refactor(pageserver): make image layer creation atomic

Signed-off-by: Alex Chi Z <chi@neon.tech>

.github/workflows/_build-and-test-locally.yml

@@ -53,20 +53,6 @@ jobs:
       BUILD_TAG: ${{ inputs.build-tag }}
 
     steps:
-      - name: Fix git ownership
-        run: |
-          # Workaround for `fatal: detected dubious ownership in repository at ...`
-          #
-          # Use both ${{ github.workspace }} and ${GITHUB_WORKSPACE} because they're different on host and in containers
-          #   Ref https://github.com/actions/checkout/issues/785
-          #
-          git config --global --add safe.directory ${{ github.workspace }}
-          git config --global --add safe.directory ${GITHUB_WORKSPACE}
-          for r in 14 15 16 17; do
-            git config --global --add safe.directory "${{ github.workspace }}/vendor/postgres-v$r"
-            git config --global --add safe.directory "${GITHUB_WORKSPACE}/vendor/postgres-v$r"
-          done
-
       - uses: actions/checkout@v4
         with:
           submodules: true

.github/workflows/benchmarking.yml

@@ -671,6 +671,10 @@ jobs:
         password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
+    # Increase timeout to 12h, default timeout is 6h
+    # we have regression in clickbench causing it to run 2-3x longer
+    timeout-minutes: 720
+
     steps:
     - uses: actions/checkout@v4
 
@@ -679,7 +683,7 @@ jobs:
       with:
         aws-region: eu-central-1
         role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
-        role-duration-seconds: 18000 # 5 hours
+        role-duration-seconds: 43200 # 12 hours
 
     - name: Download Neon artifact
       uses: ./.github/actions/download
@@ -716,7 +720,7 @@ jobs:
         test_selection: performance/test_perf_olap.py
         run_in_parallel: false
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
-        extra_params: -m remote_cluster --timeout 21600 -k test_clickbench
+        extra_params: -m remote_cluster --timeout 43200 -k test_clickbench
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
       env:
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"

.github/workflows/build_and_test.yml

@@ -839,6 +839,7 @@ jobs:
       - name: Build vm image
         run: |
           ./vm-builder \
+            -size=2G \
             -spec=compute/vm-image-spec-${{ matrix.version.debian }}.yaml \
             -src=neondatabase/compute-node-${{ matrix.version.pg }}:${{ needs.tag.outputs.build-tag }} \
             -dst=neondatabase/vm-compute-node-${{ matrix.version.pg }}:${{ needs.tag.outputs.build-tag }}
@@ -1078,20 +1079,6 @@ jobs:
     runs-on: [ self-hosted, small ]
     container: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/ansible:latest
     steps:
-      - name: Fix git ownership
-        run: |
-          # Workaround for `fatal: detected dubious ownership in repository at ...`
-          #
-          # Use both ${{ github.workspace }} and ${GITHUB_WORKSPACE} because they're different on host and in containers
-          #   Ref https://github.com/actions/checkout/issues/785
-          #
-          git config --global --add safe.directory ${{ github.workspace }}
-          git config --global --add safe.directory ${GITHUB_WORKSPACE}
-          for r in 14 15 16 17; do
-            git config --global --add safe.directory "${{ github.workspace }}/vendor/postgres-v$r"
-            git config --global --add safe.directory "${GITHUB_WORKSPACE}/vendor/postgres-v$r"
-          done
-
       - uses: actions/checkout@v4
 
       - name: Trigger deploy workflow
@@ -1130,7 +1117,10 @@ jobs:
 
             gh workflow --repo neondatabase/infra run deploy-proxy-prod.yml --ref main \
               -f deployPgSniRouter=true \
-              -f deployProxy=true \
+              -f deployProxyLink=true \
+              -f deployPrivatelinkProxy=true \
+              -f deployProxyScram=true \
+              -f deployProxyAuthBroker=true \
               -f branch=main \
               -f dockerTag=${{needs.tag.outputs.build-tag}}
           else

.gitignore

@@ -6,6 +6,8 @@ __pycache__/
 test_output/
 .vscode
 .idea
+*.swp
+tags
 neon.iml
 /.neon
 /integration_tests/.neon

Cargo.lock

@@ -3749,6 +3749,7 @@ dependencies = [
  "tracing",
  "url",
  "utils",
+ "wal_decoder",
  "walkdir",
  "workspace_hack",
 ]
@@ -4186,6 +4187,7 @@ dependencies = [
  "regex",
  "serde",
  "thiserror",
+ "tracing",
  "utils",
 ]
 
@@ -6272,7 +6274,7 @@ dependencies = [
 [[package]]
 name = "tokio-epoll-uring"
 version = "0.1.0"
-source = "git+https://github.com/neondatabase/tokio-epoll-uring.git?branch=main#08ccfa94ff5507727bf4d8d006666b5b192e04c6"
+source = "git+https://github.com/neondatabase/tokio-epoll-uring.git?branch=main#33e00106a268644d02ba0461bbd64476073b0ee1"
 dependencies = [
  "futures",
  "nix 0.26.4",
@@ -6788,7 +6790,7 @@ dependencies = [
 [[package]]
 name = "uring-common"
 version = "0.1.0"
-source = "git+https://github.com/neondatabase/tokio-epoll-uring.git?branch=main#08ccfa94ff5507727bf4d8d006666b5b192e04c6"
+source = "git+https://github.com/neondatabase/tokio-epoll-uring.git?branch=main#33e00106a268644d02ba0461bbd64476073b0ee1"
 dependencies = [
  "bytes",
  "io-uring",
@@ -6954,6 +6956,20 @@ dependencies = [
  "utils",
 ]
 
+[[package]]
+name = "wal_decoder"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "bytes",
+ "pageserver_api",
+ "postgres_ffi",
+ "serde",
+ "tracing",
+ "utils",
+ "workspace_hack",
+]
+
 [[package]]
 name = "walkdir"
 version = "2.3.3"

Cargo.toml

@@ -33,6 +33,7 @@ members = [
     "libs/postgres_ffi/wal_craft",
     "libs/vm_monitor",
     "libs/walproposer",
+    "libs/wal_decoder",
 ]
 
 [workspace.package]
@@ -238,6 +239,7 @@ tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
 vm_monitor = { version = "0.1", path = "./libs/vm_monitor/" }
 walproposer = { version = "0.1", path = "./libs/walproposer/" }
+wal_decoder = { version = "0.1", path = "./libs/wal_decoder" }
 
 ## Common library dependency
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }

build-tools.Dockerfile

@@ -57,6 +57,18 @@ RUN set -e \
         zstd \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
 
+# sql_exporter
+
+# Keep the version the same as in compute/compute-node.Dockerfile and
+# test_runner/regress/test_compute_metrics.py.
+ENV SQL_EXPORTER_VERSION=0.13.1
+RUN curl -fsSL \
+    "https://github.com/burningalchemist/sql_exporter/releases/download/${SQL_EXPORTER_VERSION}/sql_exporter-${SQL_EXPORTER_VERSION}.linux-$(case "$(uname -m)" in x86_64) echo amd64;; aarch64) echo arm64;; esac).tar.gz" \
+    --output sql_exporter.tar.gz \
+    && mkdir /tmp/sql_exporter \
+    && tar xzvf sql_exporter.tar.gz -C /tmp/sql_exporter --strip-components=1 \
+    && mv /tmp/sql_exporter/sql_exporter /usr/local/bin/sql_exporter
+
 # protobuf-compiler (protoc)
 ENV PROTOC_VERSION=25.1
 RUN curl -fsSL "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-$(uname -m | sed 's/aarch64/aarch_64/g').zip" -o "protoc.zip" \

compute/Makefile

@@ -22,6 +22,7 @@ sql_exporter.yml: $(jsonnet_files)
 		--output-file etc/$@ \
 		--tla-str collector_name=neon_collector \
 		--tla-str collector_file=neon_collector.yml \
+		--tla-str 'connection_string=postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter' \
 		etc/sql_exporter.jsonnet
 
 sql_exporter_autoscaling.yml: $(jsonnet_files)
@@ -29,7 +30,7 @@ sql_exporter_autoscaling.yml: $(jsonnet_files)
 		--output-file etc/$@ \
 		--tla-str collector_name=neon_collector_autoscaling \
 		--tla-str collector_file=neon_collector_autoscaling.yml \
-		--tla-str application_name=sql_exporter_autoscaling \
+		--tla-str 'connection_string=postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter_autoscaling' \
 		etc/sql_exporter.jsonnet
 
 .PHONY: clean

compute/compute-node.Dockerfile

@@ -431,14 +431,11 @@ COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
 COPY compute/patches/rum.patch /rum.patch
 
-# maybe version-specific
-# support for v17 is unknown
-# last release 1.3.13 - Sep 19, 2022
-RUN case "${PG_VERSION}" in "v17") \
-    echo "v17 extensions are not supported yet. Quit" && exit 0;; \
-    esac && \
-    wget https://github.com/postgrespro/rum/archive/refs/tags/1.3.13.tar.gz -O rum.tar.gz && \
-    echo "6ab370532c965568df6210bd844ac6ba649f53055e48243525b0b7e5c4d69a7d rum.tar.gz" | sha256sum --check && \
+# supports v17 since https://github.com/postgrespro/rum/commit/cb1edffc57736cd2a4455f8d0feab0d69928da25
+# doesn't use releases since 1.3.13 - Sep 19, 2022
+# use latest commit from the master branch
+RUN wget https://github.com/postgrespro/rum/archive/cb1edffc57736cd2a4455f8d0feab0d69928da25.tar.gz -O rum.tar.gz && \
+    echo "65e0a752e99f4c3226400c9b899f997049e93503db8bf5c8072efa136d32fd83 rum.tar.gz" | sha256sum --check && \
     mkdir rum-src && cd rum-src && tar xzf ../rum.tar.gz --strip-components=1 -C . && \
     patch -p1 < /rum.patch && \
     make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \
@@ -666,7 +663,7 @@ RUN apt-get update && \
 #
 # Use new version only for v17
 # because Release_2024_09_1 has some backward incompatible changes
-# https://github.com/rdkit/rdkit/releases/tag/Release_2024_09_1 
+# https://github.com/rdkit/rdkit/releases/tag/Release_2024_09_1
 ENV PATH="/usr/local/pgsql/bin/:/usr/local/pgsql/:$PATH"
 RUN case "${PG_VERSION}" in \
     "v17") \
@@ -860,18 +857,98 @@ ENV PATH="/home/nonroot/.cargo/bin:/usr/local/pgsql/bin/:$PATH"
 USER nonroot
 WORKDIR /home/nonroot
 
-RUN case "${PG_VERSION}" in "v17") \
-    echo "v17 is not supported yet by pgrx. Quit" && exit 0;; \
-    esac && \
-    curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && \
+RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && \
     chmod +x rustup-init && \
     ./rustup-init -y --no-modify-path --profile minimal --default-toolchain stable && \
     rm rustup-init && \
+    case "${PG_VERSION}" in \
+        'v17') \
+            echo 'v17 is not supported yet by pgrx. Quit' && exit 0;; \
+    esac && \
     cargo install --locked --version 0.11.3 cargo-pgrx && \
     /bin/bash -c 'cargo pgrx init --pg${PG_VERSION:1}=/usr/local/pgsql/bin/pg_config'
 
 USER root
 
+#########################################################################################
+#
+# Layer "rust extensions pgrx12"
+#
+# pgrx started to support Postgres 17 since version 12,
+# but some older extension aren't compatible with it.
+# This layer should be used as a base for new pgrx extensions,
+# and eventually get merged with `rust-extensions-build`
+#
+#########################################################################################
+FROM build-deps AS rust-extensions-build-pgrx12
+ARG PG_VERSION
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+RUN apt-get update && \
+    apt-get install --no-install-recommends -y curl libclang-dev && \
+    useradd -ms /bin/bash nonroot -b /home
+
+ENV HOME=/home/nonroot
+ENV PATH="/home/nonroot/.cargo/bin:/usr/local/pgsql/bin/:$PATH"
+USER nonroot
+WORKDIR /home/nonroot
+
+RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && \
+    chmod +x rustup-init && \
+    ./rustup-init -y --no-modify-path --profile minimal --default-toolchain stable && \
+    rm rustup-init && \
+    cargo install --locked --version 0.12.6 cargo-pgrx && \
+    /bin/bash -c 'cargo pgrx init --pg${PG_VERSION:1}=/usr/local/pgsql/bin/pg_config'
+
+USER root
+
+#########################################################################################
+#
+# Layers "pg-onnx-build" and "pgrag-pg-build"
+# Compile "pgrag" extensions
+#
+#########################################################################################
+
+FROM rust-extensions-build-pgrx12 AS pg-onnx-build
+
+# cmake 3.26 or higher is required, so installing it using pip (bullseye-backports has cmake 3.25).
+# Install it using virtual environment, because Python 3.11 (the default version on Debian 12 (Bookworm)) complains otherwise
+RUN apt-get update && apt-get install -y python3 python3-pip python3-venv && \
+    python3 -m venv venv && \
+    . venv/bin/activate && \
+    python3 -m pip install cmake==3.30.5 && \
+    wget https://github.com/microsoft/onnxruntime/archive/refs/tags/v1.18.1.tar.gz -O onnxruntime.tar.gz && \
+    mkdir onnxruntime-src && cd onnxruntime-src && tar xzf ../onnxruntime.tar.gz --strip-components=1 -C . && \
+    ./build.sh --config Release --parallel --skip_submodule_sync --skip_tests --allow_running_as_root
+
+
+FROM pg-onnx-build AS pgrag-pg-build
+
+RUN apt-get install -y protobuf-compiler && \
+    wget https://github.com/neondatabase-labs/pgrag/archive/refs/tags/v0.0.0.tar.gz -O pgrag.tar.gz &&  \
+    echo "2cbe394c1e74fc8bcad9b52d5fbbfb783aef834ca3ce44626cfd770573700bb4 pgrag.tar.gz" | sha256sum --check && \
+    mkdir pgrag-src && cd pgrag-src && tar xzf ../pgrag.tar.gz --strip-components=1 -C . && \
+    \
+    cd exts/rag && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    cargo pgrx install --release && \
+    echo "trusted = true" >> /usr/local/pgsql/share/extension/rag.control && \
+    \
+    cd ../rag_bge_small_en_v15 && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    ORT_LIB_LOCATION=/home/nonroot/onnxruntime-src/build/Linux \
+        REMOTE_ONNX_URL=http://pg-ext-s3-gateway/pgrag-data/bge_small_en_v15.onnx \
+        cargo pgrx install --release --features remote_onnx && \
+    echo "trusted = true" >> /usr/local/pgsql/share/extension/rag_bge_small_en_v15.control && \
+    \
+    cd ../rag_jina_reranker_v1_tiny_en && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    ORT_LIB_LOCATION=/home/nonroot/onnxruntime-src/build/Linux \
+        REMOTE_ONNX_URL=http://pg-ext-s3-gateway/pgrag-data/jina_reranker_v1_tiny_en.onnx \
+        cargo pgrx install --release --features remote_onnx && \
+    echo "trusted = true" >> /usr/local/pgsql/share/extension/rag_jina_reranker_v1_tiny_en.control
+
+
 #########################################################################################
 #
 # Layer "pg-jsonschema-pg-build"
@@ -879,21 +956,31 @@ USER root
 #
 #########################################################################################
 
-FROM rust-extensions-build AS pg-jsonschema-pg-build
+FROM rust-extensions-build-pgrx12 AS pg-jsonschema-pg-build
 ARG PG_VERSION
-
-RUN case "${PG_VERSION}" in "v17") \
-    echo "pg_jsonschema does not yet have a release that supports pg17" && exit 0;; \
+# version 0.3.3 supports v17
+# last release v0.3.3 - Oct 16, 2024
+#
+# there were no breaking changes
+# so we can use the same version for all postgres versions
+RUN case "${PG_VERSION}" in \
+    "v14" | "v15" | "v16" | "v17") \
+        export PG_JSONSCHEMA_VERSION=0.3.3 \
+        export PG_JSONSCHEMA_CHECKSUM=40c2cffab4187e0233cb8c3bde013be92218c282f95f4469c5282f6b30d64eac \
+    ;; \
+    *) \
+        echo "unexpected PostgreSQL version" && exit 1 \
+    ;; \
     esac && \
-    wget https://github.com/supabase/pg_jsonschema/archive/refs/tags/v0.3.1.tar.gz -O pg_jsonschema.tar.gz && \
-    echo "61df3db1ed83cf24f6aa39c826f8818bfa4f0bd33b587fd6b2b1747985642297 pg_jsonschema.tar.gz" | sha256sum --check && \
+    wget https://github.com/supabase/pg_jsonschema/archive/refs/tags/v${PG_JSONSCHEMA_VERSION}.tar.gz -O pg_jsonschema.tar.gz && \
+    echo "${PG_JSONSCHEMA_CHECKSUM} pg_jsonschema.tar.gz" | sha256sum --check && \
     mkdir pg_jsonschema-src && cd pg_jsonschema-src && tar xzf ../pg_jsonschema.tar.gz --strip-components=1 -C . && \
     # see commit 252b3685a27a0f4c31a0f91e983c6314838e89e8
     # `unsafe-postgres` feature allows to build pgx extensions
     # against postgres forks that decided to change their ABI name (like us).
     # With that we can build extensions without forking them and using stock
     # pgx. As this feature is new few manual version bumps were required.
-    sed -i 's/pgrx = "0.11.3"/pgrx = { version = "0.11.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
     cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_jsonschema.control
 
@@ -904,16 +991,27 @@ RUN case "${PG_VERSION}" in "v17") \
 #
 #########################################################################################
 
-FROM rust-extensions-build AS pg-graphql-pg-build
+FROM rust-extensions-build-pgrx12 AS pg-graphql-pg-build
 ARG PG_VERSION
 
-RUN case "${PG_VERSION}" in "v17") \
-    echo "pg_graphql does not yet have a release that supports pg17 as of now" && exit 0;; \
+# version 1.5.9 supports v17
+# last release v1.5.9 - Oct 16, 2024
+#
+# there were no breaking changes
+# so we can use the same version for all postgres versions
+RUN case "${PG_VERSION}" in \
+    "v14" | "v15" | "v16" | "v17") \
+        export PG_GRAPHQL_VERSION=1.5.9 \
+        export PG_GRAPHQL_CHECKSUM=cf768385a41278be1333472204fc0328118644ae443182cf52f7b9b23277e497 \
+    ;; \
+    *) \
+        echo "unexpected PostgreSQL version" && exit 1 \
+    ;; \
     esac && \
-    wget https://github.com/supabase/pg_graphql/archive/refs/tags/v1.5.7.tar.gz -O pg_graphql.tar.gz && \
-    echo "2b3e567a5b31019cb97ae0e33263c1bcc28580be5a444ac4c8ece5c4be2aea41 pg_graphql.tar.gz" | sha256sum --check && \
+    wget https://github.com/supabase/pg_graphql/archive/refs/tags/v${PG_GRAPHQL_VERSION}.tar.gz -O pg_graphql.tar.gz && \
+    echo "${PG_GRAPHQL_CHECKSUM} pg_graphql.tar.gz" | sha256sum --check && \
     mkdir pg_graphql-src && cd pg_graphql-src && tar xzf ../pg_graphql.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgrx = "=0.11.3"/pgrx = { version = "0.11.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx = "=0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
     cargo pgrx install --release && \
     # it's needed to enable extension because it uses untrusted C language
     sed -i 's/superuser = false/superuser = true/g' /usr/local/pgsql/share/extension/pg_graphql.control && \
@@ -926,15 +1024,13 @@ RUN case "${PG_VERSION}" in "v17") \
 #
 #########################################################################################
 
-FROM rust-extensions-build AS pg-tiktoken-pg-build
+FROM rust-extensions-build-pgrx12 AS pg-tiktoken-pg-build
 ARG PG_VERSION
 
-# 26806147b17b60763039c6a6878884c41a262318 made on 26/09/2023
-RUN case "${PG_VERSION}" in "v17") \
-    echo "pg_tiktoken does not have versions, nor support for pg17" && exit 0;; \
-    esac && \
-    wget https://github.com/kelvich/pg_tiktoken/archive/26806147b17b60763039c6a6878884c41a262318.tar.gz -O pg_tiktoken.tar.gz && \
-    echo "e64e55aaa38c259512d3e27c572da22c4637418cf124caba904cd50944e5004e pg_tiktoken.tar.gz" | sha256sum --check && \
+# doesn't use releases
+# 9118dd4549b7d8c0bbc98e04322499f7bf2fa6f7 - on Oct 29, 2024
+RUN wget https://github.com/kelvich/pg_tiktoken/archive/9118dd4549b7d8c0bbc98e04322499f7bf2fa6f7.tar.gz -O pg_tiktoken.tar.gz && \
+    echo "a5bc447e7920ee149d3c064b8b9f0086c0e83939499753178f7d35788416f628 pg_tiktoken.tar.gz" | sha256sum --check && \
     mkdir pg_tiktoken-src && cd pg_tiktoken-src && tar xzf ../pg_tiktoken.tar.gz --strip-components=1 -C . && \
     # TODO update pgrx version in the pg_tiktoken repo and remove this line
     sed -i 's/pgrx = { version = "=0.10.2",/pgrx = { version = "0.11.3",/g' Cargo.toml && \
@@ -952,6 +1048,8 @@ RUN case "${PG_VERSION}" in "v17") \
 FROM rust-extensions-build AS pg-pgx-ulid-build
 ARG PG_VERSION
 
+# doesn't support v17 yet
+# https://github.com/pksunkara/pgx_ulid/pull/52
 RUN case "${PG_VERSION}" in "v17") \
     echo "pgx_ulid does not support pg17 as of the latest version (0.1.5)" && exit 0;; \
     esac && \
@@ -969,16 +1067,16 @@ RUN case "${PG_VERSION}" in "v17") \
 #
 #########################################################################################
 
-FROM rust-extensions-build AS pg-session-jwt-build
+FROM rust-extensions-build-pgrx12 AS pg-session-jwt-build
 ARG PG_VERSION
 
-RUN case "${PG_VERSION}" in "v17") \
-    echo "pg_session_jwt does not yet have a release that supports pg17" && exit 0;; \
-    esac && \
-    wget https://github.com/neondatabase/pg_session_jwt/archive/e1310b08ba51377a19e0559e4d1194883b9b2ba2.tar.gz -O pg_session_jwt.tar.gz && \
-    echo "837932a077888d5545fd54b0abcc79e5f8e37017c2769a930afc2f5c94df6f4e pg_session_jwt.tar.gz" | sha256sum --check && \
+# NOTE: local_proxy depends on the version of pg_session_jwt
+# Do not update without approve from proxy team
+# Make sure the version is reflected in proxy/src/serverless/local_conn_pool.rs
+RUN wget https://github.com/neondatabase/pg_session_jwt/archive/refs/tags/v0.1.2-v17.tar.gz -O pg_session_jwt.tar.gz && \
+    echo "c8ecbed9cb8c6441bce5134a176002b043018adf9d05a08e457dda233090a86e pg_session_jwt.tar.gz" | sha256sum --check && \
     mkdir pg_session_jwt-src && cd pg_session_jwt-src && tar xzf ../pg_session_jwt.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgrx = "=0.11.3"/pgrx = { version = "=0.11.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "=0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
     cargo pgrx install --release
 
 #########################################################################################
@@ -1041,6 +1139,34 @@ RUN wget https://github.com/pgpartman/pg_partman/archive/refs/tags/v5.1.0.tar.gz
     make -j $(getconf _NPROCESSORS_ONLN) install && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pg_partman.control
 
+#########################################################################################
+#
+# Layer "pg_mooncake"
+# compile pg_mooncake extension
+#
+#########################################################################################
+FROM rust-extensions-build AS pg-mooncake-build
+ARG PG_VERSION
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+# The topmost commit in the `neon` branch at the time of writing this
+# https://github.com/Mooncake-Labs/pg_mooncake/commits/neon/
+# https://github.com/Mooncake-Labs/pg_mooncake/commit/568b5a82b5fc16136bdf4ca5aac3e0cc261ab48d
+ENV PG_MOONCAKE_VERSION=568b5a82b5fc16136bdf4ca5aac3e0cc261ab48d
+ENV PATH="/usr/local/pgsql/bin/:$PATH"
+
+RUN case "${PG_VERSION}" in \
+        'v14') \
+            echo "pg_mooncake is not supported on Postgres ${PG_VERSION}" && exit 0;; \
+    esac && \
+    git clone --depth 1 --branch neon https://github.com/Mooncake-Labs/pg_mooncake.git pg_mooncake-src && \
+    cd pg_mooncake-src && \
+    git checkout "${PG_MOONCAKE_VERSION}" && \
+    git submodule update --init --depth 1 --recursive && \
+    make BUILD_TYPE=release -j $(getconf _NPROCESSORS_ONLN) && \
+    make BUILD_TYPE=release -j $(getconf _NPROCESSORS_ONLN) install && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/pg_mooncake.control
+
 #########################################################################################
 #
 # Layer "neon-pg-ext-build"
@@ -1059,6 +1185,7 @@ COPY --from=h3-pg-build /h3/usr /
 COPY --from=unit-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=vector-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pgjwt-pg-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pgrag-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-jsonschema-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-graphql-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-tiktoken-pg-build /usr/local/pgsql/ /usr/local/pgsql/
@@ -1084,6 +1211,7 @@ COPY --from=wal2json-pg-build /usr/local/pgsql /usr/local/pgsql
 COPY --from=pg-anon-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-ivm-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-partman-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pg-mooncake-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY pgxn/ pgxn/
 
 RUN make -j $(getconf _NPROCESSORS_ONLN) \
@@ -1191,7 +1319,10 @@ RUN mold -run cargo build --locked --profile release-line-debug-size-lto --bin l
 #########################################################################################
 
 FROM quay.io/prometheuscommunity/postgres-exporter:v0.12.1 AS postgres-exporter
-FROM burningalchemist/sql_exporter:0.13 AS sql-exporter
+
+# Keep the version the same as in build-tools.Dockerfile and
+# test_runner/regress/test_compute_metrics.py.
+FROM burningalchemist/sql_exporter:0.13.1 AS sql-exporter
 
 #########################################################################################
 #
@@ -1247,6 +1378,7 @@ COPY --from=unit-pg-build /postgresql-unit.tar.gz /ext-src/
 COPY --from=vector-pg-build /pgvector.tar.gz /ext-src/
 COPY --from=vector-pg-build /pgvector.patch /ext-src/
 COPY --from=pgjwt-pg-build /pgjwt.tar.gz /ext-src
+#COPY --from=pgrag-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 #COPY --from=pg-jsonschema-pg-build /home/nonroot/pg_jsonschema.tar.gz /ext-src
 #COPY --from=pg-graphql-pg-build /home/nonroot/pg_graphql.tar.gz /ext-src
 #COPY --from=pg-tiktoken-pg-build /home/nonroot/pg_tiktoken.tar.gz /ext-src

compute/etc/neon_collector.jsonnet

@@ -3,6 +3,7 @@
   metrics: [
     import 'sql_exporter/checkpoints_req.libsonnet',
     import 'sql_exporter/checkpoints_timed.libsonnet',
+    import 'sql_exporter/compute_backpressure_throttling_ms.libsonnet',
     import 'sql_exporter/compute_current_lsn.libsonnet',
     import 'sql_exporter/compute_logical_snapshot_files.libsonnet',
     import 'sql_exporter/compute_receive_lsn.libsonnet',

compute/etc/sql_exporter.jsonnet

@@ -1,4 +1,4 @@
-function(collector_name, collector_file, application_name='sql_exporter') {
+function(collector_name, collector_file, connection_string) {
   // Configuration for sql_exporter for autoscaling-agent
   // Global defaults.
   global: {
@@ -23,7 +23,7 @@ function(collector_name, collector_file, application_name='sql_exporter') {
   target: {
     // Data source name always has a URI schema that matches the driver name. In some cases (e.g. MySQL)
     // the schema gets dropped or replaced to match the driver expected DSN format.
-    data_source_name: std.format('postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=%s', [application_name]),
+    data_source_name: connection_string,
 
     // Collectors (referenced by name) to execute on the target.
     // Glob patterns are supported (see <https://pkg.go.dev/path/filepath#Match> for syntax).

compute/etc/sql_exporter/compute_backpressure_throttling_ms.libsonnet

@@ -0,0 +1,10 @@
+{
+  metric_name: 'compute_backpressure_throttling_ms',
+  type: 'gauge',
+  help: 'Time compute has spent throttled',
+  key_labels: null,
+  values: [
+    'throttled',
+  ],
+  query: importstr 'sql_exporter/compute_backpressure_throttling_ms.sql',
+}

compute/etc/sql_exporter/compute_backpressure_throttling_ms.sql

@@ -0,0 +1 @@
+SELECT neon.backpressure_throttling_time() AS throttled;

compute/vm-image-spec-bookworm.yaml

@@ -18,7 +18,7 @@ commands:
   - name: pgbouncer
     user: postgres
     sysvInitAction: respawn
-    shell: '/usr/local/bin/pgbouncer /etc/pgbouncer.ini'
+    shell: '/usr/local/bin/pgbouncer /etc/pgbouncer.ini 2>&1 > /dev/virtio-ports/tech.neon.log.0'
   - name: local_proxy
     user: postgres
     sysvInitAction: respawn

compute/vm-image-spec-bullseye.yaml

@@ -18,7 +18,7 @@ commands:
   - name: pgbouncer
     user: postgres
     sysvInitAction: respawn
-    shell: '/usr/local/bin/pgbouncer /etc/pgbouncer.ini'
+    shell: '/usr/local/bin/pgbouncer /etc/pgbouncer.ini 2>&1 > /dev/virtio-ports/tech.neon.log.0'
   - name: local_proxy
     user: postgres
     sysvInitAction: respawn

compute_tools/src/compute.rs

@@ -1,7 +1,6 @@
 use std::collections::HashMap;
 use std::env;
 use std::fs;
-use std::io::BufRead;
 use std::os::unix::fs::{symlink, PermissionsExt};
 use std::path::Path;
 use std::process::{Command, Stdio};
@@ -365,8 +364,7 @@ impl ComputeNode {
         let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
 
         let basebackup_cmd = match lsn {
-            // HACK We don't use compression on first start (Lsn(0)) because there's no API for it
-            Lsn(0) => format!("basebackup {} {}", spec.tenant_id, spec.timeline_id),
+            Lsn(0) => format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id),
             _ => format!(
                 "basebackup {} {} {} --gzip",
                 spec.tenant_id, spec.timeline_id, lsn
@@ -375,38 +373,16 @@ impl ComputeNode {
 
         let copyreader = client.copy_out(basebackup_cmd.as_str())?;
         let mut measured_reader = MeasuredReader::new(copyreader);
-
-        // Check the magic number to see if it's a gzip or not. Even though
-        // we might explicitly ask for gzip, an old pageserver with no implementation
-        // of gzip compression might send us uncompressed data. After some time
-        // passes we can assume all pageservers know how to compress and we can
-        // delete this check.
-        //
-        // If the data is not gzip, it will be tar. It will not be mistakenly
-        // recognized as gzip because tar starts with an ascii encoding of a filename,
-        // and 0x1f and 0x8b are unlikely first characters for any filename. Moreover,
-        // we send the "global" directory first from the pageserver, so it definitely
-        // won't be recognized as gzip.
         let mut bufreader = std::io::BufReader::new(&mut measured_reader);
-        let gzip = {
-            let peek = bufreader.fill_buf().unwrap();
-            peek[0] == 0x1f && peek[1] == 0x8b
-        };
 
         // Read the archive directly from the `CopyOutReader`
         //
         // Set `ignore_zeros` so that unpack() reads all the Copy data and
         // doesn't stop at the end-of-archive marker. Otherwise, if the server
         // sends an Error after finishing the tarball, we will not notice it.
-        if gzip {
-            let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
-            ar.set_ignore_zeros(true);
-            ar.unpack(&self.pgdata)?;
-        } else {
-            let mut ar = tar::Archive::new(&mut bufreader);
-            ar.set_ignore_zeros(true);
-            ar.unpack(&self.pgdata)?;
-        };
+        let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
+        ar.set_ignore_zeros(true);
+        ar.unpack(&self.pgdata)?;
 
         // Report metrics
         let mut state = self.state.lock().unwrap();

control_plane/src/bin/neon_local.rs

@@ -1073,10 +1073,10 @@ async fn handle_tenant(subcmd: &TenantCmd, env: &mut local_env::LocalEnv) -> any
                     tenant_id,
                     TimelineCreateRequest {
                         new_timeline_id,
-                        ancestor_timeline_id: None,
-                        ancestor_start_lsn: None,
-                        existing_initdb_timeline_id: None,
-                        pg_version: Some(args.pg_version),
+                        mode: pageserver_api::models::TimelineCreateRequestMode::Bootstrap {
+                            existing_initdb_timeline_id: None,
+                            pg_version: Some(args.pg_version),
+                        },
                     },
                 )
                 .await?;
@@ -1133,10 +1133,10 @@ async fn handle_timeline(cmd: &TimelineCmd, env: &mut local_env::LocalEnv) -> Re
             let storage_controller = StorageController::from_env(env);
             let create_req = TimelineCreateRequest {
                 new_timeline_id,
-                ancestor_timeline_id: None,
-                existing_initdb_timeline_id: None,
-                ancestor_start_lsn: None,
-                pg_version: Some(args.pg_version),
+                mode: pageserver_api::models::TimelineCreateRequestMode::Bootstrap {
+                    existing_initdb_timeline_id: None,
+                    pg_version: Some(args.pg_version),
+                },
             };
             let timeline_info = storage_controller
                 .tenant_timeline_create(tenant_id, create_req)
@@ -1189,10 +1189,11 @@ async fn handle_timeline(cmd: &TimelineCmd, env: &mut local_env::LocalEnv) -> Re
             let storage_controller = StorageController::from_env(env);
             let create_req = TimelineCreateRequest {
                 new_timeline_id,
-                ancestor_timeline_id: Some(ancestor_timeline_id),
-                existing_initdb_timeline_id: None,
-                ancestor_start_lsn: start_lsn,
-                pg_version: None,
+                mode: pageserver_api::models::TimelineCreateRequestMode::Branch {
+                    ancestor_timeline_id,
+                    ancestor_start_lsn: start_lsn,
+                    pg_version: None,
+                },
             };
             let timeline_info = storage_controller
                 .tenant_timeline_create(tenant_id, create_req)

control_plane/src/pageserver.rs

@@ -17,7 +17,7 @@ use std::time::Duration;
 
 use anyhow::{bail, Context};
 use camino::Utf8PathBuf;
-use pageserver_api::models::{self, AuxFilePolicy, TenantInfo, TimelineInfo};
+use pageserver_api::models::{self, TenantInfo, TimelineInfo};
 use pageserver_api::shard::TenantShardId;
 use pageserver_client::mgmt_api;
 use postgres_backend::AuthType;
@@ -399,11 +399,6 @@ impl PageServerNode {
                 .map(serde_json::from_str)
                 .transpose()
                 .context("parse `timeline_get_throttle` from json")?,
-            switch_aux_file_policy: settings
-                .remove("switch_aux_file_policy")
-                .map(|x| x.parse::<AuxFilePolicy>())
-                .transpose()
-                .context("Failed to parse 'switch_aux_file_policy'")?,
             lsn_lease_length: settings.remove("lsn_lease_length").map(|x| x.to_string()),
             lsn_lease_length_for_ts: settings
                 .remove("lsn_lease_length_for_ts")
@@ -499,11 +494,6 @@ impl PageServerNode {
                     .map(serde_json::from_str)
                     .transpose()
                     .context("parse `timeline_get_throttle` from json")?,
-                switch_aux_file_policy: settings
-                    .remove("switch_aux_file_policy")
-                    .map(|x| x.parse::<AuxFilePolicy>())
-                    .transpose()
-                    .context("Failed to parse 'switch_aux_file_policy'")?,
                 lsn_lease_length: settings.remove("lsn_lease_length").map(|x| x.to_string()),
                 lsn_lease_length_for_ts: settings
                     .remove("lsn_lease_length_for_ts")
@@ -529,28 +519,6 @@ impl PageServerNode {
         Ok(self.http_client.list_timelines(*tenant_shard_id).await?)
     }
 
-    pub async fn timeline_create(
-        &self,
-        tenant_shard_id: TenantShardId,
-        new_timeline_id: TimelineId,
-        ancestor_start_lsn: Option<Lsn>,
-        ancestor_timeline_id: Option<TimelineId>,
-        pg_version: Option<u32>,
-        existing_initdb_timeline_id: Option<TimelineId>,
-    ) -> anyhow::Result<TimelineInfo> {
-        let req = models::TimelineCreateRequest {
-            new_timeline_id,
-            ancestor_start_lsn,
-            ancestor_timeline_id,
-            pg_version,
-            existing_initdb_timeline_id,
-        };
-        Ok(self
-            .http_client
-            .timeline_create(tenant_shard_id, &req)
-            .await?)
-    }
-
     /// Import a basebackup prepared using either:
     /// a) `pg_basebackup -F tar`, or
     /// b) The `fullbackup` pageserver endpoint

control_plane/storcon_cli/src/main.rs

@@ -111,6 +111,11 @@ enum Command {
         #[arg(long)]
         node: NodeId,
     },
+    /// Cancel any ongoing reconciliation for this shard
+    TenantShardCancelReconcile {
+        #[arg(long)]
+        tenant_shard_id: TenantShardId,
+    },
     /// 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 {
@@ -535,6 +540,15 @@ async fn main() -> anyhow::Result<()> {
                 )
                 .await?;
         }
+        Command::TenantShardCancelReconcile { tenant_shard_id } => {
+            storcon_client
+                .dispatch::<(), ()>(
+                    Method::PUT,
+                    format!("control/v1/tenant/{tenant_shard_id}/cancel_reconcile"),
+                    None,
+                )
+                .await?;
+        }
         Command::TenantConfig { tenant_id, config } => {
             let tenant_conf = serde_json::from_str(&config)?;
 

libs/metrics/src/lib.rs

@@ -19,6 +19,7 @@ use once_cell::sync::Lazy;
 use prometheus::core::{
     Atomic, AtomicU64, Collector, GenericCounter, GenericCounterVec, GenericGauge, GenericGaugeVec,
 };
+pub use prometheus::local::LocalHistogram;
 pub use prometheus::opts;
 pub use prometheus::register;
 pub use prometheus::Error;

libs/pageserver_api/src/config.rs

@@ -250,12 +250,6 @@ pub struct TenantConfigToml {
     // Expresed in multiples of checkpoint distance.
     pub image_layer_creation_check_threshold: u8,
 
-    /// Switch to a new aux file policy. Switching this flag requires the user has not written any aux file into
-    /// the storage before, and this flag cannot be switched back. Otherwise there will be data corruptions.
-    /// There is a `last_aux_file_policy` flag which gets persisted in `index_part.json` once the first aux
-    /// file is written.
-    pub switch_aux_file_policy: crate::models::AuxFilePolicy,
-
     /// The length for an explicit LSN lease request.
     /// Layers needed to reconstruct pages at LSN will not be GC-ed during this interval.
     #[serde(with = "humantime_serde")]
@@ -475,7 +469,6 @@ impl Default for TenantConfigToml {
             lazy_slru_download: false,
             timeline_get_throttle: crate::models::ThrottleConfig::disabled(),
             image_layer_creation_check_threshold: DEFAULT_IMAGE_LAYER_CREATION_CHECK_THRESHOLD,
-            switch_aux_file_policy: crate::models::AuxFilePolicy::default_tenant_config(),
             lsn_lease_length: LsnLease::DEFAULT_LENGTH,
             lsn_lease_length_for_ts: LsnLease::DEFAULT_LENGTH_FOR_TS,
         }

libs/pageserver_api/src/lib.rs

@@ -5,9 +5,11 @@ pub mod controller_api;
 pub mod key;
 pub mod keyspace;
 pub mod models;
+pub mod record;
 pub mod reltag;
 pub mod shard;
 /// Public API types
 pub mod upcall_api;
+pub mod value;
 
 pub mod config;

libs/pageserver_api/src/models.rs

@@ -10,7 +10,6 @@ use std::{
     io::{BufRead, Read},
     num::{NonZeroU32, NonZeroU64, NonZeroUsize},
     str::FromStr,
-    sync::atomic::AtomicUsize,
     time::{Duration, SystemTime},
 };
 
@@ -211,13 +210,30 @@ pub enum TimelineState {
 #[derive(Serialize, Deserialize, Clone)]
 pub struct TimelineCreateRequest {
     pub new_timeline_id: TimelineId,
-    #[serde(default)]
-    pub ancestor_timeline_id: Option<TimelineId>,
-    #[serde(default)]
-    pub existing_initdb_timeline_id: Option<TimelineId>,
-    #[serde(default)]
-    pub ancestor_start_lsn: Option<Lsn>,
-    pub pg_version: Option<u32>,
+    #[serde(flatten)]
+    pub mode: TimelineCreateRequestMode,
+}
+
+#[derive(Serialize, Deserialize, Clone)]
+#[serde(untagged)]
+pub enum TimelineCreateRequestMode {
+    Branch {
+        ancestor_timeline_id: TimelineId,
+        #[serde(default)]
+        ancestor_start_lsn: Option<Lsn>,
+        // TODO: cplane sets this, but, the branching code always
+        // inherits the ancestor's pg_version. Earlier code wasn't
+        // using a flattened enum, so, it was an accepted field, and
+        // we continue to accept it by having it here.
+        pg_version: Option<u32>,
+    },
+    // NB: Bootstrap is all-optional, and thus the serde(untagged) will cause serde to stop at Bootstrap.
+    // (serde picks the first matching enum variant, in declaration order).
+    Bootstrap {
+        #[serde(default)]
+        existing_initdb_timeline_id: Option<TimelineId>,
+        pg_version: Option<u32>,
+    },
 }
 
 #[derive(Serialize, Deserialize, Clone)]
@@ -292,7 +308,6 @@ pub struct TenantConfig {
     pub lazy_slru_download: Option<bool>,
     pub timeline_get_throttle: Option<ThrottleConfig>,
     pub image_layer_creation_check_threshold: Option<u8>,
-    pub switch_aux_file_policy: Option<AuxFilePolicy>,
     pub lsn_lease_length: Option<String>,
     pub lsn_lease_length_for_ts: Option<String>,
 }
@@ -333,68 +348,6 @@ pub enum AuxFilePolicy {
     CrossValidation,
 }
 
-impl AuxFilePolicy {
-    pub fn is_valid_migration_path(from: Option<Self>, to: Self) -> bool {
-        matches!(
-            (from, to),
-            (None, _) | (Some(AuxFilePolicy::CrossValidation), AuxFilePolicy::V2)
-        )
-    }
-
-    /// If a tenant writes aux files without setting `switch_aux_policy`, this value will be used.
-    pub fn default_tenant_config() -> Self {
-        Self::V2
-    }
-}
-
-/// The aux file policy memory flag. Users can store `Option<AuxFilePolicy>` into this atomic flag. 0 == unspecified.
-pub struct AtomicAuxFilePolicy(AtomicUsize);
-
-impl AtomicAuxFilePolicy {
-    pub fn new(policy: Option<AuxFilePolicy>) -> Self {
-        Self(AtomicUsize::new(
-            policy.map(AuxFilePolicy::to_usize).unwrap_or_default(),
-        ))
-    }
-
-    pub fn load(&self) -> Option<AuxFilePolicy> {
-        match self.0.load(std::sync::atomic::Ordering::Acquire) {
-            0 => None,
-            other => Some(AuxFilePolicy::from_usize(other)),
-        }
-    }
-
-    pub fn store(&self, policy: Option<AuxFilePolicy>) {
-        self.0.store(
-            policy.map(AuxFilePolicy::to_usize).unwrap_or_default(),
-            std::sync::atomic::Ordering::Release,
-        );
-    }
-}
-
-impl AuxFilePolicy {
-    pub fn to_usize(self) -> usize {
-        match self {
-            Self::V1 => 1,
-            Self::CrossValidation => 2,
-            Self::V2 => 3,
-        }
-    }
-
-    pub fn try_from_usize(this: usize) -> Option<Self> {
-        match this {
-            1 => Some(Self::V1),
-            2 => Some(Self::CrossValidation),
-            3 => Some(Self::V2),
-            _ => None,
-        }
-    }
-
-    pub fn from_usize(this: usize) -> Self {
-        Self::try_from_usize(this).unwrap()
-    }
-}
-
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(tag = "kind")]
 pub enum EvictionPolicy {
@@ -1051,6 +1004,12 @@ pub mod virtual_file {
     }
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ScanDisposableKeysResponse {
+    pub disposable_count: usize,
+    pub not_disposable_count: usize,
+}
+
 // Wrapped in libpq CopyData
 #[derive(PartialEq, Eq, Debug)]
 pub enum PagestreamFeMessage {
@@ -1610,71 +1569,6 @@ mod tests {
         }
     }
 
-    #[test]
-    fn test_aux_file_migration_path() {
-        assert!(AuxFilePolicy::is_valid_migration_path(
-            None,
-            AuxFilePolicy::V1
-        ));
-        assert!(AuxFilePolicy::is_valid_migration_path(
-            None,
-            AuxFilePolicy::V2
-        ));
-        assert!(AuxFilePolicy::is_valid_migration_path(
-            None,
-            AuxFilePolicy::CrossValidation
-        ));
-        // Self-migration is not a valid migration path, and the caller should handle it by itself.
-        assert!(!AuxFilePolicy::is_valid_migration_path(
-            Some(AuxFilePolicy::V1),
-            AuxFilePolicy::V1
-        ));
-        assert!(!AuxFilePolicy::is_valid_migration_path(
-            Some(AuxFilePolicy::V2),
-            AuxFilePolicy::V2
-        ));
-        assert!(!AuxFilePolicy::is_valid_migration_path(
-            Some(AuxFilePolicy::CrossValidation),
-            AuxFilePolicy::CrossValidation
-        ));
-        // Migrations not allowed
-        assert!(!AuxFilePolicy::is_valid_migration_path(
-            Some(AuxFilePolicy::CrossValidation),
-            AuxFilePolicy::V1
-        ));
-        assert!(!AuxFilePolicy::is_valid_migration_path(
-            Some(AuxFilePolicy::V1),
-            AuxFilePolicy::V2
-        ));
-        assert!(!AuxFilePolicy::is_valid_migration_path(
-            Some(AuxFilePolicy::V2),
-            AuxFilePolicy::V1
-        ));
-        assert!(!AuxFilePolicy::is_valid_migration_path(
-            Some(AuxFilePolicy::V2),
-            AuxFilePolicy::CrossValidation
-        ));
-        assert!(!AuxFilePolicy::is_valid_migration_path(
-            Some(AuxFilePolicy::V1),
-            AuxFilePolicy::CrossValidation
-        ));
-        // Migrations allowed
-        assert!(AuxFilePolicy::is_valid_migration_path(
-            Some(AuxFilePolicy::CrossValidation),
-            AuxFilePolicy::V2
-        ));
-    }
-
-    #[test]
-    fn test_aux_parse() {
-        assert_eq!(AuxFilePolicy::from_str("V2").unwrap(), AuxFilePolicy::V2);
-        assert_eq!(AuxFilePolicy::from_str("v2").unwrap(), AuxFilePolicy::V2);
-        assert_eq!(
-            AuxFilePolicy::from_str("cross-validation").unwrap(),
-            AuxFilePolicy::CrossValidation
-        );
-    }
-
     #[test]
     fn test_image_compression_algorithm_parsing() {
         use ImageCompressionAlgorithm::*;

libs/pageserver_api/src/record.rs

@@ -0,0 +1,113 @@
+//! This module defines the WAL record format used within the pageserver.
+
+use bytes::Bytes;
+use postgres_ffi::walrecord::{describe_postgres_wal_record, MultiXactMember};
+use postgres_ffi::{MultiXactId, MultiXactOffset, TimestampTz, TransactionId};
+use serde::{Deserialize, Serialize};
+use utils::bin_ser::DeserializeError;
+
+/// Each update to a page is represented by a NeonWalRecord. It can be a wrapper
+/// around a PostgreSQL WAL record, or a custom neon-specific "record".
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub enum NeonWalRecord {
+    /// Native PostgreSQL WAL record
+    Postgres { will_init: bool, rec: Bytes },
+
+    /// Clear bits in heap visibility map. ('flags' is bitmap of bits to clear)
+    ClearVisibilityMapFlags {
+        new_heap_blkno: Option<u32>,
+        old_heap_blkno: Option<u32>,
+        flags: u8,
+    },
+    /// Mark transaction IDs as committed on a CLOG page
+    ClogSetCommitted {
+        xids: Vec<TransactionId>,
+        timestamp: TimestampTz,
+    },
+    /// Mark transaction IDs as aborted on a CLOG page
+    ClogSetAborted { xids: Vec<TransactionId> },
+    /// Extend multixact offsets SLRU
+    MultixactOffsetCreate {
+        mid: MultiXactId,
+        moff: MultiXactOffset,
+    },
+    /// Extend multixact members SLRU.
+    MultixactMembersCreate {
+        moff: MultiXactOffset,
+        members: Vec<MultiXactMember>,
+    },
+    /// Update the map of AUX files, either writing or dropping an entry
+    AuxFile {
+        file_path: String,
+        content: Option<Bytes>,
+    },
+
+    /// A testing record for unit testing purposes. It supports append data to an existing image, or clear it.
+    #[cfg(feature = "testing")]
+    Test {
+        /// Append a string to the image.
+        append: String,
+        /// Clear the image before appending.
+        clear: bool,
+        /// Treat this record as an init record. `clear` should be set to true if this field is set
+        /// to true. This record does not need the history WALs to reconstruct. See [`NeonWalRecord::will_init`] and
+        /// its references in `timeline.rs`.
+        will_init: bool,
+    },
+}
+
+impl NeonWalRecord {
+    /// Does replaying this WAL record initialize the page from scratch, or does
+    /// it need to be applied over the previous image of the page?
+    pub fn will_init(&self) -> bool {
+        // If you change this function, you'll also need to change ValueBytes::will_init
+        match self {
+            NeonWalRecord::Postgres { will_init, rec: _ } => *will_init,
+            #[cfg(feature = "testing")]
+            NeonWalRecord::Test { will_init, .. } => *will_init,
+            // None of the special neon record types currently initialize the page
+            _ => false,
+        }
+    }
+
+    #[cfg(feature = "testing")]
+    pub fn wal_append(s: impl AsRef<str>) -> Self {
+        Self::Test {
+            append: s.as_ref().to_string(),
+            clear: false,
+            will_init: false,
+        }
+    }
+
+    #[cfg(feature = "testing")]
+    pub fn wal_clear() -> Self {
+        Self::Test {
+            append: "".to_string(),
+            clear: true,
+            will_init: false,
+        }
+    }
+
+    #[cfg(feature = "testing")]
+    pub fn wal_init() -> Self {
+        Self::Test {
+            append: "".to_string(),
+            clear: true,
+            will_init: true,
+        }
+    }
+}
+
+/// Build a human-readable string to describe a WAL record
+///
+/// For debugging purposes
+pub fn describe_wal_record(rec: &NeonWalRecord) -> Result<String, DeserializeError> {
+    match rec {
+        NeonWalRecord::Postgres { will_init, rec } => Ok(format!(
+            "will_init: {}, {}",
+            will_init,
+            describe_postgres_wal_record(rec)?
+        )),
+        _ => Ok(format!("{:?}", rec)),
+    }
+}

libs/pageserver_api/src/value.rs

@@ -1,13 +1,16 @@
-use crate::walrecord::NeonWalRecord;
-use anyhow::Result;
+//! This module defines the value type used by the storage engine.
+//!
+//! A [`Value`] represents either a completely new value for one Key ([`Value::Image`]),
+//! or a "delta" of how to get from previous version of the value to the new one
+//! ([`Value::WalRecord`]])
+//!
+//! Note that the [`Value`] type is used for the permananent storage format, so any
+//! changes to it must be backwards compatible.
+
+use crate::record::NeonWalRecord;
 use bytes::Bytes;
 use serde::{Deserialize, Serialize};
-use std::ops::AddAssign;
-use std::time::Duration;
 
-pub use pageserver_api::key::{Key, KEY_SIZE};
-
-/// A 'value' stored for a one Key.
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
 pub enum Value {
     /// An Image value contains a full copy of the value
@@ -20,10 +23,12 @@ pub enum Value {
 }
 
 impl Value {
+    #[inline(always)]
     pub fn is_image(&self) -> bool {
         matches!(self, Value::Image(_))
     }
 
+    #[inline(always)]
     pub fn will_init(&self) -> bool {
         match self {
             Value::Image(_) => true,
@@ -33,17 +38,18 @@ impl Value {
 }
 
 #[derive(Debug, PartialEq)]
-pub(crate) enum InvalidInput {
+pub enum InvalidInput {
     TooShortValue,
     TooShortPostgresRecord,
 }
 
 /// We could have a ValueRef where everything is `serde(borrow)`. Before implementing that, lets
 /// use this type for querying if a slice looks some particular way.
-pub(crate) struct ValueBytes;
+pub struct ValueBytes;
 
 impl ValueBytes {
-    pub(crate) fn will_init(raw: &[u8]) -> Result<bool, InvalidInput> {
+    #[inline(always)]
+    pub fn will_init(raw: &[u8]) -> Result<bool, InvalidInput> {
         if raw.len() < 12 {
             return Err(InvalidInput::TooShortValue);
         }
@@ -79,6 +85,7 @@ impl ValueBytes {
 mod test {
     use super::*;
 
+    use bytes::Bytes;
     use utils::bin_ser::BeSer;
 
     macro_rules! roundtrip {
@@ -229,56 +236,3 @@ mod test {
         assert!(!ValueBytes::will_init(&expected).unwrap());
     }
 }
-
-///
-/// Result of performing GC
-///
-#[derive(Default, Serialize, Debug)]
-pub struct GcResult {
-    pub layers_total: u64,
-    pub layers_needed_by_cutoff: u64,
-    pub layers_needed_by_pitr: u64,
-    pub layers_needed_by_branches: u64,
-    pub layers_needed_by_leases: u64,
-    pub layers_not_updated: u64,
-    pub layers_removed: u64, // # of layer files removed because they have been made obsolete by newer ondisk files.
-
-    #[serde(serialize_with = "serialize_duration_as_millis")]
-    pub elapsed: Duration,
-
-    /// The layers which were garbage collected.
-    ///
-    /// Used in `/v1/tenant/:tenant_id/timeline/:timeline_id/do_gc` to wait for the layers to be
-    /// dropped in tests.
-    #[cfg(feature = "testing")]
-    #[serde(skip)]
-    pub(crate) doomed_layers: Vec<crate::tenant::storage_layer::Layer>,
-}
-
-// helper function for `GcResult`, serializing a `Duration` as an integer number of milliseconds
-fn serialize_duration_as_millis<S>(d: &Duration, serializer: S) -> Result<S::Ok, S::Error>
-where
-    S: serde::Serializer,
-{
-    d.as_millis().serialize(serializer)
-}
-
-impl AddAssign for GcResult {
-    fn add_assign(&mut self, other: Self) {
-        self.layers_total += other.layers_total;
-        self.layers_needed_by_pitr += other.layers_needed_by_pitr;
-        self.layers_needed_by_cutoff += other.layers_needed_by_cutoff;
-        self.layers_needed_by_branches += other.layers_needed_by_branches;
-        self.layers_needed_by_leases += other.layers_needed_by_leases;
-        self.layers_not_updated += other.layers_not_updated;
-        self.layers_removed += other.layers_removed;
-
-        self.elapsed += other.elapsed;
-
-        #[cfg(feature = "testing")]
-        {
-            let mut other = other;
-            self.doomed_layers.append(&mut other.doomed_layers);
-        }
-    }
-}

libs/postgres_ffi/Cargo.toml

@@ -15,6 +15,7 @@ memoffset.workspace = true
 thiserror.workspace = true
 serde.workspace = true
 utils.workspace = true
+tracing.workspace = true
 
 [dev-dependencies]
 env_logger.workspace = true

libs/postgres_ffi/src/lib.rs

@@ -36,6 +36,7 @@ macro_rules! postgres_ffi {
             pub mod controlfile_utils;
             pub mod nonrelfile_utils;
             pub mod wal_craft_test_export;
+            pub mod wal_generator;
             pub mod waldecoder_handler;
             pub mod xlog_utils;
 
@@ -217,6 +218,7 @@ macro_rules! enum_pgversion {
 
 pub mod pg_constants;
 pub mod relfile_utils;
+pub mod walrecord;
 
 // Export some widely used datatypes that are unlikely to change across Postgres versions
 pub use v14::bindings::RepOriginId;

libs/postgres_ffi/src/wal_generator.rs

@@ -0,0 +1,203 @@
+use std::ffi::CStr;
+
+use bytes::{Bytes, BytesMut};
+use crc32c::crc32c_append;
+use utils::lsn::Lsn;
+
+use super::bindings::{XLogLongPageHeaderData, XLogPageHeaderData, XLOG_PAGE_MAGIC};
+use super::xlog_utils::{
+    XlLogicalMessage, XLOG_RECORD_CRC_OFFS, XLOG_SIZE_OF_XLOG_RECORD, XLP_BKP_REMOVABLE,
+    XLP_FIRST_IS_CONTRECORD,
+};
+use super::XLogRecord;
+use crate::pg_constants::{
+    RM_LOGICALMSG_ID, XLOG_LOGICAL_MESSAGE, XLP_LONG_HEADER, XLR_BLOCK_ID_DATA_LONG,
+    XLR_BLOCK_ID_DATA_SHORT,
+};
+use crate::{WAL_SEGMENT_SIZE, XLOG_BLCKSZ};
+
+/// Generates binary WAL records for use in tests and benchmarks. Currently only generates logical
+/// messages (effectively noops) with a fixed payload. It is used as an iterator which yields
+/// encoded bytes for a single WAL record, including internal page headers if it spans pages.
+/// Concatenating the bytes will yield a complete, well-formed WAL, which can be chunked at segment
+/// boundaries if desired. Not optimized for performance.
+///
+/// The WAL format is version-dependant (see e.g. `XLOG_PAGE_MAGIC`), so make sure to import this
+/// for the appropriate Postgres version (e.g. `postgres_ffi::v17::wal_generator::WalGenerator`).
+///
+/// A WAL is split into 16 MB segments. Each segment is split into 8 KB pages, with headers.
+/// Records are arbitrary length, 8-byte aligned, and may span pages. The layout is e.g.:
+///
+/// |        Segment 1         |        Segment 2         |        Segment 3         |
+/// | Page 1 | Page 2 | Page 3 | Page 4 | Page 5 | Page 6 | Page 7 | Page 8 | Page 9 |
+/// | R1 |   R2  |R3|  R4  | R5  |  R6  |                 R7            | R8  |
+///
+/// TODO: support generating actual tables and rows.
+#[derive(Default)]
+pub struct WalGenerator {
+    /// Current LSN to append the next record at.
+    ///
+    /// Callers can modify this (and prev_lsn) to restart generation at a different LSN, but should
+    /// ensure that the LSN is on a valid record boundary (i.e. we can't start appending in the
+    /// middle on an existing record or header, or beyond the end of the existing WAL).
+    pub lsn: Lsn,
+    /// The starting LSN of the previous record. Used in WAL record headers. The Safekeeper doesn't
+    /// care about this, unlike Postgres, but we include it for completeness.
+    pub prev_lsn: Lsn,
+}
+
+impl WalGenerator {
+    // For now, hardcode the message payload.
+    // TODO: support specifying the payload size.
+    const PREFIX: &CStr = c"prefix";
+    const MESSAGE: &[u8] = b"message";
+
+    // Hardcode the sys, timeline, and DB IDs. We can make them configurable if we care about them.
+    const SYS_ID: u64 = 0;
+    const TIMELINE_ID: u32 = 1;
+    const DB_ID: u32 = 0;
+
+    /// Creates a new WAL generator, which emits logical message records (noops).
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Encodes a logical message (basically a noop), with the given prefix and message.
+    pub(crate) fn encode_logical_message(prefix: &CStr, message: &[u8]) -> Bytes {
+        let prefix = prefix.to_bytes_with_nul();
+        let header = XlLogicalMessage {
+            db_id: Self::DB_ID,
+            transactional: 0,
+            prefix_size: prefix.len() as u64,
+            message_size: message.len() as u64,
+        };
+        [&header.encode(), prefix, message].concat().into()
+    }
+
+    /// Encode a WAL record with the given payload data (e.g. a logical message).
+    pub(crate) fn encode_record(data: Bytes, rmid: u8, info: u8, prev_lsn: Lsn) -> Bytes {
+        // Prefix data with block ID and length.
+        let data_header = Bytes::from(match data.len() {
+            0 => vec![],
+            1..=255 => vec![XLR_BLOCK_ID_DATA_SHORT, data.len() as u8],
+            256.. => {
+                let len_bytes = (data.len() as u32).to_le_bytes();
+                [&[XLR_BLOCK_ID_DATA_LONG], len_bytes.as_slice()].concat()
+            }
+        });
+
+        // Construct the WAL record header.
+        let mut header = XLogRecord {
+            xl_tot_len: (XLOG_SIZE_OF_XLOG_RECORD + data_header.len() + data.len()) as u32,
+            xl_xid: 0,
+            xl_prev: prev_lsn.into(),
+            xl_info: info,
+            xl_rmid: rmid,
+            __bindgen_padding_0: [0; 2],
+            xl_crc: 0, // see below
+        };
+
+        // Compute the CRC checksum for the data, and the header up to the CRC field.
+        let mut crc = 0;
+        crc = crc32c_append(crc, &data_header);
+        crc = crc32c_append(crc, &data);
+        crc = crc32c_append(crc, &header.encode().unwrap()[0..XLOG_RECORD_CRC_OFFS]);
+        header.xl_crc = crc;
+
+        // Encode the final header and record.
+        let header = header.encode().unwrap();
+
+        [header, data_header, data].concat().into()
+    }
+
+    /// Injects page headers on 8KB page boundaries. Takes the current LSN position where the record
+    /// is to be appended.
+    fn encode_pages(record: Bytes, mut lsn: Lsn) -> Bytes {
+        // Fast path: record fits in current page, and the page already has a header.
+        if lsn.remaining_in_block() as usize >= record.len() && lsn.block_offset() > 0 {
+            return record;
+        }
+
+        let mut pages = BytesMut::new();
+        let mut remaining = record.clone(); // Bytes::clone() is cheap
+        while !remaining.is_empty() {
+            // At new page boundary, inject page header.
+            if lsn.block_offset() == 0 {
+                let mut page_header = XLogPageHeaderData {
+                    xlp_magic: XLOG_PAGE_MAGIC as u16,
+                    xlp_info: XLP_BKP_REMOVABLE,
+                    xlp_tli: Self::TIMELINE_ID,
+                    xlp_pageaddr: lsn.0,
+                    xlp_rem_len: 0,
+                    __bindgen_padding_0: [0; 4],
+                };
+                // If the record was split across page boundaries, mark as continuation.
+                if remaining.len() < record.len() {
+                    page_header.xlp_rem_len = remaining.len() as u32;
+                    page_header.xlp_info |= XLP_FIRST_IS_CONTRECORD;
+                }
+                // At start of segment, use a long page header.
+                let page_header = if lsn.segment_offset(WAL_SEGMENT_SIZE) == 0 {
+                    page_header.xlp_info |= XLP_LONG_HEADER;
+                    XLogLongPageHeaderData {
+                        std: page_header,
+                        xlp_sysid: Self::SYS_ID,
+                        xlp_seg_size: WAL_SEGMENT_SIZE as u32,
+                        xlp_xlog_blcksz: XLOG_BLCKSZ as u32,
+                    }
+                    .encode()
+                    .unwrap()
+                } else {
+                    page_header.encode().unwrap()
+                };
+                pages.extend_from_slice(&page_header);
+                lsn += page_header.len() as u64;
+            }
+
+            // Append the record up to the next page boundary, if any.
+            let page_free = lsn.remaining_in_block() as usize;
+            let chunk = remaining.split_to(std::cmp::min(page_free, remaining.len()));
+            pages.extend_from_slice(&chunk);
+            lsn += chunk.len() as u64;
+        }
+        pages.freeze()
+    }
+
+    /// Records must be 8-byte aligned. Take an encoded record (including any injected page
+    /// boundaries), starting at the given LSN, and add any necessary padding at the end.
+    fn pad_record(record: Bytes, mut lsn: Lsn) -> Bytes {
+        lsn += record.len() as u64;
+        let padding = lsn.calc_padding(8u64) as usize;
+        if padding == 0 {
+            return record;
+        }
+        [record, Bytes::from(vec![0; padding])].concat().into()
+    }
+
+    /// Generates a record with an arbitrary payload at the current LSN, then increments the LSN.
+    pub fn generate_record(&mut self, data: Bytes, rmid: u8, info: u8) -> Bytes {
+        let record = Self::encode_record(data, rmid, info, self.prev_lsn);
+        let record = Self::encode_pages(record, self.lsn);
+        let record = Self::pad_record(record, self.lsn);
+        self.prev_lsn = self.lsn;
+        self.lsn += record.len() as u64;
+        record
+    }
+
+    /// Generates a logical message at the current LSN. Can be used to construct arbitrary messages.
+    pub fn generate_logical_message(&mut self, prefix: &CStr, message: &[u8]) -> Bytes {
+        let data = Self::encode_logical_message(prefix, message);
+        self.generate_record(data, RM_LOGICALMSG_ID, XLOG_LOGICAL_MESSAGE)
+    }
+}
+
+/// Generate WAL records as an iterator.
+impl Iterator for WalGenerator {
+    type Item = (Lsn, Bytes);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let lsn = self.lsn;
+        let record = self.generate_logical_message(Self::PREFIX, Self::MESSAGE);
+        Some((lsn, record))
+    }
+}

libs/postgres_ffi/src/xlog_utils.rs

@@ -7,15 +7,14 @@
 // have been named the same as the corresponding PostgreSQL functions instead.
 //
 
-use crc32c::crc32c_append;
-
 use super::super::waldecoder::WalStreamDecoder;
 use super::bindings::{
     CheckPoint, ControlFileData, DBState_DB_SHUTDOWNED, FullTransactionId, TimeLineID, TimestampTz,
     XLogLongPageHeaderData, XLogPageHeaderData, XLogRecPtr, XLogRecord, XLogSegNo, XLOG_PAGE_MAGIC,
 };
+use super::wal_generator::WalGenerator;
 use super::PG_MAJORVERSION;
-use crate::pg_constants;
+use crate::pg_constants::{self, RM_LOGICALMSG_ID, XLOG_LOGICAL_MESSAGE};
 use crate::PG_TLI;
 use crate::{uint32, uint64, Oid};
 use crate::{WAL_SEGMENT_SIZE, XLOG_BLCKSZ};
@@ -26,7 +25,7 @@ use bytes::{Buf, Bytes};
 use log::*;
 
 use serde::Serialize;
-use std::ffi::OsStr;
+use std::ffi::{CString, OsStr};
 use std::fs::File;
 use std::io::prelude::*;
 use std::io::ErrorKind;
@@ -39,6 +38,7 @@ use utils::bin_ser::SerializeError;
 use utils::lsn::Lsn;
 
 pub const XLOG_FNAME_LEN: usize = 24;
+pub const XLP_BKP_REMOVABLE: u16 = 0x0004;
 pub const XLP_FIRST_IS_CONTRECORD: u16 = 0x0001;
 pub const XLP_REM_LEN_OFFS: usize = 2 + 2 + 4 + 8;
 pub const XLOG_RECORD_CRC_OFFS: usize = 4 + 4 + 8 + 1 + 1 + 2;
@@ -489,64 +489,16 @@ impl XlLogicalMessage {
 /// Create new WAL record for non-transactional logical message.
 /// Used for creating artificial WAL for tests, as LogicalMessage
 /// record is basically no-op.
-///
-/// NOTE: This leaves the xl_prev field zero. The safekeeper and
-/// pageserver tolerate that, but PostgreSQL does not.
-pub fn encode_logical_message(prefix: &str, message: &str) -> Vec<u8> {
-    let mut prefix_bytes: Vec<u8> = Vec::with_capacity(prefix.len() + 1);
-    prefix_bytes.write_all(prefix.as_bytes()).unwrap();
-    prefix_bytes.push(0);
-
-    let message_bytes = message.as_bytes();
-
-    let logical_message = XlLogicalMessage {
-        db_id: 0,
-        transactional: 0,
-        prefix_size: prefix_bytes.len() as u64,
-        message_size: message_bytes.len() as u64,
-    };
-
-    let mainrdata = logical_message.encode();
-    let mainrdata_len: usize = mainrdata.len() + prefix_bytes.len() + message_bytes.len();
-    // only short mainrdata is supported for now
-    assert!(mainrdata_len <= 255);
-    let mainrdata_len = mainrdata_len as u8;
-
-    let mut data: Vec<u8> = vec![pg_constants::XLR_BLOCK_ID_DATA_SHORT, mainrdata_len];
-    data.extend_from_slice(&mainrdata);
-    data.extend_from_slice(&prefix_bytes);
-    data.extend_from_slice(message_bytes);
-
-    let total_len = XLOG_SIZE_OF_XLOG_RECORD + data.len();
-
-    let mut header = XLogRecord {
-        xl_tot_len: total_len as u32,
-        xl_xid: 0,
-        xl_prev: 0,
-        xl_info: 0,
-        xl_rmid: 21,
-        __bindgen_padding_0: [0u8; 2usize],
-        xl_crc: 0, // crc will be calculated later
-    };
-
-    let header_bytes = header.encode().expect("failed to encode header");
-    let crc = crc32c_append(0, &data);
-    let crc = crc32c_append(crc, &header_bytes[0..XLOG_RECORD_CRC_OFFS]);
-    header.xl_crc = crc;
-
-    let mut wal: Vec<u8> = Vec::new();
-    wal.extend_from_slice(&header.encode().expect("failed to encode header"));
-    wal.extend_from_slice(&data);
-
-    // WAL start position must be aligned at 8 bytes,
-    // this will add padding for the next WAL record.
-    const PADDING: usize = 8;
-    let padding_rem = wal.len() % PADDING;
-    if padding_rem != 0 {
-        wal.resize(wal.len() + PADDING - padding_rem, 0);
-    }
-
-    wal
+pub fn encode_logical_message(prefix: &str, message: &str) -> Bytes {
+    // This function can take untrusted input, so discard any NUL bytes in the prefix string.
+    let prefix = CString::new(prefix.replace('\0', "")).expect("no NULs");
+    let message = message.as_bytes();
+    WalGenerator::encode_record(
+        WalGenerator::encode_logical_message(&prefix, message),
+        RM_LOGICALMSG_ID,
+        XLOG_LOGICAL_MESSAGE,
+        Lsn(0),
+    )
 }
 
 #[cfg(test)]

libs/remote_storage/src/local_fs.rs

@@ -357,22 +357,20 @@ impl RemoteStorage for LocalFs {
                 .list_recursive(prefix)
                 .await
                 .map_err(DownloadError::Other)?;
-            let objects = keys
-                .into_iter()
-                .filter_map(|k| {
-                    let path = k.with_base(&self.storage_root);
-                    if path.is_dir() {
-                        None
-                    } else {
-                        Some(ListingObject {
-                            key: k.clone(),
-                            // LocalFs is just for testing, so just specify a dummy time
-                            last_modified: SystemTime::now(),
-                            size: 0,
-                        })
-                    }
-                })
-                .collect();
+            let mut objects = Vec::with_capacity(keys.len());
+            for key in keys {
+                let path = key.with_base(&self.storage_root);
+                let metadata = file_metadata(&path).await?;
+                if metadata.is_dir() {
+                    continue;
+                }
+                objects.push(ListingObject {
+                    key: key.clone(),
+                    last_modified: metadata.modified()?,
+                    size: metadata.len(),
+                });
+            }
+            let objects = objects;
 
             if let ListingMode::NoDelimiter = mode {
                 result.keys = objects;
@@ -410,9 +408,8 @@ impl RemoteStorage for LocalFs {
                     } else {
                         result.keys.push(ListingObject {
                             key: RemotePath::from_string(&relative_key).unwrap(),
-                            // LocalFs is just for testing
-                            last_modified: SystemTime::now(),
-                            size: 0,
+                            last_modified: object.last_modified,
+                            size: object.size,
                         });
                     }
                 }

libs/utils/src/lsn.rs

@@ -12,7 +12,7 @@ use crate::seqwait::MonotonicCounter;
 pub const XLOG_BLCKSZ: u32 = 8192;
 
 /// A Postgres LSN (Log Sequence Number), also known as an XLogRecPtr
-#[derive(Clone, Copy, Eq, Ord, PartialEq, PartialOrd, Hash)]
+#[derive(Clone, Copy, Default, Eq, Ord, PartialEq, PartialOrd, Hash)]
 pub struct Lsn(pub u64);
 
 impl Serialize for Lsn {

libs/wal_decoder/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "wal_decoder"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[features]
+testing = []
+
+[dependencies]
+anyhow.workspace = true
+bytes.workspace = true
+pageserver_api.workspace = true
+postgres_ffi.workspace = true
+serde.workspace = true
+tracing.workspace = true
+utils.workspace = true
+workspace_hack = { version = "0.1", path = "../../workspace_hack" }

libs/wal_decoder/src/decoder.rs

@@ -0,0 +1,970 @@
+//! This module contains logic for decoding and interpreting
+//! raw bytes which represent a raw Postgres WAL record.
+
+use crate::models::*;
+use bytes::{Buf, Bytes, BytesMut};
+use pageserver_api::key::rel_block_to_key;
+use pageserver_api::record::NeonWalRecord;
+use pageserver_api::reltag::{RelTag, SlruKind};
+use pageserver_api::shard::ShardIdentity;
+use pageserver_api::value::Value;
+use postgres_ffi::relfile_utils::VISIBILITYMAP_FORKNUM;
+use postgres_ffi::walrecord::*;
+use postgres_ffi::{page_is_new, page_set_lsn, pg_constants, BLCKSZ};
+use utils::lsn::Lsn;
+
+impl InterpretedWalRecord {
+    /// Decode and interpreted raw bytes which represent one Postgres WAL record.
+    /// Data blocks which do not match the provided shard identity are filtered out.
+    /// Shard 0 is a special case since it tracks all relation sizes. We only give it
+    /// the keys that are being written as that is enough for updating relation sizes.
+    pub fn from_bytes_filtered(
+        buf: Bytes,
+        shard: &ShardIdentity,
+        lsn: Lsn,
+        pg_version: u32,
+    ) -> anyhow::Result<InterpretedWalRecord> {
+        let mut decoded = DecodedWALRecord::default();
+        decode_wal_record(buf, &mut decoded, pg_version)?;
+
+        let flush_uncommitted = if decoded.is_dbase_create_copy(pg_version) {
+            FlushUncommittedRecords::Yes
+        } else {
+            FlushUncommittedRecords::No
+        };
+
+        let metadata_record = MetadataRecord::from_decoded(&decoded, lsn, pg_version)?;
+
+        let mut blocks = Vec::default();
+        for blk in decoded.blocks.iter() {
+            let rel = RelTag {
+                spcnode: blk.rnode_spcnode,
+                dbnode: blk.rnode_dbnode,
+                relnode: blk.rnode_relnode,
+                forknum: blk.forknum,
+            };
+
+            let key = rel_block_to_key(rel, blk.blkno);
+
+            if !key.is_valid_key_on_write_path() {
+                anyhow::bail!("Unsupported key decoded at LSN {}: {}", lsn, key);
+            }
+
+            let key_is_local = shard.is_key_local(&key);
+
+            tracing::debug!(
+                lsn=%lsn,
+                key=%key,
+                "ingest: shard decision {}",
+                if !key_is_local { "drop" } else { "keep" },
+            );
+
+            if !key_is_local {
+                if shard.is_shard_zero() {
+                    // Shard 0 tracks relation sizes.  Although we will not store this block, we will observe
+                    // its blkno in case it implicitly extends a relation.
+                    blocks.push((key.to_compact(), None));
+                }
+
+                continue;
+            }
+
+            // Instead of storing full-page-image WAL record,
+            // it is better to store extracted image: we can skip wal-redo
+            // in this case. Also some FPI records may contain multiple (up to 32) pages,
+            // so them have to be copied multiple times.
+            //
+            let value = if blk.apply_image
+                && blk.has_image
+                && decoded.xl_rmid == pg_constants::RM_XLOG_ID
+                && (decoded.xl_info == pg_constants::XLOG_FPI
+                || decoded.xl_info == pg_constants::XLOG_FPI_FOR_HINT)
+                // compression of WAL is not yet supported: fall back to storing the original WAL record
+                && !postgres_ffi::bkpimage_is_compressed(blk.bimg_info, pg_version)
+                // do not materialize null pages because them most likely be soon replaced with real data
+                && blk.bimg_len != 0
+            {
+                // Extract page image from FPI record
+                let img_len = blk.bimg_len as usize;
+                let img_offs = blk.bimg_offset as usize;
+                let mut image = BytesMut::with_capacity(BLCKSZ as usize);
+                // TODO(vlad): skip the copy
+                image.extend_from_slice(&decoded.record[img_offs..img_offs + img_len]);
+
+                if blk.hole_length != 0 {
+                    let tail = image.split_off(blk.hole_offset as usize);
+                    image.resize(image.len() + blk.hole_length as usize, 0u8);
+                    image.unsplit(tail);
+                }
+                //
+                // Match the logic of XLogReadBufferForRedoExtended:
+                // The page may be uninitialized. If so, we can't set the LSN because
+                // that would corrupt the page.
+                //
+                if !page_is_new(&image) {
+                    page_set_lsn(&mut image, lsn)
+                }
+                assert_eq!(image.len(), BLCKSZ as usize);
+
+                Value::Image(image.freeze())
+            } else {
+                Value::WalRecord(NeonWalRecord::Postgres {
+                    will_init: blk.will_init || blk.apply_image,
+                    rec: decoded.record.clone(),
+                })
+            };
+
+            blocks.push((key.to_compact(), Some(value)));
+        }
+
+        Ok(InterpretedWalRecord {
+            metadata_record,
+            blocks,
+            lsn,
+            flush_uncommitted,
+            xid: decoded.xl_xid,
+        })
+    }
+}
+
+impl MetadataRecord {
+    fn from_decoded(
+        decoded: &DecodedWALRecord,
+        lsn: Lsn,
+        pg_version: u32,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        // Note: this doesn't actually copy the bytes since
+        // the [`Bytes`] type implements it via a level of indirection.
+        let mut buf = decoded.record.clone();
+        buf.advance(decoded.main_data_offset);
+
+        match decoded.xl_rmid {
+            pg_constants::RM_HEAP_ID | pg_constants::RM_HEAP2_ID => {
+                Self::decode_heapam_record(&mut buf, decoded, pg_version)
+            }
+            pg_constants::RM_NEON_ID => Self::decode_neonmgr_record(&mut buf, decoded, pg_version),
+            // Handle other special record types
+            pg_constants::RM_SMGR_ID => Self::decode_smgr_record(&mut buf, decoded),
+            pg_constants::RM_DBASE_ID => Self::decode_dbase_record(&mut buf, decoded, pg_version),
+            pg_constants::RM_TBLSPC_ID => {
+                tracing::trace!("XLOG_TBLSPC_CREATE/DROP is not handled yet");
+                Ok(None)
+            }
+            pg_constants::RM_CLOG_ID => Self::decode_clog_record(&mut buf, decoded, pg_version),
+            pg_constants::RM_XACT_ID => Self::decode_xact_record(&mut buf, decoded, lsn),
+            pg_constants::RM_MULTIXACT_ID => {
+                Self::decode_multixact_record(&mut buf, decoded, pg_version)
+            }
+            pg_constants::RM_RELMAP_ID => Self::decode_relmap_record(&mut buf, decoded),
+            // This is an odd duck. It needs to go to all shards.
+            // Since it uses the checkpoint image (that's initialized from CHECKPOINT_KEY
+            // in WalIngest::new), we have to send the whole DecodedWalRecord::record to
+            // the pageserver and decode it there.
+            //
+            // Alternatively, one can make the checkpoint part of the subscription protocol
+            // to the pageserver. This should work fine, but can be done at a later point.
+            pg_constants::RM_XLOG_ID => Self::decode_xlog_record(&mut buf, decoded, lsn),
+            pg_constants::RM_LOGICALMSG_ID => {
+                Self::decode_logical_message_record(&mut buf, decoded)
+            }
+            pg_constants::RM_STANDBY_ID => Self::decode_standby_record(&mut buf, decoded),
+            pg_constants::RM_REPLORIGIN_ID => Self::decode_replorigin_record(&mut buf, decoded),
+            _unexpected => {
+                // TODO: consider failing here instead of blindly doing something without
+                // understanding the protocol
+                Ok(None)
+            }
+        }
+    }
+
+    fn decode_heapam_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+        pg_version: u32,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        // Handle VM bit updates that are implicitly part of heap records.
+
+        // First, look at the record to determine which VM bits need
+        // to be cleared. If either of these variables is set, we
+        // need to clear the corresponding bits in the visibility map.
+        let mut new_heap_blkno: Option<u32> = None;
+        let mut old_heap_blkno: Option<u32> = None;
+        let mut flags = pg_constants::VISIBILITYMAP_VALID_BITS;
+
+        match pg_version {
+            14 => {
+                if decoded.xl_rmid == pg_constants::RM_HEAP_ID {
+                    let info = decoded.xl_info & pg_constants::XLOG_HEAP_OPMASK;
+
+                    if info == pg_constants::XLOG_HEAP_INSERT {
+                        let xlrec = v14::XlHeapInsert::decode(buf);
+                        assert_eq!(0, buf.remaining());
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_DELETE {
+                        let xlrec = v14::XlHeapDelete::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_UPDATE
+                        || info == pg_constants::XLOG_HEAP_HOT_UPDATE
+                    {
+                        let xlrec = v14::XlHeapUpdate::decode(buf);
+                        // the size of tuple data is inferred from the size of the record.
+                        // we can't validate the remaining number of bytes without parsing
+                        // the tuple data.
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks.last().unwrap().blkno);
+                        }
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED) != 0 {
+                            // PostgreSQL only uses XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED on a
+                            // non-HOT update where the new tuple goes to different page than
+                            // the old one. Otherwise, only XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED is
+                            // set.
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_LOCK {
+                        let xlrec = v14::XlHeapLock::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_LOCK_ALL_FROZEN_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks[0].blkno);
+                            flags = pg_constants::VISIBILITYMAP_ALL_FROZEN;
+                        }
+                    }
+                } else if decoded.xl_rmid == pg_constants::RM_HEAP2_ID {
+                    let info = decoded.xl_info & pg_constants::XLOG_HEAP_OPMASK;
+                    if info == pg_constants::XLOG_HEAP2_MULTI_INSERT {
+                        let xlrec = v14::XlHeapMultiInsert::decode(buf);
+
+                        let offset_array_len =
+                            if decoded.xl_info & pg_constants::XLOG_HEAP_INIT_PAGE > 0 {
+                                // the offsets array is omitted if XLOG_HEAP_INIT_PAGE is set
+                                0
+                            } else {
+                                size_of::<u16>() * xlrec.ntuples as usize
+                            };
+                        assert_eq!(offset_array_len, buf.remaining());
+
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP2_LOCK_UPDATED {
+                        let xlrec = v14::XlHeapLockUpdated::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_LOCK_ALL_FROZEN_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks[0].blkno);
+                            flags = pg_constants::VISIBILITYMAP_ALL_FROZEN;
+                        }
+                    }
+                } else {
+                    anyhow::bail!("Unknown RMGR {} for Heap decoding", decoded.xl_rmid);
+                }
+            }
+            15 => {
+                if decoded.xl_rmid == pg_constants::RM_HEAP_ID {
+                    let info = decoded.xl_info & pg_constants::XLOG_HEAP_OPMASK;
+
+                    if info == pg_constants::XLOG_HEAP_INSERT {
+                        let xlrec = v15::XlHeapInsert::decode(buf);
+                        assert_eq!(0, buf.remaining());
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_DELETE {
+                        let xlrec = v15::XlHeapDelete::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_UPDATE
+                        || info == pg_constants::XLOG_HEAP_HOT_UPDATE
+                    {
+                        let xlrec = v15::XlHeapUpdate::decode(buf);
+                        // the size of tuple data is inferred from the size of the record.
+                        // we can't validate the remaining number of bytes without parsing
+                        // the tuple data.
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks.last().unwrap().blkno);
+                        }
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED) != 0 {
+                            // PostgreSQL only uses XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED on a
+                            // non-HOT update where the new tuple goes to different page than
+                            // the old one. Otherwise, only XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED is
+                            // set.
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_LOCK {
+                        let xlrec = v15::XlHeapLock::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_LOCK_ALL_FROZEN_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks[0].blkno);
+                            flags = pg_constants::VISIBILITYMAP_ALL_FROZEN;
+                        }
+                    }
+                } else if decoded.xl_rmid == pg_constants::RM_HEAP2_ID {
+                    let info = decoded.xl_info & pg_constants::XLOG_HEAP_OPMASK;
+                    if info == pg_constants::XLOG_HEAP2_MULTI_INSERT {
+                        let xlrec = v15::XlHeapMultiInsert::decode(buf);
+
+                        let offset_array_len =
+                            if decoded.xl_info & pg_constants::XLOG_HEAP_INIT_PAGE > 0 {
+                                // the offsets array is omitted if XLOG_HEAP_INIT_PAGE is set
+                                0
+                            } else {
+                                size_of::<u16>() * xlrec.ntuples as usize
+                            };
+                        assert_eq!(offset_array_len, buf.remaining());
+
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP2_LOCK_UPDATED {
+                        let xlrec = v15::XlHeapLockUpdated::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_LOCK_ALL_FROZEN_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks[0].blkno);
+                            flags = pg_constants::VISIBILITYMAP_ALL_FROZEN;
+                        }
+                    }
+                } else {
+                    anyhow::bail!("Unknown RMGR {} for Heap decoding", decoded.xl_rmid);
+                }
+            }
+            16 => {
+                if decoded.xl_rmid == pg_constants::RM_HEAP_ID {
+                    let info = decoded.xl_info & pg_constants::XLOG_HEAP_OPMASK;
+
+                    if info == pg_constants::XLOG_HEAP_INSERT {
+                        let xlrec = v16::XlHeapInsert::decode(buf);
+                        assert_eq!(0, buf.remaining());
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_DELETE {
+                        let xlrec = v16::XlHeapDelete::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_UPDATE
+                        || info == pg_constants::XLOG_HEAP_HOT_UPDATE
+                    {
+                        let xlrec = v16::XlHeapUpdate::decode(buf);
+                        // the size of tuple data is inferred from the size of the record.
+                        // we can't validate the remaining number of bytes without parsing
+                        // the tuple data.
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks.last().unwrap().blkno);
+                        }
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED) != 0 {
+                            // PostgreSQL only uses XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED on a
+                            // non-HOT update where the new tuple goes to different page than
+                            // the old one. Otherwise, only XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED is
+                            // set.
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_LOCK {
+                        let xlrec = v16::XlHeapLock::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_LOCK_ALL_FROZEN_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks[0].blkno);
+                            flags = pg_constants::VISIBILITYMAP_ALL_FROZEN;
+                        }
+                    }
+                } else if decoded.xl_rmid == pg_constants::RM_HEAP2_ID {
+                    let info = decoded.xl_info & pg_constants::XLOG_HEAP_OPMASK;
+                    if info == pg_constants::XLOG_HEAP2_MULTI_INSERT {
+                        let xlrec = v16::XlHeapMultiInsert::decode(buf);
+
+                        let offset_array_len =
+                            if decoded.xl_info & pg_constants::XLOG_HEAP_INIT_PAGE > 0 {
+                                // the offsets array is omitted if XLOG_HEAP_INIT_PAGE is set
+                                0
+                            } else {
+                                size_of::<u16>() * xlrec.ntuples as usize
+                            };
+                        assert_eq!(offset_array_len, buf.remaining());
+
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP2_LOCK_UPDATED {
+                        let xlrec = v16::XlHeapLockUpdated::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_LOCK_ALL_FROZEN_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks[0].blkno);
+                            flags = pg_constants::VISIBILITYMAP_ALL_FROZEN;
+                        }
+                    }
+                } else {
+                    anyhow::bail!("Unknown RMGR {} for Heap decoding", decoded.xl_rmid);
+                }
+            }
+            17 => {
+                if decoded.xl_rmid == pg_constants::RM_HEAP_ID {
+                    let info = decoded.xl_info & pg_constants::XLOG_HEAP_OPMASK;
+
+                    if info == pg_constants::XLOG_HEAP_INSERT {
+                        let xlrec = v17::XlHeapInsert::decode(buf);
+                        assert_eq!(0, buf.remaining());
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_DELETE {
+                        let xlrec = v17::XlHeapDelete::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_UPDATE
+                        || info == pg_constants::XLOG_HEAP_HOT_UPDATE
+                    {
+                        let xlrec = v17::XlHeapUpdate::decode(buf);
+                        // the size of tuple data is inferred from the size of the record.
+                        // we can't validate the remaining number of bytes without parsing
+                        // the tuple data.
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks.last().unwrap().blkno);
+                        }
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED) != 0 {
+                            // PostgreSQL only uses XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED on a
+                            // non-HOT update where the new tuple goes to different page than
+                            // the old one. Otherwise, only XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED is
+                            // set.
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP_LOCK {
+                        let xlrec = v17::XlHeapLock::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_LOCK_ALL_FROZEN_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks[0].blkno);
+                            flags = pg_constants::VISIBILITYMAP_ALL_FROZEN;
+                        }
+                    }
+                } else if decoded.xl_rmid == pg_constants::RM_HEAP2_ID {
+                    let info = decoded.xl_info & pg_constants::XLOG_HEAP_OPMASK;
+                    if info == pg_constants::XLOG_HEAP2_MULTI_INSERT {
+                        let xlrec = v17::XlHeapMultiInsert::decode(buf);
+
+                        let offset_array_len =
+                            if decoded.xl_info & pg_constants::XLOG_HEAP_INIT_PAGE > 0 {
+                                // the offsets array is omitted if XLOG_HEAP_INIT_PAGE is set
+                                0
+                            } else {
+                                size_of::<u16>() * xlrec.ntuples as usize
+                            };
+                        assert_eq!(offset_array_len, buf.remaining());
+
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    } else if info == pg_constants::XLOG_HEAP2_LOCK_UPDATED {
+                        let xlrec = v17::XlHeapLockUpdated::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_LOCK_ALL_FROZEN_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks[0].blkno);
+                            flags = pg_constants::VISIBILITYMAP_ALL_FROZEN;
+                        }
+                    }
+                } else {
+                    anyhow::bail!("Unknown RMGR {} for Heap decoding", decoded.xl_rmid);
+                }
+            }
+            _ => {}
+        }
+
+        if new_heap_blkno.is_some() || old_heap_blkno.is_some() {
+            let vm_rel = RelTag {
+                forknum: VISIBILITYMAP_FORKNUM,
+                spcnode: decoded.blocks[0].rnode_spcnode,
+                dbnode: decoded.blocks[0].rnode_dbnode,
+                relnode: decoded.blocks[0].rnode_relnode,
+            };
+
+            Ok(Some(MetadataRecord::Heapam(HeapamRecord::ClearVmBits(
+                ClearVmBits {
+                    new_heap_blkno,
+                    old_heap_blkno,
+                    vm_rel,
+                    flags,
+                },
+            ))))
+        } else {
+            Ok(None)
+        }
+    }
+
+    fn decode_neonmgr_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+        pg_version: u32,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        // Handle VM bit updates that are implicitly part of heap records.
+
+        // First, look at the record to determine which VM bits need
+        // to be cleared. If either of these variables is set, we
+        // need to clear the corresponding bits in the visibility map.
+        let mut new_heap_blkno: Option<u32> = None;
+        let mut old_heap_blkno: Option<u32> = None;
+        let mut flags = pg_constants::VISIBILITYMAP_VALID_BITS;
+
+        assert_eq!(decoded.xl_rmid, pg_constants::RM_NEON_ID);
+
+        match pg_version {
+            16 | 17 => {
+                let info = decoded.xl_info & pg_constants::XLOG_HEAP_OPMASK;
+
+                match info {
+                    pg_constants::XLOG_NEON_HEAP_INSERT => {
+                        let xlrec = v17::rm_neon::XlNeonHeapInsert::decode(buf);
+                        assert_eq!(0, buf.remaining());
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    }
+                    pg_constants::XLOG_NEON_HEAP_DELETE => {
+                        let xlrec = v17::rm_neon::XlNeonHeapDelete::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    }
+                    pg_constants::XLOG_NEON_HEAP_UPDATE
+                    | pg_constants::XLOG_NEON_HEAP_HOT_UPDATE => {
+                        let xlrec = v17::rm_neon::XlNeonHeapUpdate::decode(buf);
+                        // the size of tuple data is inferred from the size of the record.
+                        // we can't validate the remaining number of bytes without parsing
+                        // the tuple data.
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks.last().unwrap().blkno);
+                        }
+                        if (xlrec.flags & pg_constants::XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED) != 0 {
+                            // PostgreSQL only uses XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED on a
+                            // non-HOT update where the new tuple goes to different page than
+                            // the old one. Otherwise, only XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED is
+                            // set.
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    }
+                    pg_constants::XLOG_NEON_HEAP_MULTI_INSERT => {
+                        let xlrec = v17::rm_neon::XlNeonHeapMultiInsert::decode(buf);
+
+                        let offset_array_len =
+                            if decoded.xl_info & pg_constants::XLOG_HEAP_INIT_PAGE > 0 {
+                                // the offsets array is omitted if XLOG_HEAP_INIT_PAGE is set
+                                0
+                            } else {
+                                size_of::<u16>() * xlrec.ntuples as usize
+                            };
+                        assert_eq!(offset_array_len, buf.remaining());
+
+                        if (xlrec.flags & pg_constants::XLH_INSERT_ALL_VISIBLE_CLEARED) != 0 {
+                            new_heap_blkno = Some(decoded.blocks[0].blkno);
+                        }
+                    }
+                    pg_constants::XLOG_NEON_HEAP_LOCK => {
+                        let xlrec = v17::rm_neon::XlNeonHeapLock::decode(buf);
+                        if (xlrec.flags & pg_constants::XLH_LOCK_ALL_FROZEN_CLEARED) != 0 {
+                            old_heap_blkno = Some(decoded.blocks[0].blkno);
+                            flags = pg_constants::VISIBILITYMAP_ALL_FROZEN;
+                        }
+                    }
+                    info => anyhow::bail!("Unknown WAL record type for Neon RMGR: {}", info),
+                }
+            }
+            _ => anyhow::bail!(
+                "Neon RMGR has no known compatibility with PostgreSQL version {}",
+                pg_version
+            ),
+        }
+
+        if new_heap_blkno.is_some() || old_heap_blkno.is_some() {
+            let vm_rel = RelTag {
+                forknum: VISIBILITYMAP_FORKNUM,
+                spcnode: decoded.blocks[0].rnode_spcnode,
+                dbnode: decoded.blocks[0].rnode_dbnode,
+                relnode: decoded.blocks[0].rnode_relnode,
+            };
+
+            Ok(Some(MetadataRecord::Neonrmgr(NeonrmgrRecord::ClearVmBits(
+                ClearVmBits {
+                    new_heap_blkno,
+                    old_heap_blkno,
+                    vm_rel,
+                    flags,
+                },
+            ))))
+        } else {
+            Ok(None)
+        }
+    }
+
+    fn decode_smgr_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+        if info == pg_constants::XLOG_SMGR_CREATE {
+            let create = XlSmgrCreate::decode(buf);
+            let rel = RelTag {
+                spcnode: create.rnode.spcnode,
+                dbnode: create.rnode.dbnode,
+                relnode: create.rnode.relnode,
+                forknum: create.forknum,
+            };
+
+            return Ok(Some(MetadataRecord::Smgr(SmgrRecord::Create(SmgrCreate {
+                rel,
+            }))));
+        } else if info == pg_constants::XLOG_SMGR_TRUNCATE {
+            let truncate = XlSmgrTruncate::decode(buf);
+            return Ok(Some(MetadataRecord::Smgr(SmgrRecord::Truncate(truncate))));
+        }
+
+        Ok(None)
+    }
+
+    fn decode_dbase_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+        pg_version: u32,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        // TODO: Refactor this to avoid the duplication between postgres versions.
+
+        let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+        tracing::debug!(%info, %pg_version, "handle RM_DBASE_ID");
+
+        if pg_version == 14 {
+            if info == postgres_ffi::v14::bindings::XLOG_DBASE_CREATE {
+                let createdb = XlCreateDatabase::decode(buf);
+                tracing::debug!("XLOG_DBASE_CREATE v14");
+
+                let record = MetadataRecord::Dbase(DbaseRecord::Create(DbaseCreate {
+                    db_id: createdb.db_id,
+                    tablespace_id: createdb.tablespace_id,
+                    src_db_id: createdb.src_db_id,
+                    src_tablespace_id: createdb.src_tablespace_id,
+                }));
+
+                return Ok(Some(record));
+            } else if info == postgres_ffi::v14::bindings::XLOG_DBASE_DROP {
+                let dropdb = XlDropDatabase::decode(buf);
+
+                let record = MetadataRecord::Dbase(DbaseRecord::Drop(DbaseDrop {
+                    db_id: dropdb.db_id,
+                    tablespace_ids: dropdb.tablespace_ids,
+                }));
+
+                return Ok(Some(record));
+            }
+        } else if pg_version == 15 {
+            if info == postgres_ffi::v15::bindings::XLOG_DBASE_CREATE_WAL_LOG {
+                tracing::debug!("XLOG_DBASE_CREATE_WAL_LOG: noop");
+            } else if info == postgres_ffi::v15::bindings::XLOG_DBASE_CREATE_FILE_COPY {
+                // The XLOG record was renamed between v14 and v15,
+                // but the record format is the same.
+                // So we can reuse XlCreateDatabase here.
+                tracing::debug!("XLOG_DBASE_CREATE_FILE_COPY");
+
+                let createdb = XlCreateDatabase::decode(buf);
+                let record = MetadataRecord::Dbase(DbaseRecord::Create(DbaseCreate {
+                    db_id: createdb.db_id,
+                    tablespace_id: createdb.tablespace_id,
+                    src_db_id: createdb.src_db_id,
+                    src_tablespace_id: createdb.src_tablespace_id,
+                }));
+
+                return Ok(Some(record));
+            } else if info == postgres_ffi::v15::bindings::XLOG_DBASE_DROP {
+                let dropdb = XlDropDatabase::decode(buf);
+                let record = MetadataRecord::Dbase(DbaseRecord::Drop(DbaseDrop {
+                    db_id: dropdb.db_id,
+                    tablespace_ids: dropdb.tablespace_ids,
+                }));
+
+                return Ok(Some(record));
+            }
+        } else if pg_version == 16 {
+            if info == postgres_ffi::v16::bindings::XLOG_DBASE_CREATE_WAL_LOG {
+                tracing::debug!("XLOG_DBASE_CREATE_WAL_LOG: noop");
+            } else if info == postgres_ffi::v16::bindings::XLOG_DBASE_CREATE_FILE_COPY {
+                // The XLOG record was renamed between v14 and v15,
+                // but the record format is the same.
+                // So we can reuse XlCreateDatabase here.
+                tracing::debug!("XLOG_DBASE_CREATE_FILE_COPY");
+
+                let createdb = XlCreateDatabase::decode(buf);
+                let record = MetadataRecord::Dbase(DbaseRecord::Create(DbaseCreate {
+                    db_id: createdb.db_id,
+                    tablespace_id: createdb.tablespace_id,
+                    src_db_id: createdb.src_db_id,
+                    src_tablespace_id: createdb.src_tablespace_id,
+                }));
+
+                return Ok(Some(record));
+            } else if info == postgres_ffi::v16::bindings::XLOG_DBASE_DROP {
+                let dropdb = XlDropDatabase::decode(buf);
+                let record = MetadataRecord::Dbase(DbaseRecord::Drop(DbaseDrop {
+                    db_id: dropdb.db_id,
+                    tablespace_ids: dropdb.tablespace_ids,
+                }));
+
+                return Ok(Some(record));
+            }
+        } else if pg_version == 17 {
+            if info == postgres_ffi::v17::bindings::XLOG_DBASE_CREATE_WAL_LOG {
+                tracing::debug!("XLOG_DBASE_CREATE_WAL_LOG: noop");
+            } else if info == postgres_ffi::v17::bindings::XLOG_DBASE_CREATE_FILE_COPY {
+                // The XLOG record was renamed between v14 and v15,
+                // but the record format is the same.
+                // So we can reuse XlCreateDatabase here.
+                tracing::debug!("XLOG_DBASE_CREATE_FILE_COPY");
+
+                let createdb = XlCreateDatabase::decode(buf);
+                let record = MetadataRecord::Dbase(DbaseRecord::Create(DbaseCreate {
+                    db_id: createdb.db_id,
+                    tablespace_id: createdb.tablespace_id,
+                    src_db_id: createdb.src_db_id,
+                    src_tablespace_id: createdb.src_tablespace_id,
+                }));
+
+                return Ok(Some(record));
+            } else if info == postgres_ffi::v17::bindings::XLOG_DBASE_DROP {
+                let dropdb = XlDropDatabase::decode(buf);
+                let record = MetadataRecord::Dbase(DbaseRecord::Drop(DbaseDrop {
+                    db_id: dropdb.db_id,
+                    tablespace_ids: dropdb.tablespace_ids,
+                }));
+
+                return Ok(Some(record));
+            }
+        }
+
+        Ok(None)
+    }
+
+    fn decode_clog_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+        pg_version: u32,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        let info = decoded.xl_info & !pg_constants::XLR_INFO_MASK;
+
+        if info == pg_constants::CLOG_ZEROPAGE {
+            let pageno = if pg_version < 17 {
+                buf.get_u32_le()
+            } else {
+                buf.get_u64_le() as u32
+            };
+            let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+            let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+
+            Ok(Some(MetadataRecord::Clog(ClogRecord::ZeroPage(
+                ClogZeroPage { segno, rpageno },
+            ))))
+        } else {
+            assert!(info == pg_constants::CLOG_TRUNCATE);
+            let xlrec = XlClogTruncate::decode(buf, pg_version);
+
+            Ok(Some(MetadataRecord::Clog(ClogRecord::Truncate(
+                ClogTruncate {
+                    pageno: xlrec.pageno,
+                    oldest_xid: xlrec.oldest_xid,
+                    oldest_xid_db: xlrec.oldest_xid_db,
+                },
+            ))))
+        }
+    }
+
+    fn decode_xact_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+        lsn: Lsn,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        let info = decoded.xl_info & pg_constants::XLOG_XACT_OPMASK;
+        let origin_id = decoded.origin_id;
+        let xl_xid = decoded.xl_xid;
+
+        if info == pg_constants::XLOG_XACT_COMMIT {
+            let parsed = XlXactParsedRecord::decode(buf, decoded.xl_xid, decoded.xl_info);
+            return Ok(Some(MetadataRecord::Xact(XactRecord::Commit(XactCommon {
+                parsed,
+                origin_id,
+                xl_xid,
+                lsn,
+            }))));
+        } else if info == pg_constants::XLOG_XACT_ABORT {
+            let parsed = XlXactParsedRecord::decode(buf, decoded.xl_xid, decoded.xl_info);
+            return Ok(Some(MetadataRecord::Xact(XactRecord::Abort(XactCommon {
+                parsed,
+                origin_id,
+                xl_xid,
+                lsn,
+            }))));
+        } else if info == pg_constants::XLOG_XACT_COMMIT_PREPARED {
+            let parsed = XlXactParsedRecord::decode(buf, decoded.xl_xid, decoded.xl_info);
+            return Ok(Some(MetadataRecord::Xact(XactRecord::CommitPrepared(
+                XactCommon {
+                    parsed,
+                    origin_id,
+                    xl_xid,
+                    lsn,
+                },
+            ))));
+        } else if info == pg_constants::XLOG_XACT_ABORT_PREPARED {
+            let parsed = XlXactParsedRecord::decode(buf, decoded.xl_xid, decoded.xl_info);
+            return Ok(Some(MetadataRecord::Xact(XactRecord::AbortPrepared(
+                XactCommon {
+                    parsed,
+                    origin_id,
+                    xl_xid,
+                    lsn,
+                },
+            ))));
+        } else if info == pg_constants::XLOG_XACT_PREPARE {
+            return Ok(Some(MetadataRecord::Xact(XactRecord::Prepare(
+                XactPrepare {
+                    xl_xid: decoded.xl_xid,
+                    data: Bytes::copy_from_slice(&buf[..]),
+                },
+            ))));
+        }
+
+        Ok(None)
+    }
+
+    fn decode_multixact_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+        pg_version: u32,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+
+        if info == pg_constants::XLOG_MULTIXACT_ZERO_OFF_PAGE
+            || info == pg_constants::XLOG_MULTIXACT_ZERO_MEM_PAGE
+        {
+            let pageno = if pg_version < 17 {
+                buf.get_u32_le()
+            } else {
+                buf.get_u64_le() as u32
+            };
+            let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+            let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+
+            let slru_kind = match info {
+                pg_constants::XLOG_MULTIXACT_ZERO_OFF_PAGE => SlruKind::MultiXactOffsets,
+                pg_constants::XLOG_MULTIXACT_ZERO_MEM_PAGE => SlruKind::MultiXactMembers,
+                _ => unreachable!(),
+            };
+
+            return Ok(Some(MetadataRecord::MultiXact(MultiXactRecord::ZeroPage(
+                MultiXactZeroPage {
+                    slru_kind,
+                    segno,
+                    rpageno,
+                },
+            ))));
+        } else if info == pg_constants::XLOG_MULTIXACT_CREATE_ID {
+            let xlrec = XlMultiXactCreate::decode(buf);
+            return Ok(Some(MetadataRecord::MultiXact(MultiXactRecord::Create(
+                xlrec,
+            ))));
+        } else if info == pg_constants::XLOG_MULTIXACT_TRUNCATE_ID {
+            let xlrec = XlMultiXactTruncate::decode(buf);
+            return Ok(Some(MetadataRecord::MultiXact(MultiXactRecord::Truncate(
+                xlrec,
+            ))));
+        }
+
+        Ok(None)
+    }
+
+    fn decode_relmap_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        let update = XlRelmapUpdate::decode(buf);
+
+        let mut buf = decoded.record.clone();
+        buf.advance(decoded.main_data_offset);
+        // skip xl_relmap_update
+        buf.advance(12);
+
+        Ok(Some(MetadataRecord::Relmap(RelmapRecord::Update(
+            RelmapUpdate {
+                update,
+                buf: Bytes::copy_from_slice(&buf[..]),
+            },
+        ))))
+    }
+
+    fn decode_xlog_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+        lsn: Lsn,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+        Ok(Some(MetadataRecord::Xlog(XlogRecord::Raw(RawXlogRecord {
+            info,
+            lsn,
+            buf: buf.clone(),
+        }))))
+    }
+
+    fn decode_logical_message_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+        if info == pg_constants::XLOG_LOGICAL_MESSAGE {
+            let xlrec = XlLogicalMessage::decode(buf);
+            let prefix = std::str::from_utf8(&buf[0..xlrec.prefix_size - 1])?;
+
+            #[cfg(feature = "testing")]
+            if prefix == "neon-test" {
+                return Ok(Some(MetadataRecord::LogicalMessage(
+                    LogicalMessageRecord::Failpoint,
+                )));
+            }
+
+            if let Some(path) = prefix.strip_prefix("neon-file:") {
+                let buf_size = xlrec.prefix_size + xlrec.message_size;
+                let buf = Bytes::copy_from_slice(&buf[xlrec.prefix_size..buf_size]);
+                return Ok(Some(MetadataRecord::LogicalMessage(
+                    LogicalMessageRecord::Put(PutLogicalMessage {
+                        path: path.to_string(),
+                        buf,
+                    }),
+                )));
+            }
+        }
+
+        Ok(None)
+    }
+
+    fn decode_standby_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+        if info == pg_constants::XLOG_RUNNING_XACTS {
+            let xlrec = XlRunningXacts::decode(buf);
+            return Ok(Some(MetadataRecord::Standby(StandbyRecord::RunningXacts(
+                StandbyRunningXacts {
+                    oldest_running_xid: xlrec.oldest_running_xid,
+                },
+            ))));
+        }
+
+        Ok(None)
+    }
+
+    fn decode_replorigin_record(
+        buf: &mut Bytes,
+        decoded: &DecodedWALRecord,
+    ) -> anyhow::Result<Option<MetadataRecord>> {
+        let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+        if info == pg_constants::XLOG_REPLORIGIN_SET {
+            let xlrec = XlReploriginSet::decode(buf);
+            return Ok(Some(MetadataRecord::Replorigin(ReploriginRecord::Set(
+                xlrec,
+            ))));
+        } else if info == pg_constants::XLOG_REPLORIGIN_DROP {
+            let xlrec = XlReploriginDrop::decode(buf);
+            return Ok(Some(MetadataRecord::Replorigin(ReploriginRecord::Drop(
+                xlrec,
+            ))));
+        }
+
+        Ok(None)
+    }
+}

libs/wal_decoder/src/lib.rs

@@ -0,0 +1,2 @@
+pub mod decoder;
+pub mod models;

libs/wal_decoder/src/models.rs

@@ -0,0 +1,211 @@
+//! This module houses types which represent decoded PG WAL records
+//! ready for the pageserver to interpret. They are derived from the original
+//! WAL records, so that each struct corresponds closely to one WAL record of
+//! a specific kind. They contain the same information as the original WAL records,
+//! just decoded into structs and fields for easier access.
+//!
+//! The ingestion code uses these structs to help with parsing the WAL records,
+//! and it splits them into a stream of modifications to the key-value pairs that
+//! are ultimately stored in delta layers.  See also the split-out counterparts in
+//! [`postgres_ffi::walrecord`].
+//!
+//! The pipeline which processes WAL records is not super obvious, so let's follow
+//! the flow of an example XACT_COMMIT Postgres record:
+//!
+//! (Postgres XACT_COMMIT record)
+//! |
+//! |--> pageserver::walingest::WalIngest::decode_xact_record
+//!      |
+//!      |--> ([`XactRecord::Commit`])
+//!           |
+//!           |--> pageserver::walingest::WalIngest::ingest_xact_record
+//!                |
+//!                |--> (NeonWalRecord::ClogSetCommitted)
+//!                     |
+//!                     |--> write to KV store within the pageserver
+
+use bytes::Bytes;
+use pageserver_api::key::CompactKey;
+use pageserver_api::reltag::{RelTag, SlruKind};
+use pageserver_api::value::Value;
+use postgres_ffi::walrecord::{
+    XlMultiXactCreate, XlMultiXactTruncate, XlRelmapUpdate, XlReploriginDrop, XlReploriginSet,
+    XlSmgrTruncate, XlXactParsedRecord,
+};
+use postgres_ffi::{Oid, TransactionId};
+use utils::lsn::Lsn;
+
+pub enum FlushUncommittedRecords {
+    Yes,
+    No,
+}
+
+/// An interpreted Postgres WAL record, ready to be handled by the pageserver
+pub struct InterpretedWalRecord {
+    /// Optional metadata record - may cause writes to metadata keys
+    /// in the storage engine
+    pub metadata_record: Option<MetadataRecord>,
+    /// Images or deltas for blocks modified in the original WAL record.
+    /// The [`Value`] is optional to avoid sending superfluous data to
+    /// shard 0 for relation size tracking.
+    pub blocks: Vec<(CompactKey, Option<Value>)>,
+    /// Byte offset within WAL for the end of the original PG WAL record
+    pub lsn: Lsn,
+    /// Whether to flush all uncommitted modifications to the storage engine
+    /// before ingesting this record. This is currently only used for legacy PG
+    /// database creations which read pages from a template database. Such WAL
+    /// records require reading data blocks while ingesting, hence the need to flush.
+    pub flush_uncommitted: FlushUncommittedRecords,
+    /// Transaction id of the original PG WAL record
+    pub xid: TransactionId,
+}
+
+/// The interpreted part of the Postgres WAL record which requires metadata
+/// writes to the underlying storage engine.
+pub enum MetadataRecord {
+    Heapam(HeapamRecord),
+    Neonrmgr(NeonrmgrRecord),
+    Smgr(SmgrRecord),
+    Dbase(DbaseRecord),
+    Clog(ClogRecord),
+    Xact(XactRecord),
+    MultiXact(MultiXactRecord),
+    Relmap(RelmapRecord),
+    Xlog(XlogRecord),
+    LogicalMessage(LogicalMessageRecord),
+    Standby(StandbyRecord),
+    Replorigin(ReploriginRecord),
+}
+
+pub enum HeapamRecord {
+    ClearVmBits(ClearVmBits),
+}
+
+pub struct ClearVmBits {
+    pub new_heap_blkno: Option<u32>,
+    pub old_heap_blkno: Option<u32>,
+    pub vm_rel: RelTag,
+    pub flags: u8,
+}
+
+pub enum NeonrmgrRecord {
+    ClearVmBits(ClearVmBits),
+}
+
+pub enum SmgrRecord {
+    Create(SmgrCreate),
+    Truncate(XlSmgrTruncate),
+}
+
+pub struct SmgrCreate {
+    pub rel: RelTag,
+}
+
+pub enum DbaseRecord {
+    Create(DbaseCreate),
+    Drop(DbaseDrop),
+}
+
+pub struct DbaseCreate {
+    pub db_id: Oid,
+    pub tablespace_id: Oid,
+    pub src_db_id: Oid,
+    pub src_tablespace_id: Oid,
+}
+
+pub struct DbaseDrop {
+    pub db_id: Oid,
+    pub tablespace_ids: Vec<Oid>,
+}
+
+pub enum ClogRecord {
+    ZeroPage(ClogZeroPage),
+    Truncate(ClogTruncate),
+}
+
+pub struct ClogZeroPage {
+    pub segno: u32,
+    pub rpageno: u32,
+}
+
+pub struct ClogTruncate {
+    pub pageno: u32,
+    pub oldest_xid: TransactionId,
+    pub oldest_xid_db: Oid,
+}
+
+pub enum XactRecord {
+    Commit(XactCommon),
+    Abort(XactCommon),
+    CommitPrepared(XactCommon),
+    AbortPrepared(XactCommon),
+    Prepare(XactPrepare),
+}
+
+pub struct XactCommon {
+    pub parsed: XlXactParsedRecord,
+    pub origin_id: u16,
+    // Fields below are only used for logging
+    pub xl_xid: TransactionId,
+    pub lsn: Lsn,
+}
+
+pub struct XactPrepare {
+    pub xl_xid: TransactionId,
+    pub data: Bytes,
+}
+
+pub enum MultiXactRecord {
+    ZeroPage(MultiXactZeroPage),
+    Create(XlMultiXactCreate),
+    Truncate(XlMultiXactTruncate),
+}
+
+pub struct MultiXactZeroPage {
+    pub slru_kind: SlruKind,
+    pub segno: u32,
+    pub rpageno: u32,
+}
+
+pub enum RelmapRecord {
+    Update(RelmapUpdate),
+}
+
+pub struct RelmapUpdate {
+    pub update: XlRelmapUpdate,
+    pub buf: Bytes,
+}
+
+pub enum XlogRecord {
+    Raw(RawXlogRecord),
+}
+
+pub struct RawXlogRecord {
+    pub info: u8,
+    pub lsn: Lsn,
+    pub buf: Bytes,
+}
+
+pub enum LogicalMessageRecord {
+    Put(PutLogicalMessage),
+    #[cfg(feature = "testing")]
+    Failpoint,
+}
+
+pub struct PutLogicalMessage {
+    pub path: String,
+    pub buf: Bytes,
+}
+
+pub enum StandbyRecord {
+    RunningXacts(StandbyRunningXacts),
+}
+
+pub struct StandbyRunningXacts {
+    pub oldest_running_xid: TransactionId,
+}
+
+pub enum ReploriginRecord {
+    Set(XlReploriginSet),
+    Drop(XlReploriginDrop),
+}

pageserver/Cargo.toml

@@ -8,7 +8,7 @@ license.workspace = true
 default = []
 # Enables test-only APIs, incuding failpoints. In particular, enables the `fail_point!` macro,
 # which adds some runtime cost to run tests on outage conditions
-testing = ["fail/failpoints", "pageserver_api/testing" ]
+testing = ["fail/failpoints", "pageserver_api/testing", "wal_decoder/testing"]
 
 [dependencies]
 anyhow.workspace = true
@@ -83,6 +83,7 @@ enum-map.workspace = true
 enumset = { workspace = true, features = ["serde"]}
 strum.workspace = true
 strum_macros.workspace = true
+wal_decoder.workspace = true
 
 [target.'cfg(target_os = "linux")'.dependencies]
 procfs.workspace = true

pageserver/benches/bench_ingest.rs

@@ -8,13 +8,12 @@ use pageserver::{
     context::{DownloadBehavior, RequestContext},
     l0_flush::{L0FlushConfig, L0FlushGlobalState},
     page_cache,
-    repository::Value,
     task_mgr::TaskKind,
     tenant::storage_layer::inmemory_layer::SerializedBatch,
     tenant::storage_layer::InMemoryLayer,
     virtual_file,
 };
-use pageserver_api::{key::Key, shard::TenantShardId};
+use pageserver_api::{key::Key, shard::TenantShardId, value::Value};
 use utils::{
     bin_ser::BeSer,
     id::{TenantId, TimelineId},

pageserver/benches/bench_layer_map.rs

@@ -1,9 +1,9 @@
 use criterion::measurement::WallTime;
 use pageserver::keyspace::{KeyPartitioning, KeySpace};
-use pageserver::repository::Key;
 use pageserver::tenant::layer_map::LayerMap;
 use pageserver::tenant::storage_layer::LayerName;
 use pageserver::tenant::storage_layer::PersistentLayerDesc;
+use pageserver_api::key::Key;
 use pageserver_api::shard::TenantShardId;
 use rand::prelude::{SeedableRng, SliceRandom, StdRng};
 use std::cmp::{max, min};

pageserver/benches/bench_walredo.rs

@@ -60,7 +60,8 @@ use anyhow::Context;
 use bytes::{Buf, Bytes};
 use criterion::{BenchmarkId, Criterion};
 use once_cell::sync::Lazy;
-use pageserver::{config::PageServerConf, walrecord::NeonWalRecord, walredo::PostgresRedoManager};
+use pageserver::{config::PageServerConf, walredo::PostgresRedoManager};
+use pageserver_api::record::NeonWalRecord;
 use pageserver_api::{key::Key, shard::TenantShardId};
 use std::{
     future::Future,

pageserver/ctl/src/draw_timeline_dir.rs

@@ -51,7 +51,7 @@
 //!
 
 use anyhow::{Context, Result};
-use pageserver::repository::Key;
+use pageserver_api::key::Key;
 use std::cmp::Ordering;
 use std::io::{self, BufRead};
 use std::path::PathBuf;

pageserver/ctl/src/layer_map_analyzer.rs

@@ -2,7 +2,7 @@
 //!
 //! Currently it only analyzes holes, which are regions within the layer range that the layer contains no updates for. In the future it might do more analysis (maybe key quantiles?) but it should never return sensitive data.
 
-use anyhow::Result;
+use anyhow::{anyhow, Result};
 use camino::{Utf8Path, Utf8PathBuf};
 use pageserver::context::{DownloadBehavior, RequestContext};
 use pageserver::task_mgr::TaskKind;
@@ -11,15 +11,16 @@ use pageserver::virtual_file::api::IoMode;
 use std::cmp::Ordering;
 use std::collections::BinaryHeap;
 use std::ops::Range;
+use std::str::FromStr;
 use std::{fs, str};
 
 use pageserver::page_cache::{self, PAGE_SZ};
-use pageserver::repository::{Key, KEY_SIZE};
 use pageserver::tenant::block_io::FileBlockReader;
 use pageserver::tenant::disk_btree::{DiskBtreeReader, VisitDirection};
 use pageserver::tenant::storage_layer::delta_layer::{Summary, DELTA_KEY_SIZE};
-use pageserver::tenant::storage_layer::range_overlaps;
+use pageserver::tenant::storage_layer::{range_overlaps, LayerName};
 use pageserver::virtual_file::{self, VirtualFile};
+use pageserver_api::key::{Key, KEY_SIZE};
 
 use utils::{bin_ser::BeSer, lsn::Lsn};
 
@@ -74,35 +75,15 @@ impl LayerFile {
     }
 }
 
-pub(crate) fn parse_filename(name: &str) -> Option<LayerFile> {
-    let split: Vec<&str> = name.split("__").collect();
-    if split.len() != 2 {
-        return None;
-    }
-    let keys: Vec<&str> = split[0].split('-').collect();
-    let lsn_and_opt_generation: Vec<&str> = split[1].split('v').collect();
-    let lsns: Vec<&str> = lsn_and_opt_generation[0].split('-').collect();
-    let the_lsns: [&str; 2];
+pub(crate) fn parse_filename(name: &str) -> anyhow::Result<LayerFile> {
+    let layer_name =
+        LayerName::from_str(name).map_err(|e| anyhow!("failed to parse layer name: {e}"))?;
 
-    /*
-     * Generations add a -vX-XXXXXX postfix, which causes issues when we try to
-     * parse 'vX' as an LSN.
-     */
-    let is_delta = if lsns.len() == 1 || lsns[1].is_empty() {
-        the_lsns = [lsns[0], lsns[0]];
-        false
-    } else {
-        the_lsns = [lsns[0], lsns[1]];
-        true
-    };
-
-    let key_range = Key::from_hex(keys[0]).unwrap()..Key::from_hex(keys[1]).unwrap();
-    let lsn_range = Lsn::from_hex(the_lsns[0]).unwrap()..Lsn::from_hex(the_lsns[1]).unwrap();
     let holes = Vec::new();
-    Some(LayerFile {
-        key_range,
-        lsn_range,
-        is_delta,
+    Ok(LayerFile {
+        key_range: layer_name.key_range().clone(),
+        lsn_range: layer_name.lsn_as_range(),
+        is_delta: layer_name.is_delta(),
         holes,
     })
 }
@@ -179,7 +160,7 @@ pub(crate) async fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
 
             for layer in fs::read_dir(timeline.path())? {
                 let layer = layer?;
-                if let Some(mut layer_file) =
+                if let Ok(mut layer_file) =
                     parse_filename(&layer.file_name().into_string().unwrap())
                 {
                     if layer_file.is_delta {

pageserver/ctl/src/layers.rs

@@ -5,24 +5,12 @@ use camino::{Utf8Path, Utf8PathBuf};
 use clap::Subcommand;
 use pageserver::context::{DownloadBehavior, RequestContext};
 use pageserver::task_mgr::TaskKind;
-use pageserver::tenant::block_io::BlockCursor;
-use pageserver::tenant::disk_btree::DiskBtreeReader;
-use pageserver::tenant::storage_layer::delta_layer::{BlobRef, Summary};
 use pageserver::tenant::storage_layer::{delta_layer, image_layer};
 use pageserver::tenant::storage_layer::{DeltaLayer, ImageLayer};
 use pageserver::tenant::{TENANTS_SEGMENT_NAME, TIMELINES_SEGMENT_NAME};
 use pageserver::virtual_file::api::IoMode;
 use pageserver::{page_cache, virtual_file};
-use pageserver::{
-    repository::{Key, KEY_SIZE},
-    tenant::{
-        block_io::FileBlockReader, disk_btree::VisitDirection,
-        storage_layer::delta_layer::DELTA_KEY_SIZE,
-    },
-    virtual_file::VirtualFile,
-};
-use std::fs;
-use utils::bin_ser::BeSer;
+use std::fs::{self, File};
 use utils::id::{TenantId, TimelineId};
 
 use crate::layer_map_analyzer::parse_filename;
@@ -59,44 +47,30 @@ pub(crate) enum LayerCmd {
 }
 
 async fn read_delta_file(path: impl AsRef<Path>, ctx: &RequestContext) -> Result<()> {
-    let path = Utf8Path::from_path(path.as_ref()).expect("non-Unicode path");
     virtual_file::init(
         10,
         virtual_file::api::IoEngineKind::StdFs,
         IoMode::preferred(),
     );
     page_cache::init(100);
-    let file = VirtualFile::open(path, ctx).await?;
-    let file_id = page_cache::next_file_id();
-    let block_reader = FileBlockReader::new(&file, file_id);
-    let summary_blk = block_reader.read_blk(0, ctx).await?;
-    let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
-    let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
-        actual_summary.index_start_blk,
-        actual_summary.index_root_blk,
-        &block_reader,
+    let path = Utf8Path::from_path(path.as_ref()).expect("non-Unicode path");
+    let file = File::open(path)?;
+    let delta_layer = DeltaLayer::new_for_path(path, file)?;
+    delta_layer.dump(true, ctx).await?;
+    Ok(())
+}
+
+async fn read_image_file(path: impl AsRef<Path>, ctx: &RequestContext) -> Result<()> {
+    virtual_file::init(
+        10,
+        virtual_file::api::IoEngineKind::StdFs,
+        IoMode::preferred(),
     );
-    // TODO(chi): dedup w/ `delta_layer.rs` by exposing the API.
-    let mut all = vec![];
-    tree_reader
-        .visit(
-            &[0u8; DELTA_KEY_SIZE],
-            VisitDirection::Forwards,
-            |key, value_offset| {
-                let curr = Key::from_slice(&key[..KEY_SIZE]);
-                all.push((curr, BlobRef(value_offset)));
-                true
-            },
-            ctx,
-        )
-        .await?;
-    let cursor = BlockCursor::new_fileblockreader(&block_reader);
-    for (k, v) in all {
-        let value = cursor.read_blob(v.pos(), ctx).await?;
-        println!("key:{} value_len:{}", k, value.len());
-        assert!(k.is_i128_representable(), "invalid key: ");
-    }
-    // TODO(chi): special handling for last key?
+    page_cache::init(100);
+    let path = Utf8Path::from_path(path.as_ref()).expect("non-Unicode path");
+    let file = File::open(path)?;
+    let image_layer = ImageLayer::new_for_path(path, file)?;
+    image_layer.dump(true, ctx).await?;
     Ok(())
 }
 
@@ -133,8 +107,7 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
             let mut idx = 0;
             for layer in fs::read_dir(timeline_path)? {
                 let layer = layer?;
-                if let Some(layer_file) = parse_filename(&layer.file_name().into_string().unwrap())
-                {
+                if let Ok(layer_file) = parse_filename(&layer.file_name().into_string().unwrap()) {
                     println!(
                         "[{:3}]  key:{}-{}\n       lsn:{}-{}\n       delta:{}",
                         idx,
@@ -163,8 +136,7 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
             let mut idx = 0;
             for layer in fs::read_dir(timeline_path)? {
                 let layer = layer?;
-                if let Some(layer_file) = parse_filename(&layer.file_name().into_string().unwrap())
-                {
+                if let Ok(layer_file) = parse_filename(&layer.file_name().into_string().unwrap()) {
                     if *id == idx {
                         // TODO(chi): dedup code
                         println!(
@@ -180,7 +152,7 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
                         if layer_file.is_delta {
                             read_delta_file(layer.path(), &ctx).await?;
                         } else {
-                            anyhow::bail!("not supported yet :(");
+                            read_image_file(layer.path(), &ctx).await?;
                         }
 
                         break;

pageserver/pagebench/src/cmd/aux_files.rs

@@ -1,4 +1,4 @@
-use pageserver_api::models::{AuxFilePolicy, TenantConfig, TenantConfigRequest};
+use pageserver_api::models::{TenantConfig, TenantConfigRequest};
 use pageserver_api::shard::TenantShardId;
 use utils::id::TenantTimelineId;
 use utils::lsn::Lsn;
@@ -66,10 +66,7 @@ async fn main_impl(args: Args) -> anyhow::Result<()> {
     mgmt_api_client
         .tenant_config(&TenantConfigRequest {
             tenant_id: timeline.tenant_id,
-            config: TenantConfig {
-                switch_aux_file_policy: Some(AuxFilePolicy::V2),
-                ..Default::default()
-            },
+            config: TenantConfig::default(),
         })
         .await?;
 

pageserver/src/basebackup.rs

@@ -59,6 +59,7 @@ pub async fn send_basebackup_tarball<'a, W>(
     req_lsn: Option<Lsn>,
     prev_lsn: Option<Lsn>,
     full_backup: bool,
+    replica: bool,
     ctx: &'a RequestContext,
 ) -> Result<(), BasebackupError>
 where
@@ -110,8 +111,8 @@ where
     };
 
     info!(
-        "taking basebackup lsn={}, prev_lsn={} (full_backup={})",
-        backup_lsn, prev_lsn, full_backup
+        "taking basebackup lsn={}, prev_lsn={} (full_backup={}, replica={})",
+        backup_lsn, prev_lsn, full_backup, replica
     );
 
     let basebackup = Basebackup {
@@ -120,6 +121,7 @@ where
         lsn: backup_lsn,
         prev_record_lsn: prev_lsn,
         full_backup,
+        replica,
         ctx,
     };
     basebackup
@@ -140,6 +142,7 @@ where
     lsn: Lsn,
     prev_record_lsn: Lsn,
     full_backup: bool,
+    replica: bool,
     ctx: &'a RequestContext,
 }
 
@@ -372,6 +375,10 @@ where
 
         for (path, content) in aux_files {
             if path.starts_with("pg_replslot") {
+                // Do not create LR slots at standby because they are not used but prevent WAL truncation
+                if self.replica {
+                    continue;
+                }
                 let offs = pg_constants::REPL_SLOT_ON_DISK_OFFSETOF_RESTART_LSN;
                 let restart_lsn = Lsn(u64::from_le_bytes(
                     content[offs..offs + 8].try_into().unwrap(),

pageserver/src/bin/pageserver.rs

@@ -398,9 +398,7 @@ fn start_pageserver(
         ControllerUpcallClient::new(conf, &shutdown_pageserver),
         conf,
     );
-    if let Some(deletion_workers) = deletion_workers {
-        deletion_workers.spawn_with(BACKGROUND_RUNTIME.handle());
-    }
+    deletion_workers.spawn_with(BACKGROUND_RUNTIME.handle());
 
     // Up to this point no significant I/O has been done: this should have been fast.  Record
     // duration prior to starting I/O intensive phase of startup.

pageserver/src/consumption_metrics.rs

@@ -14,6 +14,7 @@ use itertools::Itertools as _;
 use pageserver_api::models::TenantState;
 use remote_storage::{GenericRemoteStorage, RemoteStorageConfig};
 use reqwest::Url;
+use serde::{Deserialize, Serialize};
 use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::{Duration, SystemTime};
@@ -35,12 +36,62 @@ const DEFAULT_HTTP_REPORTING_TIMEOUT: Duration = Duration::from_secs(60);
 /// upload attempts.
 type RawMetric = (MetricsKey, (EventType, u64));
 
+/// The new serializable metrics format
+#[derive(Serialize, Deserialize)]
+struct NewMetricsRoot {
+    version: usize,
+    metrics: Vec<NewRawMetric>,
+}
+
+impl NewMetricsRoot {
+    pub fn is_v2_metrics(json_value: &serde_json::Value) -> bool {
+        if let Some(ver) = json_value.get("version") {
+            if let Some(2) = ver.as_u64() {
+                return true;
+            }
+        }
+        false
+    }
+}
+
+/// The new serializable metrics format
+#[derive(Serialize)]
+struct NewMetricsRefRoot<'a> {
+    version: usize,
+    metrics: &'a [NewRawMetric],
+}
+
+impl<'a> NewMetricsRefRoot<'a> {
+    fn new(metrics: &'a [NewRawMetric]) -> Self {
+        Self {
+            version: 2,
+            metrics,
+        }
+    }
+}
+
+/// The new serializable metrics format
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Eq)]
+struct NewRawMetric {
+    key: MetricsKey,
+    kind: EventType,
+    value: u64,
+    // TODO: add generation field and check against generations
+}
+
+impl NewRawMetric {
+    #[cfg(test)]
+    fn to_kv_pair(&self) -> (MetricsKey, NewRawMetric) {
+        (self.key, self.clone())
+    }
+}
+
 /// Caches the [`RawMetric`]s
 ///
 /// In practice, during startup, last sent values are stored here to be used in calculating new
 /// ones. After successful uploading, the cached values are updated to cache. This used to be used
 /// for deduplication, but that is no longer needed.
-type Cache = HashMap<MetricsKey, (EventType, u64)>;
+type Cache = HashMap<MetricsKey, NewRawMetric>;
 
 pub async fn run(
     conf: &'static PageServerConf,
@@ -231,11 +282,14 @@ async fn restore_and_reschedule(
             // collect_all_metrics
             let earlier_metric_at = found_some
                 .iter()
-                .map(|(_, (et, _))| et.recorded_at())
+                .map(|item| item.kind.recorded_at())
                 .copied()
                 .next();
 
-            let cached = found_some.into_iter().collect::<Cache>();
+            let cached = found_some
+                .into_iter()
+                .map(|item| (item.key, item))
+                .collect::<Cache>();
 
             (cached, earlier_metric_at)
         }

pageserver/src/consumption_metrics/disk_cache.rs

@@ -2,11 +2,33 @@ use anyhow::Context;
 use camino::{Utf8Path, Utf8PathBuf};
 use std::sync::Arc;
 
-use super::RawMetric;
+use crate::consumption_metrics::NewMetricsRefRoot;
+
+use super::{NewMetricsRoot, NewRawMetric, RawMetric};
+
+pub(super) fn read_metrics_from_serde_value(
+    json_value: serde_json::Value,
+) -> anyhow::Result<Vec<NewRawMetric>> {
+    if NewMetricsRoot::is_v2_metrics(&json_value) {
+        let root = serde_json::from_value::<NewMetricsRoot>(json_value)?;
+        Ok(root.metrics)
+    } else {
+        let all_metrics = serde_json::from_value::<Vec<RawMetric>>(json_value)?;
+        let all_metrics = all_metrics
+            .into_iter()
+            .map(|(key, (event_type, value))| NewRawMetric {
+                key,
+                kind: event_type,
+                value,
+            })
+            .collect();
+        Ok(all_metrics)
+    }
+}
 
 pub(super) async fn read_metrics_from_disk(
     path: Arc<Utf8PathBuf>,
-) -> anyhow::Result<Vec<RawMetric>> {
+) -> anyhow::Result<Vec<NewRawMetric>> {
     // do not add context to each error, callsite will log with full path
     let span = tracing::Span::current();
     tokio::task::spawn_blocking(move || {
@@ -20,7 +42,8 @@ pub(super) async fn read_metrics_from_disk(
 
         let mut file = std::fs::File::open(&*path)?;
         let reader = std::io::BufReader::new(&mut file);
-        anyhow::Ok(serde_json::from_reader::<_, Vec<RawMetric>>(reader)?)
+        let json_value = serde_json::from_reader::<_, serde_json::Value>(reader)?;
+        read_metrics_from_serde_value(json_value)
     })
     .await
     .context("read metrics join error")
@@ -63,7 +86,7 @@ fn scan_and_delete_with_same_prefix(path: &Utf8Path) -> std::io::Result<()> {
 }
 
 pub(super) async fn flush_metrics_to_disk(
-    current_metrics: &Arc<Vec<RawMetric>>,
+    current_metrics: &Arc<Vec<NewRawMetric>>,
     path: &Arc<Utf8PathBuf>,
 ) -> anyhow::Result<()> {
     use std::io::Write;
@@ -93,8 +116,11 @@ pub(super) async fn flush_metrics_to_disk(
             // write out all of the raw metrics, to be read out later on restart as cached values
             {
                 let mut writer = std::io::BufWriter::new(&mut tempfile);
-                serde_json::to_writer(&mut writer, &*current_metrics)
-                    .context("serialize metrics")?;
+                serde_json::to_writer(
+                    &mut writer,
+                    &NewMetricsRefRoot::new(current_metrics.as_ref()),
+                )
+                .context("serialize metrics")?;
                 writer
                     .into_inner()
                     .map_err(|_| anyhow::anyhow!("flushing metrics failed"))?;

pageserver/src/consumption_metrics/metrics.rs

@@ -9,7 +9,7 @@ use utils::{
     lsn::Lsn,
 };
 
-use super::{Cache, RawMetric};
+use super::{Cache, NewRawMetric};
 
 /// Name of the metric, used by `MetricsKey` factory methods and `deserialize_cached_events`
 /// instead of static str.
@@ -64,11 +64,21 @@ impl MetricsKey {
 struct AbsoluteValueFactory(MetricsKey);
 
 impl AbsoluteValueFactory {
-    const fn at(self, time: DateTime<Utc>, val: u64) -> RawMetric {
+    #[cfg(test)]
+    const fn at_old_format(self, time: DateTime<Utc>, val: u64) -> super::RawMetric {
         let key = self.0;
         (key, (EventType::Absolute { time }, val))
     }
 
+    const fn at(self, time: DateTime<Utc>, val: u64) -> NewRawMetric {
+        let key = self.0;
+        NewRawMetric {
+            key,
+            kind: EventType::Absolute { time },
+            value: val,
+        }
+    }
+
     fn key(&self) -> &MetricsKey {
         &self.0
     }
@@ -84,7 +94,28 @@ impl IncrementalValueFactory {
         prev_end: DateTime<Utc>,
         up_to: DateTime<Utc>,
         val: u64,
-    ) -> RawMetric {
+    ) -> NewRawMetric {
+        let key = self.0;
+        // cannot assert prev_end < up_to because these are realtime clock based
+        let when = EventType::Incremental {
+            start_time: prev_end,
+            stop_time: up_to,
+        };
+        NewRawMetric {
+            key,
+            kind: when,
+            value: val,
+        }
+    }
+
+    #[allow(clippy::wrong_self_convention)]
+    #[cfg(test)]
+    const fn from_until_old_format(
+        self,
+        prev_end: DateTime<Utc>,
+        up_to: DateTime<Utc>,
+        val: u64,
+    ) -> super::RawMetric {
         let key = self.0;
         // cannot assert prev_end < up_to because these are realtime clock based
         let when = EventType::Incremental {
@@ -185,7 +216,7 @@ pub(super) async fn collect_all_metrics(
     tenant_manager: &Arc<TenantManager>,
     cached_metrics: &Cache,
     ctx: &RequestContext,
-) -> Vec<RawMetric> {
+) -> Vec<NewRawMetric> {
     use pageserver_api::models::TenantState;
 
     let started_at = std::time::Instant::now();
@@ -220,11 +251,11 @@ pub(super) async fn collect_all_metrics(
     res
 }
 
-async fn collect<S>(tenants: S, cache: &Cache, ctx: &RequestContext) -> Vec<RawMetric>
+async fn collect<S>(tenants: S, cache: &Cache, ctx: &RequestContext) -> Vec<NewRawMetric>
 where
     S: futures::stream::Stream<Item = (TenantId, Arc<crate::tenant::Tenant>)>,
 {
-    let mut current_metrics: Vec<RawMetric> = Vec::new();
+    let mut current_metrics: Vec<NewRawMetric> = Vec::new();
 
     let mut tenants = std::pin::pin!(tenants);
 
@@ -291,7 +322,7 @@ impl TenantSnapshot {
         tenant_id: TenantId,
         now: DateTime<Utc>,
         cached: &Cache,
-        metrics: &mut Vec<RawMetric>,
+        metrics: &mut Vec<NewRawMetric>,
     ) {
         let remote_size = MetricsKey::remote_storage_size(tenant_id).at(now, self.remote_size);
 
@@ -302,9 +333,9 @@ impl TenantSnapshot {
             let mut synthetic_size = self.synthetic_size;
 
             if synthetic_size == 0 {
-                if let Some((_, value)) = cached.get(factory.key()) {
-                    // use the latest value from previous session
-                    synthetic_size = *value;
+                if let Some(item) = cached.get(factory.key()) {
+                    // use the latest value from previous session, TODO: check generation number
+                    synthetic_size = item.value;
                 }
             }
 
@@ -381,37 +412,36 @@ impl TimelineSnapshot {
         tenant_id: TenantId,
         timeline_id: TimelineId,
         now: DateTime<Utc>,
-        metrics: &mut Vec<RawMetric>,
+        metrics: &mut Vec<NewRawMetric>,
         cache: &Cache,
     ) {
         let timeline_written_size = u64::from(self.last_record_lsn);
 
         let written_size_delta_key = MetricsKey::written_size_delta(tenant_id, timeline_id);
 
-        let last_stop_time = cache
-            .get(written_size_delta_key.key())
-            .map(|(until, _val)| {
-                until
-                    .incremental_timerange()
-                    .expect("never create EventType::Absolute for written_size_delta")
-                    .end
-            });
+        let last_stop_time = cache.get(written_size_delta_key.key()).map(|item| {
+            item.kind
+                .incremental_timerange()
+                .expect("never create EventType::Absolute for written_size_delta")
+                .end
+        });
 
-        let (key, written_size_now) =
+        let written_size_now =
             MetricsKey::written_size(tenant_id, timeline_id).at(now, timeline_written_size);
 
         // by default, use the last sent written_size as the basis for
         // calculating the delta. if we don't yet have one, use the load time value.
-        let prev = cache
-            .get(&key)
-            .map(|(prev_at, prev)| {
+        let prev: (DateTime<Utc>, u64) = cache
+            .get(&written_size_now.key)
+            .map(|item| {
                 // use the prev time from our last incremental update, or default to latest
                 // absolute update on the first round.
-                let prev_at = prev_at
+                let prev_at = item
+                    .kind
                     .absolute_time()
                     .expect("never create EventType::Incremental for written_size");
                 let prev_at = last_stop_time.unwrap_or(prev_at);
-                (*prev_at, *prev)
+                (*prev_at, item.value)
             })
             .unwrap_or_else(|| {
                 // if we don't have a previous point of comparison, compare to the load time
@@ -422,24 +452,28 @@ impl TimelineSnapshot {
 
         let up_to = now;
 
-        if let Some(delta) = written_size_now.1.checked_sub(prev.1) {
+        if let Some(delta) = written_size_now.value.checked_sub(prev.1) {
             let key_value = written_size_delta_key.from_until(prev.0, up_to, delta);
             // written_size_delta
             metrics.push(key_value);
             // written_size
-            metrics.push((key, written_size_now));
+            metrics.push(written_size_now);
         } else {
             // the cached value was ahead of us, report zero until we've caught up
             metrics.push(written_size_delta_key.from_until(prev.0, up_to, 0));
             // the cached value was ahead of us, report the same until we've caught up
-            metrics.push((key, (written_size_now.0, prev.1)));
+            metrics.push(NewRawMetric {
+                key: written_size_now.key,
+                kind: written_size_now.kind,
+                value: prev.1,
+            });
         }
 
         {
             let factory = MetricsKey::timeline_logical_size(tenant_id, timeline_id);
             let current_or_previous = self
                 .current_exact_logical_size
-                .or_else(|| cache.get(factory.key()).map(|(_, val)| *val));
+                .or_else(|| cache.get(factory.key()).map(|item| item.value));
 
             if let Some(size) = current_or_previous {
                 metrics.push(factory.at(now, size));
@@ -452,4 +486,4 @@ impl TimelineSnapshot {
 mod tests;
 
 #[cfg(test)]
-pub(crate) use tests::metric_examples;
+pub(crate) use tests::{metric_examples, metric_examples_old};

pageserver/src/consumption_metrics/metrics/tests.rs

@@ -1,3 +1,5 @@
+use crate::consumption_metrics::RawMetric;
+
 use super::*;
 use std::collections::HashMap;
 
@@ -50,9 +52,9 @@ fn startup_collected_timeline_metrics_second_round() {
     let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
 
     let mut metrics = Vec::new();
-    let cache = HashMap::from([
-        MetricsKey::written_size(tenant_id, timeline_id).at(before, disk_consistent_lsn.0)
-    ]);
+    let cache = HashMap::from([MetricsKey::written_size(tenant_id, timeline_id)
+        .at(before, disk_consistent_lsn.0)
+        .to_kv_pair()]);
 
     let snap = TimelineSnapshot {
         loaded_at: (disk_consistent_lsn, init),
@@ -89,9 +91,13 @@ fn startup_collected_timeline_metrics_nth_round_at_same_lsn() {
     let mut metrics = Vec::new();
     let cache = HashMap::from([
         // at t=before was the last time the last_record_lsn changed
-        MetricsKey::written_size(tenant_id, timeline_id).at(before, disk_consistent_lsn.0),
+        MetricsKey::written_size(tenant_id, timeline_id)
+            .at(before, disk_consistent_lsn.0)
+            .to_kv_pair(),
         // end time of this event is used for the next ones
-        MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(before, just_before, 0),
+        MetricsKey::written_size_delta(tenant_id, timeline_id)
+            .from_until(before, just_before, 0)
+            .to_kv_pair(),
     ]);
 
     let snap = TimelineSnapshot {
@@ -138,13 +144,17 @@ fn post_restart_written_sizes_with_rolled_back_last_record_lsn() {
     };
 
     let mut cache = HashMap::from([
-        MetricsKey::written_size(tenant_id, timeline_id).at(before_restart, 100),
-        MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(
-            way_before,
-            before_restart,
-            // not taken into account, but the timestamps are important
-            999_999_999,
-        ),
+        MetricsKey::written_size(tenant_id, timeline_id)
+            .at(before_restart, 100)
+            .to_kv_pair(),
+        MetricsKey::written_size_delta(tenant_id, timeline_id)
+            .from_until(
+                way_before,
+                before_restart,
+                // not taken into account, but the timestamps are important
+                999_999_999,
+            )
+            .to_kv_pair(),
     ]);
 
     let mut metrics = Vec::new();
@@ -163,7 +173,7 @@ fn post_restart_written_sizes_with_rolled_back_last_record_lsn() {
     );
 
     // now if we cache these metrics, and re-run while "still in recovery"
-    cache.extend(metrics.drain(..));
+    cache.extend(metrics.drain(..).map(|x| x.to_kv_pair()));
 
     // "still in recovery", because our snapshot did not change
     snap.to_metrics(tenant_id, timeline_id, later, &mut metrics, &cache);
@@ -194,14 +204,14 @@ fn post_restart_current_exact_logical_size_uses_cached() {
         current_exact_logical_size: None,
     };
 
-    let cache = HashMap::from([
-        MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(before_restart, 100)
-    ]);
+    let cache = HashMap::from([MetricsKey::timeline_logical_size(tenant_id, timeline_id)
+        .at(before_restart, 100)
+        .to_kv_pair()]);
 
     let mut metrics = Vec::new();
     snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
 
-    metrics.retain(|(key, _)| key.metric == Name::LogicalSize);
+    metrics.retain(|item| item.key.metric == Name::LogicalSize);
 
     assert_eq!(
         metrics,
@@ -224,7 +234,9 @@ fn post_restart_synthetic_size_uses_cached_if_available() {
     let before_restart = DateTime::<Utc>::from(now - std::time::Duration::from_secs(5 * 60));
     let now = DateTime::<Utc>::from(now);
 
-    let cached = HashMap::from([MetricsKey::synthetic_size(tenant_id).at(before_restart, 1000)]);
+    let cached = HashMap::from([MetricsKey::synthetic_size(tenant_id)
+        .at(before_restart, 1000)
+        .to_kv_pair()]);
 
     let mut metrics = Vec::new();
     ts.to_metrics(tenant_id, now, &cached, &mut metrics);
@@ -278,12 +290,29 @@ fn time_backwards<const N: usize>() -> [std::time::SystemTime; N] {
     times
 }
 
-pub(crate) const fn metric_examples(
+pub(crate) const fn metric_examples_old(
     tenant_id: TenantId,
     timeline_id: TimelineId,
     now: DateTime<Utc>,
     before: DateTime<Utc>,
 ) -> [RawMetric; 6] {
+    [
+        MetricsKey::written_size(tenant_id, timeline_id).at_old_format(now, 0),
+        MetricsKey::written_size_delta(tenant_id, timeline_id)
+            .from_until_old_format(before, now, 0),
+        MetricsKey::timeline_logical_size(tenant_id, timeline_id).at_old_format(now, 0),
+        MetricsKey::remote_storage_size(tenant_id).at_old_format(now, 0),
+        MetricsKey::resident_size(tenant_id).at_old_format(now, 0),
+        MetricsKey::synthetic_size(tenant_id).at_old_format(now, 1),
+    ]
+}
+
+pub(crate) const fn metric_examples(
+    tenant_id: TenantId,
+    timeline_id: TimelineId,
+    now: DateTime<Utc>,
+    before: DateTime<Utc>,
+) -> [NewRawMetric; 6] {
     [
         MetricsKey::written_size(tenant_id, timeline_id).at(now, 0),
         MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(before, now, 0),

pageserver/src/consumption_metrics/upload.rs

@@ -7,7 +7,7 @@ use tokio::io::AsyncWriteExt;
 use tokio_util::sync::CancellationToken;
 use tracing::Instrument;
 
-use super::{metrics::Name, Cache, MetricsKey, RawMetric};
+use super::{metrics::Name, Cache, MetricsKey, NewRawMetric, RawMetric};
 use utils::id::{TenantId, TimelineId};
 
 /// How the metrics from pageserver are identified.
@@ -24,7 +24,7 @@ pub(super) async fn upload_metrics_http(
     client: &reqwest::Client,
     metric_collection_endpoint: &reqwest::Url,
     cancel: &CancellationToken,
-    metrics: &[RawMetric],
+    metrics: &[NewRawMetric],
     cached_metrics: &mut Cache,
     idempotency_keys: &[IdempotencyKey<'_>],
 ) -> anyhow::Result<()> {
@@ -53,8 +53,8 @@ pub(super) async fn upload_metrics_http(
 
         match res {
             Ok(()) => {
-                for (curr_key, curr_val) in chunk {
-                    cached_metrics.insert(*curr_key, *curr_val);
+                for item in chunk {
+                    cached_metrics.insert(item.key, item.clone());
                 }
                 uploaded += chunk.len();
             }
@@ -86,7 +86,7 @@ pub(super) async fn upload_metrics_bucket(
     client: &GenericRemoteStorage,
     cancel: &CancellationToken,
     node_id: &str,
-    metrics: &[RawMetric],
+    metrics: &[NewRawMetric],
     idempotency_keys: &[IdempotencyKey<'_>],
 ) -> anyhow::Result<()> {
     if metrics.is_empty() {
@@ -140,16 +140,16 @@ pub(super) async fn upload_metrics_bucket(
 /// across different metrics sinks), and must have the same length as input.
 fn serialize_in_chunks<'a>(
     chunk_size: usize,
-    input: &'a [RawMetric],
+    input: &'a [NewRawMetric],
     idempotency_keys: &'a [IdempotencyKey<'a>],
-) -> impl ExactSizeIterator<Item = Result<(&'a [RawMetric], bytes::Bytes), serde_json::Error>> + 'a
+) -> impl ExactSizeIterator<Item = Result<(&'a [NewRawMetric], bytes::Bytes), serde_json::Error>> + 'a
 {
     use bytes::BufMut;
 
     assert_eq!(input.len(), idempotency_keys.len());
 
     struct Iter<'a> {
-        inner: std::slice::Chunks<'a, RawMetric>,
+        inner: std::slice::Chunks<'a, NewRawMetric>,
         idempotency_keys: std::slice::Iter<'a, IdempotencyKey<'a>>,
         chunk_size: usize,
 
@@ -160,7 +160,7 @@ fn serialize_in_chunks<'a>(
     }
 
     impl<'a> Iterator for Iter<'a> {
-        type Item = Result<(&'a [RawMetric], bytes::Bytes), serde_json::Error>;
+        type Item = Result<(&'a [NewRawMetric], bytes::Bytes), serde_json::Error>;
 
         fn next(&mut self) -> Option<Self::Item> {
             let chunk = self.inner.next()?;
@@ -269,6 +269,58 @@ impl RawMetricExt for RawMetric {
     }
 }
 
+impl RawMetricExt for NewRawMetric {
+    fn as_event(&self, key: &IdempotencyKey<'_>) -> Event<Ids, Name> {
+        let MetricsKey {
+            metric,
+            tenant_id,
+            timeline_id,
+        } = self.key;
+
+        let kind = self.kind;
+        let value = self.value;
+
+        Event {
+            kind,
+            metric,
+            idempotency_key: key.to_string(),
+            value,
+            extra: Ids {
+                tenant_id,
+                timeline_id,
+            },
+        }
+    }
+
+    fn update_in_place(&self, event: &mut Event<Ids, Name>, key: &IdempotencyKey<'_>) {
+        use std::fmt::Write;
+
+        let MetricsKey {
+            metric,
+            tenant_id,
+            timeline_id,
+        } = self.key;
+
+        let kind = self.kind;
+        let value = self.value;
+
+        *event = Event {
+            kind,
+            metric,
+            idempotency_key: {
+                event.idempotency_key.clear();
+                write!(event.idempotency_key, "{key}").unwrap();
+                std::mem::take(&mut event.idempotency_key)
+            },
+            value,
+            extra: Ids {
+                tenant_id,
+                timeline_id,
+            },
+        };
+    }
+}
+
 pub(crate) trait KeyGen<'a> {
     fn generate(&self) -> IdempotencyKey<'a>;
 }
@@ -381,6 +433,10 @@ async fn upload(
 
 #[cfg(test)]
 mod tests {
+    use crate::consumption_metrics::{
+        disk_cache::read_metrics_from_serde_value, NewMetricsRefRoot,
+    };
+
     use super::*;
     use chrono::{DateTime, Utc};
     use once_cell::sync::Lazy;
@@ -473,23 +529,49 @@ mod tests {
         let idempotency_key = consumption_metrics::IdempotencyKey::for_tests(*SAMPLES_NOW, "1", 0);
         let examples = examples.into_iter().zip(metric_samples());
 
-        for ((line, expected), (key, (kind, value))) in examples {
+        for ((line, expected), item) in examples {
             let e = consumption_metrics::Event {
-                kind,
-                metric: key.metric,
+                kind: item.kind,
+                metric: item.key.metric,
                 idempotency_key: idempotency_key.to_string(),
-                value,
+                value: item.value,
                 extra: Ids {
-                    tenant_id: key.tenant_id,
-                    timeline_id: key.timeline_id,
+                    tenant_id: item.key.tenant_id,
+                    timeline_id: item.key.timeline_id,
                 },
             };
             let actual = serde_json::to_string(&e).unwrap();
-            assert_eq!(expected, actual, "example for {kind:?} from line {line}");
+            assert_eq!(
+                expected, actual,
+                "example for {:?} from line {line}",
+                item.kind
+            );
         }
     }
 
-    fn metric_samples() -> [RawMetric; 6] {
+    #[test]
+    fn disk_format_upgrade() {
+        let old_samples_json = serde_json::to_value(metric_samples_old()).unwrap();
+        let new_samples =
+            serde_json::to_value(NewMetricsRefRoot::new(metric_samples().as_ref())).unwrap();
+        let upgraded_samples = read_metrics_from_serde_value(old_samples_json).unwrap();
+        let new_samples = read_metrics_from_serde_value(new_samples).unwrap();
+        assert_eq!(upgraded_samples, new_samples);
+    }
+
+    fn metric_samples_old() -> [RawMetric; 6] {
+        let tenant_id = TenantId::from_array([0; 16]);
+        let timeline_id = TimelineId::from_array([0xff; 16]);
+
+        let before = DateTime::parse_from_rfc3339("2023-09-14T00:00:00.123456789Z")
+            .unwrap()
+            .into();
+        let [now, before] = [*SAMPLES_NOW, before];
+
+        super::super::metrics::metric_examples_old(tenant_id, timeline_id, now, before)
+    }
+
+    fn metric_samples() -> [NewRawMetric; 6] {
         let tenant_id = TenantId::from_array([0; 16]);
         let timeline_id = TimelineId::from_array([0xff; 16]);
 

pageserver/src/deletion_queue.rs

@@ -618,13 +618,11 @@ impl DeletionQueue {
     /// Caller may use the returned object to construct clients with new_client.
     /// Caller should tokio::spawn the background() members of the two worker objects returned:
     /// we don't spawn those inside new() so that the caller can use their runtime/spans of choice.
-    ///
-    /// If remote_storage is None, then the returned workers will also be None.
     pub fn new<C>(
         remote_storage: GenericRemoteStorage,
         controller_upcall_client: Option<C>,
         conf: &'static PageServerConf,
-    ) -> (Self, Option<DeletionQueueWorkers<C>>)
+    ) -> (Self, DeletionQueueWorkers<C>)
     where
         C: ControlPlaneGenerationsApi + Send + Sync,
     {
@@ -656,7 +654,7 @@ impl DeletionQueue {
                 },
                 cancel: cancel.clone(),
             },
-            Some(DeletionQueueWorkers {
+            DeletionQueueWorkers {
                 frontend: ListWriter::new(conf, rx, backend_tx, cancel.clone()),
                 backend: Validator::new(
                     conf,
@@ -667,7 +665,7 @@ impl DeletionQueue {
                     cancel.clone(),
                 ),
                 executor: Deleter::new(remote_storage, executor_rx, cancel.clone()),
-            }),
+            },
         )
     }
 
@@ -696,7 +694,7 @@ impl DeletionQueue {
 mod test {
     use camino::Utf8Path;
     use hex_literal::hex;
-    use pageserver_api::{shard::ShardIndex, upcall_api::ReAttachResponseTenant};
+    use pageserver_api::{key::Key, shard::ShardIndex, upcall_api::ReAttachResponseTenant};
     use std::{io::ErrorKind, time::Duration};
     use tracing::info;
 
@@ -705,7 +703,6 @@ mod test {
 
     use crate::{
         controller_upcall_client::RetryForeverError,
-        repository::Key,
         tenant::{harness::TenantHarness, storage_layer::DeltaLayerName},
     };
 
@@ -743,9 +740,7 @@ mod test {
             );
 
             tracing::debug!("Spawning worker for new queue queue");
-            let worker_join = workers
-                .unwrap()
-                .spawn_with(&tokio::runtime::Handle::current());
+            let worker_join = workers.spawn_with(&tokio::runtime::Handle::current());
 
             let old_worker_join = std::mem::replace(&mut self.worker_join, worker_join);
             let old_deletion_queue = std::mem::replace(&mut self.deletion_queue, deletion_queue);
@@ -856,7 +851,6 @@ mod test {
             harness.conf,
         );
 
-        let worker = worker.unwrap();
         let worker_join = worker.spawn_with(&tokio::runtime::Handle::current());
 
         Ok(TestSetup {

pageserver/src/http/openapi_spec.yml

@@ -597,6 +597,10 @@ paths:
         Create a timeline. Returns new timeline id on success.
         Recreating the same timeline will succeed if the parameters match the existing timeline.
         If no pg_version is specified, assume DEFAULT_PG_VERSION hardcoded in the pageserver.
+
+        To ensure durability, the caller must retry the creation until success.
+        Just because the timeline is visible via other endpoints does not mean it is durable.
+        Future versions may stop showing timelines that are not yet durable.
       requestBody:
         content:
           application/json:

pageserver/src/http/routes.rs

@@ -38,6 +38,7 @@ use pageserver_api::models::TenantShardSplitRequest;
 use pageserver_api::models::TenantShardSplitResponse;
 use pageserver_api::models::TenantSorting;
 use pageserver_api::models::TimelineArchivalConfigRequest;
+use pageserver_api::models::TimelineCreateRequestMode;
 use pageserver_api::models::TimelinesInfoAndOffloaded;
 use pageserver_api::models::TopTenantShardItem;
 use pageserver_api::models::TopTenantShardsRequest;
@@ -79,12 +80,14 @@ use crate::tenant::size::ModelInputs;
 use crate::tenant::storage_layer::LayerAccessStatsReset;
 use crate::tenant::storage_layer::LayerName;
 use crate::tenant::timeline::offload::offload_timeline;
+use crate::tenant::timeline::offload::OffloadError;
 use crate::tenant::timeline::CompactFlags;
 use crate::tenant::timeline::CompactionError;
 use crate::tenant::timeline::Timeline;
 use crate::tenant::GetTimelineError;
 use crate::tenant::OffloadedTimeline;
 use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError};
+use crate::DEFAULT_PG_VERSION;
 use crate::{disk_usage_eviction_task, tenant};
 use pageserver_api::models::{
     StatusResponse, TenantConfigRequest, TenantInfo, TimelineCreateRequest, TimelineGcRequest,
@@ -547,6 +550,26 @@ async fn timeline_create_handler(
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
     let new_timeline_id = request_data.new_timeline_id;
+    // fill in the default pg_version if not provided & convert request into domain model
+    let params: tenant::CreateTimelineParams = match request_data.mode {
+        TimelineCreateRequestMode::Bootstrap {
+            existing_initdb_timeline_id,
+            pg_version,
+        } => tenant::CreateTimelineParams::Bootstrap(tenant::CreateTimelineParamsBootstrap {
+            new_timeline_id,
+            existing_initdb_timeline_id,
+            pg_version: pg_version.unwrap_or(DEFAULT_PG_VERSION),
+        }),
+        TimelineCreateRequestMode::Branch {
+            ancestor_timeline_id,
+            ancestor_start_lsn,
+            pg_version: _,
+        } => tenant::CreateTimelineParams::Branch(tenant::CreateTimelineParamsBranch {
+            new_timeline_id,
+            ancestor_timeline_id,
+            ancestor_start_lsn,
+        }),
+    };
 
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Error);
 
@@ -559,22 +582,12 @@ async fn timeline_create_handler(
 
         tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
-        if let Some(ancestor_id) = request_data.ancestor_timeline_id.as_ref() {
-            tracing::info!(%ancestor_id, "starting to branch");
-        } else {
-            tracing::info!("bootstrapping");
-        }
+        // earlier versions of the code had pg_version and ancestor_lsn in the span
+        // => continue to provide that information, but, through a log message that doesn't require us to destructure
+        tracing::info!(?params, "creating timeline");
 
         match tenant
-            .create_timeline(
-                new_timeline_id,
-                request_data.ancestor_timeline_id,
-                request_data.ancestor_start_lsn,
-                request_data.pg_version.unwrap_or(crate::DEFAULT_PG_VERSION),
-                request_data.existing_initdb_timeline_id,
-                state.broker_client.clone(),
-                &ctx,
-            )
+            .create_timeline(params, state.broker_client.clone(), &ctx)
             .await
         {
             Ok(new_timeline) => {
@@ -625,8 +638,6 @@ async fn timeline_create_handler(
         tenant_id = %tenant_shard_id.tenant_id,
         shard_id = %tenant_shard_id.shard_slug(),
         timeline_id = %new_timeline_id,
-        lsn=?request_data.ancestor_start_lsn,
-        pg_version=?request_data.pg_version
     ))
     .await
 }
@@ -1283,6 +1294,99 @@ async fn layer_map_info_handler(
     json_response(StatusCode::OK, layer_map_info)
 }
 
+#[instrument(skip_all, fields(tenant_id, shard_id, timeline_id, layer_name))]
+async fn timeline_layer_scan_disposable_keys(
+    request: Request<Body>,
+    cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    let layer_name: LayerName = parse_request_param(&request, "layer_name")?;
+
+    tracing::Span::current().record(
+        "tenant_id",
+        tracing::field::display(&tenant_shard_id.tenant_id),
+    );
+    tracing::Span::current().record(
+        "shard_id",
+        tracing::field::display(tenant_shard_id.shard_slug()),
+    );
+    tracing::Span::current().record("timeline_id", tracing::field::display(&timeline_id));
+    tracing::Span::current().record("layer_name", tracing::field::display(&layer_name));
+
+    let state = get_state(&request);
+
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+
+    // technically the timeline need not be active for this scan to complete
+    let timeline =
+        active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
+            .await?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+
+    let guard = timeline.layers.read().await;
+    let Some(layer) = guard.try_get_from_key(&layer_name.clone().into()) else {
+        return Err(ApiError::NotFound(
+            anyhow::anyhow!("Layer {tenant_shard_id}/{timeline_id}/{layer_name} not found").into(),
+        ));
+    };
+
+    let resident_layer = layer
+        .download_and_keep_resident()
+        .await
+        .map_err(|err| match err {
+            tenant::storage_layer::layer::DownloadError::TimelineShutdown
+            | tenant::storage_layer::layer::DownloadError::DownloadCancelled => {
+                ApiError::ShuttingDown
+            }
+            tenant::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
+            | tenant::storage_layer::layer::DownloadError::DownloadRequired
+            | tenant::storage_layer::layer::DownloadError::NotFile(_)
+            | tenant::storage_layer::layer::DownloadError::DownloadFailed
+            | tenant::storage_layer::layer::DownloadError::PreStatFailed(_) => {
+                ApiError::InternalServerError(err.into())
+            }
+            #[cfg(test)]
+            tenant::storage_layer::layer::DownloadError::Failpoint(_) => {
+                ApiError::InternalServerError(err.into())
+            }
+        })?;
+
+    let keys = resident_layer
+        .load_keys(&ctx)
+        .await
+        .map_err(ApiError::InternalServerError)?;
+
+    let shard_identity = timeline.get_shard_identity();
+
+    let mut disposable_count = 0;
+    let mut not_disposable_count = 0;
+    let cancel = cancel.clone();
+    for (i, key) in keys.into_iter().enumerate() {
+        if shard_identity.is_key_disposable(&key) {
+            disposable_count += 1;
+            tracing::debug!(key = %key, key.dbg=?key, "disposable key");
+        } else {
+            not_disposable_count += 1;
+        }
+        #[allow(clippy::collapsible_if)]
+        if i % 10000 == 0 {
+            if cancel.is_cancelled() || timeline.cancel.is_cancelled() || timeline.is_stopping() {
+                return Err(ApiError::ShuttingDown);
+            }
+        }
+    }
+
+    json_response(
+        StatusCode::OK,
+        pageserver_api::models::ScanDisposableKeysResponse {
+            disposable_count,
+            not_disposable_count,
+        },
+    )
+}
+
 async fn layer_download_handler(
     request: Request<Body>,
     _cancel: CancellationToken,
@@ -1901,7 +2005,12 @@ async fn timeline_offload_handler(
         }
         offload_timeline(&tenant, &timeline)
             .await
-            .map_err(ApiError::InternalServerError)?;
+            .map_err(|e| {
+                match e {
+                    OffloadError::Cancelled => ApiError::ResourceUnavailable("Timeline shutting down".into()),
+                    _ => ApiError::InternalServerError(anyhow!(e))
+                }
+            })?;
 
         json_response(StatusCode::OK, ())
     }
@@ -1957,6 +2066,7 @@ async fn timeline_checkpoint_handler(
                 .map_err(|e|
                     match e {
                         CompactionError::ShuttingDown => ApiError::ShuttingDown,
+                        CompactionError::Offload(e) => ApiError::InternalServerError(anyhow::anyhow!(e)),
                         CompactionError::Other(e) => ApiError::InternalServerError(e)
                     }
                 )?;
@@ -2129,13 +2239,13 @@ async fn getpage_at_lsn_handler(
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
     let state = get_state(&request);
 
-    struct Key(crate::repository::Key);
+    struct Key(pageserver_api::key::Key);
 
     impl std::str::FromStr for Key {
         type Err = anyhow::Error;
 
         fn from_str(s: &str) -> std::result::Result<Self, Self::Err> {
-            crate::repository::Key::from_hex(s).map(Key)
+            pageserver_api::key::Key::from_hex(s).map(Key)
         }
     }
 
@@ -3145,6 +3255,10 @@ pub fn make_router(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/layer/:layer_file_name",
             |r| api_handler(r, evict_timeline_layer_handler),
         )
+        .post(
+            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/layer/:layer_name/scan_disposable_keys",
+            |r| testing_api_handler("timeline_layer_scan_disposable_keys", r, timeline_layer_scan_disposable_keys),
+        )
         .post(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/block_gc",
             |r| api_handler(r, timeline_gc_blocking_handler),

pageserver/src/import_datadir.rs

@@ -12,6 +12,7 @@ use pageserver_api::key::rel_block_to_key;
 use tokio::io::{AsyncRead, AsyncReadExt};
 use tokio_tar::Archive;
 use tracing::*;
+use wal_decoder::models::InterpretedWalRecord;
 use walkdir::WalkDir;
 
 use crate::context::RequestContext;
@@ -19,8 +20,6 @@ use crate::metrics::WAL_INGEST;
 use crate::pgdatadir_mapping::*;
 use crate::tenant::Timeline;
 use crate::walingest::WalIngest;
-use crate::walrecord::decode_wal_record;
-use crate::walrecord::DecodedWALRecord;
 use pageserver_api::reltag::{RelTag, SlruKind};
 use postgres_ffi::pg_constants;
 use postgres_ffi::relfile_utils::*;
@@ -313,11 +312,15 @@ async fn import_wal(
         let mut modification = tline.begin_modification(last_lsn);
         while last_lsn <= endpoint {
             if let Some((lsn, recdata)) = waldecoder.poll_decode()? {
-                let mut decoded = DecodedWALRecord::default();
-                decode_wal_record(recdata, &mut decoded, tline.pg_version)?;
+                let interpreted = InterpretedWalRecord::from_bytes_filtered(
+                    recdata,
+                    tline.get_shard_identity(),
+                    lsn,
+                    tline.pg_version,
+                )?;
 
                 walingest
-                    .ingest_record(decoded, lsn, &mut modification, ctx)
+                    .ingest_record(interpreted, &mut modification, ctx)
                     .await?;
                 WAL_INGEST.records_committed.inc();
 
@@ -454,10 +457,15 @@ pub async fn import_wal_from_tar(
         let mut modification = tline.begin_modification(last_lsn);
         while last_lsn <= end_lsn {
             if let Some((lsn, recdata)) = waldecoder.poll_decode()? {
-                let mut decoded = DecodedWALRecord::default();
-                decode_wal_record(recdata, &mut decoded, tline.pg_version)?;
+                let interpreted = InterpretedWalRecord::from_bytes_filtered(
+                    recdata,
+                    tline.get_shard_identity(),
+                    lsn,
+                    tline.pg_version,
+                )?;
+
                 walingest
-                    .ingest_record(decoded, lsn, &mut modification, ctx)
+                    .ingest_record(interpreted, &mut modification, ctx)
                     .await?;
                 modification.commit(ctx).await?;
                 last_lsn = lsn;

pageserver/src/lib.rs

@@ -24,7 +24,6 @@ pub mod metrics;
 pub mod page_cache;
 pub mod page_service;
 pub mod pgdatadir_mapping;
-pub mod repository;
 pub mod span;
 pub(crate) mod statvfs;
 pub mod task_mgr;
@@ -32,7 +31,6 @@ pub mod tenant;
 pub mod utilization;
 pub mod virtual_file;
 pub mod walingest;
-pub mod walrecord;
 pub mod walredo;
 
 use camino::Utf8Path;

pageserver/src/metrics.rs

@@ -3040,13 +3040,111 @@ impl<F: Future<Output = Result<O, E>>, O, E> Future for MeasuredRemoteOp<F> {
 }
 
 pub mod tokio_epoll_uring {
-    use metrics::{register_int_counter, UIntGauge};
+    use std::{
+        collections::HashMap,
+        sync::{Arc, Mutex},
+    };
+
+    use metrics::{register_histogram, register_int_counter, Histogram, LocalHistogram, UIntGauge};
     use once_cell::sync::Lazy;
 
+    /// Shared storage for tokio-epoll-uring thread local metrics.
+    pub(crate) static THREAD_LOCAL_METRICS_STORAGE: Lazy<ThreadLocalMetricsStorage> =
+        Lazy::new(|| {
+            let slots_submission_queue_depth = register_histogram!(
+                "pageserver_tokio_epoll_uring_slots_submission_queue_depth",
+                "The slots waiters queue depth of each tokio_epoll_uring system",
+                vec![1.0, 2.0, 4.0, 8.0, 16.0, 32.0, 64.0, 128.0, 256.0, 512.0, 1024.0],
+            )
+            .expect("failed to define a metric");
+            ThreadLocalMetricsStorage {
+                observers: Mutex::new(HashMap::new()),
+                slots_submission_queue_depth,
+            }
+        });
+
+    pub struct ThreadLocalMetricsStorage {
+        /// List of thread local metrics observers.
+        observers: Mutex<HashMap<u64, Arc<ThreadLocalMetrics>>>,
+        /// A histogram shared between all thread local systems
+        /// for collecting slots submission queue depth.
+        slots_submission_queue_depth: Histogram,
+    }
+
+    /// Each thread-local [`tokio_epoll_uring::System`] gets one of these as its
+    /// [`tokio_epoll_uring::metrics::PerSystemMetrics`] generic.
+    ///
+    /// The System makes observations into [`Self`] and periodically, the collector
+    /// comes along and flushes [`Self`] into the shared storage [`THREAD_LOCAL_METRICS_STORAGE`].
+    ///
+    /// [`LocalHistogram`] is `!Send`, so, we need to put it behind a [`Mutex`].
+    /// But except for the periodic flush, the lock is uncontended so there's no waiting
+    /// for cache coherence protocol to get an exclusive cache line.
+    pub struct ThreadLocalMetrics {
+        /// Local observer of thread local tokio-epoll-uring system's slots waiters queue depth.
+        slots_submission_queue_depth: Mutex<LocalHistogram>,
+    }
+
+    impl ThreadLocalMetricsStorage {
+        /// Registers a new thread local system. Returns a thread local metrics observer.
+        pub fn register_system(&self, id: u64) -> Arc<ThreadLocalMetrics> {
+            let per_system_metrics = Arc::new(ThreadLocalMetrics::new(
+                self.slots_submission_queue_depth.local(),
+            ));
+            let mut g = self.observers.lock().unwrap();
+            g.insert(id, Arc::clone(&per_system_metrics));
+            per_system_metrics
+        }
+
+        /// Removes metrics observer for a thread local system.
+        /// This should be called before dropping a thread local system.
+        pub fn remove_system(&self, id: u64) {
+            let mut g = self.observers.lock().unwrap();
+            g.remove(&id);
+        }
+
+        /// Flush all thread local metrics to the shared storage.
+        pub fn flush_thread_local_metrics(&self) {
+            let g = self.observers.lock().unwrap();
+            g.values().for_each(|local| {
+                local.flush();
+            });
+        }
+    }
+
+    impl ThreadLocalMetrics {
+        pub fn new(slots_submission_queue_depth: LocalHistogram) -> Self {
+            ThreadLocalMetrics {
+                slots_submission_queue_depth: Mutex::new(slots_submission_queue_depth),
+            }
+        }
+
+        /// Flushes the thread local metrics to shared aggregator.
+        pub fn flush(&self) {
+            let Self {
+                slots_submission_queue_depth,
+            } = self;
+            slots_submission_queue_depth.lock().unwrap().flush();
+        }
+    }
+
+    impl tokio_epoll_uring::metrics::PerSystemMetrics for ThreadLocalMetrics {
+        fn observe_slots_submission_queue_depth(&self, queue_depth: u64) {
+            let Self {
+                slots_submission_queue_depth,
+            } = self;
+            slots_submission_queue_depth
+                .lock()
+                .unwrap()
+                .observe(queue_depth as f64);
+        }
+    }
+
     pub struct Collector {
         descs: Vec<metrics::core::Desc>,
         systems_created: UIntGauge,
         systems_destroyed: UIntGauge,
+        thread_local_metrics_storage: &'static ThreadLocalMetricsStorage,
     }
 
     impl metrics::core::Collector for Collector {
@@ -3056,7 +3154,7 @@ pub mod tokio_epoll_uring {
 
         fn collect(&self) -> Vec<metrics::proto::MetricFamily> {
             let mut mfs = Vec::with_capacity(Self::NMETRICS);
-            let tokio_epoll_uring::metrics::Metrics {
+            let tokio_epoll_uring::metrics::GlobalMetrics {
                 systems_created,
                 systems_destroyed,
             } = tokio_epoll_uring::metrics::global();
@@ -3064,12 +3162,21 @@ pub mod tokio_epoll_uring {
             mfs.extend(self.systems_created.collect());
             self.systems_destroyed.set(systems_destroyed);
             mfs.extend(self.systems_destroyed.collect());
+
+            self.thread_local_metrics_storage
+                .flush_thread_local_metrics();
+
+            mfs.extend(
+                self.thread_local_metrics_storage
+                    .slots_submission_queue_depth
+                    .collect(),
+            );
             mfs
         }
     }
 
     impl Collector {
-        const NMETRICS: usize = 2;
+        const NMETRICS: usize = 3;
 
         #[allow(clippy::new_without_default)]
         pub fn new() -> Self {
@@ -3101,6 +3208,7 @@ pub mod tokio_epoll_uring {
                 descs,
                 systems_created,
                 systems_destroyed,
+                thread_local_metrics_storage: &THREAD_LOCAL_METRICS_STORAGE,
             }
         }
     }
@@ -3460,6 +3568,7 @@ pub fn preinitialize_metrics() {
     Lazy::force(&RECONSTRUCT_TIME);
     Lazy::force(&BASEBACKUP_QUERY_TIME);
     Lazy::force(&COMPUTE_COMMANDS_COUNTERS);
+    Lazy::force(&tokio_epoll_uring::THREAD_LOCAL_METRICS_STORAGE);
 
     tenant_throttling::preinitialize_global_metrics();
 }

pageserver/src/page_service.rs

@@ -1080,6 +1080,7 @@ impl PageServerHandler {
         prev_lsn: Option<Lsn>,
         full_backup: bool,
         gzip: bool,
+        replica: bool,
         ctx: &RequestContext,
     ) -> Result<(), QueryError>
     where
@@ -1132,6 +1133,7 @@ impl PageServerHandler {
                 lsn,
                 prev_lsn,
                 full_backup,
+                replica,
                 ctx,
             )
             .await
@@ -1154,6 +1156,7 @@ impl PageServerHandler {
                     lsn,
                     prev_lsn,
                     full_backup,
+                    replica,
                     ctx,
                 )
                 .await
@@ -1170,6 +1173,7 @@ impl PageServerHandler {
                     lsn,
                     prev_lsn,
                     full_backup,
+                    replica,
                     ctx,
                 )
                 .await
@@ -1326,24 +1330,27 @@ where
                 .for_command(ComputeCommandKind::Basebackup)
                 .inc();
 
-            let (lsn, gzip) = match (params.get(2), params.get(3)) {
-                (None, _) => (None, false),
-                (Some(&"--gzip"), _) => (None, true),
-                (Some(lsn_str), gzip_str_opt) => {
-                    let lsn = Lsn::from_str(lsn_str)
-                        .with_context(|| format!("Failed to parse Lsn from {lsn_str}"))?;
-                    let gzip = match gzip_str_opt {
-                        Some(&"--gzip") => true,
-                        None => false,
-                        Some(third_param) => {
+            let mut lsn = None;
+            let mut replica = false;
+            let mut gzip = false;
+            for param in &params[2..] {
+                if param.starts_with("--") {
+                    match *param {
+                        "--gzip" => gzip = true,
+                        "--replica" => replica = true,
+                        _ => {
                             return Err(QueryError::Other(anyhow::anyhow!(
-                                "Parameter in position 3 unknown {third_param}",
+                                "Unknown parameter {param}",
                             )))
                         }
-                    };
-                    (Some(lsn), gzip)
+                    }
+                } else {
+                    lsn = Some(
+                        Lsn::from_str(param)
+                            .with_context(|| format!("Failed to parse Lsn from {param}"))?,
+                    );
                 }
-            };
+            }
 
             let metric_recording = metrics::BASEBACKUP_QUERY_TIME.start_recording(&ctx);
             let res = async {
@@ -1355,6 +1362,7 @@ where
                     None,
                     false,
                     gzip,
+                    replica,
                     &ctx,
                 )
                 .await?;
@@ -1415,6 +1423,7 @@ where
                 prev_lsn,
                 true,
                 false,
+                false,
                 &ctx,
             )
             .await?;

pageserver/src/pgdatadir_mapping.rs

@@ -7,14 +7,14 @@
 //! Clarify that)
 //!
 use super::tenant::{PageReconstructError, Timeline};
+use crate::aux_file;
 use crate::context::RequestContext;
 use crate::keyspace::{KeySpace, KeySpaceAccum};
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id;
-use crate::walrecord::NeonWalRecord;
-use crate::{aux_file, repository::*};
 use anyhow::{ensure, Context};
 use bytes::{Buf, Bytes, BytesMut};
 use enum_map::Enum;
+use pageserver_api::key::Key;
 use pageserver_api::key::{
     dbdir_key_range, rel_block_to_key, rel_dir_to_key, rel_key_range, rel_size_to_key,
     relmap_file_key, repl_origin_key, repl_origin_key_range, slru_block_to_key, slru_dir_to_key,
@@ -22,7 +22,9 @@ use pageserver_api::key::{
     CompactKey, AUX_FILES_KEY, CHECKPOINT_KEY, CONTROLFILE_KEY, DBDIR_KEY, TWOPHASEDIR_KEY,
 };
 use pageserver_api::keyspace::SparseKeySpace;
+use pageserver_api::record::NeonWalRecord;
 use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
+use pageserver_api::value::Value;
 use postgres_ffi::relfile_utils::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
 use postgres_ffi::BLCKSZ;
 use postgres_ffi::{Oid, RepOriginId, TimestampTz, TransactionId};
@@ -1506,35 +1508,42 @@ impl<'a> DatadirModification<'a> {
         Ok(())
     }
 
-    /// Drop a relation.
-    pub async fn put_rel_drop(&mut self, rel: RelTag, ctx: &RequestContext) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
+    /// Drop some relations
+    pub(crate) async fn put_rel_drops(
+        &mut self,
+        drop_relations: HashMap<(u32, u32), Vec<RelTag>>,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        for ((spc_node, db_node), rel_tags) in drop_relations {
+            let dir_key = rel_dir_to_key(spc_node, db_node);
+            let buf = self.get(dir_key, ctx).await?;
+            let mut dir = RelDirectory::des(&buf)?;
 
-        // Remove it from the directory entry
-        let dir_key = rel_dir_to_key(rel.spcnode, rel.dbnode);
-        let buf = self.get(dir_key, ctx).await?;
-        let mut dir = RelDirectory::des(&buf)?;
+            let mut dirty = false;
+            for rel_tag in rel_tags {
+                if dir.rels.remove(&(rel_tag.relnode, rel_tag.forknum)) {
+                    dirty = true;
 
-        self.pending_directory_entries
-            .push((DirectoryKind::Rel, dir.rels.len()));
+                    // update logical size
+                    let size_key = rel_size_to_key(rel_tag);
+                    let old_size = self.get(size_key, ctx).await?.get_u32_le();
+                    self.pending_nblocks -= old_size as i64;
 
-        if dir.rels.remove(&(rel.relnode, rel.forknum)) {
-            self.put(dir_key, Value::Image(Bytes::from(RelDirectory::ser(&dir)?)));
-        } else {
-            warn!("dropped rel {} did not exist in rel directory", rel);
+                    // Remove entry from relation size cache
+                    self.tline.remove_cached_rel_size(&rel_tag);
+
+                    // Delete size entry, as well as all blocks
+                    self.delete(rel_key_range(rel_tag));
+                }
+            }
+
+            if dirty {
+                self.put(dir_key, Value::Image(Bytes::from(RelDirectory::ser(&dir)?)));
+                self.pending_directory_entries
+                    .push((DirectoryKind::Rel, dir.rels.len()));
+            }
         }
 
-        // update logical size
-        let size_key = rel_size_to_key(rel);
-        let old_size = self.get(size_key, ctx).await?.get_u32_le();
-        self.pending_nblocks -= old_size as i64;
-
-        // Remove enty from relation size cache
-        self.tline.remove_cached_rel_size(&rel);
-
-        // Delete size entry, as well as all blocks
-        self.delete(rel_key_range(rel));
-
         Ok(())
     }
 

pageserver/src/tenant/config.rs

@@ -9,7 +9,6 @@
 //! may lead to a data loss.
 //!
 pub(crate) use pageserver_api::config::TenantConfigToml as TenantConf;
-use pageserver_api::models::AuxFilePolicy;
 use pageserver_api::models::CompactionAlgorithmSettings;
 use pageserver_api::models::EvictionPolicy;
 use pageserver_api::models::{self, ThrottleConfig};
@@ -341,10 +340,6 @@ pub struct TenantConfOpt {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub image_layer_creation_check_threshold: Option<u8>,
 
-    #[serde(skip_serializing_if = "Option::is_none")]
-    #[serde(default)]
-    pub switch_aux_file_policy: Option<AuxFilePolicy>,
-
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(with = "humantime_serde")]
     #[serde(default)]
@@ -410,9 +405,6 @@ impl TenantConfOpt {
             image_layer_creation_check_threshold: self
                 .image_layer_creation_check_threshold
                 .unwrap_or(global_conf.image_layer_creation_check_threshold),
-            switch_aux_file_policy: self
-                .switch_aux_file_policy
-                .unwrap_or(global_conf.switch_aux_file_policy),
             lsn_lease_length: self
                 .lsn_lease_length
                 .unwrap_or(global_conf.lsn_lease_length),
@@ -470,7 +462,6 @@ impl From<TenantConfOpt> for models::TenantConfig {
             lazy_slru_download: value.lazy_slru_download,
             timeline_get_throttle: value.timeline_get_throttle.map(ThrottleConfig::from),
             image_layer_creation_check_threshold: value.image_layer_creation_check_threshold,
-            switch_aux_file_policy: value.switch_aux_file_policy,
             lsn_lease_length: value.lsn_lease_length.map(humantime),
             lsn_lease_length_for_ts: value.lsn_lease_length_for_ts.map(humantime),
         }

pageserver/src/tenant/gc_result.rs

@@ -0,0 +1,57 @@
+use anyhow::Result;
+use serde::Serialize;
+use std::ops::AddAssign;
+use std::time::Duration;
+
+///
+/// Result of performing GC
+///
+#[derive(Default, Serialize, Debug)]
+pub struct GcResult {
+    pub layers_total: u64,
+    pub layers_needed_by_cutoff: u64,
+    pub layers_needed_by_pitr: u64,
+    pub layers_needed_by_branches: u64,
+    pub layers_needed_by_leases: u64,
+    pub layers_not_updated: u64,
+    pub layers_removed: u64, // # of layer files removed because they have been made obsolete by newer ondisk files.
+
+    #[serde(serialize_with = "serialize_duration_as_millis")]
+    pub elapsed: Duration,
+
+    /// The layers which were garbage collected.
+    ///
+    /// Used in `/v1/tenant/:tenant_id/timeline/:timeline_id/do_gc` to wait for the layers to be
+    /// dropped in tests.
+    #[cfg(feature = "testing")]
+    #[serde(skip)]
+    pub(crate) doomed_layers: Vec<crate::tenant::storage_layer::Layer>,
+}
+
+// helper function for `GcResult`, serializing a `Duration` as an integer number of milliseconds
+fn serialize_duration_as_millis<S>(d: &Duration, serializer: S) -> Result<S::Ok, S::Error>
+where
+    S: serde::Serializer,
+{
+    d.as_millis().serialize(serializer)
+}
+
+impl AddAssign for GcResult {
+    fn add_assign(&mut self, other: Self) {
+        self.layers_total += other.layers_total;
+        self.layers_needed_by_pitr += other.layers_needed_by_pitr;
+        self.layers_needed_by_cutoff += other.layers_needed_by_cutoff;
+        self.layers_needed_by_branches += other.layers_needed_by_branches;
+        self.layers_needed_by_leases += other.layers_needed_by_leases;
+        self.layers_not_updated += other.layers_not_updated;
+        self.layers_removed += other.layers_removed;
+
+        self.elapsed += other.elapsed;
+
+        #[cfg(feature = "testing")]
+        {
+            let mut other = other;
+            self.doomed_layers.append(&mut other.doomed_layers);
+        }
+    }
+}

pageserver/src/tenant/layer_map.rs

@@ -48,9 +48,9 @@ mod layer_coverage;
 
 use crate::context::RequestContext;
 use crate::keyspace::KeyPartitioning;
-use crate::repository::Key;
 use crate::tenant::storage_layer::InMemoryLayer;
 use anyhow::Result;
+use pageserver_api::key::Key;
 use pageserver_api::keyspace::{KeySpace, KeySpaceAccum};
 use range_set_blaze::{CheckSortedDisjoint, RangeSetBlaze};
 use std::collections::{HashMap, VecDeque};

pageserver/src/tenant/mgr.rs

@@ -2811,7 +2811,7 @@ where
 }
 
 use {
-    crate::repository::GcResult, pageserver_api::models::TimelineGcRequest,
+    crate::tenant::gc_result::GcResult, pageserver_api::models::TimelineGcRequest,
     utils::http::error::ApiError,
 };
 

pageserver/src/tenant/remote_timeline_client.rs

@@ -190,6 +190,7 @@ use chrono::{NaiveDateTime, Utc};
 pub(crate) use download::download_initdb_tar_zst;
 use pageserver_api::models::TimelineArchivalState;
 use pageserver_api::shard::{ShardIndex, TenantShardId};
+use regex::Regex;
 use scopeguard::ScopeGuard;
 use tokio_util::sync::CancellationToken;
 use utils::backoff::{
@@ -199,7 +200,7 @@ use utils::pausable_failpoint;
 
 use std::collections::{HashMap, VecDeque};
 use std::sync::atomic::{AtomicU32, Ordering};
-use std::sync::{Arc, Mutex};
+use std::sync::{Arc, Mutex, OnceLock};
 use std::time::Duration;
 
 use remote_storage::{
@@ -245,11 +246,11 @@ use super::upload_queue::{NotInitialized, SetDeletedFlagProgress};
 use super::Generation;
 
 pub(crate) use download::{
-    do_download_tenant_manifest, download_index_part, is_temp_download_file,
+    download_index_part, download_tenant_manifest, is_temp_download_file,
     list_remote_tenant_shards, list_remote_timelines,
 };
 pub(crate) use index::LayerFileMetadata;
-pub(crate) use upload::{upload_initdb_dir, upload_tenant_manifest};
+pub(crate) use upload::upload_initdb_dir;
 
 // Occasional network issues and such can cause remote operations to fail, and
 // that's expected. If a download fails, we log it at info-level, and retry.
@@ -274,12 +275,6 @@ pub(crate) const BUFFER_SIZE: usize = 32 * 1024;
 /// which we warn and skip.
 const DELETION_QUEUE_FLUSH_TIMEOUT: Duration = Duration::from_secs(10);
 
-/// Hardcode a generation for the tenant manifest for now so that we don't
-/// need to deal with generation-less manifests in the future.
-///
-/// TODO: add proper generation support to all the places that use this.
-pub(crate) const TENANT_MANIFEST_GENERATION: Generation = Generation::new(1);
-
 pub enum MaybeDeletedIndexPart {
     IndexPart(IndexPart),
     Deleted(IndexPart),
@@ -1278,10 +1273,14 @@ impl RemoteTimelineClient {
         let fut = {
             let mut guard = self.upload_queue.lock().unwrap();
             let upload_queue = match &mut *guard {
-                UploadQueue::Stopped(_) => return,
+                UploadQueue::Stopped(_) => {
+                    scopeguard::ScopeGuard::into_inner(sg);
+                    return;
+                }
                 UploadQueue::Uninitialized => {
                     // transition into Stopped state
                     self.stop_impl(&mut guard);
+                    scopeguard::ScopeGuard::into_inner(sg);
                     return;
                 }
                 UploadQueue::Initialized(ref mut init) => init,
@@ -2235,6 +2234,12 @@ pub fn remote_tenant_manifest_path(
     RemotePath::from_string(&path).expect("Failed to construct path")
 }
 
+/// Prefix to all generations' manifest objects in a tenant shard
+pub fn remote_tenant_manifest_prefix(tenant_shard_id: &TenantShardId) -> RemotePath {
+    let path = format!("tenants/{tenant_shard_id}/tenant-manifest",);
+    RemotePath::from_string(&path).expect("Failed to construct path")
+}
+
 pub fn remote_timelines_path(tenant_shard_id: &TenantShardId) -> RemotePath {
     let path = format!("tenants/{tenant_shard_id}/{TIMELINES_SEGMENT_NAME}");
     RemotePath::from_string(&path).expect("Failed to construct path")
@@ -2329,6 +2334,15 @@ pub fn parse_remote_index_path(path: RemotePath) -> Option<Generation> {
     }
 }
 
+/// Given the key of a tenant manifest, parse out the generation number
+pub(crate) fn parse_remote_tenant_manifest_path(path: RemotePath) -> Option<Generation> {
+    static RE: OnceLock<Regex> = OnceLock::new();
+    let re = RE.get_or_init(|| Regex::new(r".+tenant-manifest-([0-9a-f]{8}).json").unwrap());
+    re.captures(path.get_path().as_str())
+        .and_then(|c| c.get(1))
+        .and_then(|m| Generation::parse_suffix(m.as_str()))
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -20,7 +20,9 @@ use utils::backoff;
 
 use crate::config::PageServerConf;
 use crate::context::RequestContext;
-use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
+use crate::span::{
+    debug_assert_current_span_has_tenant_and_timeline_id, debug_assert_current_span_has_tenant_id,
+};
 use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerName;
 use crate::tenant::Generation;
@@ -36,9 +38,10 @@ use utils::pausable_failpoint;
 use super::index::{IndexPart, LayerFileMetadata};
 use super::manifest::TenantManifest;
 use super::{
-    parse_remote_index_path, remote_index_path, remote_initdb_archive_path,
-    remote_initdb_preserved_archive_path, remote_tenant_manifest_path, remote_tenant_path,
-    FAILED_DOWNLOAD_WARN_THRESHOLD, FAILED_REMOTE_OP_RETRIES, INITDB_PATH,
+    parse_remote_index_path, parse_remote_tenant_manifest_path, remote_index_path,
+    remote_initdb_archive_path, remote_initdb_preserved_archive_path, remote_tenant_manifest_path,
+    remote_tenant_manifest_prefix, remote_tenant_path, FAILED_DOWNLOAD_WARN_THRESHOLD,
+    FAILED_REMOTE_OP_RETRIES, INITDB_PATH,
 };
 
 ///
@@ -365,32 +368,34 @@ async fn do_download_remote_path_retry_forever(
     .await
 }
 
-pub async fn do_download_tenant_manifest(
+async fn do_download_tenant_manifest(
     storage: &GenericRemoteStorage,
     tenant_shard_id: &TenantShardId,
+    _timeline_id: Option<&TimelineId>,
+    generation: Generation,
     cancel: &CancellationToken,
-) -> Result<(TenantManifest, Generation), DownloadError> {
-    // TODO: generation support
-    let generation = super::TENANT_MANIFEST_GENERATION;
+) -> Result<(TenantManifest, Generation, SystemTime), DownloadError> {
     let remote_path = remote_tenant_manifest_path(tenant_shard_id, generation);
 
-    let (manifest_bytes, _manifest_bytes_mtime) =
+    let (manifest_bytes, manifest_bytes_mtime) =
         do_download_remote_path_retry_forever(storage, &remote_path, cancel).await?;
 
     let tenant_manifest = TenantManifest::from_json_bytes(&manifest_bytes)
         .with_context(|| format!("deserialize tenant manifest file at {remote_path:?}"))
         .map_err(DownloadError::Other)?;
 
-    Ok((tenant_manifest, generation))
+    Ok((tenant_manifest, generation, manifest_bytes_mtime))
 }
 
 async fn do_download_index_part(
     storage: &GenericRemoteStorage,
     tenant_shard_id: &TenantShardId,
-    timeline_id: &TimelineId,
+    timeline_id: Option<&TimelineId>,
     index_generation: Generation,
     cancel: &CancellationToken,
 ) -> Result<(IndexPart, Generation, SystemTime), DownloadError> {
+    let timeline_id =
+        timeline_id.expect("A timeline ID is always provided when downloading an index");
     let remote_path = remote_index_path(tenant_shard_id, timeline_id, index_generation);
 
     let (index_part_bytes, index_part_mtime) =
@@ -403,59 +408,79 @@ async fn do_download_index_part(
     Ok((index_part, index_generation, index_part_mtime))
 }
 
-/// index_part.json objects are suffixed with a generation number, so we cannot
-/// directly GET the latest index part without doing some probing.
+/// Metadata objects are "generationed", meaning that they include a generation suffix.  This
+/// function downloads the object with the highest generation <= `my_generation`.
 ///
-/// In this function we probe for the most recent index in a generation <= our current generation.
-/// See "Finding the remote indices for timelines" in docs/rfcs/025-generation-numbers.md
+/// Data objects (layer files) also include a generation in their path, but there is no equivalent
+/// search process, because their reference from an index includes the generation.
+///
+/// An expensive object listing operation is only done if necessary: the typical fast path is to issue two
+/// GET operations, one to our own generation (stale attachment case), and one to the immediately preceding
+/// generation (normal case when migrating/restarting).  Only if both of these return 404 do we fall back
+/// to listing objects.
+///
+/// * `my_generation`: the value of `[crate::tenant::Tenant::generation]`
+/// * `what`: for logging, what object are we downloading
+/// * `prefix`: when listing objects, use this prefix (i.e. the part of the object path before the generation)
+/// * `do_download`: a GET of the object in a particular generation, which should **retry indefinitely** unless
+///                  `cancel`` has fired.  This function does not do its own retries of GET operations, and relies
+///                  on the function passed in to do so.
+/// * `parse_path`: parse a fully qualified remote storage path to get the generation of the object.
+#[allow(clippy::too_many_arguments)]
 #[tracing::instrument(skip_all, fields(generation=?my_generation))]
-pub(crate) async fn download_index_part(
-    storage: &GenericRemoteStorage,
-    tenant_shard_id: &TenantShardId,
-    timeline_id: &TimelineId,
+pub(crate) async fn download_generation_object<'a, T, DF, DFF, PF>(
+    storage: &'a GenericRemoteStorage,
+    tenant_shard_id: &'a TenantShardId,
+    timeline_id: Option<&'a TimelineId>,
     my_generation: Generation,
-    cancel: &CancellationToken,
-) -> Result<(IndexPart, Generation, SystemTime), DownloadError> {
-    debug_assert_current_span_has_tenant_and_timeline_id();
+    what: &str,
+    prefix: RemotePath,
+    do_download: DF,
+    parse_path: PF,
+    cancel: &'a CancellationToken,
+) -> Result<(T, Generation, SystemTime), DownloadError>
+where
+    DF: Fn(
+        &'a GenericRemoteStorage,
+        &'a TenantShardId,
+        Option<&'a TimelineId>,
+        Generation,
+        &'a CancellationToken,
+    ) -> DFF,
+    DFF: Future<Output = Result<(T, Generation, SystemTime), DownloadError>>,
+    PF: Fn(RemotePath) -> Option<Generation>,
+    T: 'static,
+{
+    debug_assert_current_span_has_tenant_id();
 
     if my_generation.is_none() {
         // Operating without generations: just fetch the generation-less path
-        return do_download_index_part(
-            storage,
-            tenant_shard_id,
-            timeline_id,
-            my_generation,
-            cancel,
-        )
-        .await;
+        return do_download(storage, tenant_shard_id, timeline_id, my_generation, cancel).await;
     }
 
-    // Stale case: If we were intentionally attached in a stale generation, there may already be a remote
-    // index in our generation.
+    // Stale case: If we were intentionally attached in a stale generation, the remote object may already
+    // exist in our generation.
     //
     // This is an optimization to avoid doing the listing for the general case below.
-    let res =
-        do_download_index_part(storage, tenant_shard_id, timeline_id, my_generation, cancel).await;
+    let res = do_download(storage, tenant_shard_id, timeline_id, my_generation, cancel).await;
     match res {
-        Ok(index_part) => {
-            tracing::debug!(
-                "Found index_part from current generation (this is a stale attachment)"
-            );
-            return Ok(index_part);
+        Ok(decoded) => {
+            tracing::debug!("Found {what} from current generation (this is a stale attachment)");
+            return Ok(decoded);
         }
         Err(DownloadError::NotFound) => {}
         Err(e) => return Err(e),
     };
 
-    // Typical case: the previous generation of this tenant was running healthily, and had uploaded
-    // and index part.  We may safely start from this index without doing a listing, because:
+    // Typical case: the previous generation of this tenant was running healthily, and had uploaded the object
+    // we are seeking in that generation.  We may safely start from this index without doing a listing, because:
     //  - We checked for current generation case above
     //  - generations > my_generation are to be ignored
-    //  - any other indices that exist would have an older generation than `previous_gen`, and
-    //    we want to find the most recent index from a previous generation.
+    //  - any other objects that exist would have an older generation than `previous_gen`, and
+    //    we want to find the most recent object from a previous generation.
     //
     // This is an optimization to avoid doing the listing for the general case below.
-    let res = do_download_index_part(
+    let res = do_download(
         storage,
         tenant_shard_id,
         timeline_id,
@@ -464,14 +489,12 @@ pub(crate) async fn download_index_part(
     )
     .await;
     match res {
-        Ok(index_part) => {
-            tracing::debug!("Found index_part from previous generation");
-            return Ok(index_part);
+        Ok(decoded) => {
+            tracing::debug!("Found {what} from previous generation");
+            return Ok(decoded);
         }
         Err(DownloadError::NotFound) => {
-            tracing::debug!(
-                "No index_part found from previous generation, falling back to listing"
-            );
+            tracing::debug!("No {what} found from previous generation, falling back to listing");
         }
         Err(e) => {
             return Err(e);
@@ -481,12 +504,10 @@ pub(crate) async fn download_index_part(
     // General case/fallback: if there is no index at my_generation or prev_generation, then list all index_part.json
     // objects, and select the highest one with a generation <= my_generation.  Constructing the prefix is equivalent
     // to constructing a full index path with no generation, because the generation is a suffix.
-    let index_prefix = remote_index_path(tenant_shard_id, timeline_id, Generation::none());
-
-    let indices = download_retry(
+    let paths = download_retry(
         || async {
             storage
-                .list(Some(&index_prefix), ListingMode::NoDelimiter, None, cancel)
+                .list(Some(&prefix), ListingMode::NoDelimiter, None, cancel)
                 .await
         },
         "list index_part files",
@@ -497,22 +518,22 @@ pub(crate) async fn download_index_part(
 
     // General case logic for which index to use: the latest index whose generation
     // is <= our own.  See "Finding the remote indices for timelines" in docs/rfcs/025-generation-numbers.md
-    let max_previous_generation = indices
+    let max_previous_generation = paths
         .into_iter()
-        .filter_map(|o| parse_remote_index_path(o.key))
+        .filter_map(|o| parse_path(o.key))
         .filter(|g| g <= &my_generation)
         .max();
 
     match max_previous_generation {
         Some(g) => {
-            tracing::debug!("Found index_part in generation {g:?}");
-            do_download_index_part(storage, tenant_shard_id, timeline_id, g, cancel).await
+            tracing::debug!("Found {what} in generation {g:?}");
+            do_download(storage, tenant_shard_id, timeline_id, g, cancel).await
         }
         None => {
             // Migration from legacy pre-generation state: we have a generation but no prior
             // attached pageservers did.  Try to load from a no-generation path.
-            tracing::debug!("No index_part.json* found");
-            do_download_index_part(
+            tracing::debug!("No {what}* found");
+            do_download(
                 storage,
                 tenant_shard_id,
                 timeline_id,
@@ -524,6 +545,57 @@ pub(crate) async fn download_index_part(
     }
 }
 
+/// index_part.json objects are suffixed with a generation number, so we cannot
+/// directly GET the latest index part without doing some probing.
+///
+/// In this function we probe for the most recent index in a generation <= our current generation.
+/// See "Finding the remote indices for timelines" in docs/rfcs/025-generation-numbers.md
+pub(crate) async fn download_index_part(
+    storage: &GenericRemoteStorage,
+    tenant_shard_id: &TenantShardId,
+    timeline_id: &TimelineId,
+    my_generation: Generation,
+    cancel: &CancellationToken,
+) -> Result<(IndexPart, Generation, SystemTime), DownloadError> {
+    debug_assert_current_span_has_tenant_and_timeline_id();
+
+    let index_prefix = remote_index_path(tenant_shard_id, timeline_id, Generation::none());
+    download_generation_object(
+        storage,
+        tenant_shard_id,
+        Some(timeline_id),
+        my_generation,
+        "index_part",
+        index_prefix,
+        do_download_index_part,
+        parse_remote_index_path,
+        cancel,
+    )
+    .await
+}
+
+pub(crate) async fn download_tenant_manifest(
+    storage: &GenericRemoteStorage,
+    tenant_shard_id: &TenantShardId,
+    my_generation: Generation,
+    cancel: &CancellationToken,
+) -> Result<(TenantManifest, Generation, SystemTime), DownloadError> {
+    let manifest_prefix = remote_tenant_manifest_prefix(tenant_shard_id);
+
+    download_generation_object(
+        storage,
+        tenant_shard_id,
+        None,
+        my_generation,
+        "tenant-manifest",
+        manifest_prefix,
+        do_download_tenant_manifest,
+        parse_remote_tenant_manifest_path,
+        cancel,
+    )
+    .await
+}
+
 pub(crate) async fn download_initdb_tar_zst(
     conf: &'static PageServerConf,
     storage: &GenericRemoteStorage,

pageserver/src/tenant/remote_timeline_client/manifest.rs

@@ -3,7 +3,7 @@ use serde::{Deserialize, Serialize};
 use utils::{id::TimelineId, lsn::Lsn};
 
 /// Tenant-shard scoped manifest
-#[derive(Clone, Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize, PartialEq, Eq)]
 pub struct TenantManifest {
     /// Debugging aid describing the version of this manifest.
     /// Can also be used for distinguishing breaking changes later on.
@@ -23,7 +23,7 @@ pub struct TenantManifest {
 /// Very similar to [`pageserver_api::models::OffloadedTimelineInfo`],
 /// but the two datastructures serve different needs, this is for a persistent disk format
 /// that must be backwards compatible, while the other is only for informative purposes.
-#[derive(Clone, Serialize, Deserialize, Copy)]
+#[derive(Clone, Serialize, Deserialize, Copy, PartialEq, Eq)]
 pub struct OffloadedTimelineManifest {
     pub timeline_id: TimelineId,
     /// Whether the timeline has a parent it has been branched off from or not

pageserver/src/tenant/size.rs

@@ -187,6 +187,8 @@ pub(super) async fn gather_inputs(
     // but it is unlikely to cause any issues. In the worst case,
     // the calculation will error out.
     timelines.retain(|t| t.is_active());
+    // Also filter out archived timelines.
+    timelines.retain(|t| t.is_archived() != Some(true));
 
     // Build a map of branch points.
     let mut branchpoints: HashMap<TimelineId, HashSet<Lsn>> = HashMap::new();

pageserver/src/tenant/storage_layer.rs

@@ -1,5 +1,6 @@
 //! Common traits and structs for layers
 
+pub mod batch_split_writer;
 pub mod delta_layer;
 pub mod filter_iterator;
 pub mod image_layer;
@@ -8,14 +9,13 @@ pub(crate) mod layer;
 mod layer_desc;
 mod layer_name;
 pub mod merge_iterator;
-pub mod split_writer;
 
 use crate::context::{AccessStatsBehavior, RequestContext};
-use crate::repository::Value;
-use crate::walrecord::NeonWalRecord;
 use bytes::Bytes;
 use pageserver_api::key::Key;
 use pageserver_api::keyspace::{KeySpace, KeySpaceRandomAccum};
+use pageserver_api::record::NeonWalRecord;
+use pageserver_api::value::Value;
 use std::cmp::{Ordering, Reverse};
 use std::collections::hash_map::Entry;
 use std::collections::{BinaryHeap, HashMap};

pageserver/src/tenant/storage_layer/batch_split_writer.rs

@@ -5,48 +5,164 @@ use pageserver_api::key::{Key, KEY_SIZE};
 use utils::{id::TimelineId, lsn::Lsn, shard::TenantShardId};
 
 use crate::tenant::storage_layer::Layer;
-use crate::{config::PageServerConf, context::RequestContext, repository::Value, tenant::Timeline};
+use crate::{config::PageServerConf, context::RequestContext, tenant::Timeline};
+use pageserver_api::value::Value;
 
 use super::layer::S3_UPLOAD_LIMIT;
 use super::{
     DeltaLayerWriter, ImageLayerWriter, PersistentLayerDesc, PersistentLayerKey, ResidentLayer,
 };
 
-pub(crate) enum SplitWriterResult {
+pub(crate) enum BatchWriterResult {
     Produced(ResidentLayer),
     Discarded(PersistentLayerKey),
 }
 
 #[cfg(test)]
-impl SplitWriterResult {
+impl BatchWriterResult {
     fn into_resident_layer(self) -> ResidentLayer {
         match self {
-            SplitWriterResult::Produced(layer) => layer,
-            SplitWriterResult::Discarded(_) => panic!("unexpected discarded layer"),
+            BatchWriterResult::Produced(layer) => layer,
+            BatchWriterResult::Discarded(_) => panic!("unexpected discarded layer"),
         }
     }
 
     fn into_discarded_layer(self) -> PersistentLayerKey {
         match self {
-            SplitWriterResult::Produced(_) => panic!("unexpected produced layer"),
-            SplitWriterResult::Discarded(layer) => layer,
+            BatchWriterResult::Produced(_) => panic!("unexpected produced layer"),
+            BatchWriterResult::Discarded(layer) => layer,
         }
     }
 }
 
+enum LayerWriterWrapper {
+    Image(ImageLayerWriter),
+    Delta(DeltaLayerWriter),
+}
+
+/// An layer writer that takes unfinished layers and finish them atomically.
+#[must_use]
+pub struct BatchLayerWriter {
+    generated_layer_writers: Vec<(LayerWriterWrapper, PersistentLayerKey)>,
+    conf: &'static PageServerConf,
+}
+
+impl BatchLayerWriter {
+    pub async fn new(conf: &'static PageServerConf) -> anyhow::Result<Self> {
+        Ok(Self {
+            generated_layer_writers: Vec::new(),
+            conf,
+        })
+    }
+
+    pub fn add_unfinished_image_writer(
+        &mut self,
+        writer: ImageLayerWriter,
+        key_range: Range<Key>,
+        lsn: Lsn,
+    ) {
+        self.generated_layer_writers.push((
+            LayerWriterWrapper::Image(writer),
+            PersistentLayerKey {
+                key_range,
+                lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
+                is_delta: false,
+            },
+        ));
+    }
+
+    pub fn add_unfinished_delta_writer(
+        &mut self,
+        writer: DeltaLayerWriter,
+        key_range: Range<Key>,
+        lsn_range: Range<Lsn>,
+    ) {
+        self.generated_layer_writers.push((
+            LayerWriterWrapper::Delta(writer),
+            PersistentLayerKey {
+                key_range,
+                lsn_range,
+                is_delta: true,
+            },
+        ));
+    }
+
+    pub(crate) async fn finish_with_discard_fn<D, F>(
+        self,
+        tline: &Arc<Timeline>,
+        ctx: &RequestContext,
+        discard_fn: D,
+    ) -> anyhow::Result<Vec<BatchWriterResult>>
+    where
+        D: Fn(&PersistentLayerKey) -> F,
+        F: Future<Output = bool>,
+    {
+        let Self {
+            generated_layer_writers,
+            ..
+        } = self;
+        let clean_up_layers = |generated_layers: Vec<BatchWriterResult>| {
+            for produced_layer in generated_layers {
+                if let BatchWriterResult::Produced(resident_layer) = produced_layer {
+                    let layer: Layer = resident_layer.into();
+                    layer.delete_on_drop();
+                }
+            }
+        };
+        // BEGIN: catch every error and do the recovery in the below section
+        let mut generated_layers: Vec<BatchWriterResult> = Vec::new();
+        for (inner, layer_key) in generated_layer_writers {
+            if discard_fn(&layer_key).await {
+                generated_layers.push(BatchWriterResult::Discarded(layer_key));
+            } else {
+                let res = match inner {
+                    LayerWriterWrapper::Delta(writer) => {
+                        writer.finish(layer_key.key_range.end, ctx).await
+                    }
+                    LayerWriterWrapper::Image(writer) => {
+                        assert_eq!(writer.key_range().start, layer_key.key_range.start);
+                        assert_eq!(writer.lsn(), layer_key.lsn_range.start);
+                        writer
+                            .finish_with_end_key(layer_key.key_range.end, ctx)
+                            .await
+                    }
+                };
+                let layer = match res {
+                    Ok((desc, path)) => {
+                        match Layer::finish_creating(self.conf, tline, desc, &path) {
+                            Ok(layer) => layer,
+                            Err(e) => {
+                                tokio::fs::remove_file(&path).await.ok();
+                                clean_up_layers(generated_layers);
+                                return Err(e);
+                            }
+                        }
+                    }
+                    Err(e) => {
+                        // Image/DeltaLayerWriter::finish will clean up the temporary layer if anything goes wrong,
+                        // so we don't need to remove the layer we just failed to create by ourselves.
+                        clean_up_layers(generated_layers);
+                        return Err(e);
+                    }
+                };
+                generated_layers.push(BatchWriterResult::Produced(layer));
+            }
+        }
+        // END: catch every error and do the recovery in the above section
+        Ok(generated_layers)
+    }
+}
+
 /// An image writer that takes images and produces multiple image layers.
-///
-/// The interface does not guarantee atomicity (i.e., if the image layer generation
-/// fails, there might be leftover files to be cleaned up)
 #[must_use]
 pub struct SplitImageLayerWriter {
     inner: ImageLayerWriter,
     target_layer_size: u64,
-    generated_layer_writers: Vec<(ImageLayerWriter, PersistentLayerKey)>,
+    lsn: Lsn,
     conf: &'static PageServerConf,
     timeline_id: TimelineId,
     tenant_shard_id: TenantShardId,
-    lsn: Lsn,
+    batches: BatchLayerWriter,
     start_key: Key,
 }
 
@@ -71,10 +187,10 @@ impl SplitImageLayerWriter {
                 ctx,
             )
             .await?,
-            generated_layer_writers: Vec::new(),
             conf,
             timeline_id,
             tenant_shard_id,
+            batches: BatchLayerWriter::new(conf).await?,
             lsn,
             start_key,
         })
@@ -102,16 +218,13 @@ impl SplitImageLayerWriter {
                 ctx,
             )
             .await?;
-            let layer_key = PersistentLayerKey {
-                key_range: self.start_key..key,
-                lsn_range: PersistentLayerDesc::image_layer_lsn_range(self.lsn),
-                is_delta: false,
-            };
             let prev_image_writer = std::mem::replace(&mut self.inner, next_image_writer);
+            self.batches.add_unfinished_image_writer(
+                prev_image_writer,
+                self.start_key..key,
+                self.lsn,
+            );
             self.start_key = key;
-
-            self.generated_layer_writers
-                .push((prev_image_writer, layer_key));
         }
         self.inner.put_image(key, img, ctx).await
     }
@@ -122,64 +235,18 @@ impl SplitImageLayerWriter {
         ctx: &RequestContext,
         end_key: Key,
         discard_fn: D,
-    ) -> anyhow::Result<Vec<SplitWriterResult>>
+    ) -> anyhow::Result<Vec<BatchWriterResult>>
     where
         D: Fn(&PersistentLayerKey) -> F,
         F: Future<Output = bool>,
     {
         let Self {
-            mut generated_layer_writers,
-            inner,
-            ..
+            mut batches, inner, ..
         } = self;
         if inner.num_keys() != 0 {
-            let layer_key = PersistentLayerKey {
-                key_range: self.start_key..end_key,
-                lsn_range: PersistentLayerDesc::image_layer_lsn_range(self.lsn),
-                is_delta: false,
-            };
-            generated_layer_writers.push((inner, layer_key));
+            batches.add_unfinished_image_writer(inner, self.start_key..end_key, self.lsn);
         }
-        let clean_up_layers = |generated_layers: Vec<SplitWriterResult>| {
-            for produced_layer in generated_layers {
-                if let SplitWriterResult::Produced(image_layer) = produced_layer {
-                    let layer: Layer = image_layer.into();
-                    layer.delete_on_drop();
-                }
-            }
-        };
-        // BEGIN: catch every error and do the recovery in the below section
-        let mut generated_layers = Vec::new();
-        for (inner, layer_key) in generated_layer_writers {
-            if discard_fn(&layer_key).await {
-                generated_layers.push(SplitWriterResult::Discarded(layer_key));
-            } else {
-                let layer = match inner
-                    .finish_with_end_key(layer_key.key_range.end, ctx)
-                    .await
-                {
-                    Ok((desc, path)) => {
-                        match Layer::finish_creating(self.conf, tline, desc, &path) {
-                            Ok(layer) => layer,
-                            Err(e) => {
-                                tokio::fs::remove_file(&path).await.ok();
-                                clean_up_layers(generated_layers);
-                                return Err(e);
-                            }
-                        }
-                    }
-                    Err(e) => {
-                        // ImageLayerWriter::finish will clean up the temporary layer if anything goes wrong,
-                        // so we don't need to remove the layer we just failed to create by ourselves.
-                        clean_up_layers(generated_layers);
-                        return Err(e);
-                    }
-                };
-                generated_layers.push(SplitWriterResult::Produced(layer));
-            }
-        }
-        // END: catch every error and do the recovery in the above section
-        Ok(generated_layers)
+        batches.finish_with_discard_fn(tline, ctx, discard_fn).await
     }
 
     #[cfg(test)]
@@ -188,7 +255,7 @@ impl SplitImageLayerWriter {
         tline: &Arc<Timeline>,
         ctx: &RequestContext,
         end_key: Key,
-    ) -> anyhow::Result<Vec<SplitWriterResult>> {
+    ) -> anyhow::Result<Vec<BatchWriterResult>> {
         self.finish_with_discard_fn(tline, ctx, end_key, |_| async { false })
             .await
     }
@@ -196,9 +263,6 @@ impl SplitImageLayerWriter {
 
 /// A delta writer that takes key-lsn-values and produces multiple delta layers.
 ///
-/// The interface does not guarantee atomicity (i.e., if the delta layer generation fails,
-/// there might be leftover files to be cleaned up).
-///
 /// Note that if updates of a single key exceed the target size limit, all of the updates will be batched
 /// into a single file. This behavior might change in the future. For reference, the legacy compaction algorithm
 /// will split them into multiple files based on size.
@@ -206,12 +270,12 @@ impl SplitImageLayerWriter {
 pub struct SplitDeltaLayerWriter {
     inner: Option<(Key, DeltaLayerWriter)>,
     target_layer_size: u64,
-    generated_layer_writers: Vec<(DeltaLayerWriter, PersistentLayerKey)>,
     conf: &'static PageServerConf,
     timeline_id: TimelineId,
     tenant_shard_id: TenantShardId,
     lsn_range: Range<Lsn>,
     last_key_written: Key,
+    batches: BatchLayerWriter,
 }
 
 impl SplitDeltaLayerWriter {
@@ -225,12 +289,12 @@ impl SplitDeltaLayerWriter {
         Ok(Self {
             target_layer_size,
             inner: None,
-            generated_layer_writers: Vec::new(),
             conf,
             timeline_id,
             tenant_shard_id,
             lsn_range,
             last_key_written: Key::MIN,
+            batches: BatchLayerWriter::new(conf).await?,
         })
     }
 
@@ -279,13 +343,11 @@ impl SplitDeltaLayerWriter {
                 .await?;
                 let (start_key, prev_delta_writer) =
                     std::mem::replace(&mut self.inner, Some((key, next_delta_writer))).unwrap();
-                let layer_key = PersistentLayerKey {
-                    key_range: start_key..key,
-                    lsn_range: self.lsn_range.clone(),
-                    is_delta: true,
-                };
-                self.generated_layer_writers
-                    .push((prev_delta_writer, layer_key));
+                self.batches.add_unfinished_delta_writer(
+                    prev_delta_writer,
+                    start_key..key,
+                    self.lsn_range.clone(),
+                );
             } else if inner.estimated_size() >= S3_UPLOAD_LIMIT {
                 // We have to produce a very large file b/c a key is updated too often.
                 anyhow::bail!(
@@ -305,64 +367,25 @@ impl SplitDeltaLayerWriter {
         tline: &Arc<Timeline>,
         ctx: &RequestContext,
         discard_fn: D,
-    ) -> anyhow::Result<Vec<SplitWriterResult>>
+    ) -> anyhow::Result<Vec<BatchWriterResult>>
     where
         D: Fn(&PersistentLayerKey) -> F,
         F: Future<Output = bool>,
     {
         let Self {
-            mut generated_layer_writers,
-            inner,
-            ..
+            mut batches, inner, ..
         } = self;
         if let Some((start_key, writer)) = inner {
             if writer.num_keys() != 0 {
                 let end_key = self.last_key_written.next();
-                let layer_key = PersistentLayerKey {
-                    key_range: start_key..end_key,
-                    lsn_range: self.lsn_range.clone(),
-                    is_delta: true,
-                };
-                generated_layer_writers.push((writer, layer_key));
+                batches.add_unfinished_delta_writer(
+                    writer,
+                    start_key..end_key,
+                    self.lsn_range.clone(),
+                );
             }
         }
-        let clean_up_layers = |generated_layers: Vec<SplitWriterResult>| {
-            for produced_layer in generated_layers {
-                if let SplitWriterResult::Produced(delta_layer) = produced_layer {
-                    let layer: Layer = delta_layer.into();
-                    layer.delete_on_drop();
-                }
-            }
-        };
-        // BEGIN: catch every error and do the recovery in the below section
-        let mut generated_layers = Vec::new();
-        for (inner, layer_key) in generated_layer_writers {
-            if discard_fn(&layer_key).await {
-                generated_layers.push(SplitWriterResult::Discarded(layer_key));
-            } else {
-                let layer = match inner.finish(layer_key.key_range.end, ctx).await {
-                    Ok((desc, path)) => {
-                        match Layer::finish_creating(self.conf, tline, desc, &path) {
-                            Ok(layer) => layer,
-                            Err(e) => {
-                                tokio::fs::remove_file(&path).await.ok();
-                                clean_up_layers(generated_layers);
-                                return Err(e);
-                            }
-                        }
-                    }
-                    Err(e) => {
-                        // DeltaLayerWriter::finish will clean up the temporary layer if anything goes wrong,
-                        // so we don't need to remove the layer we just failed to create by ourselves.
-                        clean_up_layers(generated_layers);
-                        return Err(e);
-                    }
-                };
-                generated_layers.push(SplitWriterResult::Produced(layer));
-            }
-        }
-        // END: catch every error and do the recovery in the above section
-        Ok(generated_layers)
+        batches.finish_with_discard_fn(tline, ctx, discard_fn).await
     }
 
     #[cfg(test)]
@@ -370,7 +393,7 @@ impl SplitDeltaLayerWriter {
         self,
         tline: &Arc<Timeline>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<Vec<SplitWriterResult>> {
+    ) -> anyhow::Result<Vec<BatchWriterResult>> {
         self.finish_with_discard_fn(tline, ctx, |_| async { false })
             .await
     }

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -30,7 +30,6 @@
 use crate::config::PageServerConf;
 use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
 use crate::page_cache::{self, FileId, PAGE_SZ};
-use crate::repository::{Key, Value, KEY_SIZE};
 use crate::tenant::blob_io::BlobWriter;
 use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{
@@ -46,7 +45,7 @@ use crate::tenant::PageReconstructError;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
 use crate::virtual_file::IoBufferMut;
 use crate::virtual_file::{self, MaybeFatalIo, VirtualFile};
-use crate::{walrecord, TEMP_FILE_SUFFIX};
+use crate::TEMP_FILE_SUFFIX;
 use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
 use anyhow::{anyhow, bail, ensure, Context, Result};
 use camino::{Utf8Path, Utf8PathBuf};
@@ -54,9 +53,11 @@ use futures::StreamExt;
 use itertools::Itertools;
 use pageserver_api::config::MaxVectoredReadBytes;
 use pageserver_api::key::DBDIR_KEY;
+use pageserver_api::key::{Key, KEY_SIZE};
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::ImageCompressionAlgorithm;
 use pageserver_api::shard::TenantShardId;
+use pageserver_api::value::Value;
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
 use std::collections::VecDeque;
@@ -269,7 +270,7 @@ impl AsLayerDesc for DeltaLayer {
 }
 
 impl DeltaLayer {
-    pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
         self.desc.dump();
 
         if !verbose {
@@ -1084,7 +1085,7 @@ impl DeltaLayerInner {
         }
     }
 
-    pub(super) async fn load_keys<'a>(
+    pub(crate) async fn index_entries<'a>(
         &'a self,
         ctx: &RequestContext,
     ) -> Result<Vec<DeltaEntry<'a>>> {
@@ -1293,7 +1294,7 @@ impl DeltaLayerInner {
                     // is it an image or will_init walrecord?
                     // FIXME: this could be handled by threading the BlobRef to the
                     // VectoredReadBuilder
-                    let will_init = crate::repository::ValueBytes::will_init(&data)
+                    let will_init = pageserver_api::value::ValueBytes::will_init(&data)
                         .inspect_err(|_e| {
                             #[cfg(feature = "testing")]
                             tracing::error!(data=?utils::Hex(&data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
@@ -1346,7 +1347,7 @@ impl DeltaLayerInner {
 
         tree_reader.dump().await?;
 
-        let keys = self.load_keys(ctx).await?;
+        let keys = self.index_entries(ctx).await?;
 
         async fn dump_blob(val: &ValueRef<'_>, ctx: &RequestContext) -> anyhow::Result<String> {
             let buf = val.load_raw(ctx).await?;
@@ -1356,7 +1357,7 @@ impl DeltaLayerInner {
                     format!(" img {} bytes", img.len())
                 }
                 Value::WalRecord(rec) => {
-                    let wal_desc = walrecord::describe_wal_record(&rec)?;
+                    let wal_desc = pageserver_api::record::describe_wal_record(&rec)?;
                     format!(
                         " rec {} bytes will_init: {} {}",
                         buf.len(),
@@ -1437,7 +1438,7 @@ impl DeltaLayerInner {
         offset
     }
 
-    pub(crate) fn iter<'a>(&'a self, ctx: &'a RequestContext) -> DeltaLayerIterator<'a> {
+    pub fn iter<'a>(&'a self, ctx: &'a RequestContext) -> DeltaLayerIterator<'a> {
         let block_reader = FileBlockReader::new(&self.file, self.file_id);
         let tree_reader =
             DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
@@ -1453,6 +1454,16 @@ impl DeltaLayerInner {
             ),
         }
     }
+
+    /// NB: not super efficient, but not terrible either. Should prob be an iterator.
+    //
+    // We're reusing the index traversal logical in plan_reads; would be nice to
+    // factor that out.
+    pub(crate) async fn load_keys(&self, ctx: &RequestContext) -> anyhow::Result<Vec<Key>> {
+        self.index_entries(ctx)
+            .await
+            .map(|entries| entries.into_iter().map(|entry| entry.key).collect())
+    }
 }
 
 /// A set of data associated with a delta layer key and its value
@@ -1600,7 +1611,6 @@ pub(crate) mod test {
     use rand::RngCore;
 
     use super::*;
-    use crate::repository::Value;
     use crate::tenant::harness::TIMELINE_ID;
     use crate::tenant::storage_layer::{Layer, ResidentLayer};
     use crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner;
@@ -1612,6 +1622,7 @@ pub(crate) mod test {
         DEFAULT_PG_VERSION,
     };
     use bytes::Bytes;
+    use pageserver_api::value::Value;
 
     /// Construct an index for a fictional delta layer and and then
     /// traverse in order to plan vectored reads for a query. Finally,
@@ -1964,8 +1975,8 @@ pub(crate) mod test {
 
     #[tokio::test]
     async fn copy_delta_prefix_smoke() {
-        use crate::walrecord::NeonWalRecord;
         use bytes::Bytes;
+        use pageserver_api::record::NeonWalRecord;
 
         let h = crate::tenant::harness::TenantHarness::create("truncate_delta_smoke")
             .await
@@ -2188,6 +2199,7 @@ pub(crate) mod test {
         (k1, l1).cmp(&(k2, l2))
     }
 
+    #[cfg(feature = "testing")]
     pub(crate) fn sort_delta_value(
         (k1, l1, v1): &(Key, Lsn, Value),
         (k2, l2, v2): &(Key, Lsn, Value),

pageserver/src/tenant/storage_layer/filter_iterator.rs

@@ -7,7 +7,7 @@ use pageserver_api::{
 };
 use utils::lsn::Lsn;
 
-use crate::repository::Value;
+use pageserver_api::value::Value;
 
 use super::merge_iterator::MergeIterator;
 
@@ -121,8 +121,8 @@ mod tests {
 
     #[tokio::test]
     async fn filter_keyspace_iterator() {
-        use crate::repository::Value;
         use bytes::Bytes;
+        use pageserver_api::value::Value;
 
         let harness = TenantHarness::create("filter_iterator_filter_keyspace_iterator")
             .await

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -28,7 +28,6 @@
 use crate::config::PageServerConf;
 use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
 use crate::page_cache::{self, FileId, PAGE_SZ};
-use crate::repository::{Key, Value, KEY_SIZE};
 use crate::tenant::blob_io::BlobWriter;
 use crate::tenant::block_io::{BlockBuf, FileBlockReader};
 use crate::tenant::disk_btree::{
@@ -51,8 +50,10 @@ use hex;
 use itertools::Itertools;
 use pageserver_api::config::MaxVectoredReadBytes;
 use pageserver_api::key::DBDIR_KEY;
+use pageserver_api::key::{Key, KEY_SIZE};
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::shard::{ShardIdentity, TenantShardId};
+use pageserver_api::value::Value;
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
 use std::collections::VecDeque;
@@ -230,7 +231,7 @@ impl AsLayerDesc for ImageLayer {
 }
 
 impl ImageLayer {
-    pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
         self.desc.dump();
 
         if !verbose {
@@ -673,6 +674,21 @@ impl ImageLayerInner {
             ),
         }
     }
+
+    /// NB: not super efficient, but not terrible either. Should prob be an iterator.
+    //
+    // We're reusing the index traversal logical in plan_reads; would be nice to
+    // factor that out.
+    pub(crate) async fn load_keys(&self, ctx: &RequestContext) -> anyhow::Result<Vec<Key>> {
+        let plan = self
+            .plan_reads(KeySpace::single(self.key_range.clone()), None, ctx)
+            .await?;
+        Ok(plan
+            .into_iter()
+            .flat_map(|read| read.blobs_at)
+            .map(|(_, blob_meta)| blob_meta.key)
+            .collect())
+    }
 }
 
 /// A builder object for constructing a new image layer.
@@ -869,6 +885,7 @@ impl ImageLayerWriterInner {
         }
 
         let final_key_range = if let Some(end_key) = end_key {
+            assert!(end_key <= self.key_range.end);
             self.key_range.start..end_key
         } else {
             self.key_range.clone()
@@ -1009,7 +1026,7 @@ impl ImageLayerWriter {
         self.inner.take().unwrap().finish(ctx, None).await
     }
 
-    /// Finish writing the image layer with an end key, used in [`super::split_writer::SplitImageLayerWriter`]. The end key determines the end of the image layer's covered range and is exclusive.
+    /// Finish writing the image layer with an end key, used in [`super::batch_split_writer::SplitImageLayerWriter`]. The end key determines the end of the image layer's covered range and is exclusive.
     pub(super) async fn finish_with_end_key(
         mut self,
         end_key: Key,
@@ -1017,6 +1034,14 @@ impl ImageLayerWriter {
     ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
         self.inner.take().unwrap().finish(ctx, Some(end_key)).await
     }
+
+    pub(crate) fn key_range(&self) -> Range<Key> {
+        self.inner.as_ref().unwrap().key_range.clone()
+    }
+
+    pub(crate) fn lsn(&self) -> Lsn {
+        self.inner.as_ref().unwrap().lsn
+    }
 }
 
 impl Drop for ImageLayerWriter {
@@ -1110,6 +1135,7 @@ mod test {
     use pageserver_api::{
         key::Key,
         shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize},
+        value::Value,
     };
     use utils::{
         generation::Generation,
@@ -1119,7 +1145,6 @@ mod test {
 
     use crate::{
         context::RequestContext,
-        repository::Value,
         tenant::{
             config::TenantConf,
             harness::{TenantHarness, TIMELINE_ID},

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -7,7 +7,6 @@
 use crate::assert_u64_eq_usize::{u64_to_usize, U64IsUsize, UsizeIsU64};
 use crate::config::PageServerConf;
 use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
-use crate::repository::{Key, Value};
 use crate::tenant::ephemeral_file::EphemeralFile;
 use crate::tenant::timeline::GetVectoredError;
 use crate::tenant::PageReconstructError;
@@ -16,9 +15,11 @@ use crate::{l0_flush, page_cache};
 use anyhow::{anyhow, Context, Result};
 use camino::Utf8PathBuf;
 use pageserver_api::key::CompactKey;
+use pageserver_api::key::Key;
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::InMemoryLayerInfo;
 use pageserver_api::shard::TenantShardId;
+use pageserver_api::value::Value;
 use std::collections::{BTreeMap, HashMap};
 use std::sync::{Arc, OnceLock};
 use std::time::Instant;

pageserver/src/tenant/storage_layer/layer.rs

@@ -19,7 +19,7 @@ use crate::task_mgr::TaskKind;
 use crate::tenant::timeline::{CompactionError, GetVectoredError};
 use crate::tenant::{remote_timeline_client::LayerFileMetadata, Timeline};
 
-use super::delta_layer::{self, DeltaEntry};
+use super::delta_layer::{self};
 use super::image_layer::{self};
 use super::{
     AsLayerDesc, ImageLayerWriter, LayerAccessStats, LayerAccessStatsReset, LayerName,
@@ -1841,23 +1841,22 @@ impl ResidentLayer {
     pub(crate) async fn load_keys<'a>(
         &'a self,
         ctx: &RequestContext,
-    ) -> anyhow::Result<Vec<DeltaEntry<'a>>> {
+    ) -> anyhow::Result<Vec<pageserver_api::key::Key>> {
         use LayerKind::*;
 
         let owner = &self.owner.0;
-        match self.downloaded.get(owner, ctx).await? {
-            Delta(ref d) => {
-                // this is valid because the DownloadedLayer::kind is a OnceCell, not a
-                // Mutex<OnceCell>, so we cannot go and deinitialize the value with OnceCell::take
-                // while it's being held.
-                self.owner.record_access(ctx);
+        let inner = self.downloaded.get(owner, ctx).await?;
 
-                delta_layer::DeltaLayerInner::load_keys(d, ctx)
-                    .await
-                    .with_context(|| format!("Layer index is corrupted for {self}"))
-            }
-            Image(_) => anyhow::bail!(format!("cannot load_keys on a image layer {self}")),
-        }
+        // this is valid because the DownloadedLayer::kind is a OnceCell, not a
+        // Mutex<OnceCell>, so we cannot go and deinitialize the value with OnceCell::take
+        // while it's being held.
+        self.owner.record_access(ctx);
+
+        let res = match inner {
+            Delta(ref d) => delta_layer::DeltaLayerInner::load_keys(d, ctx).await,
+            Image(ref i) => image_layer::ImageLayerInner::load_keys(i, ctx).await,
+        };
+        res.with_context(|| format!("Layer index is corrupted for {self}"))
     }
 
     /// Read all they keys in this layer which match the ShardIdentity, and write them all to

pageserver/src/tenant/storage_layer/layer/tests.rs

@@ -760,8 +760,8 @@ async fn evict_and_wait_does_not_wait_for_download() {
 /// Also checks that the same does not happen on a non-evicted layer (regression test).
 #[tokio::test(start_paused = true)]
 async fn eviction_cancellation_on_drop() {
-    use crate::repository::Value;
     use bytes::Bytes;
+    use pageserver_api::value::Value;
 
     // this is the runtime on which Layer spawns the blocking tasks on
     let handle = tokio::runtime::Handle::current();
@@ -782,7 +782,7 @@ async fn eviction_cancellation_on_drop() {
         let mut writer = timeline.writer().await;
         writer
             .put(
-                crate::repository::Key::from_i128(5),
+                pageserver_api::key::Key::from_i128(5),
                 Lsn(0x20),
                 &Value::Image(Bytes::from_static(b"this does not matter either")),
                 &ctx,

pageserver/src/tenant/storage_layer/layer_desc.rs

@@ -3,7 +3,7 @@ use pageserver_api::shard::TenantShardId;
 use std::ops::Range;
 use utils::{id::TimelineId, lsn::Lsn};
 
-use crate::repository::Key;
+use pageserver_api::key::Key;
 
 use super::{DeltaLayerName, ImageLayerName, LayerName};
 
@@ -57,6 +57,34 @@ impl std::fmt::Display for PersistentLayerKey {
     }
 }
 
+impl From<ImageLayerName> for PersistentLayerKey {
+    fn from(image_layer_name: ImageLayerName) -> Self {
+        Self {
+            key_range: image_layer_name.key_range,
+            lsn_range: PersistentLayerDesc::image_layer_lsn_range(image_layer_name.lsn),
+            is_delta: false,
+        }
+    }
+}
+
+impl From<DeltaLayerName> for PersistentLayerKey {
+    fn from(delta_layer_name: DeltaLayerName) -> Self {
+        Self {
+            key_range: delta_layer_name.key_range,
+            lsn_range: delta_layer_name.lsn_range,
+            is_delta: true,
+        }
+    }
+}
+
+impl From<LayerName> for PersistentLayerKey {
+    fn from(layer_name: LayerName) -> Self {
+        match layer_name {
+            LayerName::Image(i) => i.into(),
+            LayerName::Delta(d) => d.into(),
+        }
+    }
+}
 impl PersistentLayerDesc {
     pub fn key(&self) -> PersistentLayerKey {
         PersistentLayerKey {

pageserver/src/tenant/storage_layer/layer_name.rs

@@ -1,14 +1,12 @@
 //!
 //! Helper functions for dealing with filenames of the image and delta layer files.
 //!
-use crate::repository::Key;
-use std::borrow::Cow;
+use pageserver_api::key::Key;
 use std::cmp::Ordering;
 use std::fmt;
 use std::ops::Range;
 use std::str::FromStr;
 
-use regex::Regex;
 use utils::lsn::Lsn;
 
 use super::PersistentLayerDesc;
@@ -60,32 +58,31 @@ impl Ord for DeltaLayerName {
 /// Represents the region of the LSN-Key space covered by a DeltaLayer
 ///
 /// ```text
-///    <key start>-<key end>__<LSN start>-<LSN end>
+///    <key start>-<key end>__<LSN start>-<LSN end>-<generation>
 /// ```
 impl DeltaLayerName {
     /// Parse the part of a delta layer's file name that represents the LayerName. Returns None
     /// if the filename does not match the expected pattern.
     pub fn parse_str(fname: &str) -> Option<Self> {
-        let mut parts = fname.split("__");
-        let mut key_parts = parts.next()?.split('-');
-        let mut lsn_parts = parts.next()?.split('-');
-
-        let key_start_str = key_parts.next()?;
-        let key_end_str = key_parts.next()?;
-        let lsn_start_str = lsn_parts.next()?;
-        let lsn_end_str = lsn_parts.next()?;
-
-        if parts.next().is_some() || key_parts.next().is_some() || key_parts.next().is_some() {
-            return None;
-        }
-
-        if key_start_str.len() != 36
-            || key_end_str.len() != 36
-            || lsn_start_str.len() != 16
-            || lsn_end_str.len() != 16
+        let (key_parts, lsn_generation_parts) = fname.split_once("__")?;
+        let (key_start_str, key_end_str) = key_parts.split_once('-')?;
+        let (lsn_start_str, lsn_end_generation_parts) = lsn_generation_parts.split_once('-')?;
+        let lsn_end_str = if let Some((lsn_end_str, maybe_generation)) =
+            lsn_end_generation_parts.split_once('-')
         {
-            return None;
-        }
+            if maybe_generation.starts_with("v") {
+                // vY-XXXXXXXX
+                lsn_end_str
+            } else if maybe_generation.len() == 8 {
+                // XXXXXXXX
+                lsn_end_str
+            } else {
+                // no idea what this is
+                return None;
+            }
+        } else {
+            lsn_end_generation_parts
+        };
 
         let key_start = Key::from_hex(key_start_str).ok()?;
         let key_end = Key::from_hex(key_end_str).ok()?;
@@ -173,25 +170,29 @@ impl ImageLayerName {
 /// Represents the part of the Key-LSN space covered by an ImageLayer
 ///
 /// ```text
-///    <key start>-<key end>__<LSN>
+///    <key start>-<key end>__<LSN>-<generation>
 /// ```
 impl ImageLayerName {
     /// Parse a string as then LayerName part of an image layer file name. Returns None if the
     /// filename does not match the expected pattern.
     pub fn parse_str(fname: &str) -> Option<Self> {
-        let mut parts = fname.split("__");
-        let mut key_parts = parts.next()?.split('-');
-
-        let key_start_str = key_parts.next()?;
-        let key_end_str = key_parts.next()?;
-        let lsn_str = parts.next()?;
-        if parts.next().is_some() || key_parts.next().is_some() {
-            return None;
-        }
-
-        if key_start_str.len() != 36 || key_end_str.len() != 36 || lsn_str.len() != 16 {
-            return None;
-        }
+        let (key_parts, lsn_generation_parts) = fname.split_once("__")?;
+        let (key_start_str, key_end_str) = key_parts.split_once('-')?;
+        let lsn_str =
+            if let Some((lsn_str, maybe_generation)) = lsn_generation_parts.split_once('-') {
+                if maybe_generation.starts_with("v") {
+                    // vY-XXXXXXXX
+                    lsn_str
+                } else if maybe_generation.len() == 8 {
+                    // XXXXXXXX
+                    lsn_str
+                } else {
+                    // likely a delta layer
+                    return None;
+                }
+            } else {
+                lsn_generation_parts
+            };
 
         let key_start = Key::from_hex(key_start_str).ok()?;
         let key_end = Key::from_hex(key_end_str).ok()?;
@@ -258,6 +259,14 @@ impl LayerName {
         }
     }
 
+    /// Gets the LSN range encoded in the layer name.
+    pub fn lsn_as_range(&self) -> Range<Lsn> {
+        match &self {
+            LayerName::Image(layer) => layer.lsn_as_range(),
+            LayerName::Delta(layer) => layer.lsn_range.clone(),
+        }
+    }
+
     pub fn is_delta(&self) -> bool {
         matches!(self, LayerName::Delta(_))
     }
@@ -290,18 +299,8 @@ impl FromStr for LayerName {
     /// Self. When loading a physical layer filename, we drop any extra information
     /// not needed to build Self.
     fn from_str(value: &str) -> Result<Self, Self::Err> {
-        let gen_suffix_regex = Regex::new("^(?<base>.+)(?<gen>-v1-[0-9a-f]{8})$").unwrap();
-        let file_name: Cow<str> = match gen_suffix_regex.captures(value) {
-            Some(captures) => captures
-                .name("base")
-                .expect("Non-optional group")
-                .as_str()
-                .into(),
-            None => value.into(),
-        };
-
-        let delta = DeltaLayerName::parse_str(&file_name);
-        let image = ImageLayerName::parse_str(&file_name);
+        let delta = DeltaLayerName::parse_str(value);
+        let image = ImageLayerName::parse_str(value);
         let ok = match (delta, image) {
             (None, None) => {
                 return Err(format!(
@@ -367,11 +366,14 @@ mod test {
             lsn: Lsn::from_hex("00000000014FED58").unwrap(),
         });
         let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-v1-00000001").unwrap();
-        assert_eq!(parsed, expected,);
+        assert_eq!(parsed, expected);
+
+        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-00000001").unwrap();
+        assert_eq!(parsed, expected);
 
         // Omitting generation suffix is valid
         let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58").unwrap();
-        assert_eq!(parsed, expected,);
+        assert_eq!(parsed, expected);
     }
 
     #[test]
@@ -385,6 +387,9 @@ mod test {
         let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-000000000154C481-v1-00000001").unwrap();
         assert_eq!(parsed, expected);
 
+        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-000000000154C481-00000001").unwrap();
+        assert_eq!(parsed, expected);
+
         // Omitting generation suffix is valid
         let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-000000000154C481").unwrap();
         assert_eq!(parsed, expected);

pageserver/src/tenant/storage_layer/merge_iterator.rs

@@ -7,7 +7,8 @@ use anyhow::bail;
 use pageserver_api::key::Key;
 use utils::lsn::Lsn;
 
-use crate::{context::RequestContext, repository::Value};
+use crate::context::RequestContext;
+use pageserver_api::value::Value;
 
 use super::{
     delta_layer::{DeltaLayerInner, DeltaLayerIterator},
@@ -291,12 +292,16 @@ mod tests {
     use crate::{
         tenant::{
             harness::{TenantHarness, TIMELINE_ID},
-            storage_layer::delta_layer::test::{produce_delta_layer, sort_delta, sort_delta_value},
+            storage_layer::delta_layer::test::{produce_delta_layer, sort_delta},
         },
-        walrecord::NeonWalRecord,
         DEFAULT_PG_VERSION,
     };
 
+    #[cfg(feature = "testing")]
+    use crate::tenant::storage_layer::delta_layer::test::sort_delta_value;
+    #[cfg(feature = "testing")]
+    use pageserver_api::record::NeonWalRecord;
+
     async fn assert_merge_iter_equal(
         merge_iter: &mut MergeIterator<'_>,
         expect: &[(Key, Lsn, Value)],
@@ -319,8 +324,8 @@ mod tests {
 
     #[tokio::test]
     async fn merge_in_between() {
-        use crate::repository::Value;
         use bytes::Bytes;
+        use pageserver_api::value::Value;
 
         let harness = TenantHarness::create("merge_iterator_merge_in_between")
             .await
@@ -384,8 +389,8 @@ mod tests {
 
     #[tokio::test]
     async fn delta_merge() {
-        use crate::repository::Value;
         use bytes::Bytes;
+        use pageserver_api::value::Value;
 
         let harness = TenantHarness::create("merge_iterator_delta_merge")
             .await
@@ -458,10 +463,11 @@ mod tests {
         // TODO: test layers are loaded only when needed, reducing num of active iterators in k-merge
     }
 
+    #[cfg(feature = "testing")]
     #[tokio::test]
     async fn delta_image_mixed_merge() {
-        use crate::repository::Value;
         use bytes::Bytes;
+        use pageserver_api::value::Value;
 
         let harness = TenantHarness::create("merge_iterator_delta_image_mixed_merge")
             .await
@@ -586,5 +592,6 @@ mod tests {
         is_send(merge_iter);
     }
 
+    #[cfg(feature = "testing")]
     fn is_send(_: impl Send) {}
 }

pageserver/src/tenant/tasks.rs

@@ -279,6 +279,7 @@ fn log_compaction_error(
 
     let decision = match e {
         ShuttingDown => None,
+        Offload(_) => Some(LooksLike::Error),
         _ if task_cancelled => Some(LooksLike::Info),
         Other(e) => {
             let root_cause = e.root_cause();

pageserver/src/tenant/timeline.rs

@@ -20,6 +20,8 @@ use chrono::{DateTime, Utc};
 use enumset::EnumSet;
 use fail::fail_point;
 use handle::ShardTimelineId;
+use itertools::Itertools;
+use offload::OffloadError;
 use once_cell::sync::Lazy;
 use pageserver_api::{
     key::{
@@ -64,13 +66,14 @@ use std::{
 };
 use std::{pin::pin, sync::OnceLock};
 
+use crate::pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS;
 use crate::{
     aux_file::AuxFileSizeEstimator,
     tenant::{
         config::AttachmentMode,
         layer_map::{LayerMap, SearchResult},
         metadata::TimelineMetadata,
-        storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
+        storage_layer::inmemory_layer::IndexEntry,
     },
     walingest::WalLagCooldown,
     walredo,
@@ -103,7 +106,6 @@ use crate::{
     virtual_file::{MaybeFatalIo, VirtualFile},
 };
 use crate::{pgdatadir_mapping::LsnForTimestamp, tenant::tasks::BackgroundLoopKind};
-use crate::{pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS, tenant::storage_layer::PersistentLayerKey};
 use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
 
 use crate::config::PageServerConf;
@@ -125,11 +127,12 @@ use utils::{
     simple_rcu::{Rcu, RcuReadGuard},
 };
 
-use crate::repository::GcResult;
-use crate::repository::{Key, Value};
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
+use crate::tenant::gc_result::GcResult;
 use crate::ZERO_PAGE;
+use pageserver_api::key::Key;
+use pageserver_api::value::Value;
 
 use self::delete::DeleteTimelineFlow;
 pub(super) use self::eviction_task::EvictionTaskTenantState;
@@ -140,7 +143,10 @@ use self::walreceiver::{WalReceiver, WalReceiverConf};
 
 use super::{
     config::TenantConf,
-    storage_layer::{inmemory_layer, LayerVisibilityHint},
+    storage_layer::{
+        batch_split_writer::{BatchLayerWriter, BatchWriterResult},
+        inmemory_layer, LayerVisibilityHint,
+    },
     upload_queue::NotInitialized,
     MaybeOffloaded,
 };
@@ -424,6 +430,9 @@ pub struct Timeline {
     pub(crate) handles: handle::PerTimelineState<crate::page_service::TenantManagerTypes>,
 
     pub(crate) attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
+
+    /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
+    pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
 }
 
 pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
@@ -851,7 +860,7 @@ pub(crate) enum ShutdownMode {
 }
 
 struct ImageLayerCreationOutcome {
-    image: Option<ResidentLayer>,
+    image: Option<ImageLayerWriter>,
     next_start_key: Key,
 }
 
@@ -2136,6 +2145,7 @@ impl Timeline {
         pg_version: u32,
         state: TimelineState,
         attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
+        create_idempotency: crate::tenant::CreateTimelineIdempotency,
         cancel: CancellationToken,
     ) -> Arc<Self> {
         let disk_consistent_lsn = metadata.disk_consistent_lsn();
@@ -2274,6 +2284,8 @@ impl Timeline {
                 handles: Default::default(),
 
                 attach_wal_lag_cooldown,
+
+                create_idempotency,
             };
 
             result.repartition_threshold =
@@ -4045,11 +4057,14 @@ impl Timeline {
         if wrote_keys {
             // Normal path: we have written some data into the new image layer for this
             // partition, so flush it to disk.
-            let (desc, path) = image_layer_writer.finish(ctx).await?;
-            let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
-            info!("created image layer for rel {}", image_layer.local_path());
+            info!(
+                "flushed image layer for rel key_start={} key_end={} lsn={}",
+                image_layer_writer.key_range().start,
+                image_layer_writer.key_range().end,
+                image_layer_writer.lsn()
+            );
             Ok(ImageLayerCreationOutcome {
-                image: Some(image_layer),
+                image: Some(image_layer_writer),
                 next_start_key: img_range.end,
             })
         } else {
@@ -4135,14 +4150,14 @@ impl Timeline {
         if wrote_any_image {
             // Normal path: we have written some data into the new image layer for this
             // partition, so flush it to disk.
-            let (desc, path) = image_layer_writer.finish(ctx).await?;
-            let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
             info!(
-                "created image layer for metadata {}",
-                image_layer.local_path()
+                "flushed image layer for metadata key_start={} key_end={} lsn={}",
+                image_layer_writer.key_range().start,
+                image_layer_writer.key_range().end,
+                image_layer_writer.lsn()
             );
             Ok(ImageLayerCreationOutcome {
-                image: Some(image_layer),
+                image: Some(image_layer_writer),
                 next_start_key: img_range.end,
             })
         } else {
@@ -4219,7 +4234,6 @@ impl Timeline {
         ctx: &RequestContext,
     ) -> Result<Vec<ResidentLayer>, CreateImageLayersError> {
         let timer = self.metrics.create_images_time_histo.start_timer();
-        let mut image_layers = Vec::new();
 
         // We need to avoid holes between generated image layers.
         // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
@@ -4234,6 +4248,8 @@ impl Timeline {
 
         let check_for_image_layers = self.should_check_if_image_layers_required(lsn);
 
+        let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
+
         for partition in partitioning.parts.iter() {
             if self.cancel.is_cancelled() {
                 return Err(CreateImageLayersError::Cancelled);
@@ -4264,24 +4280,6 @@ impl Timeline {
                     continue;
                 }
             }
-            if let ImageLayerCreationMode::Force = mode {
-                // When forced to create image layers, we might try and create them where they already
-                // exist.  This mode is only used in tests/debug.
-                let layers = self.layers.read().await;
-                if layers.contains_key(&PersistentLayerKey {
-                    key_range: img_range.clone(),
-                    lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
-                    is_delta: false,
-                }) {
-                    tracing::info!(
-                        "Skipping image layer at {lsn} {}..{}, already exists",
-                        img_range.start,
-                        img_range.end
-                    );
-                    start = img_range.end;
-                    continue;
-                }
-            }
 
             let image_layer_writer = ImageLayerWriter::new(
                 self.conf,
@@ -4315,7 +4313,11 @@ impl Timeline {
                     .await?;
 
                 start = next_start_key;
-                image_layers.extend(image);
+                if let Some(image) = image {
+                    let key_range = image.key_range();
+                    let lsn = image.lsn();
+                    batch_image_writer.add_unfinished_image_writer(image, key_range, lsn);
+                }
             } else {
                 let ImageLayerCreationOutcome {
                     image,
@@ -4332,10 +4334,48 @@ impl Timeline {
                     )
                     .await?;
                 start = next_start_key;
-                image_layers.extend(image);
+                if let Some(image) = image {
+                    let key_range = image.key_range();
+                    let lsn = image.lsn();
+                    batch_image_writer.add_unfinished_image_writer(image, key_range, lsn);
+                }
             }
         }
 
+        let image_layers = batch_image_writer
+            .finish_with_discard_fn(self, ctx, |key| {
+                // TODO: remove this clone when Rust Edition 2024 is available, closure should capture this
+                // lifetime.
+                let key = key.clone();
+                async move {
+                    // When forced to create image layers, we might try and create them where they already
+                    // exist.  The force mode is only used in tests/debug.
+                    let layers = self.layers.read().await;
+                    if layers.contains_key(&key) {
+                        tracing::info!(
+                            "Skipping image layer at {} {}..{}, already exists",
+                            key.lsn_range.start,
+                            key.key_range.start,
+                            key.key_range.end
+                        );
+                        true
+                    } else {
+                        false
+                    }
+                }
+            })
+            .await?;
+        let image_layers = image_layers
+            .into_iter()
+            .filter_map(|x| {
+                if let BatchWriterResult::Produced(x) = x {
+                    Some(x)
+                } else {
+                    None
+                }
+            })
+            .collect_vec();
+
         let mut guard = self.layers.write().await;
 
         // FIXME: we could add the images to be uploaded *before* returning from here, but right
@@ -4468,11 +4508,23 @@ impl Drop for Timeline {
 pub(crate) enum CompactionError {
     #[error("The timeline or pageserver is shutting down")]
     ShuttingDown,
+    /// Compaction tried to offload a timeline and failed
+    #[error("Failed to offload timeline: {0}")]
+    Offload(OffloadError),
     /// Compaction cannot be done right now; page reconstruction and so on.
     #[error(transparent)]
     Other(anyhow::Error),
 }
 
+impl From<OffloadError> for CompactionError {
+    fn from(e: OffloadError) -> Self {
+        match e {
+            OffloadError::Cancelled => Self::ShuttingDown,
+            _ => Self::Offload(e),
+        }
+    }
+}
+
 impl CompactionError {
     pub fn is_cancelled(&self) -> bool {
         matches!(self, CompactionError::ShuttingDown)
@@ -5816,17 +5868,15 @@ fn is_send() {
 #[cfg(test)]
 mod tests {
     use pageserver_api::key::Key;
+    use pageserver_api::value::Value;
     use utils::{id::TimelineId, lsn::Lsn};
 
-    use crate::{
-        repository::Value,
-        tenant::{
-            harness::{test_img, TenantHarness},
-            layer_map::LayerMap,
-            storage_layer::{Layer, LayerName},
-            timeline::{DeltaLayerTestDesc, EvictionError},
-            Timeline,
-        },
+    use crate::tenant::{
+        harness::{test_img, TenantHarness},
+        layer_map::LayerMap,
+        storage_layer::{Layer, LayerName},
+        timeline::{DeltaLayerTestDesc, EvictionError},
+        Timeline,
     };
 
     #[tokio::test]

pageserver/src/tenant/timeline/compaction.rs

@@ -32,11 +32,11 @@ use crate::page_cache;
 use crate::statvfs::Statvfs;
 use crate::tenant::checks::check_valid_layermap;
 use crate::tenant::remote_timeline_client::WaitCompletionError;
+use crate::tenant::storage_layer::batch_split_writer::{
+    BatchWriterResult, SplitDeltaLayerWriter, SplitImageLayerWriter,
+};
 use crate::tenant::storage_layer::filter_iterator::FilterIterator;
 use crate::tenant::storage_layer::merge_iterator::MergeIterator;
-use crate::tenant::storage_layer::split_writer::{
-    SplitDeltaLayerWriter, SplitImageLayerWriter, SplitWriterResult,
-};
 use crate::tenant::storage_layer::{
     AsLayerDesc, PersistentLayerDesc, PersistentLayerKey, ValueReconstructState,
 };
@@ -49,9 +49,10 @@ use pageserver_api::config::tenant_conf_defaults::{
     DEFAULT_CHECKPOINT_DISTANCE, DEFAULT_COMPACTION_THRESHOLD,
 };
 
-use crate::keyspace::KeySpace;
-use crate::repository::{Key, Value};
-use crate::walrecord::NeonWalRecord;
+use pageserver_api::key::Key;
+use pageserver_api::keyspace::KeySpace;
+use pageserver_api::record::NeonWalRecord;
+use pageserver_api::value::Value;
 
 use utils::lsn::Lsn;
 
@@ -834,7 +835,12 @@ impl Timeline {
                 if self.cancel.is_cancelled() {
                     return Err(CompactionError::ShuttingDown);
                 }
-                all_keys.extend(l.load_keys(ctx).await.map_err(CompactionError::Other)?);
+                let delta = l.get_as_delta(ctx).await.map_err(CompactionError::Other)?;
+                let keys = delta
+                    .index_entries(ctx)
+                    .await
+                    .map_err(CompactionError::Other)?;
+                all_keys.extend(keys);
             }
             // The current stdlib sorting implementation is designed in a way where it is
             // particularly fast where the slice is made up of sorted sub-ranges.
@@ -1710,20 +1716,32 @@ impl Timeline {
         Ok(())
     }
 
-    /// An experimental compaction building block that combines compaction with garbage collection.
-    ///
-    /// The current implementation picks all delta + image layers that are below or intersecting with
-    /// the GC horizon without considering retain_lsns. Then, it does a full compaction over all these delta
-    /// layers and image layers, which generates image layers on the gc horizon, drop deltas below gc horizon,
-    /// and create delta layers with all deltas >= gc horizon.
     pub(crate) async fn compact_with_gc(
         self: &Arc<Self>,
         cancel: &CancellationToken,
         flags: EnumSet<CompactFlags>,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
-        use std::collections::BTreeSet;
+        self.partial_compact_with_gc(None, cancel, flags, ctx).await
+    }
 
+    /// An experimental compaction building block that combines compaction with garbage collection.
+    ///
+    /// The current implementation picks all delta + image layers that are below or intersecting with
+    /// the GC horizon without considering retain_lsns. Then, it does a full compaction over all these delta
+    /// layers and image layers, which generates image layers on the gc horizon, drop deltas below gc horizon,
+    /// and create delta layers with all deltas >= gc horizon.
+    ///
+    /// If `key_range`, it will only compact the keys within the range, aka partial compaction. This functionality
+    /// is not complete yet, and if it is set, only image layers will be generated.
+    ///
+    pub(crate) async fn partial_compact_with_gc(
+        self: &Arc<Self>,
+        compaction_key_range: Option<Range<Key>>,
+        cancel: &CancellationToken,
+        flags: EnumSet<CompactFlags>,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
         // Block other compaction/GC tasks from running for now. GC-compaction could run along
         // with legacy compaction tasks in the future. Always ensure the lock order is compaction -> gc.
         // Note that we already acquired the compaction lock when the outer `compact` function gets called.
@@ -1744,8 +1762,13 @@ impl Timeline {
         .await?;
 
         let dry_run = flags.contains(CompactFlags::DryRun);
+        let partial_compaction = compaction_key_range.is_some();
 
-        info!("running enhanced gc bottom-most compaction, dry_run={dry_run}");
+        if let Some(ref compaction_key_range) = compaction_key_range {
+            info!("running enhanced gc bottom-most compaction, dry_run={dry_run}, compaction_key_range={}..{}", compaction_key_range.start, compaction_key_range.end);
+        } else {
+            info!("running enhanced gc bottom-most compaction, dry_run={dry_run}");
+        }
 
         scopeguard::defer! {
             info!("done enhanced gc bottom-most compaction");
@@ -1757,7 +1780,7 @@ impl Timeline {
         // The layer selection has the following properties:
         // 1. If a layer is in the selection, all layers below it are in the selection.
         // 2. Inferred from (1), for each key in the layer selection, the value can be reconstructed only with the layers in the layer selection.
-        let (layer_selection, gc_cutoff, retain_lsns_below_horizon) = {
+        let (layer_selection, gc_cutoff, retain_lsns_below_horizon) = if !partial_compaction {
             let guard = self.layers.read().await;
             let layers = guard.layer_map()?;
             let gc_info = self.gc_info.read().unwrap();
@@ -1773,7 +1796,7 @@ impl Timeline {
                     retain_lsns_below_horizon.push(*lsn);
                 }
             }
-            let mut selected_layers = Vec::new();
+            let mut selected_layers: Vec<Layer> = Vec::new();
             drop(gc_info);
             // Pick all the layers intersect or below the gc_cutoff, get the largest LSN in the selected layers.
             let Some(max_layer_lsn) = layers
@@ -1798,8 +1821,52 @@ impl Timeline {
             }
             retain_lsns_below_horizon.sort();
             (selected_layers, gc_cutoff, retain_lsns_below_horizon)
+        } else {
+            // In case of partial compaction, we currently only support generating image layers, and therefore,
+            // we pick all layers that are below the lowest retain_lsn and does not intersect with any of the layers.
+            let guard = self.layers.read().await;
+            let layers = guard.layer_map()?;
+            let gc_info = self.gc_info.read().unwrap();
+            let mut min_lsn = gc_info.cutoffs.select_min();
+            for (lsn, _, _) in &gc_info.retain_lsns {
+                if lsn < &min_lsn {
+                    min_lsn = *lsn;
+                }
+            }
+            for lsn in gc_info.leases.keys() {
+                if lsn < &min_lsn {
+                    min_lsn = *lsn;
+                }
+            }
+            let mut selected_layers = Vec::new();
+            drop(gc_info);
+            // |-------| |-------| |-------|
+            // | Delta | | Delta | | Delta | -- min_lsn could be intersecting with the layers
+            // |-------| |-------| |-------| <- we want to pick all the layers below min_lsn, so that
+            // | Delta | | Delta | | Delta |    ...we can remove them after compaction
+            // |-------| |-------| |-------|
+            // Pick all the layers intersect or below the min_lsn, get the largest LSN in the selected layers.
+            let Some(compaction_key_range) = compaction_key_range.as_ref() else {
+                unreachable!()
+            };
+            for desc in layers.iter_historic_layers() {
+                if desc.get_lsn_range().end <= min_lsn
+                    && overlaps_with(&desc.key_range, compaction_key_range)
+                {
+                    selected_layers.push(guard.get_from_desc(&desc));
+                }
+            }
+            if selected_layers.is_empty() {
+                info!("no layers to compact with gc");
+                return Ok(());
+            }
+            (selected_layers, min_lsn, Vec::new())
         };
         let lowest_retain_lsn = if self.ancestor_timeline.is_some() {
+            if partial_compaction {
+                warn!("partial compaction cannot run on child branches (for now)");
+                return Ok(());
+            }
             Lsn(self.ancestor_lsn.0 + 1)
         } else {
             let res = retain_lsns_below_horizon
@@ -1827,23 +1894,18 @@ impl Timeline {
 
         self.check_compaction_space(&layer_selection).await?;
 
-        // Step 1: (In the future) construct a k-merge iterator over all layers. For now, simply collect all keys + LSNs.
-        // Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
-        let mut lsn_split_point = BTreeSet::new(); // TODO: use a better data structure (range tree / range set?)
+        // Generate statistics for the compaction
         for layer in &layer_selection {
             let desc = layer.layer_desc();
             if desc.is_delta() {
-                // ignore single-key layer files
-                if desc.key_range.start.next() != desc.key_range.end {
-                    let lsn_range = &desc.lsn_range;
-                    lsn_split_point.insert(lsn_range.start);
-                    lsn_split_point.insert(lsn_range.end);
-                }
                 stat.visit_delta_layer(desc.file_size());
             } else {
                 stat.visit_image_layer(desc.file_size());
             }
         }
+
+        // Step 1: construct a k-merge iterator over all layers.
+        // Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
         let layer_names: Vec<crate::tenant::storage_layer::LayerName> = layer_selection
             .iter()
             .map(|layer| layer.layer_desc().layer_name())
@@ -1894,7 +1956,10 @@ impl Timeline {
                     self.conf,
                     self.timeline_id,
                     self.tenant_shard_id,
-                    Key::MIN,
+                    compaction_key_range
+                        .as_ref()
+                        .map(|x| x.start)
+                        .unwrap_or(Key::MIN),
                     lowest_retain_lsn,
                     self.get_compaction_target_size(),
                     ctx,
@@ -1955,55 +2020,71 @@ impl Timeline {
             } else {
                 let last_key = last_key.as_mut().unwrap();
                 stat.on_unique_key_visited();
-                let retention = self
-                    .generate_key_retention(
-                        *last_key,
-                        &accumulated_values,
-                        gc_cutoff,
-                        &retain_lsns_below_horizon,
-                        COMPACTION_DELTA_THRESHOLD,
-                        get_ancestor_image(self, *last_key, ctx).await?,
-                    )
-                    .await?;
-                // Put the image into the image layer. Currently we have a single big layer for the compaction.
-                retention
-                    .pipe_to(
-                        *last_key,
-                        &mut delta_layer_writer,
-                        image_layer_writer.as_mut(),
-                        &mut stat,
-                        ctx,
-                    )
-                    .await?;
+                let skip_adding_key = if let Some(ref compaction_key_range) = compaction_key_range {
+                    !compaction_key_range.contains(last_key)
+                } else {
+                    false
+                };
+                if !skip_adding_key {
+                    let retention = self
+                        .generate_key_retention(
+                            *last_key,
+                            &accumulated_values,
+                            gc_cutoff,
+                            &retain_lsns_below_horizon,
+                            COMPACTION_DELTA_THRESHOLD,
+                            get_ancestor_image(self, *last_key, ctx).await?,
+                        )
+                        .await?;
+                    // Put the image into the image layer. Currently we have a single big layer for the compaction.
+                    retention
+                        .pipe_to(
+                            *last_key,
+                            &mut delta_layer_writer,
+                            image_layer_writer.as_mut(),
+                            &mut stat,
+                            ctx,
+                        )
+                        .await?;
+                }
                 accumulated_values.clear();
                 *last_key = key;
                 accumulated_values.push((key, lsn, val));
             }
         }
 
+        // TODO: move the below part to the loop body
         let last_key = last_key.expect("no keys produced during compaction");
-        // TODO: move this part to the loop body
         stat.on_unique_key_visited();
-        let retention = self
-            .generate_key_retention(
-                last_key,
-                &accumulated_values,
-                gc_cutoff,
-                &retain_lsns_below_horizon,
-                COMPACTION_DELTA_THRESHOLD,
-                get_ancestor_image(self, last_key, ctx).await?,
-            )
-            .await?;
-        // Put the image into the image layer. Currently we have a single big layer for the compaction.
-        retention
-            .pipe_to(
-                last_key,
-                &mut delta_layer_writer,
-                image_layer_writer.as_mut(),
-                &mut stat,
-                ctx,
-            )
-            .await?;
+
+        let skip_adding_key = if let Some(ref compaction_key_range) = compaction_key_range {
+            !compaction_key_range.contains(&last_key)
+        } else {
+            false
+        };
+        if !skip_adding_key {
+            let retention = self
+                .generate_key_retention(
+                    last_key,
+                    &accumulated_values,
+                    gc_cutoff,
+                    &retain_lsns_below_horizon,
+                    COMPACTION_DELTA_THRESHOLD,
+                    get_ancestor_image(self, last_key, ctx).await?,
+                )
+                .await?;
+            // Put the image into the image layer. Currently we have a single big layer for the compaction.
+            retention
+                .pipe_to(
+                    last_key,
+                    &mut delta_layer_writer,
+                    image_layer_writer.as_mut(),
+                    &mut stat,
+                    ctx,
+                )
+                .await?;
+        }
+        // end: move the above part to the loop body
 
         let discard = |key: &PersistentLayerKey| {
             let key = key.clone();
@@ -2012,8 +2093,12 @@ impl Timeline {
 
         let produced_image_layers = if let Some(writer) = image_layer_writer {
             if !dry_run {
+                let end_key = compaction_key_range
+                    .as_ref()
+                    .map(|x| x.end)
+                    .unwrap_or(Key::MAX);
                 writer
-                    .finish_with_discard_fn(self, ctx, Key::MAX, discard)
+                    .finish_with_discard_fn(self, ctx, end_key, discard)
                     .await?
             } else {
                 drop(writer);
@@ -2032,17 +2117,21 @@ impl Timeline {
             Vec::new()
         };
 
+        if partial_compaction && !produced_delta_layers.is_empty() {
+            bail!("implementation error: partial compaction should not be producing delta layers (for now)");
+        }
+
         let mut compact_to = Vec::new();
         let mut keep_layers = HashSet::new();
         let produced_delta_layers_len = produced_delta_layers.len();
         let produced_image_layers_len = produced_image_layers.len();
         for action in produced_delta_layers {
             match action {
-                SplitWriterResult::Produced(layer) => {
+                BatchWriterResult::Produced(layer) => {
                     stat.produce_delta_layer(layer.layer_desc().file_size());
                     compact_to.push(layer);
                 }
-                SplitWriterResult::Discarded(l) => {
+                BatchWriterResult::Discarded(l) => {
                     keep_layers.insert(l);
                     stat.discard_delta_layer();
                 }
@@ -2050,11 +2139,11 @@ impl Timeline {
         }
         for action in produced_image_layers {
             match action {
-                SplitWriterResult::Produced(layer) => {
+                BatchWriterResult::Produced(layer) => {
                     stat.produce_image_layer(layer.layer_desc().file_size());
                     compact_to.push(layer);
                 }
-                SplitWriterResult::Discarded(l) => {
+                BatchWriterResult::Discarded(l) => {
                     keep_layers.insert(l);
                     stat.discard_image_layer();
                 }
@@ -2062,6 +2151,28 @@ impl Timeline {
         }
         let mut layer_selection = layer_selection;
         layer_selection.retain(|x| !keep_layers.contains(&x.layer_desc().key()));
+        if let Some(ref compaction_key_range) = compaction_key_range {
+            // Partial compaction might select more data than it processes, e.g., if
+            // the compaction_key_range only partially overlaps:
+            //
+            //         [---compaction_key_range---]
+            //   [---A----][----B----][----C----][----D----]
+            //
+            // A,B,C,D are all in the `layer_selection`. The created image layers contain
+            // whatever is needed from B, C, and from `----]` of A, and from  `[--` of D.
+            //
+            // In contrast, `[--A-` and `--D----]` have not been processed, so, we must
+            // keep that data.
+            //
+            // The solution for now is to keep A and D completely.
+            // (layer_selection is what we'll remove from the layer map, so,
+            //  retain what is _not_ fully covered by compaction_key_range).
+            layer_selection.retain(|x| {
+                let key_range = &x.layer_desc().key_range;
+                key_range.start >= compaction_key_range.start
+                    && key_range.end <= compaction_key_range.end
+            });
+        }
 
         info!(
             "gc-compaction statistics: {}",
@@ -2143,7 +2254,7 @@ struct ResidentDeltaLayer(ResidentLayer);
 struct ResidentImageLayer(ResidentLayer);
 
 impl CompactionJobExecutor for TimelineAdaptor {
-    type Key = crate::repository::Key;
+    type Key = pageserver_api::key::Key;
 
     type Layer = OwnArc<PersistentLayerDesc>;
     type DeltaLayer = ResidentDeltaLayer;
@@ -2350,7 +2461,10 @@ impl TimelineAdaptor {
             )
             .await?;
 
-        if let Some(image_layer) = image {
+        if let Some(image_layer_writer) = image {
+            let (desc, path) = image_layer_writer.finish(ctx).await?;
+            let image_layer =
+                Layer::finish_creating(self.timeline.conf, &self.timeline, desc, &path)?;
             self.new_images.push(image_layer);
         }
 
@@ -2438,7 +2552,7 @@ impl CompactionDeltaLayer<TimelineAdaptor> for ResidentDeltaLayer {
     type DeltaEntry<'a> = DeltaEntry<'a>;
 
     async fn load_keys<'a>(&self, ctx: &RequestContext) -> anyhow::Result<Vec<DeltaEntry<'_>>> {
-        self.0.load_keys(ctx).await
+        self.0.get_as_delta(ctx).await?.index_entries(ctx).await
     }
 }
 

pageserver/src/tenant/timeline/delete.rs

@@ -6,7 +6,7 @@ use std::{
 use anyhow::Context;
 use pageserver_api::{models::TimelineState, shard::TenantShardId};
 use tokio::sync::OwnedMutexGuard;
-use tracing::{error, info, instrument, Instrument};
+use tracing::{error, info, info_span, instrument, Instrument};
 use utils::{crashsafe, fs_ext, id::TimelineId, pausable_failpoint};
 
 use crate::{
@@ -14,10 +14,9 @@ use crate::{
     task_mgr::{self, TaskKind},
     tenant::{
         metadata::TimelineMetadata,
-        remote_timeline_client::{
-            self, PersistIndexPartWithDeletedFlagError, RemoteTimelineClient,
-        },
-        CreateTimelineCause, DeleteTimelineError, Tenant, TimelineOrOffloaded,
+        remote_timeline_client::{PersistIndexPartWithDeletedFlagError, RemoteTimelineClient},
+        CreateTimelineCause, DeleteTimelineError, MaybeDeletedIndexPart, Tenant,
+        TimelineOrOffloaded,
     },
 };
 
@@ -176,32 +175,6 @@ async fn remove_maybe_offloaded_timeline_from_tenant(
     Ok(())
 }
 
-/// It is important that this gets called when DeletionGuard is being held.
-/// For more context see comments in [`DeleteTimelineFlow::prepare`]
-async fn upload_new_tenant_manifest(
-    tenant: &Tenant,
-    _: &DeletionGuard, // using it as a witness
-) -> anyhow::Result<()> {
-    // This is susceptible to race conditions, i.e. we won't continue deletions if there is a crash
-    // between the deletion of the index-part.json and reaching of this code.
-    // So indeed, the tenant manifest might refer to an offloaded timeline which has already been deleted.
-    // However, we handle this case in tenant loading code so the next time we attach, the issue is
-    // resolved.
-    let manifest = tenant.tenant_manifest();
-    // TODO: generation support
-    let generation = remote_timeline_client::TENANT_MANIFEST_GENERATION;
-    remote_timeline_client::upload_tenant_manifest(
-        &tenant.remote_storage,
-        &tenant.tenant_shard_id,
-        generation,
-        &manifest,
-        &tenant.cancel,
-    )
-    .await?;
-
-    Ok(())
-}
-
 /// Orchestrates timeline shut down of all timeline tasks, removes its in-memory structures,
 /// and deletes its data from both disk and s3.
 /// The sequence of steps:
@@ -241,7 +214,8 @@ impl DeleteTimelineFlow {
     ) -> Result<(), DeleteTimelineError> {
         super::debug_assert_current_span_has_tenant_and_timeline_id();
 
-        let (timeline, mut guard) = Self::prepare(tenant, timeline_id)?;
+        let allow_offloaded_children = false;
+        let (timeline, mut guard) = Self::prepare(tenant, timeline_id, allow_offloaded_children)?;
 
         guard.mark_in_progress()?;
 
@@ -258,7 +232,32 @@ impl DeleteTimelineFlow {
             ))?
         });
 
-        let remote_client = timeline.remote_client_maybe_construct(tenant);
+        let remote_client = match timeline.maybe_remote_client() {
+            Some(remote_client) => remote_client,
+            None => {
+                let remote_client = tenant
+                    .build_timeline_client(timeline.timeline_id(), tenant.remote_storage.clone());
+                let result = remote_client
+                    .download_index_file(&tenant.cancel)
+                    .instrument(info_span!("download_index_file"))
+                    .await
+                    .map_err(|e| DeleteTimelineError::Other(anyhow::anyhow!("error: {:?}", e)))?;
+                let index_part = match result {
+                    MaybeDeletedIndexPart::Deleted(p) => {
+                        tracing::info!("Timeline already set as deleted in remote index");
+                        p
+                    }
+                    MaybeDeletedIndexPart::IndexPart(p) => p,
+                };
+                let remote_client = Arc::new(remote_client);
+
+                remote_client
+                    .init_upload_queue(&index_part)
+                    .map_err(DeleteTimelineError::Other)?;
+                remote_client.shutdown().await;
+                remote_client
+            }
+        };
         set_deleted_in_remote_index(&remote_client).await?;
 
         fail::fail_point!("timeline-delete-before-schedule", |_| {
@@ -313,6 +312,7 @@ impl DeleteTimelineFlow {
                 // Important. We dont pass ancestor above because it can be missing.
                 // Thus we need to skip the validation here.
                 CreateTimelineCause::Delete,
+                crate::tenant::CreateTimelineIdempotency::FailWithConflict, // doesn't matter what we put here
             )
             .context("create_timeline_struct")?;
 
@@ -341,6 +341,7 @@ impl DeleteTimelineFlow {
     pub(super) fn prepare(
         tenant: &Tenant,
         timeline_id: TimelineId,
+        allow_offloaded_children: bool,
     ) -> Result<(TimelineOrOffloaded, DeletionGuard), DeleteTimelineError> {
         // Note the interaction between this guard and deletion guard.
         // Here we attempt to lock deletion guard when we're holding a lock on timelines.
@@ -353,30 +354,27 @@ impl DeleteTimelineFlow {
         // T1: acquire deletion lock, do another `DeleteTimelineFlow::run`
         // For more context see this discussion: `https://github.com/neondatabase/neon/pull/4552#discussion_r1253437346`
         let timelines = tenant.timelines.lock().unwrap();
+        let timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
 
         let timeline = match timelines.get(&timeline_id) {
             Some(t) => TimelineOrOffloaded::Timeline(Arc::clone(t)),
-            None => {
-                let offloaded_timelines = tenant.timelines_offloaded.lock().unwrap();
-                match offloaded_timelines.get(&timeline_id) {
-                    Some(t) => TimelineOrOffloaded::Offloaded(Arc::clone(t)),
-                    None => return Err(DeleteTimelineError::NotFound),
-                }
-            }
+            None => match timelines_offloaded.get(&timeline_id) {
+                Some(t) => TimelineOrOffloaded::Offloaded(Arc::clone(t)),
+                None => return Err(DeleteTimelineError::NotFound),
+            },
         };
 
-        // Ensure that there are no child timelines **attached to that pageserver**,
-        // because detach removes files, which will break child branches
-        let children: Vec<TimelineId> = timelines
-            .iter()
-            .filter_map(|(id, entry)| {
-                if entry.get_ancestor_timeline_id() == Some(timeline_id) {
-                    Some(*id)
-                } else {
-                    None
-                }
-            })
-            .collect();
+        // Ensure that there are no child timelines, because we are about to remove files,
+        // which will break child branches
+        let mut children = Vec::new();
+        if !allow_offloaded_children {
+            children.extend(timelines_offloaded.iter().filter_map(|(id, entry)| {
+                (entry.ancestor_timeline_id == Some(timeline_id)).then_some(*id)
+            }));
+        }
+        children.extend(timelines.iter().filter_map(|(id, entry)| {
+            (entry.get_ancestor_timeline_id() == Some(timeline_id)).then_some(*id)
+        }));
 
         if !children.is_empty() {
             return Err(DeleteTimelineError::HasChildren(children));
@@ -454,7 +452,15 @@ impl DeleteTimelineFlow {
 
         remove_maybe_offloaded_timeline_from_tenant(tenant, timeline, &guard).await?;
 
-        upload_new_tenant_manifest(tenant, &guard).await?;
+        // This is susceptible to race conditions, i.e. we won't continue deletions if there is a crash
+        // between the deletion of the index-part.json and reaching of this code.
+        // So indeed, the tenant manifest might refer to an offloaded timeline which has already been deleted.
+        // However, we handle this case in tenant loading code so the next time we attach, the issue is
+        // resolved.
+        tenant
+            .store_tenant_manifest()
+            .await
+            .map_err(|e| DeleteTimelineError::Other(anyhow::anyhow!(e)))?;
 
         *guard = Self::Finished;
 

pageserver/src/tenant/timeline/detach_ancestor.rs

@@ -29,6 +29,9 @@ pub(crate) enum Error {
     #[error("shutting down, please retry later")]
     ShuttingDown,
 
+    #[error("archived: {}", .0)]
+    Archived(TimelineId),
+
     #[error(transparent)]
     NotFound(crate::tenant::GetTimelineError),
 
@@ -79,8 +82,9 @@ impl From<Error> for ApiError {
     fn from(value: Error) -> Self {
         match value {
             Error::NoAncestor => ApiError::Conflict(value.to_string()),
-            Error::TooManyAncestors => ApiError::BadRequest(anyhow::anyhow!("{}", value)),
+            Error::TooManyAncestors => ApiError::BadRequest(anyhow::anyhow!("{value}")),
             Error::ShuttingDown => ApiError::ShuttingDown,
+            Error::Archived(_) => ApiError::BadRequest(anyhow::anyhow!("{value}")),
             Error::OtherTimelineDetachOngoing(_) | Error::FailedToReparentAll => {
                 ApiError::ResourceUnavailable(value.to_string().into())
             }
@@ -201,12 +205,18 @@ pub(super) async fn prepare(
         }));
     };
 
+    if detached.is_archived() != Some(false) {
+        return Err(Archived(detached.timeline_id));
+    }
+
     if !ancestor_lsn.is_valid() {
         // rare case, probably wouldn't even load
         tracing::error!("ancestor is set, but ancestor_lsn is invalid, this timeline needs fixing");
         return Err(NoAncestor);
     }
 
+    check_no_archived_children_of_ancestor(tenant, detached, &ancestor, ancestor_lsn)?;
+
     if ancestor.ancestor_timeline.is_some() {
         // non-technical requirement; we could flatten N ancestors just as easily but we chose
         // not to, at least initially
@@ -950,3 +960,36 @@ where
         }
     })
 }
+
+fn check_no_archived_children_of_ancestor(
+    tenant: &Tenant,
+    detached: &Arc<Timeline>,
+    ancestor: &Arc<Timeline>,
+    ancestor_lsn: Lsn,
+) -> Result<(), Error> {
+    let timelines = tenant.timelines.lock().unwrap();
+    let timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
+    for timeline in reparentable_timelines(timelines.values(), detached, ancestor, ancestor_lsn) {
+        if timeline.is_archived() == Some(true) {
+            return Err(Error::Archived(timeline.timeline_id));
+        }
+    }
+    for timeline_offloaded in timelines_offloaded.values() {
+        if timeline_offloaded.ancestor_timeline_id != Some(ancestor.timeline_id) {
+            continue;
+        }
+        // This forbids the detach ancestor feature if flattened timelines are present,
+        // even if the ancestor_lsn is from after the branchpoint of the detached timeline.
+        // But as per current design, we don't record the ancestor_lsn of flattened timelines.
+        // This is a bit unfortunate, but as of writing this we don't support flattening
+        // anyway. Maybe we can evolve the data model in the future.
+        if let Some(retain_lsn) = timeline_offloaded.ancestor_retain_lsn {
+            let is_earlier = retain_lsn <= ancestor_lsn;
+            if !is_earlier {
+                continue;
+            }
+        }
+        return Err(Error::Archived(timeline_offloaded.timeline_id));
+    }
+    Ok(())
+}

pageserver/src/tenant/timeline/layer_manager.rs

@@ -45,13 +45,16 @@ impl LayerManager {
     pub(crate) fn get_from_key(&self, key: &PersistentLayerKey) -> Layer {
         // The assumption for the `expect()` is that all code maintains the following invariant:
         // A layer's descriptor is present in the LayerMap => the LayerFileManager contains a layer for the descriptor.
-        self.layers()
-            .get(key)
+        self.try_get_from_key(key)
             .with_context(|| format!("get layer from key: {key}"))
             .expect("not found")
             .clone()
     }
 
+    pub(crate) fn try_get_from_key(&self, key: &PersistentLayerKey) -> Option<&Layer> {
+        self.layers().get(key)
+    }
+
     pub(crate) fn get_from_desc(&self, desc: &PersistentLayerDesc) -> Layer {
         self.get_from_key(&desc.key())
     }

pageserver/src/tenant/timeline/offload.rs

@@ -3,16 +3,40 @@ use std::sync::Arc;
 use super::delete::{delete_local_timeline_directory, DeleteTimelineFlow, DeletionGuard};
 use super::Timeline;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
-use crate::tenant::{remote_timeline_client, OffloadedTimeline, Tenant, TimelineOrOffloaded};
+use crate::tenant::{OffloadedTimeline, Tenant, TenantManifestError, TimelineOrOffloaded};
+
+#[derive(thiserror::Error, Debug)]
+pub(crate) enum OffloadError {
+    #[error("Cancelled")]
+    Cancelled,
+    #[error("Timeline is not archived")]
+    NotArchived,
+    #[error(transparent)]
+    RemoteStorage(anyhow::Error),
+    #[error("Unexpected offload error: {0}")]
+    Other(anyhow::Error),
+}
+
+impl From<TenantManifestError> for OffloadError {
+    fn from(e: TenantManifestError) -> Self {
+        match e {
+            TenantManifestError::Cancelled => Self::Cancelled,
+            TenantManifestError::RemoteStorage(e) => Self::RemoteStorage(e),
+        }
+    }
+}
 
 pub(crate) async fn offload_timeline(
     tenant: &Tenant,
     timeline: &Arc<Timeline>,
-) -> anyhow::Result<()> {
+) -> Result<(), OffloadError> {
     debug_assert_current_span_has_tenant_and_timeline_id();
     tracing::info!("offloading archived timeline");
 
-    let (timeline, guard) = DeleteTimelineFlow::prepare(tenant, timeline.timeline_id)?;
+    let allow_offloaded_children = true;
+    let (timeline, guard) =
+        DeleteTimelineFlow::prepare(tenant, timeline.timeline_id, allow_offloaded_children)
+            .map_err(|e| OffloadError::Other(anyhow::anyhow!(e)))?;
 
     let TimelineOrOffloaded::Timeline(timeline) = timeline else {
         tracing::error!("timeline already offloaded, but given timeline object");
@@ -24,14 +48,15 @@ pub(crate) async fn offload_timeline(
         Some(true) => (),
         Some(false) => {
             tracing::warn!(?is_archived, "tried offloading a non-archived timeline");
-            anyhow::bail!("timeline isn't archived");
+            return Err(OffloadError::NotArchived);
         }
         None => {
-            tracing::warn!(
+            // This is legal: calls to this function can race with the timeline shutting down
+            tracing::info!(
                 ?is_archived,
-                "tried offloading a timeline where manifest is not yet available"
+                "tried offloading a timeline whose remote storage is not initialized"
             );
-            anyhow::bail!("timeline manifest hasn't been loaded yet");
+            return Err(OffloadError::Cancelled);
         }
     }
 
@@ -42,9 +67,11 @@ pub(crate) async fn offload_timeline(
     // to make deletions possible while offloading is in progress
 
     let conf = &tenant.conf;
-    delete_local_timeline_directory(conf, tenant.tenant_shard_id, &timeline).await?;
+    delete_local_timeline_directory(conf, tenant.tenant_shard_id, &timeline)
+        .await
+        .map_err(OffloadError::Other)?;
 
-    remove_timeline_from_tenant(tenant, &timeline, &guard).await?;
+    remove_timeline_from_tenant(tenant, &timeline, &guard);
 
     {
         let mut offloaded_timelines = tenant.timelines_offloaded.lock().unwrap();
@@ -63,28 +90,18 @@ pub(crate) async fn offload_timeline(
     // at the next restart attach it again.
     // For that to happen, we'd need to make the manifest reflect our *intended* state,
     // not our actual state of offloaded timelines.
-    let manifest = tenant.tenant_manifest();
-    // TODO: generation support
-    let generation = remote_timeline_client::TENANT_MANIFEST_GENERATION;
-    remote_timeline_client::upload_tenant_manifest(
-        &tenant.remote_storage,
-        &tenant.tenant_shard_id,
-        generation,
-        &manifest,
-        &tenant.cancel,
-    )
-    .await?;
+    tenant.store_tenant_manifest().await?;
 
     Ok(())
 }
 
 /// It is important that this gets called when DeletionGuard is being held.
 /// For more context see comments in [`DeleteTimelineFlow::prepare`]
-async fn remove_timeline_from_tenant(
+fn remove_timeline_from_tenant(
     tenant: &Tenant,
     timeline: &Timeline,
     _: &DeletionGuard, // using it as a witness
-) -> anyhow::Result<()> {
+) {
     // Remove the timeline from the map.
     let mut timelines = tenant.timelines.lock().unwrap();
     let children_exist = timelines
@@ -100,8 +117,4 @@ async fn remove_timeline_from_tenant(
     timelines
         .remove(&timeline.timeline_id)
         .expect("timeline that we were deleting was concurrently removed from 'timelines' map");
-
-    drop(timelines);
-
-    Ok(())
 }

pageserver/src/tenant/timeline/uninit.rs

@@ -5,7 +5,11 @@ use camino::Utf8PathBuf;
 use tracing::{error, info, info_span};
 use utils::{fs_ext, id::TimelineId, lsn::Lsn};
 
-use crate::{context::RequestContext, import_datadir, tenant::Tenant};
+use crate::{
+    context::RequestContext,
+    import_datadir,
+    tenant::{CreateTimelineIdempotency, Tenant, TimelineOrOffloaded},
+};
 
 use super::Timeline;
 
@@ -137,7 +141,9 @@ impl Drop for UninitializedTimeline<'_> {
     fn drop(&mut self) {
         if let Some((_, create_guard)) = self.raw_timeline.take() {
             let _entered = info_span!("drop_uninitialized_timeline", tenant_id = %self.owning_tenant.tenant_shard_id.tenant_id, shard_id = %self.owning_tenant.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id).entered();
-            error!("Timeline got dropped without initializing, cleaning its files");
+            // This is unusual, but can happen harmlessly if the pageserver is stopped while
+            // creating a timeline.
+            info!("Timeline got dropped without initializing, cleaning its files");
             cleanup_timeline_directory(create_guard);
         }
     }
@@ -165,13 +171,17 @@ pub(crate) struct TimelineCreateGuard<'t> {
     owning_tenant: &'t Tenant,
     timeline_id: TimelineId,
     pub(crate) timeline_path: Utf8PathBuf,
+    pub(crate) idempotency: CreateTimelineIdempotency,
 }
 
 /// Errors when acquiring exclusive access to a timeline ID for creation
 #[derive(thiserror::Error, Debug)]
 pub(crate) enum TimelineExclusionError {
     #[error("Already exists")]
-    AlreadyExists(Arc<Timeline>),
+    AlreadyExists {
+        existing: TimelineOrOffloaded,
+        arg: CreateTimelineIdempotency,
+    },
     #[error("Already creating")]
     AlreadyCreating,
 
@@ -185,27 +195,42 @@ impl<'t> TimelineCreateGuard<'t> {
         owning_tenant: &'t Tenant,
         timeline_id: TimelineId,
         timeline_path: Utf8PathBuf,
+        idempotency: CreateTimelineIdempotency,
+        allow_offloaded: bool,
     ) -> Result<Self, TimelineExclusionError> {
         // Lock order: this is the only place we take both locks.  During drop() we only
         // lock creating_timelines
         let timelines = owning_tenant.timelines.lock().unwrap();
+        let timelines_offloaded = owning_tenant.timelines_offloaded.lock().unwrap();
         let mut creating_timelines: std::sync::MutexGuard<
             '_,
             std::collections::HashSet<TimelineId>,
         > = owning_tenant.timelines_creating.lock().unwrap();
 
         if let Some(existing) = timelines.get(&timeline_id) {
-            Err(TimelineExclusionError::AlreadyExists(existing.clone()))
-        } else if creating_timelines.contains(&timeline_id) {
-            Err(TimelineExclusionError::AlreadyCreating)
-        } else {
-            creating_timelines.insert(timeline_id);
-            Ok(Self {
-                owning_tenant,
-                timeline_id,
-                timeline_path,
-            })
+            return Err(TimelineExclusionError::AlreadyExists {
+                existing: TimelineOrOffloaded::Timeline(existing.clone()),
+                arg: idempotency,
+            });
         }
+        if !allow_offloaded {
+            if let Some(existing) = timelines_offloaded.get(&timeline_id) {
+                return Err(TimelineExclusionError::AlreadyExists {
+                    existing: TimelineOrOffloaded::Offloaded(existing.clone()),
+                    arg: idempotency,
+                });
+            }
+        }
+        if creating_timelines.contains(&timeline_id) {
+            return Err(TimelineExclusionError::AlreadyCreating);
+        }
+        creating_timelines.insert(timeline_id);
+        Ok(Self {
+            owning_tenant,
+            timeline_id,
+            timeline_path,
+            idempotency,
+        })
     }
 }
 

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

@@ -22,6 +22,7 @@ use tokio::{select, sync::watch, time};
 use tokio_postgres::{replication::ReplicationStream, Client};
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info, trace, warn, Instrument};
+use wal_decoder::models::{FlushUncommittedRecords, InterpretedWalRecord};
 
 use super::TaskStateUpdate;
 use crate::{
@@ -31,7 +32,6 @@ use crate::{
     task_mgr::{TaskKind, WALRECEIVER_RUNTIME},
     tenant::{debug_assert_current_span_has_tenant_and_timeline_id, Timeline, WalReceiverInfo},
     walingest::WalIngest,
-    walrecord::{decode_wal_record, DecodedWALRecord},
 };
 use postgres_backend::is_expected_io_error;
 use postgres_connection::PgConnectionConfig;
@@ -339,11 +339,15 @@ pub(super) async fn handle_walreceiver_connection(
                             return Err(WalReceiverError::Other(anyhow!("LSN not aligned")));
                         }
 
-                        // Deserialize WAL record
-                        let mut decoded = DecodedWALRecord::default();
-                        decode_wal_record(recdata, &mut decoded, modification.tline.pg_version)?;
+                        // Deserialize and interpret WAL record
+                        let interpreted = InterpretedWalRecord::from_bytes_filtered(
+                            recdata,
+                            modification.tline.get_shard_identity(),
+                            lsn,
+                            modification.tline.pg_version,
+                        )?;
 
-                        if decoded.is_dbase_create_copy(timeline.pg_version)
+                        if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes)
                             && uncommitted_records > 0
                         {
                             // Special case: legacy PG database creations operate by reading pages from a 'template' database:
@@ -360,7 +364,7 @@ pub(super) async fn handle_walreceiver_connection(
 
                         // Ingest the records without immediately committing them.
                         let ingested = walingest
-                            .ingest_record(decoded, lsn, &mut modification, &ctx)
+                            .ingest_record(interpreted, &mut modification, &ctx)
                             .await
                             .with_context(|| format!("could not ingest record at {lsn}"))?;
                         if !ingested {

pageserver/src/virtual_file/io_engine/tokio_epoll_uring_ext.rs

@@ -16,18 +16,24 @@ use tokio_epoll_uring::{System, SystemHandle};
 
 use crate::virtual_file::on_fatal_io_error;
 
-use crate::metrics::tokio_epoll_uring as metrics;
+use crate::metrics::tokio_epoll_uring::{self as metrics, THREAD_LOCAL_METRICS_STORAGE};
 
 #[derive(Clone)]
 struct ThreadLocalState(Arc<ThreadLocalStateInner>);
 
 struct ThreadLocalStateInner {
-    cell: tokio::sync::OnceCell<SystemHandle>,
+    cell: tokio::sync::OnceCell<SystemHandle<metrics::ThreadLocalMetrics>>,
     launch_attempts: AtomicU32,
     /// populated through fetch_add from [`THREAD_LOCAL_STATE_ID`]
     thread_local_state_id: u64,
 }
 
+impl Drop for ThreadLocalStateInner {
+    fn drop(&mut self) {
+        THREAD_LOCAL_METRICS_STORAGE.remove_system(self.thread_local_state_id);
+    }
+}
+
 impl ThreadLocalState {
     pub fn new() -> Self {
         Self(Arc::new(ThreadLocalStateInner {
@@ -71,7 +77,8 @@ pub async fn thread_local_system() -> Handle {
                         &fake_cancel,
                     )
                     .await;
-                    let res = System::launch()
+                    let per_system_metrics = metrics::THREAD_LOCAL_METRICS_STORAGE.register_system(inner.thread_local_state_id);
+                    let res = System::launch_with_metrics(per_system_metrics)
                     // this might move us to another executor thread => loop outside the get_or_try_init, not inside it
                     .await;
                     match res {
@@ -86,6 +93,7 @@ pub async fn thread_local_system() -> Handle {
                                 emit_launch_failure_process_stats();
                             });
                             metrics::THREAD_LOCAL_LAUNCH_FAILURES.inc();
+                            metrics::THREAD_LOCAL_METRICS_STORAGE.remove_system(inner.thread_local_state_id);
                             Err(())
                         }
                         // abort the process instead of panicking because pageserver usually becomes half-broken if we panic somewhere.
@@ -115,7 +123,7 @@ fn emit_launch_failure_process_stats() {
     // number of threads
     // rss / system memory usage generally
 
-    let tokio_epoll_uring::metrics::Metrics {
+    let tokio_epoll_uring::metrics::GlobalMetrics {
         systems_created,
         systems_destroyed,
     } = tokio_epoll_uring::metrics::global();
@@ -182,7 +190,7 @@ fn emit_launch_failure_process_stats() {
 pub struct Handle(ThreadLocalState);
 
 impl std::ops::Deref for Handle {
-    type Target = SystemHandle;
+    type Target = SystemHandle<metrics::ThreadLocalMetrics>;
 
     fn deref(&self) -> &Self::Target {
         self.0

pageserver/src/walredo.rs

@@ -29,11 +29,11 @@ use crate::metrics::{
     WAL_REDO_BYTES_HISTOGRAM, WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM,
     WAL_REDO_RECORDS_HISTOGRAM, WAL_REDO_TIME,
 };
-use crate::repository::Key;
-use crate::walrecord::NeonWalRecord;
 use anyhow::Context;
 use bytes::{Bytes, BytesMut};
+use pageserver_api::key::Key;
 use pageserver_api::models::{WalRedoManagerProcessStatus, WalRedoManagerStatus};
+use pageserver_api::record::NeonWalRecord;
 use pageserver_api::shard::TenantShardId;
 use std::future::Future;
 use std::sync::Arc;
@@ -548,9 +548,10 @@ impl PostgresRedoManager {
 #[cfg(test)]
 mod tests {
     use super::PostgresRedoManager;
-    use crate::repository::Key;
-    use crate::{config::PageServerConf, walrecord::NeonWalRecord};
+    use crate::config::PageServerConf;
     use bytes::Bytes;
+    use pageserver_api::key::Key;
+    use pageserver_api::record::NeonWalRecord;
     use pageserver_api::shard::TenantShardId;
     use std::str::FromStr;
     use tracing::Instrument;

pageserver/src/walredo/apply_neon.rs

@@ -1,8 +1,8 @@
-use crate::walrecord::NeonWalRecord;
 use anyhow::Context;
 use byteorder::{ByteOrder, LittleEndian};
 use bytes::BytesMut;
 use pageserver_api::key::Key;
+use pageserver_api::record::NeonWalRecord;
 use pageserver_api::reltag::SlruKind;
 use postgres_ffi::pg_constants;
 use postgres_ffi::relfile_utils::VISIBILITYMAP_FORKNUM;
@@ -238,7 +238,7 @@ pub(crate) fn apply_in_neon(
             // No-op: this record will never be created in aux v2.
             warn!("AuxFile record should not be created in aux v2");
         }
-        #[cfg(test)]
+        #[cfg(feature = "testing")]
         NeonWalRecord::Test {
             append,
             clear,

pageserver/src/walredo/process.rs

@@ -8,10 +8,10 @@ use crate::{
     metrics::{WalRedoKillCause, WAL_REDO_PROCESS_COUNTERS, WAL_REDO_RECORD_COUNTER},
     page_cache::PAGE_SZ,
     span::debug_assert_current_span_has_tenant_id,
-    walrecord::NeonWalRecord,
 };
 use anyhow::Context;
 use bytes::Bytes;
+use pageserver_api::record::NeonWalRecord;
 use pageserver_api::{reltag::RelTag, shard::TenantShardId};
 use postgres_ffi::BLCKSZ;
 #[cfg(feature = "testing")]

pgxn/neon/Makefile

@@ -8,6 +8,7 @@ OBJS = \
 	file_cache.o \
 	hll.o \
 	libpagestore.o \
+	logical_replication_monitor.o \
 	neon.o \
 	neon_pgversioncompat.o \
 	neon_perf_counters.o \
@@ -15,6 +16,7 @@ OBJS = \
 	neon_walreader.o \
 	pagestore_smgr.o \
 	relsize_cache.o \
+	unstable_extensions.o \
 	walproposer.o \
 	walproposer_pg.o \
 	control_plane_connector.o \

pgxn/neon/control_plane_connector.c

@@ -18,6 +18,7 @@
  *
  *-------------------------------------------------------------------------
  */
+
 #include "postgres.h"
 
 #include <curl/curl.h>
@@ -508,6 +509,8 @@ NeonXactCallback(XactEvent event, void *arg)
 static bool
 RoleIsNeonSuperuser(const char *role_name)
 {
+	Assert(role_name);
+
 	return strcmp(role_name, "neon_superuser") == 0;
 }
 
@@ -670,7 +673,7 @@ HandleCreateRole(CreateRoleStmt *stmt)
 static void
 HandleAlterRole(AlterRoleStmt *stmt)
 {
-	const char *role_name = stmt->role->rolename;
+	char	   *role_name;
 	DefElem    *dpass;
 	ListCell   *option;
 	bool		found = false;
@@ -678,6 +681,7 @@ HandleAlterRole(AlterRoleStmt *stmt)
 
 	InitRoleTableIfNeeded();
 
+	role_name = get_rolespec_name(stmt->role);
 	if (RoleIsNeonSuperuser(role_name) && !superuser())
 		elog(ERROR, "can't ALTER neon_superuser");
 
@@ -689,9 +693,13 @@ HandleAlterRole(AlterRoleStmt *stmt)
 		if (strcmp(defel->defname, "password") == 0)
 			dpass = defel;
 	}
+
 	/* We only care about updates to the password */
 	if (!dpass)
+	{
+		pfree(role_name);
 		return;
+	}
 
 	entry = hash_search(CurrentDdlTable->role_table,
 						role_name,
@@ -704,6 +712,8 @@ HandleAlterRole(AlterRoleStmt *stmt)
 	else
 		entry->password = NULL;
 	entry->type = Op_Set;
+
+	pfree(role_name);
 }
 
 static void

pgxn/neon/logical_replication_monitor.c

@@ -0,0 +1,253 @@
+#include <limits.h>
+#include <string.h>
+#include <dirent.h>
+#include <signal.h>
+
+#include "postgres.h"
+
+#include "miscadmin.h"
+#include "postmaster/bgworker.h"
+#include "postmaster/interrupt.h"
+#include "replication/slot.h"
+#include "storage/fd.h"
+#include "storage/procsignal.h"
+#include "tcop/tcopprot.h"
+#include "utils/guc.h"
+#include "utils/wait_event.h"
+
+#include "logical_replication_monitor.h"
+
+#define LS_MONITOR_CHECK_INTERVAL 10000 /* ms */
+
+static int	logical_replication_max_snap_files = 300;
+
+PGDLLEXPORT void LogicalSlotsMonitorMain(Datum main_arg);
+
+static int
+LsnDescComparator(const void *a, const void *b)
+{
+	XLogRecPtr	lsn1 = *((const XLogRecPtr *) a);
+	XLogRecPtr	lsn2 = *((const XLogRecPtr *) b);
+
+	if (lsn1 < lsn2)
+		return 1;
+	else if (lsn1 == lsn2)
+		return 0;
+	else
+		return -1;
+}
+
+/*
+ * Look at .snap files and calculate minimum allowed restart_lsn of slot so that
+ * next gc would leave not more than logical_replication_max_snap_files; all
+ * slots having lower restart_lsn should be dropped.
+ */
+static XLogRecPtr
+get_num_snap_files_lsn_threshold(void)
+{
+	DIR		   *dirdesc;
+	struct dirent *de;
+	char	   *snap_path = "pg_logical/snapshots/";
+	int			lsns_allocated = 1024;
+	int			lsns_num = 0;
+	XLogRecPtr *lsns;
+	XLogRecPtr	cutoff;
+
+	if (logical_replication_max_snap_files < 0)
+		return 0;
+
+	lsns = palloc(sizeof(XLogRecPtr) * lsns_allocated);
+
+	/* find all .snap files and get their lsns */
+	dirdesc = AllocateDir(snap_path);
+	while ((de = ReadDir(dirdesc, snap_path)) != NULL)
+	{
+		XLogRecPtr	lsn;
+		uint32		hi;
+		uint32		lo;
+
+		if (strcmp(de->d_name, ".") == 0 ||
+			strcmp(de->d_name, "..") == 0)
+			continue;
+
+		if (sscanf(de->d_name, "%X-%X.snap", &hi, &lo) != 2)
+		{
+			ereport(LOG,
+					(errmsg("could not parse file name as .snap file \"%s\"", de->d_name)));
+			continue;
+		}
+
+		lsn = ((uint64) hi) << 32 | lo;
+		elog(DEBUG5, "found snap file %X/%X", LSN_FORMAT_ARGS(lsn));
+		if (lsns_allocated == lsns_num)
+		{
+			lsns_allocated *= 2;
+			lsns = repalloc(lsns, sizeof(XLogRecPtr) * lsns_allocated);
+		}
+		lsns[lsns_num++] = lsn;
+	}
+	/* sort by lsn desc */
+	qsort(lsns, lsns_num, sizeof(XLogRecPtr), LsnDescComparator);
+	/* and take cutoff at logical_replication_max_snap_files */
+	if (logical_replication_max_snap_files > lsns_num)
+		cutoff = 0;
+	/* have less files than cutoff */
+	else
+	{
+		cutoff = lsns[logical_replication_max_snap_files - 1];
+		elog(LOG, "ls_monitor: dropping logical slots with restart_lsn lower %X/%X, found %d .snap files, limit is %d",
+			 LSN_FORMAT_ARGS(cutoff), lsns_num, logical_replication_max_snap_files);
+	}
+	pfree(lsns);
+	FreeDir(dirdesc);
+	return cutoff;
+}
+
+void
+InitLogicalReplicationMonitor(void)
+{
+	BackgroundWorker bgw;
+
+	DefineCustomIntVariable(
+							"neon.logical_replication_max_snap_files",
+							"Maximum allowed logical replication .snap files. When exceeded, slots are dropped until the limit is met. -1 disables the limit.",
+							NULL,
+							&logical_replication_max_snap_files,
+							300, -1, INT_MAX,
+							PGC_SIGHUP,
+							0,
+							NULL, NULL, NULL);
+
+	memset(&bgw, 0, sizeof(bgw));
+	bgw.bgw_flags = BGWORKER_SHMEM_ACCESS;
+	bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
+	snprintf(bgw.bgw_library_name, BGW_MAXLEN, "neon");
+	snprintf(bgw.bgw_function_name, BGW_MAXLEN, "LogicalSlotsMonitorMain");
+	snprintf(bgw.bgw_name, BGW_MAXLEN, "Logical replication monitor");
+	snprintf(bgw.bgw_type, BGW_MAXLEN, "Logical replication monitor");
+	bgw.bgw_restart_time = 5;
+	bgw.bgw_notify_pid = 0;
+	bgw.bgw_main_arg = (Datum) 0;
+
+	RegisterBackgroundWorker(&bgw);
+}
+
+/*
+ * Unused logical replication slots pins WAL and prevents deletion of snapshots.
+ * WAL bloat is guarded by max_slot_wal_keep_size; this bgw removes slots which
+ * need too many .snap files.
+ */
+void
+LogicalSlotsMonitorMain(Datum main_arg)
+{
+	/* Establish signal handlers. */
+	pqsignal(SIGUSR1, procsignal_sigusr1_handler);
+	pqsignal(SIGHUP, SignalHandlerForConfigReload);
+	pqsignal(SIGTERM, die);
+
+	BackgroundWorkerUnblockSignals();
+
+	for (;;)
+	{
+		XLogRecPtr	cutoff_lsn;
+
+		/* In case of a SIGHUP, just reload the configuration. */
+		if (ConfigReloadPending)
+		{
+			ConfigReloadPending = false;
+			ProcessConfigFile(PGC_SIGHUP);
+		}
+
+		/*
+		 * If there are too many .snap files, just drop all logical slots to
+		 * prevent aux files bloat.
+		 */
+		cutoff_lsn = get_num_snap_files_lsn_threshold();
+		if (cutoff_lsn > 0)
+		{
+			for (int i = 0; i < max_replication_slots; i++)
+			{
+				char		slot_name[NAMEDATALEN];
+				ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
+				XLogRecPtr	restart_lsn;
+
+				/* find the name */
+				LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
+				/* Consider only logical repliction slots */
+				if (!s->in_use || !SlotIsLogical(s))
+				{
+					LWLockRelease(ReplicationSlotControlLock);
+					continue;
+				}
+
+				/* do we need to drop it? */
+				SpinLockAcquire(&s->mutex);
+				restart_lsn = s->data.restart_lsn;
+				SpinLockRelease(&s->mutex);
+				if (restart_lsn >= cutoff_lsn)
+				{
+					LWLockRelease(ReplicationSlotControlLock);
+					continue;
+				}
+
+				strlcpy(slot_name, s->data.name.data, NAMEDATALEN);
+				elog(LOG, "ls_monitor: dropping slot %s with restart_lsn %X/%X below horizon %X/%X",
+					 slot_name, LSN_FORMAT_ARGS(restart_lsn), LSN_FORMAT_ARGS(cutoff_lsn));
+				LWLockRelease(ReplicationSlotControlLock);
+
+				/* now try to drop it, killing owner before if any */
+				for (;;)
+				{
+					pid_t		active_pid;
+
+					SpinLockAcquire(&s->mutex);
+					active_pid = s->active_pid;
+					SpinLockRelease(&s->mutex);
+
+					if (active_pid == 0)
+					{
+						/*
+						 * Slot is releasted, try to drop it. Though of course
+						 * it could have been reacquired, so drop can ERROR
+						 * out. Similarly it could have been dropped in the
+						 * meanwhile.
+						 *
+						 * In principle we could remove pg_try/pg_catch, that
+						 * would restart the whole bgworker.
+						 */
+						ConditionVariableCancelSleep();
+						PG_TRY();
+						{
+							ReplicationSlotDrop(slot_name, true);
+							elog(LOG, "ls_monitor: slot %s dropped", slot_name);
+						}
+						PG_CATCH();
+						{
+							/* log ERROR and reset elog stack */
+							EmitErrorReport();
+							FlushErrorState();
+							elog(LOG, "ls_monitor: failed to drop slot %s", slot_name);
+						}
+						PG_END_TRY();
+						break;
+					}
+					else
+					{
+						/* kill the owner and wait for release */
+						elog(LOG, "ls_monitor: killing slot %s owner %d", slot_name, active_pid);
+						(void) kill(active_pid, SIGTERM);
+						/* We shouldn't get stuck, but to be safe add timeout. */
+						ConditionVariableTimedSleep(&s->active_cv, 1000, WAIT_EVENT_REPLICATION_SLOT_DROP);
+					}
+				}
+			}
+		}
+
+		(void) WaitLatch(MyLatch,
+						 WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT,
+						 LS_MONITOR_CHECK_INTERVAL,
+						 PG_WAIT_EXTENSION);
+		ResetLatch(MyLatch);
+		CHECK_FOR_INTERRUPTS();
+	}
+}

pgxn/neon/logical_replication_monitor.h

@@ -0,0 +1,6 @@
+#ifndef __NEON_LOGICAL_REPLICATION_MONITOR_H__
+#define __NEON_LOGICAL_REPLICATION_MONITOR_H__
+
+void InitLogicalReplicationMonitor(void);
+
+#endif

pgxn/neon/neon.c

@@ -14,32 +14,23 @@
 #include "miscadmin.h"
 #include "access/subtrans.h"
 #include "access/twophase.h"
-#include "access/xact.h"
 #include "access/xlog.h"
-#include "storage/buf_internals.h"
-#include "storage/bufmgr.h"
-#include "catalog/pg_type.h"
-#include "postmaster/bgworker.h"
-#include "postmaster/interrupt.h"
 #include "replication/logical.h"
 #include "replication/slot.h"
 #include "replication/walsender.h"
 #include "storage/proc.h"
-#include "storage/procsignal.h"
-#include "tcop/tcopprot.h"
 #include "funcapi.h"
 #include "access/htup_details.h"
 #include "utils/builtins.h"
 #include "utils/pg_lsn.h"
 #include "utils/guc.h"
 #include "utils/guc_tables.h"
-#include "utils/wait_event.h"
 
 #include "extension_server.h"
 #include "neon.h"
-#include "walproposer.h"
-#include "pagestore_client.h"
 #include "control_plane_connector.h"
+#include "logical_replication_monitor.h"
+#include "unstable_extensions.h"
 #include "walsender_hooks.h"
 #if PG_MAJORVERSION_NUM >= 16
 #include "storage/ipc.h"
@@ -48,7 +39,6 @@
 PG_MODULE_MAGIC;
 void		_PG_init(void);
 
-static int	logical_replication_max_snap_files = 300;
 
 static int  running_xacts_overflow_policy;
 
@@ -82,237 +72,6 @@ static const struct config_enum_entry running_xacts_overflow_policies[] = {
 	{NULL, 0, false}
 };
 
-static void
-InitLogicalReplicationMonitor(void)
-{
-	BackgroundWorker bgw;
-
-	DefineCustomIntVariable(
-							"neon.logical_replication_max_snap_files",
-							"Maximum allowed logical replication .snap files. When exceeded, slots are dropped until the limit is met. -1 disables the limit.",
-							NULL,
-							&logical_replication_max_snap_files,
-							300, -1, INT_MAX,
-							PGC_SIGHUP,
-							0,
-							NULL, NULL, NULL);
-
-	memset(&bgw, 0, sizeof(bgw));
-	bgw.bgw_flags = BGWORKER_SHMEM_ACCESS;
-	bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
-	snprintf(bgw.bgw_library_name, BGW_MAXLEN, "neon");
-	snprintf(bgw.bgw_function_name, BGW_MAXLEN, "LogicalSlotsMonitorMain");
-	snprintf(bgw.bgw_name, BGW_MAXLEN, "Logical replication monitor");
-	snprintf(bgw.bgw_type, BGW_MAXLEN, "Logical replication monitor");
-	bgw.bgw_restart_time = 5;
-	bgw.bgw_notify_pid = 0;
-	bgw.bgw_main_arg = (Datum) 0;
-
-	RegisterBackgroundWorker(&bgw);
-}
-
-static int
-LsnDescComparator(const void *a, const void *b)
-{
-	XLogRecPtr	lsn1 = *((const XLogRecPtr *) a);
-	XLogRecPtr	lsn2 = *((const XLogRecPtr *) b);
-
-	if (lsn1 < lsn2)
-		return 1;
-	else if (lsn1 == lsn2)
-		return 0;
-	else
-		return -1;
-}
-
-/*
- * Look at .snap files and calculate minimum allowed restart_lsn of slot so that
- * next gc would leave not more than logical_replication_max_snap_files; all
- * slots having lower restart_lsn should be dropped.
- */
-static XLogRecPtr
-get_num_snap_files_lsn_threshold(void)
-{
-	DIR		   *dirdesc;
-	struct dirent *de;
-	char	   *snap_path = "pg_logical/snapshots/";
-	int			lsns_allocated = 1024;
-	int			lsns_num = 0;
-	XLogRecPtr *lsns;
-	XLogRecPtr	cutoff;
-
-	if (logical_replication_max_snap_files < 0)
-		return 0;
-
-	lsns = palloc(sizeof(XLogRecPtr) * lsns_allocated);
-
-	/* find all .snap files and get their lsns */
-	dirdesc = AllocateDir(snap_path);
-	while ((de = ReadDir(dirdesc, snap_path)) != NULL)
-	{
-		XLogRecPtr	lsn;
-		uint32		hi;
-		uint32		lo;
-
-		if (strcmp(de->d_name, ".") == 0 ||
-			strcmp(de->d_name, "..") == 0)
-			continue;
-
-		if (sscanf(de->d_name, "%X-%X.snap", &hi, &lo) != 2)
-		{
-			ereport(LOG,
-					(errmsg("could not parse file name as .snap file \"%s\"", de->d_name)));
-			continue;
-		}
-
-		lsn = ((uint64) hi) << 32 | lo;
-		elog(DEBUG5, "found snap file %X/%X", LSN_FORMAT_ARGS(lsn));
-		if (lsns_allocated == lsns_num)
-		{
-			lsns_allocated *= 2;
-			lsns = repalloc(lsns, sizeof(XLogRecPtr) * lsns_allocated);
-		}
-		lsns[lsns_num++] = lsn;
-	}
-	/* sort by lsn desc */
-	qsort(lsns, lsns_num, sizeof(XLogRecPtr), LsnDescComparator);
-	/* and take cutoff at logical_replication_max_snap_files */
-	if (logical_replication_max_snap_files > lsns_num)
-		cutoff = 0;
-	/* have less files than cutoff */
-	else
-	{
-		cutoff = lsns[logical_replication_max_snap_files - 1];
-		elog(LOG, "ls_monitor: dropping logical slots with restart_lsn lower %X/%X, found %d .snap files, limit is %d",
-			 LSN_FORMAT_ARGS(cutoff), lsns_num, logical_replication_max_snap_files);
-	}
-	pfree(lsns);
-	FreeDir(dirdesc);
-	return cutoff;
-}
-
-#define LS_MONITOR_CHECK_INTERVAL 10000 /* ms */
-
-/*
- * Unused logical replication slots pins WAL and prevents deletion of snapshots.
- * WAL bloat is guarded by max_slot_wal_keep_size; this bgw removes slots which
- * need too many .snap files.
- */
-PGDLLEXPORT void
-LogicalSlotsMonitorMain(Datum main_arg)
-{
-	/* Establish signal handlers. */
-	pqsignal(SIGUSR1, procsignal_sigusr1_handler);
-	pqsignal(SIGHUP, SignalHandlerForConfigReload);
-	pqsignal(SIGTERM, die);
-
-	BackgroundWorkerUnblockSignals();
-
-	for (;;)
-	{
-		XLogRecPtr	cutoff_lsn;
-
-		/* In case of a SIGHUP, just reload the configuration. */
-		if (ConfigReloadPending)
-		{
-			ConfigReloadPending = false;
-			ProcessConfigFile(PGC_SIGHUP);
-		}
-
-		/*
-		 * If there are too many .snap files, just drop all logical slots to
-		 * prevent aux files bloat.
-		 */
-		cutoff_lsn = get_num_snap_files_lsn_threshold();
-		if (cutoff_lsn > 0)
-		{
-			for (int i = 0; i < max_replication_slots; i++)
-			{
-				char		slot_name[NAMEDATALEN];
-				ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
-				XLogRecPtr	restart_lsn;
-
-				/* find the name */
-				LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
-				/* Consider only logical repliction slots */
-				if (!s->in_use || !SlotIsLogical(s))
-				{
-					LWLockRelease(ReplicationSlotControlLock);
-					continue;
-				}
-
-				/* do we need to drop it? */
-				SpinLockAcquire(&s->mutex);
-				restart_lsn = s->data.restart_lsn;
-				SpinLockRelease(&s->mutex);
-				if (restart_lsn >= cutoff_lsn)
-				{
-					LWLockRelease(ReplicationSlotControlLock);
-					continue;
-				}
-
-				strlcpy(slot_name, s->data.name.data, NAMEDATALEN);
-				elog(LOG, "ls_monitor: dropping slot %s with restart_lsn %X/%X below horizon %X/%X",
-					 slot_name, LSN_FORMAT_ARGS(restart_lsn), LSN_FORMAT_ARGS(cutoff_lsn));
-				LWLockRelease(ReplicationSlotControlLock);
-
-				/* now try to drop it, killing owner before if any */
-				for (;;)
-				{
-					pid_t		active_pid;
-
-					SpinLockAcquire(&s->mutex);
-					active_pid = s->active_pid;
-					SpinLockRelease(&s->mutex);
-
-					if (active_pid == 0)
-					{
-						/*
-						 * Slot is releasted, try to drop it. Though of course
-						 * it could have been reacquired, so drop can ERROR
-						 * out. Similarly it could have been dropped in the
-						 * meanwhile.
-						 *
-						 * In principle we could remove pg_try/pg_catch, that
-						 * would restart the whole bgworker.
-						 */
-						ConditionVariableCancelSleep();
-						PG_TRY();
-						{
-							ReplicationSlotDrop(slot_name, true);
-							elog(LOG, "ls_monitor: slot %s dropped", slot_name);
-						}
-						PG_CATCH();
-						{
-							/* log ERROR and reset elog stack */
-							EmitErrorReport();
-							FlushErrorState();
-							elog(LOG, "ls_monitor: failed to drop slot %s", slot_name);
-						}
-						PG_END_TRY();
-						break;
-					}
-					else
-					{
-						/* kill the owner and wait for release */
-						elog(LOG, "ls_monitor: killing slot %s owner %d", slot_name, active_pid);
-						(void) kill(active_pid, SIGTERM);
-						/* We shouldn't get stuck, but to be safe add timeout. */
-						ConditionVariableTimedSleep(&s->active_cv, 1000, WAIT_EVENT_REPLICATION_SLOT_DROP);
-					}
-				}
-			}
-		}
-
-		(void) WaitLatch(MyLatch,
-						 WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT,
-						 LS_MONITOR_CHECK_INTERVAL,
-						 PG_WAIT_EXTENSION);
-		ResetLatch(MyLatch);
-		CHECK_FOR_INTERRUPTS();
-	}
-}
-
 /*
  * XXX: These private to procarray.c, but we need them here.
  */
@@ -666,8 +425,8 @@ _PG_init(void)
 	LogicalFuncs_Custom_XLogReaderRoutines = NeonOnDemandXLogReaderRoutines;
 	SlotFuncs_Custom_XLogReaderRoutines = NeonOnDemandXLogReaderRoutines;
 
+	InitUnstableExtensionsSupport();
 	InitLogicalReplicationMonitor();
-
 	InitControlPlaneConnector();
 
 	pg_init_extension_server();

pgxn/neon/neon_pgversioncompat.c

@@ -42,3 +42,4 @@ InitMaterializedSRF(FunctionCallInfo fcinfo, bits32 flags)
 	MemoryContextSwitchTo(old_context);
 }
 #endif
+

pgxn/neon/unstable_extensions.c

@@ -0,0 +1,129 @@
+#include <stdlib.h>
+#include <string.h>
+
+#include "postgres.h"
+
+#include "nodes/plannodes.h"
+#include "nodes/parsenodes.h"
+#include "tcop/utility.h"
+#include "utils/errcodes.h"
+#include "utils/guc.h"
+
+#include "neon_pgversioncompat.h"
+#include "unstable_extensions.h"
+
+static bool					allow_unstable_extensions = false;
+static char				   *unstable_extensions = NULL;
+
+static ProcessUtility_hook_type PreviousProcessUtilityHook = NULL;
+
+static bool
+list_contains(char const* comma_separated_list, char const* val)
+{
+	char const* occ = comma_separated_list;
+	size_t val_len = strlen(val);
+
+	if (val_len == 0)
+		return false;
+
+	while ((occ = strstr(occ, val)) != NULL)
+	{
+		if ((occ == comma_separated_list || occ[-1] == ',')
+			&& (occ[val_len] == '\0' || occ[val_len] == ','))
+		{
+			return true;
+		}
+		occ += val_len;
+	}
+
+	return false;
+}
+
+
+static void
+CheckUnstableExtension(
+	PlannedStmt *pstmt,
+	const char *queryString,
+	bool readOnlyTree,
+	ProcessUtilityContext context,
+	ParamListInfo params,
+	QueryEnvironment *queryEnv,
+	DestReceiver *dest,
+	QueryCompletion *qc)
+{
+	Node	   *parseTree = pstmt->utilityStmt;
+
+	if (allow_unstable_extensions || unstable_extensions == NULL)
+		goto process;
+
+	switch (nodeTag(parseTree))
+	{
+		case T_CreateExtensionStmt:
+		{
+			CreateExtensionStmt *stmt = castNode(CreateExtensionStmt, parseTree);
+			if (list_contains(unstable_extensions, stmt->extname))
+			{
+				ereport(ERROR,
+						(errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
+						 errmsg("%s extension is in beta and may be unstable or introduce backward-incompatible changes.\nWe recommend testing it in a separate, dedicated Neon project.", stmt->extname),
+						 errhint("to proceed with installation, run SET neon.allow_unstable_extensions='true'")));
+			}
+			break;
+		}
+		default:
+			goto process;
+	}
+
+process:
+	if (PreviousProcessUtilityHook)
+	{
+		PreviousProcessUtilityHook(
+			pstmt,
+			queryString,
+			readOnlyTree,
+			context,
+			params,
+			queryEnv,
+			dest,
+			qc);
+	}
+	else
+	{
+		standard_ProcessUtility(
+			pstmt,
+			queryString,
+			readOnlyTree,
+			context,
+			params,
+			queryEnv,
+			dest,
+			qc);
+	}
+}
+
+void
+InitUnstableExtensionsSupport(void)
+{
+	DefineCustomBoolVariable(
+		"neon.allow_unstable_extensions",
+		"Allow unstable extensions to be installed and used",
+		NULL,
+		&allow_unstable_extensions,
+		false,
+		PGC_USERSET,
+		0,
+		NULL, NULL, NULL);
+
+	DefineCustomStringVariable(
+		"neon.unstable_extensions",
+		"List of unstable extensions",
+		NULL,
+		&unstable_extensions,
+		NULL,
+		PGC_SUSET,
+		0,
+		NULL, NULL, NULL);
+
+	PreviousProcessUtilityHook = ProcessUtility_hook;
+	ProcessUtility_hook = CheckUnstableExtension;
+}