta_size)
+        Ok(())
     }
 
     /// Instrument an operation on the underlying [`Vec`].
     /// Will panic if the operation decreases capacity.
     /// Returns the increase in memory usage caused by the op.
     fn instrument_vec_op(&mut self, op: impl FnOnce(&mut Vec<(K, V)>)) -> usize {
-        let old_cap = self.0.capacity();
-        op(&mut self.0);
-        let new_cap = self.0.capacity();
+        let old_cap = self.data.capacity();
+        op(&mut self.data);
+        let new_cap = self.data.capacity();
 
         match old_cap.cmp(&new_cap) {
             Ordering::Less => {
@@ -145,6 +201,36 @@ impl<K: Ord, V> VecMap<K, V> {
             Ordering::Greater => panic!("VecMap capacity shouldn't ever decrease"),
         }
     }
+
+    /// Similar to `from_iter` defined in `FromIter` trait except
+    /// that it accepts an [`VecMapOrdering`]
+    pub fn from_iter<I: IntoIterator<Item = (K, V)>>(iter: I, ordering: VecMapOrdering) -> Self {
+        let iter = iter.into_iter();
+        let initial_capacity = {
+            match iter.size_hint() {
+                (lower_bound, None) => lower_bound,
+                (_, Some(upper_bound)) => upper_bound,
+            }
+        };
+
+        let mut vec_map = VecMap::with_capacity(initial_capacity, ordering);
+        for (key, value) in iter {
+            vec_map
+                .append(key, value)
+                .expect("The passed collection needs to be sorted!");
+        }
+
+        vec_map
+    }
+}
+
+impl<K: Ord, V> IntoIterator for VecMap<K, V> {
+    type Item = (K, V);
+    type IntoIter = std::vec::IntoIter<(K, V)>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        self.data.into_iter()
+    }
 }
 
 fn extract_key<K, V>(entry: &(K, V)) -> &K {
@@ -155,7 +241,7 @@ fn extract_key<K, V>(entry: &(K, V)) -> &K {
 mod tests {
     use std::{collections::BTreeMap, ops::Bound};
 
-    use super::VecMap;
+    use super::{VecMap, VecMapOrdering};
 
     #[test]
     fn unbounded_range() {
@@ -310,5 +396,59 @@ mod tests {
         left.extend(&mut one_map).unwrap_err();
         assert_eq!(left.as_slice(), &[(0, ()), (1, ())]);
         assert_eq!(one_map.as_slice(), &[(1, ())]);
+
+        let mut map_greater_or_equal = VecMap::new(VecMapOrdering::GreaterOrEqual);
+        map_greater_or_equal.append(2, ()).unwrap();
+        map_greater_or_equal.append(2, ()).unwrap();
+
+        left.extend(&mut map_greater_or_equal).unwrap_err();
+        assert_eq!(left.as_slice(), &[(0, ()), (1, ())]);
+        assert_eq!(map_greater_or_equal.as_slice(), &[(2, ()), (2, ())]);
+    }
+
+    #[test]
+    fn extend_with_ordering() {
+        let mut left = VecMap::new(VecMapOrdering::GreaterOrEqual);
+        left.append(0, ()).unwrap();
+        assert_eq!(left.as_slice(), &[(0, ())]);
+
+        let mut greater_right = VecMap::new(VecMapOrdering::Greater);
+        greater_right.append(0, ()).unwrap();
+        left.extend(&mut greater_right).unwrap_err();
+        assert_eq!(left.as_slice(), &[(0, ())]);
+
+        let mut greater_or_equal_right = VecMap::new(VecMapOrdering::GreaterOrEqual);
+        greater_or_equal_right.append(2, ()).unwrap();
+        greater_or_equal_right.append(2, ()).unwrap();
+        left.extend(&mut greater_or_equal_right).unwrap();
+        assert_eq!(left.as_slice(), &[(0, ()), (2, ()), (2, ())]);
+    }
+
+    #[test]
+    fn vec_map_from_sorted() {
+        let vec = vec![(1, ()), (2, ()), (3, ()), (6, ())];
+        let vec_map = VecMap::from_iter(vec, VecMapOrdering::Greater);
+        assert_eq!(vec_map.as_slice(), &[(1, ()), (2, ()), (3, ()), (6, ())]);
+
+        let vec = vec![(1, ()), (2, ()), (3, ()), (3, ()), (6, ()), (6, ())];
+        let vec_map = VecMap::from_iter(vec, VecMapOrdering::GreaterOrEqual);
+        assert_eq!(
+            vec_map.as_slice(),
+            &[(1, ()), (2, ()), (3, ()), (3, ()), (6, ()), (6, ())]
+        );
+    }
+
+    #[test]
+    #[should_panic]
+    fn vec_map_from_unsorted_greater() {
+        let vec = vec![(1, ()), (2, ()), (2, ()), (3, ()), (6, ())];
+        let _ = VecMap::from_iter(vec, VecMapOrdering::Greater);
+    }
+
+    #[test]
+    #[should_panic]
+    fn vec_map_from_unsorted_greater_or_equal() {
+        let vec = vec![(1, ()), (2, ()), (3, ()), (6, ()), (5, ())];
+        let _ = VecMap::from_iter(vec, VecMapOrdering::GreaterOrEqual);
     }
 }

libs/utils/src/zstd.rs

@@ -0,0 +1,78 @@
+use std::io::SeekFrom;
+
+use anyhow::{Context, Result};
+use async_compression::{
+    tokio::{bufread::ZstdDecoder, write::ZstdEncoder},
+    zstd::CParameter,
+    Level,
+};
+use camino::Utf8Path;
+use nix::NixPath;
+use tokio::{
+    fs::{File, OpenOptions},
+    io::AsyncBufRead,
+    io::AsyncSeekExt,
+    io::AsyncWriteExt,
+};
+use tokio_tar::{Archive, Builder, HeaderMode};
+use walkdir::WalkDir;
+
+/// Creates a Zstandard tarball.
+pub async fn create_zst_tarball(path: &Utf8Path, tarball: &Utf8Path) -> Result<(File, u64)> {
+    let file = OpenOptions::new()
+        .create(true)
+        .truncate(true)
+        .read(true)
+        .write(true)
+        .open(&tarball)
+        .await
+        .with_context(|| format!("tempfile creation {tarball}"))?;
+
+    let mut paths = Vec::new();
+    for entry in WalkDir::new(path) {
+        let entry = entry?;
+        let metadata = entry.metadata().expect("error getting dir entry metadata");
+        // Also allow directories so that we also get empty directories
+        if !(metadata.is_file() || metadata.is_dir()) {
+            continue;
+        }
+        let path = entry.into_path();
+        paths.push(path);
+    }
+    // Do a sort to get a more consistent listing
+    paths.sort_unstable();
+    let zstd = ZstdEncoder::with_quality_and_params(
+        file,
+        Level::Default,
+        &[CParameter::enable_long_distance_matching(true)],
+    );
+    let mut builder = Builder::new(zstd);
+    // Use reproducible header mode
+    builder.mode(HeaderMode::Deterministic);
+    for p in paths {
+        let rel_path = p.strip_prefix(path)?;
+        if rel_path.is_empty() {
+            // The top directory should not be compressed,
+            // the tar crate doesn't like that
+            continue;
+        }
+        builder.append_path_with_name(&p, rel_path).await?;
+    }
+    let mut zstd = builder.into_inner().await?;
+    zstd.shutdown().await?;
+    let mut compressed = zstd.into_inner();
+    let compressed_len = compressed.metadata().await?.len();
+    compressed.seek(SeekFrom::Start(0)).await?;
+    Ok((compressed, compressed_len))
+}
+
+/// Creates a Zstandard tarball.
+pub async fn extract_zst_tarball(
+    path: &Utf8Path,
+    tarball: impl AsyncBufRead + Unpin,
+) -> Result<()> {
+    let decoder = Box::pin(ZstdDecoder::new(tarball));
+    let mut archive = Archive::new(decoder);
+    archive.unpack(path).await?;
+    Ok(())
+}

libs/vm_monitor/src/runner.rs

@@ -69,7 +69,7 @@ pub struct Config {
     /// should be removed once we have a better solution there.
     sys_buffer_bytes: u64,
 
-    /// Minimum fraction of total system memory reserved *before* the the cgroup threshold; in
+    /// Minimum fraction of total system memory reserved *before* the cgroup threshold; in
     /// other words, providing a ceiling for the highest value of the threshold by enforcing that
     /// there's at least `cgroup_min_overhead_fraction` of the total memory remaining beyond the
     /// threshold.

libs/walproposer/src/api_bindings.rs

@@ -50,6 +50,14 @@ extern "C" fn get_flush_rec_ptr(wp: *mut WalProposer) -> XLogRecPtr {
     }
 }
 
+extern "C" fn update_donor(wp: *mut WalProposer, donor: *mut Safekeeper, donor_lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).update_donor(&mut (*donor), donor_lsn)
+    }
+}
+
 extern "C" fn get_current_timestamp(wp: *mut WalProposer) -> TimestampTz {
     unsafe {
         let callback_data = (*(*wp).config).callback_data;
@@ -324,11 +332,11 @@ extern "C" fn finish_sync_safekeepers(wp: *mut WalProposer, lsn: XLogRecPtr) {
     }
 }
 
-extern "C" fn process_safekeeper_feedback(wp: *mut WalProposer) {
+extern "C" fn process_safekeeper_feedback(wp: *mut WalProposer, sk: *mut Safekeeper) {
     unsafe {
         let callback_data = (*(*wp).config).callback_data;
         let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).process_safekeeper_feedback(&mut (*wp))
+        (*api).process_safekeeper_feedback(&mut (*wp), &mut (*sk));
     }
 }
 
@@ -391,6 +399,7 @@ pub(crate) fn create_api() -> walproposer_api {
         get_shmem_state: Some(get_shmem_state),
         start_streaming: Some(start_streaming),
         get_flush_rec_ptr: Some(get_flush_rec_ptr),
+        update_donor: Some(update_donor),
         get_current_timestamp: Some(get_current_timestamp),
         conn_error_message: Some(conn_error_message),
         conn_status: Some(conn_status),
@@ -421,6 +430,32 @@ pub(crate) fn create_api() -> walproposer_api {
     }
 }
 
+pub fn empty_shmem() -> crate::bindings::WalproposerShmemState {
+    let empty_feedback = crate::bindings::PageserverFeedback {
+        present: false,
+        currentClusterSize: 0,
+        last_received_lsn: 0,
+        disk_consistent_lsn: 0,
+        remote_consistent_lsn: 0,
+        replytime: 0,
+        shard_number: 0,
+    };
+
+    crate::bindings::WalproposerShmemState {
+        propEpochStartLsn: crate::bindings::pg_atomic_uint64 { value: 0 },
+        donor_name: [0; 64],
+        donor_conninfo: [0; 1024],
+        donor_lsn: 0,
+        mutex: 0,
+        mineLastElectedTerm: crate::bindings::pg_atomic_uint64 { value: 0 },
+        backpressureThrottlingTime: crate::bindings::pg_atomic_uint64 { value: 0 },
+        currentClusterSize: crate::bindings::pg_atomic_uint64 { value: 0 },
+        shard_ps_feedback: [empty_feedback; 128],
+        num_shards: 0,
+        min_ps_feedback: empty_feedback,
+    }
+}
+
 impl std::fmt::Display for Level {
     fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
         write!(f, "{:?}", self)

libs/walproposer/src/walproposer.rs

@@ -1,8 +1,5 @@
 use std::ffi::CString;
 
-use postgres_ffi::WAL_SEGMENT_SIZE;
-use utils::{id::TenantTimelineId, lsn::Lsn};
-
 use crate::{
     api_bindings::{create_api, take_vec_u8, Level},
     bindings::{
@@ -10,6 +7,8 @@ use crate::{
         WalProposerCreate, WalProposerFree, WalProposerPoll, WalProposerStart,
     },
 };
+use postgres_ffi::WAL_SEGMENT_SIZE;
+use utils::{id::TenantTimelineId, lsn::Lsn};
 
 /// Rust high-level wrapper for C walproposer API. Many methods are not required
 /// for simple cases, hence todo!() in default implementations.
@@ -28,6 +27,10 @@ pub trait ApiImpl {
         todo!()
     }
 
+    fn update_donor(&self, _donor: &mut Safekeeper, _donor_lsn: u64) {
+        todo!()
+    }
+
     fn get_current_timestamp(&self) -> i64 {
         todo!()
     }
@@ -142,7 +145,7 @@ pub trait ApiImpl {
         todo!()
     }
 
-    fn process_safekeeper_feedback(&mut self, _wp: &mut WalProposer) {
+    fn process_safekeeper_feedback(&mut self, _wp: &mut WalProposer, _sk: &mut Safekeeper) {
         todo!()
     }
 
@@ -274,6 +277,7 @@ mod tests {
         sync::{atomic::AtomicUsize, mpsc::sync_channel},
     };
 
+    use std::cell::UnsafeCell;
     use utils::id::TenantTimelineId;
 
     use crate::{api_bindings::Level, bindings::NeonWALReadResult, walproposer::Wrapper};
@@ -297,6 +301,8 @@ mod tests {
         replies_ptr: AtomicUsize,
         // channel to send LSN to the main thread
         sync_channel: std::sync::mpsc::SyncSender<u64>,
+        // Shmem state, used for storing donor info
+        shmem: UnsafeCell<crate::bindings::WalproposerShmemState>,
     }
 
     impl MockImpl {
@@ -327,11 +333,22 @@ mod tests {
     }
 
     impl ApiImpl for MockImpl {
+        fn get_shmem_state(&self) -> *mut crate::bindings::WalproposerShmemState {
+            self.shmem.get()
+        }
+
         fn get_current_timestamp(&self) -> i64 {
             println!("get_current_timestamp");
             0
         }
 
+        fn update_donor(&self, donor: &mut crate::bindings::Safekeeper, donor_lsn: u64) {
+            let mut shmem = unsafe { *self.get_shmem_state() };
+            shmem.propEpochStartLsn.value = donor_lsn;
+            shmem.donor_conninfo = donor.conninfo;
+            shmem.donor_lsn = donor_lsn;
+        }
+
         fn conn_status(
             &self,
             _: &mut crate::bindings::Safekeeper,
@@ -507,6 +524,7 @@ mod tests {
             ],
             replies_ptr: AtomicUsize::new(0),
             sync_channel: sender,
+            shmem: UnsafeCell::new(crate::api_bindings::empty_shmem()),
         });
         let config = crate::walproposer::Config {
             ttid,

pageserver/Cargo.toml

@@ -37,7 +37,6 @@ humantime-serde.workspace = true
 hyper.workspace = true
 itertools.workspace = true
 leaky-bucket.workspace = true
-lz4_flex.workspace = true
 md5.workspace = true
 nix.workspace = true
 # hack to get the number of worker threads tokio uses
@@ -60,6 +59,7 @@ signal-hook.workspace = true
 smallvec = { workspace = true, features = ["write"] }
 svg_fmt.workspace = true
 sync_wrapper.workspace = true
+sysinfo.workspace = true
 tokio-tar.workspace = true
 thiserror.workspace = true
 tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time"] }
@@ -70,6 +70,7 @@ tokio-stream.workspace = true
 tokio-util.workspace = true
 toml_edit = { workspace = true, features = [ "serde" ] }
 tracing.workspace = true
+twox-hash.workspace = true
 url.workspace = true
 walkdir.workspace = true
 metrics.workspace = true
@@ -90,6 +91,9 @@ enumset = { workspace = true, features = ["serde"]}
 strum.workspace = true
 strum_macros.workspace = true
 
+[target.'cfg(target_os = "linux")'.dependencies]
+procfs.workspace = true
+
 [dev-dependencies]
 criterion.workspace = true
 hex-literal.workspace = true

pageserver/benches/bench_walredo.rs

@@ -1,160 +1,203 @@
-//! Simple benchmarking around walredo.
+//! Quantify a single walredo manager's throughput under N concurrent callers.
 //!
-//! Right now they hope to just set a baseline. Later we can try to expand into latency and
-//! throughput after figuring out the coordinated omission problems below.
+//! The benchmark implementation ([`bench_impl`]) is parametrized by
+//! - `redo_work` => [`Request::short_request`] or [`Request::medium_request`]
+//! - `n_redos` => number of times the benchmark shell execute the `redo_work`
+//! - `nclients` => number of clients (more on this shortly).
 //!
-//! There are two sets of inputs; `short` and `medium`. They were collected on postgres v14 by
-//! logging what happens when a sequential scan is requested on a small table, then picking out two
-//! suitable from logs.
+//! The benchmark impl sets up a multi-threaded tokio runtime with default parameters.
+//! It spawns `nclients` times [`client`] tokio tasks.
+//! Each task executes the `redo_work` `n_redos/nclients` times.
 //!
+//! We exercise the following combinations:
+//! - `redo_work = short / medium``
+//! - `nclients = [1, 2, 4, 8, 16, 32, 64, 128]`
 //!
-//! Reference data (git blame to see commit) on an i3en.3xlarge
-// ```text
-//! short/short/1           time:   [39.175 µs 39.348 µs 39.536 µs]
-//! short/short/2           time:   [51.227 µs 51.487 µs 51.755 µs]
-//! short/short/4           time:   [76.048 µs 76.362 µs 76.674 µs]
-//! short/short/8           time:   [128.94 µs 129.82 µs 130.74 µs]
-//! short/short/16          time:   [227.84 µs 229.00 µs 230.28 µs]
-//! short/short/32          time:   [455.97 µs 457.81 µs 459.90 µs]
-//! short/short/64          time:   [902.46 µs 904.84 µs 907.32 µs]
-//! short/short/128         time:   [1.7416 ms 1.7487 ms 1.7561 ms]
-//! ``
-
-use std::sync::Arc;
+//! We let `criterion` determine the `n_redos` using `iter_custom`.
+//! The idea is that for each `(redo_work, nclients)` combination,
+//! criterion will run the `bench_impl` multiple times with different `n_redos`.
+//! The `bench_impl` reports the aggregate wall clock time from the clients' perspective.
+//! Criterion will divide that by `n_redos` to compute the "time per iteration".
+//! In our case, "time per iteration" means "time per redo_work execution".
+//!
+//! NB: the way by which `iter_custom` determines the "number of iterations"
+//! is called sampling. Apparently the idea here is to detect outliers.
+//! We're not sure whether the current choice of sampling method makes sense.
+//! See https://bheisler.github.io/criterion.rs/book/user_guide/command_line_output.html#collecting-samples
+//!
+//! # Reference Numbers
+//!
+//! 2024-04-15 on i3en.3xlarge
+//!
+//! ```text
+//! async-short/1           time:   [24.584 µs 24.737 µs 24.922 µs]
+//! async-short/2           time:   [33.479 µs 33.660 µs 33.888 µs]
+//! async-short/4           time:   [42.713 µs 43.046 µs 43.440 µs]
+//! async-short/8           time:   [71.814 µs 72.478 µs 73.240 µs]
+//! async-short/16          time:   [132.73 µs 134.45 µs 136.22 µs]
+//! async-short/32          time:   [258.31 µs 260.73 µs 263.27 µs]
+//! async-short/64          time:   [511.61 µs 514.44 µs 517.51 µs]
+//! async-short/128         time:   [992.64 µs 998.23 µs 1.0042 ms]
+//! async-medium/1          time:   [110.11 µs 110.50 µs 110.96 µs]
+//! async-medium/2          time:   [153.06 µs 153.85 µs 154.99 µs]
+//! async-medium/4          time:   [317.51 µs 319.92 µs 322.85 µs]
+//! async-medium/8          time:   [638.30 µs 644.68 µs 652.12 µs]
+//! async-medium/16         time:   [1.2651 ms 1.2773 ms 1.2914 ms]
+//! async-medium/32         time:   [2.5117 ms 2.5410 ms 2.5720 ms]
+//! async-medium/64         time:   [4.8088 ms 4.8555 ms 4.9047 ms]
+//! async-medium/128        time:   [8.8311 ms 8.9849 ms 9.1263 ms]
+//! sync-short/1            time:   [25.503 µs 25.626 µs 25.771 µs]
+//! sync-short/2            time:   [30.850 µs 31.013 µs 31.208 µs]
+//! sync-short/4            time:   [45.543 µs 45.856 µs 46.193 µs]
+//! sync-short/8            time:   [84.114 µs 84.639 µs 85.220 µs]
+//! sync-short/16           time:   [185.22 µs 186.15 µs 187.13 µs]
+//! sync-short/32           time:   [377.43 µs 378.87 µs 380.46 µs]
+//! sync-short/64           time:   [756.49 µs 759.04 µs 761.70 µs]
+//! sync-short/128          time:   [1.4825 ms 1.4874 ms 1.4923 ms]
+//! sync-medium/1           time:   [105.66 µs 106.01 µs 106.43 µs]
+//! sync-medium/2           time:   [153.10 µs 153.84 µs 154.72 µs]
+//! sync-medium/4           time:   [327.13 µs 329.44 µs 332.27 µs]
+//! sync-medium/8           time:   [654.26 µs 658.73 µs 663.63 µs]
+//! sync-medium/16          time:   [1.2682 ms 1.2748 ms 1.2816 ms]
+//! sync-medium/32          time:   [2.4456 ms 2.4595 ms 2.4731 ms]
+//! sync-medium/64          time:   [4.6523 ms 4.6890 ms 4.7256 ms]
+//! sync-medium/128         time:   [8.7215 ms 8.8323 ms 8.9344 ms]
+//! ```
 
 use bytes::{Buf, Bytes};
+use criterion::{BenchmarkId, Criterion};
 use pageserver::{
-    config::PageServerConf, repository::Key, walrecord::NeonWalRecord, walredo::PostgresRedoManager,
+    config::PageServerConf,
+    walrecord::NeonWalRecord,
+    walredo::{PostgresRedoManager, ProcessKind},
 };
-use pageserver_api::shard::TenantShardId;
-use tokio::task::JoinSet;
+use pageserver_api::{key::Key, shard::TenantShardId};
+use std::{
+    sync::Arc,
+    time::{Duration, Instant},
+};
+use tokio::{sync::Barrier, task::JoinSet};
 use utils::{id::TenantId, lsn::Lsn};
 
-use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion};
+fn bench(c: &mut Criterion) {
+    for process_kind in &[ProcessKind::Async, ProcessKind::Sync] {
+        {
+            let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
+            for nclients in nclients {
+                let mut group = c.benchmark_group(format!("{process_kind}-short"));
+                group.bench_with_input(
+                    BenchmarkId::from_parameter(nclients),
+                    &nclients,
+                    |b, nclients| {
+                        let redo_work = Arc::new(Request::short_input());
+                        b.iter_custom(|iters| {
+                            bench_impl(*process_kind, Arc::clone(&redo_work), iters, *nclients)
+                        });
+                    },
+                );
+            }
+        }
 
-fn redo_scenarios(c: &mut Criterion) {
-    // logging should be enabled when adding more inputs, since walredo will only report malformed
-    // input to the stderr.
-    // utils::logging::init(utils::logging::LogFormat::Plain).unwrap();
+        {
+            let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
+            for nclients in nclients {
+                let mut group = c.benchmark_group(format!("{process_kind}-medium"));
+                group.bench_with_input(
+                    BenchmarkId::from_parameter(nclients),
+                    &nclients,
+                    |b, nclients| {
+                        let redo_work = Arc::new(Request::medium_input());
+                        b.iter_custom(|iters| {
+                            bench_impl(*process_kind, Arc::clone(&redo_work), iters, *nclients)
+                        });
+                    },
+                );
+            }
+        }
+    }
+}
+criterion::criterion_group!(benches, bench);
+criterion::criterion_main!(benches);
 
+// Returns the sum of each client's wall-clock time spent executing their share of the n_redos.
+fn bench_impl(
+    process_kind: ProcessKind,
+    redo_work: Arc<Request>,
+    n_redos: u64,
+    nclients: u64,
+) -> Duration {
     let repo_dir = camino_tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
 
-    let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
+    let mut conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
+    conf.walredo_process_kind = process_kind;
     let conf = Box::leak(Box::new(conf));
     let tenant_shard_id = TenantShardId::unsharded(TenantId::generate());
 
-    let manager = PostgresRedoManager::new(conf, tenant_shard_id);
-
-    let manager = Arc::new(manager);
-
-    {
-        let rt = tokio::runtime::Builder::new_current_thread()
-            .enable_all()
-            .build()
-            .unwrap();
-        tracing::info!("executing first");
-        rt.block_on(short().execute(&manager)).unwrap();
-        tracing::info!("first executed");
-    }
-
-    let thread_counts = [1, 2, 4, 8, 16, 32, 64, 128];
-
-    let mut group = c.benchmark_group("short");
-    group.sampling_mode(criterion::SamplingMode::Flat);
-
-    for thread_count in thread_counts {
-        group.bench_with_input(
-            BenchmarkId::new("short", thread_count),
-            &thread_count,
-            |b, thread_count| {
-                add_multithreaded_walredo_requesters(b, *thread_count, &manager, short);
-            },
-        );
-    }
-    drop(group);
-
-    let mut group = c.benchmark_group("medium");
-    group.sampling_mode(criterion::SamplingMode::Flat);
-
-    for thread_count in thread_counts {
-        group.bench_with_input(
-            BenchmarkId::new("medium", thread_count),
-            &thread_count,
-            |b, thread_count| {
-                add_multithreaded_walredo_requesters(b, *thread_count, &manager, medium);
-            },
-        );
-    }
-    drop(group);
-}
-
-/// Sets up a multi-threaded tokio runtime with default worker thread count,
-/// then, spawn `requesters` tasks that repeatedly:
-/// - get input from `input_factor()`
-/// - call `manager.request_redo()` with their input
-///
-/// This stress-tests the scalability of a single walredo manager at high tokio-level concurrency.
-///
-/// Using tokio's default worker thread count means the results will differ on machines
-/// with different core countrs. We don't care about that, the performance will always
-/// be different on different hardware. To compare performance of different software versions,
-/// use the same hardware.
-fn add_multithreaded_walredo_requesters(
-    b: &mut criterion::Bencher,
-    nrequesters: usize,
-    manager: &Arc<PostgresRedoManager>,
-    input_factory: fn() -> Request,
-) {
-    assert_ne!(nrequesters, 0);
-
     let rt = tokio::runtime::Builder::new_multi_thread()
         .enable_all()
         .build()
         .unwrap();
 
-    let barrier = Arc::new(tokio::sync::Barrier::new(nrequesters + 1));
+    let start = Arc::new(Barrier::new(nclients as usize));
 
-    let mut requesters = JoinSet::new();
-    for _ in 0..nrequesters {
-        let _entered = rt.enter();
-        let manager = manager.clone();
-        let barrier = barrier.clone();
-        requesters.spawn(async move {
-            loop {
-                let input = input_factory();
-                barrier.wait().await;
-                let page = input.execute(&manager).await.unwrap();
-                assert_eq!(page.remaining(), 8192);
-                barrier.wait().await;
-            }
+    let mut tasks = JoinSet::new();
+
+    let manager = PostgresRedoManager::new(conf, tenant_shard_id);
+    let manager = Arc::new(manager);
+
+    // divide the amount of work equally among the clients.
+    let nredos_per_client = n_redos / nclients;
+    for _ in 0..nclients {
+        rt.block_on(async {
+            tasks.spawn(client(
+                Arc::clone(&manager),
+                Arc::clone(&start),
+                Arc::clone(&redo_work),
+                nredos_per_client,
+            ))
         });
     }
 
-    let do_one_iteration = || {
-        rt.block_on(async {
-            barrier.wait().await;
-            // wait for work to complete
-            barrier.wait().await;
-        })
-    };
+    let elapsed = rt.block_on(async move {
+        let mut total_wallclock_time = Duration::ZERO;
+        while let Some(res) = tasks.join_next().await {
+            total_wallclock_time += res.unwrap();
+        }
+        total_wallclock_time
+    });
 
-    b.iter_batched(
-        || {
-            // warmup
-            do_one_iteration();
-        },
-        |()| {
-            // work loop
-            do_one_iteration();
-        },
-        criterion::BatchSize::PerIteration,
-    );
+    // consistency check to ensure process kind setting worked
+    if nredos_per_client > 0 {
+        assert_eq!(
+            manager
+                .status()
+                .process
+                .map(|p| p.kind)
+                .expect("the benchmark work causes a walredo process to be spawned"),
+            std::borrow::Cow::Borrowed(process_kind.into())
+        );
+    }
 
-    rt.block_on(requesters.shutdown());
+    elapsed
 }
 
-criterion_group!(benches, redo_scenarios);
-criterion_main!(benches);
+async fn client(
+    mgr: Arc<PostgresRedoManager>,
+    start: Arc<Barrier>,
+    redo_work: Arc<Request>,
+    n_redos: u64,
+) -> Duration {
+    start.wait().await;
+    let start = Instant::now();
+    for _ in 0..n_redos {
+        let page = redo_work.execute(&mgr).await.unwrap();
+        assert_eq!(page.remaining(), 8192);
+        // The real pageserver will rarely if ever do 2 walredos in a row without
+        // yielding to the executor.
+        tokio::task::yield_now().await;
+    }
+    start.elapsed()
+}
 
 macro_rules! lsn {
     ($input:expr) => {{
@@ -166,12 +209,46 @@ macro_rules! lsn {
     }};
 }
 
-/// Short payload, 1132 bytes.
-// pg_records are copypasted from log, where they are put with Debug impl of Bytes, which uses \0
-// for null bytes.
-#[allow(clippy::octal_escapes)]
-fn short() -> Request {
-    Request {
+/// Simple wrapper around `WalRedoManager::request_redo`.
+///
+/// In benchmarks this is cloned around.
+#[derive(Clone)]
+struct Request {
+    key: Key,
+    lsn: Lsn,
+    base_img: Option<(Lsn, Bytes)>,
+    records: Vec<(Lsn, NeonWalRecord)>,
+    pg_version: u32,
+}
+
+impl Request {
+    async fn execute(&self, manager: &PostgresRedoManager) -> anyhow::Result<Bytes> {
+        let Request {
+            key,
+            lsn,
+            base_img,
+            records,
+            pg_version,
+        } = self;
+
+        // TODO: avoid these clones
+        manager
+            .request_redo(*key, *lsn, base_img.clone(), records.clone(), *pg_version)
+            .await
+    }
+
+    fn pg_record(will_init: bool, bytes: &'static [u8]) -> NeonWalRecord {
+        let rec = Bytes::from_static(bytes);
+        NeonWalRecord::Postgres { will_init, rec }
+    }
+
+    /// Short payload, 1132 bytes.
+    // pg_records are copypasted from log, where they are put with Debug impl of Bytes, which uses \0
+    // for null bytes.
+    #[allow(clippy::octal_escapes)]
+    pub fn short_input() -> Request {
+        let pg_record = Self::pg_record;
+        Request {
         key: Key {
             field1: 0,
             field2: 1663,
@@ -194,13 +271,14 @@ fn short() -> Request {
         ],
         pg_version: 14,
     }
-}
+    }
 
-/// Medium sized payload, serializes as 26393 bytes.
-// see [`short`]
-#[allow(clippy::octal_escapes)]
-fn medium() -> Request {
-    Request {
+    /// Medium sized payload, serializes as 26393 bytes.
+    // see [`short`]
+    #[allow(clippy::octal_escapes)]
+    pub fn medium_input() -> Request {
+        let pg_record = Self::pg_record;
+        Request {
         key: Key {
             field1: 0,
             field2: 1663,
@@ -442,37 +520,5 @@ fn medium() -> Request {
         ],
         pg_version: 14,
     }
-}
-
-fn pg_record(will_init: bool, bytes: &'static [u8]) -> NeonWalRecord {
-    let rec = Bytes::from_static(bytes);
-    NeonWalRecord::Postgres { will_init, rec }
-}
-
-/// Simple wrapper around `WalRedoManager::request_redo`.
-///
-/// In benchmarks this is cloned around.
-#[derive(Clone)]
-struct Request {
-    key: Key,
-    lsn: Lsn,
-    base_img: Option<(Lsn, Bytes)>,
-    records: Vec<(Lsn, NeonWalRecord)>,
-    pg_version: u32,
-}
-
-impl Request {
-    async fn execute(self, manager: &PostgresRedoManager) -> anyhow::Result<Bytes> {
-        let Request {
-            key,
-            lsn,
-            base_img,
-            records,
-            pg_version,
-        } = self;
-
-        manager
-            .request_redo(key, lsn, base_img, records, pg_version)
-            .await
     }
 }

pageserver/client/src/mgmt_api.rs

@@ -128,12 +128,12 @@ impl Client {
 
     pub async fn timeline_info(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         force_await_logical_size: ForceAwaitLogicalSize,
     ) -> Result<pageserver_api::models::TimelineInfo> {
         let uri = format!(
-            "{}/v1/tenant/{tenant_id}/timeline/{timeline_id}",
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}",
             self.mgmt_api_endpoint
         );
 
@@ -151,11 +151,11 @@ impl Client {
 
     pub async fn keyspace(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
     ) -> Result<pageserver_api::models::partitioning::Partitioning> {
         let uri = format!(
-            "{}/v1/tenant/{tenant_id}/timeline/{timeline_id}/keyspace",
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/keyspace",
             self.mgmt_api_endpoint
         );
         self.get(&uri)
@@ -169,7 +169,7 @@ impl Client {
         self.request(Method::GET, uri, ()).await
     }
 
-    async fn request<B: serde::Serialize, U: reqwest::IntoUrl>(
+    async fn request_noerror<B: serde::Serialize, U: reqwest::IntoUrl>(
         &self,
         method: Method,
         uri: U,
@@ -181,7 +181,16 @@ impl Client {
         } else {
             req
         };
-        let res = req.json(&body).send().await.map_err(Error::ReceiveBody)?;
+        req.json(&body).send().await.map_err(Error::ReceiveBody)
+    }
+
+    async fn request<B: serde::Serialize, U: reqwest::IntoUrl>(
+        &self,
+        method: Method,
+        uri: U,
+        body: B,
+    ) -> Result<reqwest::Response> {
+        let res = self.request_noerror(method, uri, body).await?;
         let response = res.error_from_body().await?;
         Ok(response)
     }
@@ -234,18 +243,72 @@ impl Client {
         Ok(())
     }
 
+    pub async fn tenant_scan_remote_storage(
+        &self,
+        tenant_id: TenantId,
+    ) -> Result<TenantScanRemoteStorageResponse> {
+        let uri = format!(
+            "{}/v1/tenant/{tenant_id}/scan_remote_storage",
+            self.mgmt_api_endpoint
+        );
+        let response = self.request(Method::GET, &uri, ()).await?;
+        let body = response.json().await.map_err(Error::ReceiveBody)?;
+        Ok(body)
+    }
+
     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 tenant_secondary_download(&self, tenant_id: TenantShardId) -> Result<()> {
-        let uri = format!(
+    pub async fn tenant_secondary_download(
+        &self,
+        tenant_id: TenantShardId,
+        wait: Option<std::time::Duration>,
+    ) -> Result<(StatusCode, SecondaryProgress)> {
+        let mut path = reqwest::Url::parse(&format!(
             "{}/v1/tenant/{}/secondary/download",
             self.mgmt_api_endpoint, tenant_id
-        );
-        self.request(Method::POST, &uri, ()).await?;
+        ))
+        .expect("Cannot build URL");
+
+        if let Some(wait) = wait {
+            path.query_pairs_mut()
+                .append_pair("wait_ms", &format!("{}", wait.as_millis()));
+        }
+
+        let response = self.request(Method::POST, path, ()).await?;
+        let status = response.status();
+        let progress: SecondaryProgress = response.json().await.map_err(Error::ReceiveBody)?;
+        Ok((status, progress))
+    }
+
+    pub async fn tenant_secondary_status(
+        &self,
+        tenant_shard_id: TenantShardId,
+    ) -> Result<SecondaryProgress> {
+        let path = reqwest::Url::parse(&format!(
+            "{}/v1/tenant/{}/secondary/status",
+            self.mgmt_api_endpoint, tenant_shard_id
+        ))
+        .expect("Cannot build URL");
+
+        self.request(Method::GET, path, ())
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
+    pub async fn tenant_heatmap_upload(&self, tenant_id: TenantShardId) -> Result<()> {
+        let path = reqwest::Url::parse(&format!(
+            "{}/v1/tenant/{}/heatmap_upload",
+            self.mgmt_api_endpoint, tenant_id
+        ))
+        .expect("Cannot build URL");
+
+        self.request(Method::POST, path, ()).await?;
         Ok(())
     }
 
@@ -256,10 +319,7 @@ impl Client {
         flush_ms: Option<std::time::Duration>,
         lazy: bool,
     ) -> Result<()> {
-        let req_body = TenantLocationConfigRequest {
-            tenant_id: Some(tenant_shard_id),
-            config,
-        };
+        let req_body = TenantLocationConfigRequest { config };
 
         let mut path = reqwest::Url::parse(&format!(
             "{}/v1/tenant/{}/location_config",
@@ -416,4 +476,77 @@ impl Client {
             .await
             .map_err(Error::ReceiveBody)
     }
+
+    pub async fn get_utilization(&self) -> Result<PageserverUtilization> {
+        let uri = format!("{}/v1/utilization", self.mgmt_api_endpoint);
+        self.get(uri)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
+    pub async fn layer_map_info(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+    ) -> Result<LayerMapInfo> {
+        let uri = format!(
+            "{}/v1/tenant/{}/timeline/{}/layer",
+            self.mgmt_api_endpoint, tenant_shard_id, timeline_id,
+        );
+        self.get(&uri)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
+    pub async fn layer_evict(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        layer_file_name: &str,
+    ) -> Result<bool> {
+        let uri = format!(
+            "{}/v1/tenant/{}/timeline/{}/layer/{}",
+            self.mgmt_api_endpoint, tenant_shard_id, timeline_id, layer_file_name
+        );
+        let resp = self.request_noerror(Method::DELETE, &uri, ()).await?;
+        match resp.status() {
+            StatusCode::OK => Ok(true),
+            StatusCode::NOT_MODIFIED => Ok(false),
+            // TODO: dedupe this pattern / introduce separate error variant?
+            status => Err(match resp.json::<HttpErrorBody>().await {
+                Ok(HttpErrorBody { msg }) => Error::ApiError(status, msg),
+                Err(_) => {
+                    Error::ReceiveErrorBody(format!("Http error ({}) at {}.", status.as_u16(), uri))
+                }
+            }),
+        }
+    }
+
+    pub async fn layer_ondemand_download(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        layer_file_name: &str,
+    ) -> Result<bool> {
+        let uri = format!(
+            "{}/v1/tenant/{}/timeline/{}/layer/{}",
+            self.mgmt_api_endpoint, tenant_shard_id, timeline_id, layer_file_name
+        );
+        let resp = self.request_noerror(Method::GET, &uri, ()).await?;
+        match resp.status() {
+            StatusCode::OK => Ok(true),
+            StatusCode::NOT_MODIFIED => Ok(false),
+            // TODO: dedupe this pattern / introduce separate error variant?
+            status => Err(match resp.json::<HttpErrorBody>().await {
+                Ok(HttpErrorBody { msg }) => Error::ApiError(status, msg),
+                Err(_) => {
+                    Error::ReceiveErrorBody(format!("Http error ({}) at {}.", status.as_u16(), uri))
+                }
+            }),
+        }
+    }
 }

pageserver/client/src/page_service.rs

@@ -60,7 +60,7 @@ impl Client {
     ) -> anyhow::Result<PagestreamClient> {
         let copy_both: tokio_postgres::CopyBothDuplex<bytes::Bytes> = self
             .client
-            .copy_both_simple(&format!("pagestream {tenant_id} {timeline_id}"))
+            .copy_both_simple(&format!("pagestream_v2 {tenant_id} {timeline_id}"))
             .await?;
         let Client {
             cancel_on_client_drop,

pageserver/compaction/Cargo.toml

@@ -11,7 +11,6 @@ default = []
 anyhow.workspace = true
 async-compression.workspace = true
 async-stream.workspace = true
-async-trait.workspace = true
 byteorder.workspace = true
 bytes.workspace = true
 chrono = { workspace = true, features = ["serde"] }

pageserver/compaction/src/compact_tiered.rs

@@ -18,6 +18,7 @@
 //! database size. For example, if the logical database size is 10 GB, we would
 //! generate new image layers every 10 GB of WAL.
 use futures::StreamExt;
+use pageserver_api::shard::ShardIdentity;
 use tracing::{debug, info};
 
 use std::collections::{HashSet, VecDeque};
@@ -43,7 +44,8 @@ pub async fn compact_tiered<E: CompactionJobExecutor>(
     fanout: u64,
     ctx: &E::RequestContext,
 ) -> anyhow::Result<()> {
-    assert!(fanout >= 2);
+    assert!(fanout >= 1, "fanout needs to be at least 1 but is {fanout}");
+    let exp_base = fanout.max(2);
     // Start at L0
     let mut current_level_no = 0;
     let mut current_level_target_height = target_file_size;
@@ -106,7 +108,7 @@ pub async fn compact_tiered<E: CompactionJobExecutor>(
             break;
         }
         current_level_no += 1;
-        current_level_target_height = current_level_target_height.saturating_mul(fanout);
+        current_level_target_height = current_level_target_height.saturating_mul(exp_base);
     }
     Ok(())
 }
@@ -124,6 +126,7 @@ async fn compact_level<E: CompactionJobExecutor>(
     }
 
     let mut state = LevelCompactionState {
+        shard_identity: *executor.get_shard_identity(),
         target_file_size,
         _lsn_range: lsn_range.clone(),
         layers: layer_fragments,
@@ -163,6 +166,8 @@ struct LevelCompactionState<'a, E>
 where
     E: CompactionJobExecutor,
 {
+    shard_identity: ShardIdentity,
+
     // parameters
     target_file_size: u64,
 
@@ -365,6 +370,7 @@ where
                 .executor
                 .get_keyspace(&job.key_range, job.lsn_range.end, ctx)
                 .await?,
+            &self.shard_identity,
         ) * 8192;
 
         let wal_size = job
@@ -429,7 +435,7 @@ where
             keyspace,
             self.target_file_size / 8192,
         );
-        while let Some(key_range) = window.choose_next_image() {
+        while let Some(key_range) = window.choose_next_image(&self.shard_identity) {
             new_jobs.push(CompactionJob::<E> {
                 key_range,
                 lsn_range: job.lsn_range.clone(),
@@ -622,7 +628,12 @@ impl<K: CompactionKey> KeyspaceWindowPos<K> {
     }
 
     // Advance the cursor until it reaches 'target_keysize'.
-    fn advance_until_size(&mut self, w: &KeyspaceWindowHead<K>, max_size: u64) {
+    fn advance_until_size(
+        &mut self,
+        w: &KeyspaceWindowHead<K>,
+        max_size: u64,
+        shard_identity: &ShardIdentity,
+    ) {
         while self.accum_keysize < max_size && !self.reached_end(w) {
             let curr_range = &w.keyspace[self.keyspace_idx];
             if self.end_key < curr_range.start {
@@ -631,7 +642,7 @@ impl<K: CompactionKey> KeyspaceWindowPos<K> {
             }
 
             // We're now within 'curr_range'. Can we advance past it completely?
-            let distance = K::key_range_size(&(self.end_key..curr_range.end));
+            let distance = K::key_range_size(&(self.end_key..curr_range.end), shard_identity);
             if (self.accum_keysize + distance as u64) < max_size {
                 // oh yeah, it fits
                 self.end_key = curr_range.end;
@@ -640,7 +651,7 @@ impl<K: CompactionKey> KeyspaceWindowPos<K> {
             } else {
                 // advance within the range
                 let skip_key = self.end_key.skip_some();
-                let distance = K::key_range_size(&(self.end_key..skip_key));
+                let distance = K::key_range_size(&(self.end_key..skip_key), shard_identity);
                 if (self.accum_keysize + distance as u64) < max_size {
                     self.end_key = skip_key;
                     self.accum_keysize += distance as u64;
@@ -676,7 +687,7 @@ where
         }
     }
 
-    fn choose_next_image(&mut self) -> Option<Range<K>> {
+    fn choose_next_image(&mut self, shard_identity: &ShardIdentity) -> Option<Range<K>> {
         if self.start_pos.keyspace_idx == self.head.keyspace.len() {
             // we've reached the end
             return None;
@@ -686,6 +697,7 @@ where
         next_pos.advance_until_size(
             &self.head,
             self.start_pos.accum_keysize + self.head.target_keysize,
+            shard_identity,
         );
 
         // See if we can gobble up the rest of the keyspace if we stretch out the layer, up to
@@ -694,6 +706,7 @@ where
         end_pos.advance_until_size(
             &self.head,
             self.start_pos.accum_keysize + (self.head.target_keysize * 5 / 4),
+            shard_identity,
         );
         if end_pos.reached_end(&self.head) {
             // gobble up any unused keyspace between the last used key and end of the range

pageserver/compaction/src/helpers.rs

@@ -5,6 +5,7 @@ use crate::interface::*;
 use futures::future::BoxFuture;
 use futures::{Stream, StreamExt};
 use itertools::Itertools;
+use pageserver_api::shard::ShardIdentity;
 use pin_project_lite::pin_project;
 use std::collections::BinaryHeap;
 use std::collections::VecDeque;
@@ -13,11 +14,17 @@ use std::ops::{DerefMut, Range};
 use std::pin::Pin;
 use std::task::{ready, Poll};
 
-pub fn keyspace_total_size<K>(keyspace: &CompactionKeySpace<K>) -> u64
+pub fn keyspace_total_size<K>(
+    keyspace: &CompactionKeySpace<K>,
+    shard_identity: &ShardIdentity,
+) -> u64
 where
     K: CompactionKey,
 {
-    keyspace.iter().map(|r| K::key_range_size(r) as u64).sum()
+    keyspace
+        .iter()
+        .map(|r| K::key_range_size(r, shard_identity) as u64)
+        .sum()
 }
 
 pub fn overlaps_with<T: Ord>(a: &Range<T>, b: &Range<T>) -> bool {
@@ -180,7 +187,7 @@ where
                 match top.deref_mut() {
                     LazyLoadLayer::Unloaded(ref mut l) => {
                         let fut = l.load_keys(this.ctx);
-                        this.load_future.set(Some(fut));
+                        this.load_future.set(Some(Box::pin(fut)));
                         continue;
                     }
                     LazyLoadLayer::Loaded(ref mut entries) => {