Set NEON_MOTD to null to fix regression tests

NEON_MOTD env variable is set for psql greetings to be printed
alogn with the session_id. Disable this for regression tests

.config/hakari.toml

@@ -21,13 +21,14 @@ platforms = [
     # "x86_64-apple-darwin",
     # "x86_64-pc-windows-msvc",
 ]
-
 [final-excludes]
 workspace-members = [
     # vm_monitor benefits from the same Cargo.lock as the rest of our artifacts, but
     # it is built primarly in separate repo neondatabase/autoscaling and thus is excluded
     # from depending on workspace-hack because most of the dependencies are not used.
     "vm_monitor",
+    # subzero-core is a stub crate that should be excluded from workspace-hack
+    "subzero-core",
     # All of these exist in libs and are not usually built independently.
     # Putting workspace hack there adds a bottleneck for cargo builds.
     "compute_api",

.github/actions/prepare-for-subzero/action.yml

@@ -0,0 +1,28 @@
+name: 'Prepare current job for subzero'
+description: >
+  Set git token to access `neondatabase/subzero` from cargo build,
+  and set `CARGO_NET_GIT_FETCH_WITH_CLI=true` env variable to use git CLI
+
+inputs:
+  token:
+    description: 'GitHub token with access to neondatabase/subzero'
+    required: true
+
+runs:
+  using: "composite"
+
+  steps:
+    - name: Set git token for neondatabase/subzero
+      uses: pyTooling/Actions/with-post-step@2307b526df64d55e95884e072e49aac2a00a9afa # v5.1.0
+      env:
+        SUBZERO_ACCESS_TOKEN: ${{ inputs.token }}
+      with:
+        main: |
+          git config --global url."https://x-access-token:${SUBZERO_ACCESS_TOKEN}@github.com/neondatabase/subzero".insteadOf "https://github.com/neondatabase/subzero"
+          cargo add -p proxy subzero-core --git https://github.com/neondatabase/subzero --rev 396264617e78e8be428682f87469bb25429af88a
+        post: |
+          git config --global --unset url."https://x-access-token:${SUBZERO_ACCESS_TOKEN}@github.com/neondatabase/subzero".insteadOf "https://github.com/neondatabase/subzero"
+
+    - name: Set `CARGO_NET_GIT_FETCH_WITH_CLI=true` env variable
+      shell: bash -euxo pipefail {0}
+      run: echo "CARGO_NET_GIT_FETCH_WITH_CLI=true" >> ${GITHUB_ENV}

.github/actions/run-python-test-set/action.yml

@@ -181,6 +181,8 @@ runs:
           # Ref https://github.com/neondatabase/neon/issues/4540
           # cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
           cov_prefix=()
+          # Explicitly set LLVM_PROFILE_FILE to /dev/null to avoid writing *.profraw files
+          export LLVM_PROFILE_FILE=/dev/null
         else
           cov_prefix=()
         fi

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

@@ -86,6 +86,10 @@ jobs:
         with:
           submodules: true
 
+      - uses: ./.github/actions/prepare-for-subzero
+        with:
+          token: ${{ secrets.CI_ACCESS_TOKEN }}
+
       - name: Set pg 14 revision for caching
         id: pg_v14_rev
         run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-v14) >> $GITHUB_OUTPUT
@@ -116,7 +120,7 @@ jobs:
           ARCH: ${{ inputs.arch }}
           SANITIZERS: ${{ inputs.sanitizers }}
         run: |
-          CARGO_FLAGS="--locked --features testing"
+          CARGO_FLAGS="--locked --features testing,rest_broker"
           if [[ $BUILD_TYPE == "debug" && $ARCH == 'x64' ]]; then
             cov_prefix="scripts/coverage --profraw-prefix=$GITHUB_JOB --dir=/tmp/coverage run"
             CARGO_PROFILE=""

.github/workflows/_check-codestyle-rust.yml

@@ -46,6 +46,10 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
         with:
           submodules: true
+      
+      - uses: ./.github/actions/prepare-for-subzero
+        with:
+          token: ${{ secrets.CI_ACCESS_TOKEN }}
 
       - name: Cache cargo deps
         uses: tespkg/actions-cache@b7bf5fcc2f98a52ac6080eb0fd282c2f752074b1  # v1.8.0

.github/workflows/build-macos.yml

@@ -54,6 +54,10 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
         with:
           submodules: true
+      
+      - uses: ./.github/actions/prepare-for-subzero
+        with:
+          token: ${{ secrets.CI_ACCESS_TOKEN }}
 
       - name: Install build dependencies
         run: |

.github/workflows/build_and_test.yml

@@ -87,22 +87,27 @@ jobs:
     uses: ./.github/workflows/build-build-tools-image.yml
     secrets: inherit
 
-  lint-openapi-spec:
-    runs-on: ubuntu-22.04
-    needs: [ meta, check-permissions ]
+  lint-yamls:
+    needs: [ meta, check-permissions, build-build-tools-image ]
     # We do need to run this in `.*-rc-pr` because of hotfixes.
     if: ${{ contains(fromJSON('["pr", "push-main", "storage-rc-pr", "proxy-rc-pr", "compute-rc-pr"]'), needs.meta.outputs.run-kind) }}
+    runs-on: [ self-hosted, small ]
+    container:
+      image: ${{ needs.build-build-tools-image.outputs.image }}
+      credentials:
+        username: ${{ github.actor }}
+        password: ${{ secrets.GITHUB_TOKEN }}
+      options: --init
+
     steps:
       - name: Harden the runner (Audit all outbound calls)
         uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
         with:
           egress-policy: audit
+
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
-        with:
-          registry: ghcr.io
-          username: ${{ github.actor }}
-          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - run: make -C compute manifest-schema-validation
       - run: make lint-openapi-spec
 
   check-codestyle-python:
@@ -217,28 +222,6 @@ jobs:
       build-tools-image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm
     secrets: inherit
 
-  validate-compute-manifest:
-    runs-on: ubuntu-22.04
-    needs: [ meta, check-permissions ]
-    # We do need to run this in `.*-rc-pr` because of hotfixes.
-    if: ${{ contains(fromJSON('["pr", "push-main", "storage-rc-pr", "proxy-rc-pr", "compute-rc-pr"]'), needs.meta.outputs.run-kind) }}
-    steps:
-      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
-        with:
-          egress-policy: audit
-
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-      - name: Set up Node.js
-        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
-        with:
-          node-version: '24'
-
-      - name: Validate manifest against schema
-        run: |
-          make -C compute manifest-schema-validation
-
   build-and-test-locally:
     needs: [ meta, build-build-tools-image ]
     # We do need to run this in `.*-rc-pr` because of hotfixes.
@@ -649,6 +632,8 @@ jobs:
             BUILD_TAG=${{ needs.meta.outputs.release-tag || needs.meta.outputs.build-tag }}
             TAG=${{ needs.build-build-tools-image.outputs.image-tag }}-bookworm
             DEBIAN_VERSION=bookworm
+          secrets: |
+            SUBZERO_ACCESS_TOKEN=${{ secrets.CI_ACCESS_TOKEN }}
           provenance: false
           push: true
           pull: true

.github/workflows/neon_extra_builds.yml

@@ -72,6 +72,7 @@ jobs:
   check-macos-build:
     needs: [ check-permissions, files-changed ]
     uses: ./.github/workflows/build-macos.yml
+    secrets: inherit
     with:
       pg_versions: ${{ needs.files-changed.outputs.postgres_changes }}
       rebuild_rust_code: ${{ fromJSON(needs.files-changed.outputs.rebuild_rust_code) }}

.gitignore

@@ -26,6 +26,14 @@ docker-compose/docker-compose-parallel.yml
 *.o
 *.so
 *.Po
+*.pid
 
 # pgindent typedef lists
 *.list
+
+# Node
+**/node_modules/
+
+# various files for local testing
+/proxy/.subzero
+local_proxy.json

Cargo.toml

@@ -49,6 +49,7 @@ members = [
     "libs/proxy/tokio-postgres2",
     "endpoint_storage",
     "pgxn/neon/communicator",
+    "proxy/subzero_core",
 ]
 
 [workspace.package]
@@ -130,6 +131,7 @@ jemalloc_pprof = { version = "0.7", features = ["symbolize", "flamegraph"] }
 jsonwebtoken = "9"
 lasso = "0.7"
 libc = "0.2"
+lock_api = "0.4.13"
 md5 = "0.7.0"
 measured = { version = "0.0.22", features=["lasso"] }
 measured-process = { version = "0.0.22" }
@@ -141,10 +143,10 @@ notify = "6.0.0"
 num_cpus = "1.15"
 num-traits = "0.2.19"
 once_cell = "1.13"
-opentelemetry = "0.27"
-opentelemetry_sdk = "0.27"
-opentelemetry-otlp = { version = "0.27", default-features = false, features = ["http-proto", "trace", "http", "reqwest-client"] }
-opentelemetry-semantic-conventions = "0.27"
+opentelemetry = "0.30"
+opentelemetry_sdk = "0.30"
+opentelemetry-otlp = { version = "0.30", default-features = false, features = ["http-proto", "trace", "http", "reqwest-blocking-client"] }
+opentelemetry-semantic-conventions = "0.30"
 parking_lot = "0.12"
 parquet = { version = "53", default-features = false, features = ["zstd"] }
 parquet_derive = "53"
@@ -156,16 +158,18 @@ procfs = "0.16"
 prometheus = {version = "0.13", default-features=false, features = ["process"]} # removes protobuf dependency
 prost = "0.13.5"
 prost-types = "0.13.5"
-rand = "0.8"
+rand = "0.9"
+# Remove after p256 is updated to 0.14.
+rand_core = "=0.6"
 redis = { version = "0.29.2", features = ["tokio-rustls-comp", "keep-alive"] }
 regex = "1.10.2"
 reqwest = { version = "0.12", default-features = false, features = ["rustls-tls"] }
-reqwest-tracing = { version = "0.5", features = ["opentelemetry_0_27"] }
+reqwest-tracing = { version = "0.5", features = ["opentelemetry_0_30"] }
 reqwest-middleware = "0.4"
 reqwest-retry = "0.7"
 routerify = "3"
 rpds = "0.13"
-rustc-hash = "1.1.0"
+rustc-hash = "2.1.1"
 rustls = { version = "0.23.16", default-features = false }
 rustls-pemfile = "2"
 rustls-pki-types = "1.11"
@@ -210,17 +214,15 @@ tonic = { version = "0.13.1", default-features = false, features = ["channel", "
 tonic-reflection = { version = "0.13.1", features = ["server"] }
 tower = { version = "0.5.2", default-features = false }
 tower-http = { version = "0.6.2", features = ["auth", "request-id", "trace"] }
-
-# This revision uses opentelemetry 0.27. There's no tag for it.
-tower-otel = { git = "https://github.com/mattiapenati/tower-otel", rev = "56a7321053bcb72443888257b622ba0d43a11fcd" }
-
+tower-otel = { version = "0.6", features = ["axum"] }
 tower-service = "0.3.3"
 tracing = "0.1"
 tracing-error = "0.2"
 tracing-log = "0.2"
-tracing-opentelemetry = "0.28"
+tracing-opentelemetry = "0.31"
 tracing-serde = "0.2.0"
 tracing-subscriber = { version = "0.3", default-features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json"] }
+tracing-appender = "0.2.3"
 try-lock = "0.2.5"
 test-log = { version = "0.2.17", default-features = false, features = ["log"] }
 twox-hash = { version = "1.6.3", default-features = false }

Dockerfile

@@ -63,7 +63,14 @@ WORKDIR /home/nonroot
 
 COPY --chown=nonroot . .
 
-RUN cargo chef prepare --recipe-path recipe.json
+RUN --mount=type=secret,uid=1000,id=SUBZERO_ACCESS_TOKEN \
+    set -e \
+    && if [ -s /run/secrets/SUBZERO_ACCESS_TOKEN ]; then \
+        export CARGO_NET_GIT_FETCH_WITH_CLI=true && \
+        git config --global url."https://$(cat /run/secrets/SUBZERO_ACCESS_TOKEN)@github.com/neondatabase/subzero".insteadOf "https://github.com/neondatabase/subzero" && \
+        cargo add -p proxy subzero-core --git https://github.com/neondatabase/subzero --rev 396264617e78e8be428682f87469bb25429af88a; \
+    fi \
+    && cargo chef prepare --recipe-path recipe.json
 
 # Main build image
 FROM $REPOSITORY/$IMAGE:$TAG AS build
@@ -71,20 +78,33 @@ WORKDIR /home/nonroot
 ARG GIT_VERSION=local
 ARG BUILD_TAG
 ARG ADDITIONAL_RUSTFLAGS=""
+ENV CARGO_FEATURES="default"
 
 # 3. Build cargo dependencies. Note that this step doesn't depend on anything else than
 # `recipe.json`, so the layer can be reused as long as none of the dependencies change.
 COPY --from=plan     /home/nonroot/recipe.json                              recipe.json
-RUN set -e \
+RUN --mount=type=secret,uid=1000,id=SUBZERO_ACCESS_TOKEN \
+    set -e \
+    && if [ -s /run/secrets/SUBZERO_ACCESS_TOKEN ]; then \
+        export CARGO_NET_GIT_FETCH_WITH_CLI=true && \
+        git config --global url."https://$(cat /run/secrets/SUBZERO_ACCESS_TOKEN)@github.com/neondatabase/subzero".insteadOf "https://github.com/neondatabase/subzero"; \
+    fi \
     && RUSTFLAGS="-Clinker=clang -Clink-arg=-fuse-ld=mold -Clink-arg=-Wl,--no-rosegment -Cforce-frame-pointers=yes ${ADDITIONAL_RUSTFLAGS}" cargo chef cook --locked --release --recipe-path recipe.json
 
 # Perform the main build. We reuse the Postgres build artifacts from the intermediate 'pg-build'
 # layer, and the cargo dependencies built in the previous step.
 COPY --chown=nonroot --from=pg-build /home/nonroot/pg_install/ pg_install
 COPY --chown=nonroot . .
+COPY --chown=nonroot --from=plan     /home/nonroot/proxy/Cargo.toml         proxy/Cargo.toml
+COPY --chown=nonroot --from=plan     /home/nonroot/Cargo.lock               Cargo.lock
 
-RUN set -e \
+RUN  --mount=type=secret,uid=1000,id=SUBZERO_ACCESS_TOKEN \
+    set -e \
+    && if [ -s /run/secrets/SUBZERO_ACCESS_TOKEN ]; then \
+        export CARGO_FEATURES="rest_broker"; \
+    fi \
     && RUSTFLAGS="-Clinker=clang -Clink-arg=-fuse-ld=mold -Clink-arg=-Wl,--no-rosegment -Cforce-frame-pointers=yes ${ADDITIONAL_RUSTFLAGS}" cargo build \
+      --features $CARGO_FEATURES \
       --bin pg_sni_router  \
       --bin pageserver  \
       --bin pagectl  \

Makefile

@@ -2,7 +2,7 @@ ROOT_PROJECT_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
 
 # Where to install Postgres, default is ./pg_install, maybe useful for package
 # managers.
-POSTGRES_INSTALL_DIR ?= $(ROOT_PROJECT_DIR)/pg_install/
+POSTGRES_INSTALL_DIR ?= $(ROOT_PROJECT_DIR)/pg_install
 
 # Supported PostgreSQL versions
 POSTGRES_VERSIONS = v17 v16 v15 v14
@@ -14,7 +14,7 @@ POSTGRES_VERSIONS = v17 v16 v15 v14
 # it is derived from BUILD_TYPE.
 
 # All intermediate build artifacts are stored here.
-BUILD_DIR := build
+BUILD_DIR := $(ROOT_PROJECT_DIR)/build
 
 ICU_PREFIX_DIR := /usr/local/icu
 
@@ -212,7 +212,7 @@ neon-pgindent: postgres-v17-pg-bsd-indent neon-pg-ext-v17
 		FIND_TYPEDEF=$(ROOT_PROJECT_DIR)/vendor/postgres-v17/src/tools/find_typedef \
 		INDENT=$(BUILD_DIR)/v17/src/tools/pg_bsd_indent/pg_bsd_indent \
 		PGINDENT_SCRIPT=$(ROOT_PROJECT_DIR)/vendor/postgres-v17/src/tools/pgindent/pgindent \
-		-C $(BUILD_DIR)/neon-v17 \
+		-C $(BUILD_DIR)/pgxn-v17/neon \
 		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile pgindent
 
 
@@ -220,11 +220,15 @@ neon-pgindent: postgres-v17-pg-bsd-indent neon-pg-ext-v17
 setup-pre-commit-hook:
 	ln -s -f $(ROOT_PROJECT_DIR)/pre-commit.py .git/hooks/pre-commit
 
+build-tools/node_modules: build-tools/package.json
+	cd build-tools && $(if $(CI),npm ci,npm install)
+	touch build-tools/node_modules
+
 .PHONY: lint-openapi-spec
-lint-openapi-spec:
+lint-openapi-spec: build-tools/node_modules
 	# operation-2xx-response: pageserver timeline delete returns 404 on success
 	find . -iname "openapi_spec.y*ml" -exec\
-		docker run --rm -v ${PWD}:/spec ghcr.io/redocly/cli:1.34.4\
+		npx --prefix=build-tools/ redocly\
 			--skip-rule=operation-operationId --skip-rule=operation-summary --extends=minimal\
 			--skip-rule=no-server-example.com --skip-rule=operation-2xx-response\
 			lint {} \+

build-tools/Dockerfile

@@ -188,6 +188,12 @@ RUN curl -fsSL 'https://apt.llvm.org/llvm-snapshot.gpg.key' | apt-key add - \
     && bash -c 'for f in /usr/bin/clang*-${LLVM_VERSION} /usr/bin/llvm*-${LLVM_VERSION}; do ln -s "${f}" "${f%-${LLVM_VERSION}}"; done' \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
 
+# Install node
+ENV NODE_VERSION=24
+RUN curl -fsSL https://deb.nodesource.com/setup_${NODE_VERSION}.x | bash - \
+    && apt install -y nodejs \
+    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+
 # Install docker
 RUN curl -fsSL https://download.docker.com/linux/ubuntu/gpg | gpg --dearmor -o /usr/share/keyrings/docker-archive-keyring.gpg \
     && echo "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/docker-archive-keyring.gpg] https://download.docker.com/linux/debian ${DEBIAN_VERSION} stable" > /etc/apt/sources.list.d/docker.list \
@@ -311,14 +317,14 @@ RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux
     . "$HOME/.cargo/env" && \
     cargo --version && rustup --version && \
     rustup component add llvm-tools rustfmt clippy && \
-    cargo install rustfilt            --version ${RUSTFILT_VERSION} --locked && \
-    cargo install cargo-hakari        --version ${CARGO_HAKARI_VERSION} --locked && \
-    cargo install cargo-deny          --version ${CARGO_DENY_VERSION} --locked && \
-    cargo install cargo-hack          --version ${CARGO_HACK_VERSION} --locked && \
-    cargo install cargo-nextest       --version ${CARGO_NEXTEST_VERSION} --locked && \
-    cargo install cargo-chef          --version ${CARGO_CHEF_VERSION} --locked && \
-    cargo install diesel_cli          --version ${CARGO_DIESEL_CLI_VERSION} --locked \
-                                      --features postgres-bundled --no-default-features && \
+    cargo install rustfilt      --locked --version ${RUSTFILT_VERSION} && \
+    cargo install cargo-hakari  --locked --version ${CARGO_HAKARI_VERSION} && \
+    cargo install cargo-deny    --locked --version ${CARGO_DENY_VERSION} && \
+    cargo install cargo-hack    --locked --version ${CARGO_HACK_VERSION} && \
+    cargo install cargo-nextest --locked --version ${CARGO_NEXTEST_VERSION} && \
+    cargo install cargo-chef    --locked --version ${CARGO_CHEF_VERSION} && \
+    cargo install diesel_cli    --locked --version ${CARGO_DIESEL_CLI_VERSION} \
+                                --features postgres-bundled --no-default-features && \
     rm -rf /home/nonroot/.cargo/registry && \
     rm -rf /home/nonroot/.cargo/git
 

build-tools/package.json

@@ -0,0 +1,8 @@
+{
+  "name": "build-tools",
+  "private": true,
+  "devDependencies": {
+    "@redocly/cli": "1.34.4",
+    "@sourcemeta/jsonschema": "10.0.0"
+  }
+}

compute/Makefile

@@ -50,9 +50,9 @@ jsonnetfmt-format:
 	jsonnetfmt --in-place $(jsonnet_files)
 
 .PHONY: manifest-schema-validation
-manifest-schema-validation: node_modules
-	node_modules/.bin/jsonschema validate -d https://json-schema.org/draft/2020-12/schema manifest.schema.json manifest.yaml
+manifest-schema-validation: ../build-tools/node_modules
+	npx --prefix=../build-tools/ jsonschema validate -d https://json-schema.org/draft/2020-12/schema manifest.schema.json manifest.yaml
 
-node_modules: package.json
-	npm install
-	touch node_modules
+../build-tools/node_modules: ../build-tools/package.json
+	cd ../build-tools && $(if $(CI),npm ci,npm install)
+	touch ../build-tools/node_modules

compute/compute-node.Dockerfile

@@ -133,7 +133,7 @@ RUN case $DEBIAN_VERSION in \
       # Install newer version (3.25) from backports.
       # libstdc++-10-dev is required for plv8
       bullseye) \
-        echo "deb http://deb.debian.org/debian bullseye-backports main" > /etc/apt/sources.list.d/bullseye-backports.list; \
+        echo "deb http://archive.debian.org/debian bullseye-backports main" > /etc/apt/sources.list.d/bullseye-backports.list; \
         VERSION_INSTALLS="cmake/bullseye-backports cmake-data/bullseye-backports libstdc++-10-dev"; \
       ;; \
       # Version-specific installs for Bookworm (PG17):
@@ -170,7 +170,29 @@ RUN case $DEBIAN_VERSION in \
 FROM build-deps AS pg-build
 ARG PG_VERSION
 COPY vendor/postgres-${PG_VERSION:?} postgres
+COPY compute/patches/postgres_fdw.patch .
+COPY compute/patches/pg_stat_statements_pg14-16.patch .
+COPY compute/patches/pg_stat_statements_pg17.patch .
 RUN cd postgres && \
+    # Apply patches to some contrib extensions
+    # For example, we need to grant EXECUTE on pg_stat_statements_reset() to {privileged_role_name}.
+    # In vanilla Postgres this function is limited to Postgres role superuser.
+    # In Neon we have {privileged_role_name} role that is not a superuser but replaces superuser in some cases.
+    # We could add the additional grant statements to the Postgres repository but it would be hard to maintain,
+    # whenever we need to pick up a new Postgres version and we want to limit the changes in our Postgres fork,
+    # so we do it here.
+    case "${PG_VERSION}" in \
+    "v14" | "v15" | "v16") \
+    patch -p1 < /pg_stat_statements_pg14-16.patch; \
+    ;; \
+    "v17") \
+    patch -p1 < /pg_stat_statements_pg17.patch; \
+    ;; \
+    *) \
+    # To do not forget to migrate patches to the next major version
+    echo "No contrib patches for this PostgreSQL version" && exit 1;; \
+    esac && \
+    patch -p1 < /postgres_fdw.patch && \
     export CONFIGURE_CMD="./configure CFLAGS='-O2 -g3 -fsigned-char' --enable-debug --with-openssl --with-uuid=ossp \
     --with-icu --with-libxml --with-libxslt --with-lz4" && \
     if [ "${PG_VERSION:?}" != "v14" ]; then \
@@ -184,8 +206,6 @@ RUN cd postgres && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/autoinc.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/dblink.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/postgres_fdw.control && \
-    file=/usr/local/pgsql/share/extension/postgres_fdw--1.0.sql && [ -e $file ] && \
-    echo 'GRANT USAGE ON FOREIGN DATA WRAPPER postgres_fdw TO neon_superuser;' >> $file && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/bloom.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/earthdistance.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/insert_username.control && \
@@ -195,34 +215,7 @@ RUN cd postgres && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgrowlocks.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgstattuple.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/refint.control && \
-    echo 'trusted = true' >> /usr/local/pgsql/share/extension/xml2.control && \
-    # We need to grant EXECUTE on pg_stat_statements_reset() to neon_superuser.
-    # In vanilla postgres this function is limited to Postgres role superuser.
-    # In neon we have neon_superuser role that is not a superuser but replaces superuser in some cases.
-    # We could add the additional grant statements to the postgres repository but it would be hard to maintain,
-    # whenever we need to pick up a new postgres version and we want to limit the changes in our postgres fork,
-    # so we do it here.
-    for file in /usr/local/pgsql/share/extension/pg_stat_statements--*.sql; do \
-        filename=$(basename "$file"); \
-        # Note that there are no downgrade scripts for pg_stat_statements, so we \
-        # don't have to modify any downgrade paths or (much) older versions: we only \
-        # have to make sure every creation of the pg_stat_statements_reset function \
-        # also adds execute permissions to the neon_superuser.
-        case $filename in \
-          pg_stat_statements--1.4.sql) \
-            # pg_stat_statements_reset is first created with 1.4
-            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO neon_superuser;' >> $file; \
-            ;; \
-          pg_stat_statements--1.6--1.7.sql) \
-            # Then with the 1.6-1.7 migration it is re-created with a new signature, thus add the permissions back
-            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) TO neon_superuser;' >> $file; \
-            ;; \
-          pg_stat_statements--1.10--1.11.sql) \
-            # Then with the 1.10-1.11 migration it is re-created with a new signature again, thus add the permissions back
-            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint, boolean) TO neon_superuser;' >> $file; \
-            ;; \
-        esac; \
-    done;
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/xml2.control
 
 # Set PATH for all the subsequent build steps
 ENV PATH="/usr/local/pgsql/bin:$PATH"
@@ -1524,7 +1517,7 @@ WORKDIR /ext-src
 COPY compute/patches/pg_duckdb_v031.patch .
 COPY compute/patches/duckdb_v120.patch .
 # pg_duckdb build requires source dir to be a git repo to get submodules
-# allow neon_superuser to execute some functions that in pg_duckdb are available to superuser only:
+# allow {privileged_role_name} to execute some functions that in pg_duckdb are available to superuser only:
 # - extension management function duckdb.install_extension()
 # - access to duckdb.extensions table and its sequence
 RUN git clone --depth 1 --branch v0.3.1 https://github.com/duckdb/pg_duckdb.git pg_duckdb-src && \

compute/package.json

@@ -1,7 +0,0 @@
-{
-  "name": "neon-compute",
-  "private": true,
-  "dependencies": {
-    "@sourcemeta/jsonschema": "9.3.4"
-  }
-} 

compute/patches/anon_v2.patch

@@ -1,22 +1,26 @@
 diff --git a/sql/anon.sql b/sql/anon.sql
-index 0cdc769..b450327 100644
+index 0cdc769..5eab1d6 100644
 --- a/sql/anon.sql
 +++ b/sql/anon.sql
-@@ -1141,3 +1141,15 @@ $$
+@@ -1141,3 +1141,19 @@ $$
  -- TODO : https://en.wikipedia.org/wiki/L-diversity
  
  -- TODO : https://en.wikipedia.org/wiki/T-closeness
 +
 +-- NEON Patches
 +
-+GRANT ALL ON SCHEMA anon to neon_superuser;
-+GRANT ALL ON ALL TABLES IN SCHEMA anon TO neon_superuser;
-+
 +DO $$
++DECLARE
++  privileged_role_name text;
 +BEGIN
-+    IF current_setting('server_version_num')::int >= 150000 THEN
-+        GRANT SET ON PARAMETER anon.transparent_dynamic_masking TO neon_superuser;
-+    END IF;
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT ALL ON SCHEMA anon to %I', privileged_role_name);
++  EXECUTE format('GRANT ALL ON ALL TABLES IN SCHEMA anon TO %I', privileged_role_name);
++
++  IF current_setting('server_version_num')::int >= 150000 THEN
++    EXECUTE format('GRANT SET ON PARAMETER anon.transparent_dynamic_masking TO %I', privileged_role_name);
++  END IF;
 +END $$;
 diff --git a/sql/init.sql b/sql/init.sql
 index 7da6553..9b6164b 100644

compute/patches/pg_duckdb_v031.patch

@@ -21,13 +21,21 @@ index 3235cc8..6b892bc 100644
  include Makefile.global
  
 diff --git a/sql/pg_duckdb--0.2.0--0.3.0.sql b/sql/pg_duckdb--0.2.0--0.3.0.sql
-index d777d76..af60106 100644
+index d777d76..3b54396 100644
 --- a/sql/pg_duckdb--0.2.0--0.3.0.sql
 +++ b/sql/pg_duckdb--0.2.0--0.3.0.sql
-@@ -1056,3 +1056,6 @@ GRANT ALL ON FUNCTION duckdb.cache(TEXT, TEXT) TO PUBLIC;
+@@ -1056,3 +1056,14 @@ GRANT ALL ON FUNCTION duckdb.cache(TEXT, TEXT) TO PUBLIC;
  GRANT ALL ON FUNCTION duckdb.cache_info() TO PUBLIC;
  GRANT ALL ON FUNCTION duckdb.cache_delete(TEXT) TO PUBLIC;
  GRANT ALL ON PROCEDURE duckdb.recycle_ddb() TO PUBLIC;
-+GRANT ALL ON FUNCTION duckdb.install_extension(TEXT) TO neon_superuser;
-+GRANT ALL ON TABLE duckdb.extensions TO neon_superuser;
-+GRANT ALL ON SEQUENCE duckdb.extensions_table_seq TO neon_superuser;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT ALL ON FUNCTION duckdb.install_extension(TEXT) TO %I', privileged_role_name);
++  EXECUTE format('GRANT ALL ON TABLE duckdb.extensions TO %I', privileged_role_name);
++  EXECUTE format('GRANT ALL ON SEQUENCE duckdb.extensions_table_seq TO %I', privileged_role_name);
++END $$;

compute/patches/pg_stat_statements_pg14-16.patch

@@ -0,0 +1,34 @@
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.4.sql b/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
+index 58cdf600fce..8be57a996f6 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
+@@ -46,3 +46,12 @@ GRANT SELECT ON pg_stat_statements TO PUBLIC;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset() FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO %I', privileged_role_name);
++END $$;
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql b/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
+index 6fc3fed4c93..256345a8f79 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
+@@ -20,3 +20,12 @@ LANGUAGE C STRICT PARALLEL SAFE;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) TO %I', privileged_role_name);
++END $$;

compute/patches/pg_stat_statements_pg17.patch

@@ -0,0 +1,52 @@
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.10--1.11.sql b/contrib/pg_stat_statements/pg_stat_statements--1.10--1.11.sql
+index 0bb2c397711..32764db1d8b 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.10--1.11.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.10--1.11.sql
+@@ -80,3 +80,12 @@ LANGUAGE C STRICT PARALLEL SAFE;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint, boolean) FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint, boolean) TO %I', privileged_role_name);
++END $$;
+\ No newline at end of file
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.4.sql b/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
+index 58cdf600fce..8be57a996f6 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
+@@ -46,3 +46,12 @@ GRANT SELECT ON pg_stat_statements TO PUBLIC;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset() FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO %I', privileged_role_name);
++END $$;
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql b/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
+index 6fc3fed4c93..256345a8f79 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
+@@ -20,3 +20,12 @@ LANGUAGE C STRICT PARALLEL SAFE;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) TO %I', privileged_role_name);
++END $$;

compute/patches/postgres_fdw.patch

@@ -0,0 +1,17 @@
+diff --git a/contrib/postgres_fdw/postgres_fdw--1.0.sql b/contrib/postgres_fdw/postgres_fdw--1.0.sql
+index a0f0fc1bf45..ee077f2eea6 100644
+--- a/contrib/postgres_fdw/postgres_fdw--1.0.sql
++++ b/contrib/postgres_fdw/postgres_fdw--1.0.sql
+@@ -16,3 +16,12 @@ LANGUAGE C STRICT;
+ CREATE FOREIGN DATA WRAPPER postgres_fdw
+   HANDLER postgres_fdw_handler
+   VALIDATOR postgres_fdw_validator;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT USAGE ON FOREIGN DATA WRAPPER postgres_fdw TO %I', privileged_role_name);
++END $$;

compute_tools/Cargo.toml

@@ -27,7 +27,10 @@ fail.workspace = true
 flate2.workspace = true
 futures.workspace = true
 http.workspace = true
+http-body-util.workspace = true
 hostname-validator = "1.1"
+hyper.workspace = true
+hyper-util.workspace = true
 indexmap.workspace = true
 itertools.workspace = true
 jsonwebtoken.workspace = true
@@ -44,6 +47,7 @@ postgres.workspace = true
 regex.workspace = true
 reqwest = { workspace = true, features = ["json"] }
 ring = "0.17"
+scopeguard.workspace = true
 serde.workspace = true
 serde_with.workspace = true
 serde_json.workspace = true
@@ -58,6 +62,7 @@ tokio-stream.workspace = true
 tonic.workspace = true
 tower-otel.workspace = true
 tracing.workspace = true
+tracing-appender.workspace = true
 tracing-opentelemetry.workspace = true
 tracing-subscriber.workspace = true
 tracing-utils.workspace = true

compute_tools/src/bin/compute_ctl.rs

@@ -51,6 +51,7 @@ use compute_tools::compute::{
 use compute_tools::extension_server::get_pg_version_string;
 use compute_tools::logger::*;
 use compute_tools::params::*;
+use compute_tools::pg_isready::get_pg_isready_bin;
 use compute_tools::spec::*;
 use rlimit::{Resource, setrlimit};
 use signal_hook::consts::{SIGINT, SIGQUIT, SIGTERM};
@@ -87,6 +88,14 @@ struct Cli {
     #[arg(short = 'C', long, value_name = "DATABASE_URL")]
     pub connstr: String,
 
+    #[arg(
+        long,
+        default_value = "neon_superuser",
+        value_name = "PRIVILEGED_ROLE_NAME",
+        value_parser = Self::parse_privileged_role_name
+    )]
+    pub privileged_role_name: String,
+
     #[cfg(target_os = "linux")]
     #[arg(long, default_value = "neon-postgres")]
     pub cgroup: String,
@@ -130,6 +139,12 @@ struct Cli {
     /// Run in development mode, skipping VM-specific operations like process termination
     #[arg(long, action = clap::ArgAction::SetTrue)]
     pub dev: bool,
+
+    #[arg(long)]
+    pub pg_init_timeout: Option<u64>,
+
+    #[arg(long, default_value_t = false, action = clap::ArgAction::Set)]
+    pub lakebase_mode: bool,
 }
 
 impl Cli {
@@ -149,6 +164,21 @@ impl Cli {
 
         Ok(url)
     }
+
+    /// For simplicity, we do not escape `privileged_role_name` anywhere in the code.
+    /// Since it's a system role, which we fully control, that's fine. Still, let's
+    /// validate it to avoid any surprises.
+    fn parse_privileged_role_name(value: &str) -> Result<String> {
+        use regex::Regex;
+
+        let pattern = Regex::new(r"^[a-z_]+$").unwrap();
+
+        if !pattern.is_match(value) {
+            bail!("--privileged-role-name can only contain lowercase letters and underscores")
+        }
+
+        Ok(value.to_string())
+    }
 }
 
 fn main() -> Result<()> {
@@ -165,7 +195,12 @@ fn main() -> Result<()> {
         .build()?;
     let _rt_guard = runtime.enter();
 
-    runtime.block_on(init(cli.dev))?;
+    let mut log_dir = None;
+    if cli.lakebase_mode {
+        log_dir = std::env::var("COMPUTE_CTL_LOG_DIRECTORY").ok();
+    }
+
+    let (tracing_provider, _file_logs_guard) = init(cli.dev, log_dir)?;
 
     // enable core dumping for all child processes
     setrlimit(Resource::CORE, rlimit::INFINITY, rlimit::INFINITY)?;
@@ -178,6 +213,7 @@ fn main() -> Result<()> {
         ComputeNodeParams {
             compute_id: cli.compute_id,
             connstr,
+            privileged_role_name: cli.privileged_role_name.clone(),
             pgdata: cli.pgdata.clone(),
             pgbin: cli.pgbin.clone(),
             pgversion: get_pg_version_string(&cli.pgbin),
@@ -195,6 +231,10 @@ fn main() -> Result<()> {
             installed_extensions_collection_interval: Arc::new(AtomicU64::new(
                 cli.installed_extensions_collection_interval,
             )),
+            pg_init_timeout: cli.pg_init_timeout.map(Duration::from_secs),
+            pg_isready_bin: get_pg_isready_bin(&cli.pgbin),
+            instance_id: std::env::var("INSTANCE_ID").ok(),
+            lakebase_mode: cli.lakebase_mode,
         },
         config,
     )?;
@@ -203,11 +243,17 @@ fn main() -> Result<()> {
 
     scenario.teardown();
 
-    deinit_and_exit(exit_code);
+    deinit_and_exit(tracing_provider, exit_code);
 }
 
-async fn init(dev_mode: bool) -> Result<()> {
-    init_tracing_and_logging(DEFAULT_LOG_LEVEL).await?;
+fn init(
+    dev_mode: bool,
+    log_dir: Option<String>,
+) -> Result<(
+    Option<tracing_utils::Provider>,
+    Option<tracing_appender::non_blocking::WorkerGuard>,
+)> {
+    let (provider, file_logs_guard) = init_tracing_and_logging(DEFAULT_LOG_LEVEL, &log_dir)?;
 
     let mut signals = Signals::new([SIGINT, SIGTERM, SIGQUIT])?;
     thread::spawn(move || {
@@ -218,7 +264,7 @@ async fn init(dev_mode: bool) -> Result<()> {
 
     info!("compute build_tag: {}", &BUILD_TAG.to_string());
 
-    Ok(())
+    Ok((provider, file_logs_guard))
 }
 
 fn get_config(cli: &Cli) -> Result<ComputeConfig> {
@@ -243,25 +289,27 @@ fn get_config(cli: &Cli) -> Result<ComputeConfig> {
     }
 }
 
-fn deinit_and_exit(exit_code: Option<i32>) -> ! {
-    // Shutdown trace pipeline gracefully, so that it has a chance to send any
-    // pending traces before we exit. Shutting down OTEL tracing provider may
-    // hang for quite some time, see, for example:
-    // - https://github.com/open-telemetry/opentelemetry-rust/issues/868
-    // - and our problems with staging https://github.com/neondatabase/cloud/issues/3707#issuecomment-1493983636
-    //
-    // Yet, we want computes to shut down fast enough, as we may need a new one
-    // for the same timeline ASAP. So wait no longer than 2s for the shutdown to
-    // complete, then just error out and exit the main thread.
-    info!("shutting down tracing");
-    let (sender, receiver) = mpsc::channel();
-    let _ = thread::spawn(move || {
-        tracing_utils::shutdown_tracing();
-        sender.send(()).ok()
-    });
-    let shutdown_res = receiver.recv_timeout(Duration::from_millis(2000));
-    if shutdown_res.is_err() {
-        error!("timed out while shutting down tracing, exiting anyway");
+fn deinit_and_exit(tracing_provider: Option<tracing_utils::Provider>, exit_code: Option<i32>) -> ! {
+    if let Some(p) = tracing_provider {
+        // Shutdown trace pipeline gracefully, so that it has a chance to send any
+        // pending traces before we exit. Shutting down OTEL tracing provider may
+        // hang for quite some time, see, for example:
+        // - https://github.com/open-telemetry/opentelemetry-rust/issues/868
+        // - and our problems with staging https://github.com/neondatabase/cloud/issues/3707#issuecomment-1493983636
+        //
+        // Yet, we want computes to shut down fast enough, as we may need a new one
+        // for the same timeline ASAP. So wait no longer than 2s for the shutdown to
+        // complete, then just error out and exit the main thread.
+        info!("shutting down tracing");
+        let (sender, receiver) = mpsc::channel();
+        let _ = thread::spawn(move || {
+            _ = p.shutdown();
+            sender.send(()).ok()
+        });
+        let shutdown_res = receiver.recv_timeout(Duration::from_millis(2000));
+        if shutdown_res.is_err() {
+            error!("timed out while shutting down tracing, exiting anyway");
+        }
     }
 
     info!("shutting down");
@@ -327,4 +375,49 @@ mod test {
         ])
         .expect_err("URL parameters are not allowed");
     }
+
+    #[test]
+    fn verify_privileged_role_name() {
+        // Valid name
+        let cli = Cli::parse_from([
+            "compute_ctl",
+            "--pgdata=test",
+            "--connstr=test",
+            "--compute-id=test",
+            "--privileged-role-name",
+            "my_superuser",
+        ]);
+        assert_eq!(cli.privileged_role_name, "my_superuser");
+
+        // Invalid names
+        Cli::try_parse_from([
+            "compute_ctl",
+            "--pgdata=test",
+            "--connstr=test",
+            "--compute-id=test",
+            "--privileged-role-name",
+            "NeonSuperuser",
+        ])
+        .expect_err("uppercase letters are not allowed");
+
+        Cli::try_parse_from([
+            "compute_ctl",
+            "--pgdata=test",
+            "--connstr=test",
+            "--compute-id=test",
+            "--privileged-role-name",
+            "$'neon_superuser",
+        ])
+        .expect_err("special characters are not allowed");
+
+        Cli::try_parse_from([
+            "compute_ctl",
+            "--pgdata=test",
+            "--connstr=test",
+            "--compute-id=test",
+            "--privileged-role-name",
+            "",
+        ])
+        .expect_err("empty name is not allowed");
+    }
 }

compute_tools/src/communicator_socket_client.rs

@@ -0,0 +1,98 @@
+//! Client for making request to a running Postgres server's communicator control socket.
+//!
+//! The storage communicator process that runs inside Postgres exposes an HTTP endpoint in
+//! a Unix Domain Socket in the Postgres data directory. This provides access to it.
+
+use std::path::Path;
+
+use anyhow::Context;
+use hyper::client::conn::http1::SendRequest;
+use hyper_util::rt::TokioIo;
+
+/// Name of the socket within the Postgres data directory. This better match that in
+/// `pgxn/neon/communicator/src/lib.rs`.
+const NEON_COMMUNICATOR_SOCKET_NAME: &str = "neon-communicator.socket";
+
+/// Open a connection to the communicator's control socket, prepare to send requests to it
+/// with hyper.
+pub async fn connect_communicator_socket<B>(pgdata: &Path) -> anyhow::Result<SendRequest<B>>
+where
+    B: hyper::body::Body + 'static + Send,
+    B::Data: Send,
+    B::Error: Into<Box<dyn std::error::Error + Send + Sync>>,
+{
+    let socket_path = pgdata.join(NEON_COMMUNICATOR_SOCKET_NAME);
+    let socket_path_len = socket_path.display().to_string().len();
+
+    // There is a limit of around 100 bytes (108 on Linux?) on the length of the path to a
+    // Unix Domain socket. The limit is on the connect(2) function used to open the
+    // socket, not on the absolute path itself. Postgres changes the current directory to
+    // the data directory and uses a relative path to bind to the socket, and the relative
+    // path "./neon-communicator.socket" is always short, but when compute_ctl needs to
+    // open the socket, we need to use a full path, which can be arbitrarily long.
+    //
+    // There are a few ways we could work around this:
+    //
+    // 1. Change the current directory to the Postgres data directory and use a relative
+    //    path in the connect(2) call. That's problematic because the current directory
+    //    applies to the whole process. We could change the current directory early in
+    //    compute_ctl startup, and that might be a good idea anyway for other reasons too:
+    //    it would be more robust if the data directory is moved around or unlinked for
+    //    some reason, and you would be less likely to accidentally litter other parts of
+    //    the filesystem with e.g. temporary files. However, that's a pretty invasive
+    //    change.
+    //
+    // 2. On Linux, you could open() the data directory, and refer to the the socket
+    //    inside it as "/proc/self/fd/<fd>/neon-communicator.socket". But that's
+    //    Linux-only.
+    //
+    // 3. Create a symbolic link to the socket with a shorter path, and use that.
+    //
+    // We use the symbolic link approach here. Hopefully the paths we use in production
+    // are shorter, so that we can open the socket directly, so that this hack is needed
+    // only in development.
+    let connect_result = if socket_path_len < 100 {
+        // We can open the path directly with no hacks.
+        tokio::net::UnixStream::connect(socket_path).await
+    } else {
+        // The path to the socket is too long. Create a symlink to it with a shorter path.
+        let short_path = std::env::temp_dir().join(format!(
+            "compute_ctl.short-socket.{}.{}",
+            std::process::id(),
+            tokio::task::id()
+        ));
+        std::os::unix::fs::symlink(&socket_path, &short_path)?;
+
+        // Delete the symlink as soon as we have connected to it. There's a small chance
+        // of leaking if the process dies before we remove it, so try to keep that window
+        // as small as possible.
+        scopeguard::defer! {
+            if let Err(err) = std::fs::remove_file(&short_path) {
+                tracing::warn!("could not remove symlink \"{}\" created for socket: {}",
+                               short_path.display(), err);
+            }
+        }
+
+        tracing::info!(
+            "created symlink \"{}\" for socket \"{}\", opening it now",
+            short_path.display(),
+            socket_path.display()
+        );
+
+        tokio::net::UnixStream::connect(&short_path).await
+    };
+
+    let stream = connect_result.context("connecting to communicator control socket")?;
+
+    let io = TokioIo::new(stream);
+    let (request_sender, connection) = hyper::client::conn::http1::handshake(io).await?;
+
+    // spawn a task to poll the connection and drive the HTTP state
+    tokio::spawn(async move {
+        if let Err(err) = connection.await {
+            eprintln!("Error in connection: {err}");
+        }
+    });
+
+    Ok(request_sender)
+}

compute_tools/src/compute.rs

@@ -74,12 +74,20 @@ const DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL: u64 = 3600;
 
 /// Static configuration params that don't change after startup. These mostly
 /// come from the CLI args, or are derived from them.
+#[derive(Clone, Debug)]
 pub struct ComputeNodeParams {
     /// The ID of the compute
     pub compute_id: String,
-    // Url type maintains proper escaping
+
+    /// Url type maintains proper escaping
     pub connstr: url::Url,
 
+    /// The name of the 'weak' superuser role, which we give to the users.
+    /// It follows the allow list approach, i.e., we take a standard role
+    /// and grant it extra permissions with explicit GRANTs here and there,
+    /// and core patches.
+    pub privileged_role_name: String,
+
     pub resize_swap_on_bind: bool,
     pub set_disk_quota_for_fs: Option<String>,
 
@@ -105,6 +113,13 @@ pub struct ComputeNodeParams {
 
     /// Interval for installed extensions collection
     pub installed_extensions_collection_interval: Arc<AtomicU64>,
+    /// Hadron instance ID of the compute node.
+    pub instance_id: Option<String>,
+    /// Timeout of PG compute startup in the Init state.
+    pub pg_init_timeout: Option<Duration>,
+    // Path to the `pg_isready` binary.
+    pub pg_isready_bin: String,
+    pub lakebase_mode: bool,
 }
 
 type TaskHandle = Mutex<Option<JoinHandle<()>>>;
@@ -146,6 +161,7 @@ pub struct RemoteExtensionMetrics {
 #[derive(Clone, Debug)]
 pub struct ComputeState {
     pub start_time: DateTime<Utc>,
+    pub pg_start_time: Option<DateTime<Utc>>,
     pub status: ComputeStatus,
     /// Timestamp of the last Postgres activity. It could be `None` if
     /// compute wasn't used since start.
@@ -183,6 +199,7 @@ impl ComputeState {
     pub fn new() -> Self {
         Self {
             start_time: Utc::now(),
+            pg_start_time: None,
             status: ComputeStatus::Empty,
             last_active: None,
             error: None,
@@ -471,6 +488,7 @@ impl ComputeNode {
             port: this.params.external_http_port,
             config: this.compute_ctl_config.clone(),
             compute_id: this.params.compute_id.clone(),
+            instance_id: this.params.instance_id.clone(),
         }
         .launch(&this);
 
@@ -640,6 +658,9 @@ impl ComputeNode {
             };
             _this_entered = start_compute_span.enter();
 
+            // Hadron: Record postgres start time (used to enforce pg_init_timeout).
+            state_guard.pg_start_time.replace(Utc::now());
+
             state_guard.set_status(ComputeStatus::Init, &self.state_changed);
             compute_state = state_guard.clone()
         }
@@ -1286,9 +1307,7 @@ impl ComputeNode {
 
         // In case of error, log and fail the check, but don't crash.
         // We're playing it safe because these errors could be transient
-        // and we don't yet retry. Also being careful here allows us to
-        // be backwards compatible with safekeepers that don't have the
-        // TIMELINE_STATUS API yet.
+        // and we don't yet retry.
         if responses.len() < quorum {
             error!(
                 "failed sync safekeepers check {:?} {:?} {:?}",
@@ -1391,6 +1410,7 @@ impl ComputeNode {
         self.create_pgdata()?;
         config::write_postgres_conf(
             pgdata_path,
+            &self.params,
             &pspec.spec,
             self.params.internal_http_port,
             tls_config,
@@ -1434,7 +1454,7 @@ impl ComputeNode {
         })?;
 
         // Update pg_hba.conf received with basebackup.
-        update_pg_hba(pgdata_path)?;
+        update_pg_hba(pgdata_path, None)?;
 
         // Place pg_dynshmem under /dev/shm. This allows us to use
         // 'dynamic_shared_memory_type = mmap' so that the files are placed in
@@ -1739,6 +1759,8 @@ impl ComputeNode {
         }
 
         // Run migrations separately to not hold up cold starts
+        let lakebase_mode = self.params.lakebase_mode;
+        let params = self.params.clone();
         tokio::spawn(async move {
             let mut conf = conf.as_ref().clone();
             conf.application_name("compute_ctl:migrations");
@@ -1750,7 +1772,7 @@ impl ComputeNode {
                             eprintln!("connection error: {e}");
                         }
                     });
-                    if let Err(e) = handle_migrations(&mut client).await {
+                    if let Err(e) = handle_migrations(params, &mut client, lakebase_mode).await {
                         error!("Failed to run migrations: {}", e);
                     }
                 }
@@ -1766,6 +1788,34 @@ impl ComputeNode {
         Ok::<(), anyhow::Error>(())
     }
 
+    // Signal to the configurator to refresh the configuration by pulling a new spec from the HCC.
+    // Note that this merely triggers a notification on a condition variable the configurator thread
+    // waits on. The configurator thread (in configurator.rs) pulls the new spec from the HCC and
+    // applies it.
+    pub async fn signal_refresh_configuration(&self) -> Result<()> {
+        let states_allowing_configuration_refresh = [
+            ComputeStatus::Running,
+            ComputeStatus::Failed,
+            // ComputeStatus::RefreshConfigurationPending,
+        ];
+
+        let state = self.state.lock().expect("state lock poisoned");
+        if states_allowing_configuration_refresh.contains(&state.status) {
+            // state.status = ComputeStatus::RefreshConfigurationPending;
+            self.state_changed.notify_all();
+            Ok(())
+        } else if state.status == ComputeStatus::Init {
+            // If the compute is in Init state, we can't refresh the configuration immediately,
+            // but we should be able to do that soon.
+            Ok(())
+        } else {
+            Err(anyhow::anyhow!(
+                "Cannot refresh compute configuration in state {:?}",
+                state.status
+            ))
+        }
+    }
+
     // Wrapped this around `pg_ctl reload`, but right now we don't use
     // `pg_ctl` for start / stop.
     #[instrument(skip_all)]
@@ -1829,6 +1879,7 @@ impl ComputeNode {
         let pgdata_path = Path::new(&self.params.pgdata);
         config::write_postgres_conf(
             pgdata_path,
+            &self.params,
             &spec,
             self.params.internal_http_port,
             tls_config,
@@ -2441,14 +2492,31 @@ LIMIT 100",
     pub fn spawn_lfc_offload_task(self: &Arc<Self>, interval: Duration) {
         self.terminate_lfc_offload_task();
         let secs = interval.as_secs();
-        info!("spawning lfc offload worker with {secs}s interval");
         let this = self.clone();
+
+        info!("spawning LFC offload worker with {secs}s interval");
         let handle = spawn(async move {
             let mut interval = time::interval(interval);
             interval.tick().await; // returns immediately
             loop {
                 interval.tick().await;
-                this.offload_lfc_async().await;
+
+                let prewarm_state = this.state.lock().unwrap().lfc_prewarm_state.clone();
+                // Do not offload LFC state if we are currently prewarming or any issue occurred.
+                // If we'd do that, we might override the LFC state in endpoint storage with some
+                // incomplete state. Imagine a situation:
+                // 1. Endpoint started with `autoprewarm: true`
+                // 2. While prewarming is not completed, we upload the new incomplete state
+                // 3. Compute gets interrupted and restarts
+                // 4. We start again and try to prewarm with the state from 2. instead of the previous complete state
+                if matches!(
+                    prewarm_state,
+                    LfcPrewarmState::Completed
+                        | LfcPrewarmState::NotPrewarmed
+                        | LfcPrewarmState::Skipped
+                ) {
+                    this.offload_lfc_async().await;
+                }
             }
         });
         *self.lfc_offload_task.lock().unwrap() = Some(handle);

compute_tools/src/compute_prewarm.rs

@@ -89,7 +89,8 @@ impl ComputeNode {
         self.state.lock().unwrap().lfc_offload_state.clone()
     }
 
-    /// If there is a prewarm request ongoing, return false, true otherwise
+    /// If there is a prewarm request ongoing, return `false`, `true` otherwise.
+    /// Has a failpoint "compute-prewarm"
     pub fn prewarm_lfc(self: &Arc<Self>, from_endpoint: Option<String>) -> bool {
         {
             let state = &mut self.state.lock().unwrap().lfc_prewarm_state;
@@ -101,15 +102,24 @@ impl ComputeNode {
 
         let cloned = self.clone();
         spawn(async move {
-            let Err(err) = cloned.prewarm_impl(from_endpoint).await else {
-                cloned.state.lock().unwrap().lfc_prewarm_state = LfcPrewarmState::Completed;
-                return;
-            };
-            crate::metrics::LFC_PREWARM_ERRORS.inc();
-            error!(%err, "prewarming lfc");
-            cloned.state.lock().unwrap().lfc_prewarm_state = LfcPrewarmState::Failed {
-                error: err.to_string(),
+            let state = match cloned.prewarm_impl(from_endpoint).await {
+                Ok(true) => LfcPrewarmState::Completed,
+                Ok(false) => {
+                    info!(
+                        "skipping LFC prewarm because LFC state is not found in endpoint storage"
+                    );
+                    LfcPrewarmState::Skipped
+                }
+                Err(err) => {
+                    crate::metrics::LFC_PREWARM_ERRORS.inc();
+                    error!(%err, "could not prewarm LFC");
+                    LfcPrewarmState::Failed {
+                        error: format!("{err:#}"),
+                    }
+                }
             };
+
+            cloned.state.lock().unwrap().lfc_prewarm_state = state;
         });
         true
     }
@@ -120,15 +130,25 @@ impl ComputeNode {
         EndpointStoragePair::from_spec_and_endpoint(state.pspec.as_ref().unwrap(), from_endpoint)
     }
 
-    async fn prewarm_impl(&self, from_endpoint: Option<String>) -> Result<()> {
+    /// Request LFC state from endpoint storage and load corresponding pages into Postgres.
+    /// Returns a result with `false` if the LFC state is not found in endpoint storage.
+    async fn prewarm_impl(&self, from_endpoint: Option<String>) -> Result<bool> {
         let EndpointStoragePair { url, token } = self.endpoint_storage_pair(from_endpoint)?;
-        info!(%url, "requesting LFC state from endpoint storage");
 
+        #[cfg(feature = "testing")]
+        fail::fail_point!("compute-prewarm", |_| {
+            bail!("prewarm configured to fail because of a failpoint")
+        });
+
+        info!(%url, "requesting LFC state from endpoint storage");
         let request = Client::new().get(&url).bearer_auth(token);
         let res = request.send().await.context("querying endpoint storage")?;
-        let status = res.status();
-        if status != StatusCode::OK {
-            bail!("{status} querying endpoint storage")
+        match res.status() {
+            StatusCode::OK => (),
+            StatusCode::NOT_FOUND => {
+                return Ok(false);
+            }
+            status => bail!("{status} querying endpoint storage"),
         }
 
         let mut uncompressed = Vec::new();
@@ -141,7 +161,8 @@ impl ComputeNode {
             .await
             .context("decoding LFC state")?;
         let uncompressed_len = uncompressed.len();
-        info!(%url, "downloaded LFC state, uncompressed size {uncompressed_len}, loading into postgres");
+
+        info!(%url, "downloaded LFC state, uncompressed size {uncompressed_len}, loading into Postgres");
 
         ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
@@ -149,7 +170,9 @@ impl ComputeNode {
             .query_one("select neon.prewarm_local_cache($1)", &[&uncompressed])
             .await
             .context("loading LFC state into postgres")
-            .map(|_| ())
+            .map(|_| ())?;
+
+        Ok(true)
     }
 
     /// If offload request is ongoing, return false, true otherwise
@@ -177,41 +200,53 @@ impl ComputeNode {
 
     async fn offload_lfc_with_state_update(&self) {
         crate::metrics::LFC_OFFLOADS.inc();
+
         let Err(err) = self.offload_lfc_impl().await else {
             self.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Completed;
             return;
         };
+
         crate::metrics::LFC_OFFLOAD_ERRORS.inc();
-        error!(%err, "offloading lfc");
+        error!(%err, "could not offload LFC state to endpoint storage");
         self.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Failed {
-            error: err.to_string(),
+            error: format!("{err:#}"),
         };
     }
 
     async fn offload_lfc_impl(&self) -> Result<()> {
         let EndpointStoragePair { url, token } = self.endpoint_storage_pair(None)?;
-        info!(%url, "requesting LFC state from postgres");
+        info!(%url, "requesting LFC state from Postgres");
 
-        let mut compressed = Vec::new();
-        ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
+        let row = ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
             .context("connecting to postgres")?
             .query_one("select neon.get_local_cache_state()", &[])
             .await
-            .context("querying LFC state")?
-            .try_get::<usize, &[u8]>(0)
-            .context("deserializing LFC state")
-            .map(ZstdEncoder::new)?
+            .context("querying LFC state")?;
+        let state = row
+            .try_get::<usize, Option<&[u8]>>(0)
+            .context("deserializing LFC state")?;
+        let Some(state) = state else {
+            info!(%url, "empty LFC state, not exporting");
+            return Ok(());
+        };
+
+        let mut compressed = Vec::new();
+        ZstdEncoder::new(state)
             .read_to_end(&mut compressed)
             .await
             .context("compressing LFC state")?;
+
         let compressed_len = compressed.len();
         info!(%url, "downloaded LFC state, compressed size {compressed_len}, writing to endpoint storage");
 
         let request = Client::new().put(url).bearer_auth(token).body(compressed);
         match request.send().await {
             Ok(res) if res.status() == StatusCode::OK => Ok(()),
-            Ok(res) => bail!("Error writing to endpoint storage: {}", res.status()),
+            Ok(res) => bail!(
+                "Request to endpoint storage failed with status: {}",
+                res.status()
+            ),
             Err(err) => Err(err).context("writing to endpoint storage"),
         }
     }

compute_tools/src/compute_promote.rs

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

compute_tools/src/config.rs

@@ -9,6 +9,7 @@ use std::path::Path;
 use compute_api::responses::TlsConfig;
 use compute_api::spec::{ComputeAudit, ComputeMode, ComputeSpec, GenericOption};
 
+use crate::compute::ComputeNodeParams;
 use crate::pg_helpers::{
     GenericOptionExt, GenericOptionsSearch, PgOptionsSerialize, escape_conf_value,
 };
@@ -41,6 +42,7 @@ pub fn line_in_file(path: &Path, line: &str) -> Result<bool> {
 /// Create or completely rewrite configuration file specified by `path`
 pub fn write_postgres_conf(
     pgdata_path: &Path,
+    params: &ComputeNodeParams,
     spec: &ComputeSpec,
     extension_server_port: u16,
     tls_config: &Option<TlsConfig>,
@@ -54,14 +56,15 @@ pub fn write_postgres_conf(
         writeln!(file, "{conf}")?;
     }
 
+    // Stripe size GUC should be defined prior to connection string
+    if let Some(stripe_size) = spec.shard_stripe_size {
+        writeln!(file, "neon.stripe_size={stripe_size}")?;
+    }
     // Add options for connecting to storage
     writeln!(file, "# Neon storage settings")?;
     if let Some(s) = &spec.pageserver_connstring {
         writeln!(file, "neon.pageserver_connstring={}", escape_conf_value(s))?;
     }
-    if let Some(stripe_size) = spec.shard_stripe_size {
-        writeln!(file, "neon.stripe_size={stripe_size}")?;
-    }
     if !spec.safekeeper_connstrings.is_empty() {
         let mut neon_safekeepers_value = String::new();
         tracing::info!(
@@ -161,6 +164,12 @@ pub fn write_postgres_conf(
         }
     }
 
+    writeln!(
+        file,
+        "neon.privileged_role_name={}",
+        escape_conf_value(params.privileged_role_name.as_str())
+    )?;
+
     // If there are any extra options in the 'settings' field, append those
     if spec.cluster.settings.is_some() {
         writeln!(file, "# Managed by compute_ctl: begin")?;

compute_tools/src/hadron_metrics.rs

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

compute_tools/src/http/middleware/authorize.rs

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

compute_tools/src/http/openapi_spec.yaml

@@ -96,7 +96,7 @@ paths:
         content:
           application/json:
             schema:
-                $ref: "#/components/schemas/SafekeepersLsn"
+              $ref: "#/components/schemas/ComputeSchemaWithLsn"
       responses:
         200:
           description: Promote succeeded or wasn't started
@@ -297,14 +297,7 @@ paths:
         content:
           application/json:
             schema:
-              type: object
-              required:
-                - spec
-              properties:
-                spec:
-                  # XXX: I don't want to explain current spec in the OpenAPI format,
-                  # as it could be changed really soon. Consider doing it later.
-                  type: object
+              $ref: "#/components/schemas/ComputeSchema"
       responses:
         200:
           description: Compute configuration finished.
@@ -591,18 +584,25 @@ components:
           type: string
           example: "1.0.0"
 
-    SafekeepersLsn:
+    ComputeSchema:
       type: object
       required:
-        - safekeepers
+        - spec
+      properties:
+        spec:
+          type: object
+    ComputeSchemaWithLsn:
+      type: object
+      required:
+        - spec
         - wal_flush_lsn
       properties:
-        safekeepers:
-          description: Primary replica safekeepers
-          type: string
+        spec:
+          $ref: "#/components/schemas/ComputeState"
         wal_flush_lsn:
-          description: Primary last WAL flush LSN
           type: string
+          description: "last WAL flush LSN"
+          example: "0/028F10D8"
 
     LfcPrewarmState:
       type: object
@@ -613,11 +613,11 @@ components:
         - skipped
       properties:
         status:
-          description: Lfc prewarm status
-          enum: [not_prewarmed, prewarming, completed, failed]
+          description: LFC prewarm status
+          enum: [not_prewarmed, prewarming, completed, failed, skipped]
           type: string
         error:
-          description: Lfc prewarm error, if any
+          description: LFC prewarm error, if any
           type: string
         total:
           description: Total pages processed
@@ -635,11 +635,11 @@ components:
         - status
       properties:
         status:
-          description: Lfc offload status
+          description: LFC offload status
           enum: [not_offloaded, offloading, completed, failed]
           type: string
         error:
-          description: Lfc offload error, if any
+          description: LFC offload error, if any
           type: string
 
     PromoteState:

compute_tools/src/http/routes/hadron_liveness_probe.rs

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

compute_tools/src/http/routes/metrics.rs

@@ -1,10 +1,18 @@
+use std::path::Path;
+use std::sync::Arc;
+
+use anyhow::Context;
 use axum::body::Body;
+use axum::extract::State;
 use axum::response::Response;
-use http::StatusCode;
 use http::header::CONTENT_TYPE;
+use http_body_util::BodyExt;
+use hyper::{Request, StatusCode};
 use metrics::proto::MetricFamily;
 use metrics::{Encoder, TextEncoder};
 
+use crate::communicator_socket_client::connect_communicator_socket;
+use crate::compute::ComputeNode;
 use crate::http::JsonResponse;
 use crate::metrics::collect;
 
@@ -31,3 +39,42 @@ pub(in crate::http) async fn get_metrics() -> Response {
         .body(Body::from(buffer))
         .unwrap()
 }
+
+/// Fetch and forward metrics from the Postgres neon extension's metrics
+/// exporter that are used by autoscaling-agent.
+///
+/// The neon extension exposes these metrics over a Unix domain socket
+/// in the data directory. That's not accessible directly from the outside
+/// world, so we have this endpoint in compute_ctl to expose it
+pub(in crate::http) async fn get_autoscaling_metrics(
+    State(compute): State<Arc<ComputeNode>>,
+) -> Result<Response, Response> {
+    let pgdata = Path::new(&compute.params.pgdata);
+
+    // Connect to the communicator process's metrics socket
+    let mut metrics_client = connect_communicator_socket(pgdata)
+        .await
+        .map_err(|e| JsonResponse::error(StatusCode::INTERNAL_SERVER_ERROR, format!("{e:#}")))?;
+
+    // Make a request for /autoscaling_metrics
+    let request = Request::builder()
+        .method("GET")
+        .uri("/autoscaling_metrics")
+        .header("Host", "localhost") // hyper requires Host, even though the server won't care
+        .body(Body::from(""))
+        .unwrap();
+    let resp = metrics_client
+        .send_request(request)
+        .await
+        .context("fetching metrics from Postgres metrics service")
+        .map_err(|e| JsonResponse::error(StatusCode::INTERNAL_SERVER_ERROR, format!("{e:#}")))?;
+
+    // Build a response that just forwards the response we got.
+    let mut response = Response::builder();
+    response = response.status(resp.status());
+    if let Some(content_type) = resp.headers().get(CONTENT_TYPE) {
+        response = response.header(CONTENT_TYPE, content_type);
+    }
+    let body = tonic::service::AxumBody::from_stream(resp.into_body().into_data_stream());
+    Ok(response.body(body).unwrap())
+}

compute_tools/src/http/routes/mod.rs

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

compute_tools/src/http/routes/promote.rs

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

compute_tools/src/http/routes/refresh_configuration.rs

@@ -0,0 +1,34 @@
+// This file is added by Hadron
+
+use std::sync::Arc;
+
+use axum::{
+    extract::State,
+    response::{IntoResponse, Response},
+};
+use http::StatusCode;
+use tracing::debug;
+
+use crate::compute::ComputeNode;
+// use crate::hadron_metrics::POSTGRES_PAGESTREAM_REQUEST_ERRORS;
+use crate::http::JsonResponse;
+
+// The /refresh_configuration POST method is used to nudge compute_ctl to pull a new spec
+// from the HCC and attempt to reconfigure Postgres with the new spec. The method does not wait
+// for the reconfiguration to complete. Rather, it simply delivers a signal that will cause
+// configuration to be reloaded in a best effort manner. Invocation of this method does not
+// guarantee that a reconfiguration will occur. The caller should consider keep sending this
+// request while it believes that the compute configuration is out of date.
+pub(in crate::http) async fn refresh_configuration(
+    State(compute): State<Arc<ComputeNode>>,
+) -> Response {
+    debug!("serving /refresh_configuration POST request");
+    // POSTGRES_PAGESTREAM_REQUEST_ERRORS.inc();
+    match compute.signal_refresh_configuration().await {
+        Ok(_) => StatusCode::OK.into_response(),
+        Err(e) => {
+            tracing::error!("error handling /refresh_configuration request: {}", e);
+            JsonResponse::error(StatusCode::INTERNAL_SERVER_ERROR, e)
+        }
+    }
+}

compute_tools/src/http/server.rs

@@ -27,6 +27,7 @@ use super::{
     },
 };
 use crate::compute::ComputeNode;
+use crate::http::routes::{hadron_liveness_probe, refresh_configuration};
 
 /// `compute_ctl` has two servers: internal and external. The internal server
 /// binds to the loopback interface and handles communication from clients on
@@ -43,6 +44,7 @@ pub enum Server {
         port: u16,
         config: ComputeCtlConfig,
         compute_id: String,
+        instance_id: Option<String>,
     },
 }
 
@@ -67,7 +69,12 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
                         post(extension_server::download_extension),
                     )
                     .route("/extensions", post(extensions::install_extension))
-                    .route("/grants", post(grants::add_grant));
+                    .route("/grants", post(grants::add_grant))
+                    // Hadron: Compute-initiated configuration refresh
+                    .route(
+                        "/refresh_configuration",
+                        post(refresh_configuration::refresh_configuration),
+                    );
 
                 // Add in any testing support
                 if cfg!(feature = "testing") {
@@ -79,10 +86,17 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
                 router
             }
             Server::External {
-                config, compute_id, ..
+                config,
+                compute_id,
+                instance_id,
+                ..
             } => {
-                let unauthenticated_router =
-                    Router::<Arc<ComputeNode>>::new().route("/metrics", get(metrics::get_metrics));
+                let unauthenticated_router = Router::<Arc<ComputeNode>>::new()
+                    .route("/metrics", get(metrics::get_metrics))
+                    .route(
+                        "/autoscaling_metrics",
+                        get(metrics::get_autoscaling_metrics),
+                    );
 
                 let authenticated_router = Router::<Arc<ComputeNode>>::new()
                     .route("/lfc/prewarm", get(lfc::prewarm_state).post(lfc::prewarm))
@@ -96,8 +110,13 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
                     .route("/metrics.json", get(metrics_json::get_metrics))
                     .route("/status", get(status::get_status))
                     .route("/terminate", post(terminate::terminate))
+                    .route(
+                        "/hadron_liveness_probe",
+                        get(hadron_liveness_probe::hadron_liveness_probe),
+                    )
                     .layer(AsyncRequireAuthorizationLayer::new(Authorize::new(
                         compute_id.clone(),
+                        instance_id.clone(),
                         config.jwks.clone(),
                     )));
 

compute_tools/src/installed_extensions.rs

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

compute_tools/src/lib.rs

@@ -4,6 +4,7 @@
 #![deny(clippy::undocumented_unsafe_blocks)]
 
 pub mod checker;
+pub mod communicator_socket_client;
 pub mod config;
 pub mod configurator;
 pub mod http;
@@ -15,6 +16,7 @@ pub mod compute_prewarm;
 pub mod compute_promote;
 pub mod disk_quota;
 pub mod extension_server;
+pub mod hadron_metrics;
 pub mod installed_extensions;
 pub mod local_proxy;
 pub mod lsn_lease;
@@ -23,6 +25,7 @@ mod migration;
 pub mod monitor;
 pub mod params;
 pub mod pg_helpers;
+pub mod pg_isready;
 pub mod pgbouncer;
 pub mod rsyslog;
 pub mod spec;

compute_tools/src/logger.rs

@@ -1,7 +1,10 @@
 use std::collections::HashMap;
+use std::sync::{LazyLock, RwLock};
+use tracing::Subscriber;
 use tracing::info;
-use tracing_subscriber::layer::SubscriberExt;
+use tracing_appender;
 use tracing_subscriber::prelude::*;
+use tracing_subscriber::{fmt, layer::SubscriberExt, registry::LookupSpan};
 
 /// Initialize logging to stderr, and OpenTelemetry tracing and exporter.
 ///
@@ -13,31 +16,63 @@ use tracing_subscriber::prelude::*;
 /// set `OTEL_EXPORTER_OTLP_ENDPOINT=http://jaeger:4318`. See
 /// `tracing-utils` package description.
 ///
-pub async fn init_tracing_and_logging(default_log_level: &str) -> anyhow::Result<()> {
+pub fn init_tracing_and_logging(
+    default_log_level: &str,
+    log_dir_opt: &Option<String>,
+) -> anyhow::Result<(
+    Option<tracing_utils::Provider>,
+    Option<tracing_appender::non_blocking::WorkerGuard>,
+)> {
     // Initialize Logging
     let env_filter = tracing_subscriber::EnvFilter::try_from_default_env()
         .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new(default_log_level));
 
+    // Standard output streams
     let fmt_layer = tracing_subscriber::fmt::layer()
         .with_ansi(false)
         .with_target(false)
         .with_writer(std::io::stderr);
 
+    // Logs with file rotation. Files in `$log_dir/pgcctl.yyyy-MM-dd`
+    let (json_to_file_layer, _file_logs_guard) = if let Some(log_dir) = log_dir_opt {
+        std::fs::create_dir_all(log_dir)?;
+        let file_logs_appender = tracing_appender::rolling::RollingFileAppender::builder()
+            .rotation(tracing_appender::rolling::Rotation::DAILY)
+            .filename_prefix("pgcctl")
+            // Lib appends to existing files, so we will keep files for up to 2 days even on restart loops.
+            // At minimum, log-daemon will have 1 day to detect and upload a file (if created right before midnight).
+            .max_log_files(2)
+            .build(log_dir)
+            .expect("Initializing rolling file appender should succeed");
+        let (file_logs_writer, _file_logs_guard) =
+            tracing_appender::non_blocking(file_logs_appender);
+        let json_to_file_layer = tracing_subscriber::fmt::layer()
+            .with_ansi(false)
+            .with_target(false)
+            .event_format(PgJsonLogShapeFormatter)
+            .with_writer(file_logs_writer);
+        (Some(json_to_file_layer), Some(_file_logs_guard))
+    } else {
+        (None, None)
+    };
+
     // Initialize OpenTelemetry
-    let otlp_layer =
-        tracing_utils::init_tracing("compute_ctl", tracing_utils::ExportConfig::default()).await;
+    let provider =
+        tracing_utils::init_tracing("compute_ctl", tracing_utils::ExportConfig::default());
+    let otlp_layer = provider.as_ref().map(tracing_utils::layer);
 
     // Put it all together
     tracing_subscriber::registry()
         .with(env_filter)
         .with(otlp_layer)
         .with(fmt_layer)
+        .with(json_to_file_layer)
         .init();
     tracing::info!("logging and tracing started");
 
     utils::logging::replace_panic_hook_with_tracing_panic_hook().forget();
 
-    Ok(())
+    Ok((provider, _file_logs_guard))
 }
 
 /// Replace all newline characters with a special character to make it
@@ -92,3 +127,157 @@ pub fn startup_context_from_env() -> Option<opentelemetry::Context> {
         None
     }
 }
+
+/// Track relevant id's
+const UNKNOWN_IDS: &str = r#""pg_instance_id": "", "pg_compute_id": """#;
+static IDS: LazyLock<RwLock<String>> = LazyLock::new(|| RwLock::new(UNKNOWN_IDS.to_string()));
+
+pub fn update_ids(instance_id: &Option<String>, compute_id: &Option<String>) -> anyhow::Result<()> {
+    let ids = format!(
+        r#""pg_instance_id": "{}", "pg_compute_id": "{}""#,
+        instance_id.as_ref().map(|s| s.as_str()).unwrap_or_default(),
+        compute_id.as_ref().map(|s| s.as_str()).unwrap_or_default()
+    );
+    let mut guard = IDS
+        .write()
+        .map_err(|e| anyhow::anyhow!("Log set id's rwlock poisoned: {}", e))?;
+    *guard = ids;
+    Ok(())
+}
+
+/// Massage compute_ctl logs into PG json log shape so we can use the same Lumberjack setup.
+struct PgJsonLogShapeFormatter;
+impl<S, N> fmt::format::FormatEvent<S, N> for PgJsonLogShapeFormatter
+where
+    S: Subscriber + for<'a> LookupSpan<'a>,
+    N: for<'a> fmt::format::FormatFields<'a> + 'static,
+{
+    fn format_event(
+        &self,
+        ctx: &fmt::FmtContext<'_, S, N>,
+        mut writer: fmt::format::Writer<'_>,
+        event: &tracing::Event<'_>,
+    ) -> std::fmt::Result {
+        // Format values from the event's metadata, and open message string
+        let metadata = event.metadata();
+        {
+            let ids_guard = IDS.read();
+            let ids = ids_guard
+                .as_ref()
+                .map(|guard| guard.as_str())
+                // Surpress so that we don't lose all uploaded/ file logs if something goes super wrong. We would notice the missing id's.
+                .unwrap_or(UNKNOWN_IDS);
+            write!(
+                &mut writer,
+                r#"{{"timestamp": "{}", "error_severity": "{}", "file_name": "{}", "backend_type": "compute_ctl_self", {}, "message": "#,
+                chrono::Utc::now().format("%Y-%m-%d %H:%M:%S%.3f GMT"),
+                metadata.level(),
+                metadata.target(),
+                ids
+            )?;
+        }
+
+        let mut message = String::new();
+        let message_writer = fmt::format::Writer::new(&mut message);
+
+        // Gather the message
+        ctx.field_format().format_fields(message_writer, event)?;
+
+        // TODO: any better options than to copy-paste this OSS span formatter?
+        // impl<S, N, T> FormatEvent<S, N> for Format<Full, T>
+        // https://docs.rs/tracing-subscriber/latest/tracing_subscriber/fmt/trait.FormatEvent.html#impl-FormatEvent%3CS,+N%3E-for-Format%3CFull,+T%3E
+
+        // write message, close bracket, and new line
+        writeln!(writer, "{}}}", serde_json::to_string(&message).unwrap())
+    }
+}
+
+#[cfg(feature = "testing")]
+#[cfg(test)]
+mod test {
+    use super::*;
+    use std::{cell::RefCell, io};
+
+    // Use thread_local! instead of Mutex for test isolation
+    thread_local! {
+        static WRITER_OUTPUT: RefCell<String> = const { RefCell::new(String::new()) };
+    }
+
+    #[derive(Clone, Default)]
+    struct StaticStringWriter;
+
+    impl io::Write for StaticStringWriter {
+        fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+            let output = String::from_utf8(buf.to_vec()).expect("Invalid UTF-8 in test output");
+            WRITER_OUTPUT.with(|s| s.borrow_mut().push_str(&output));
+            Ok(buf.len())
+        }
+
+        fn flush(&mut self) -> io::Result<()> {
+            Ok(())
+        }
+    }
+
+    impl fmt::MakeWriter<'_> for StaticStringWriter {
+        type Writer = Self;
+
+        fn make_writer(&self) -> Self::Writer {
+            Self
+        }
+    }
+
+    #[test]
+    fn test_log_pg_json_shape_formatter() {
+        // Use a scoped subscriber to prevent global state pollution
+        let subscriber = tracing_subscriber::registry().with(
+            tracing_subscriber::fmt::layer()
+                .with_ansi(false)
+                .with_target(false)
+                .event_format(PgJsonLogShapeFormatter)
+                .with_writer(StaticStringWriter),
+        );
+
+        let _ = update_ids(&Some("000".to_string()), &Some("111".to_string()));
+
+        // Clear any previous test state
+        WRITER_OUTPUT.with(|s| s.borrow_mut().clear());
+
+        let messages = [
+            "test message",
+            r#"json escape check:  name="BatchSpanProcessor.Flush.ExportError" reason="Other(reqwest::Error { kind: Request, url: \"http://localhost:4318/v1/traces\", source: hyper_
+            util::client::legacy::Error(Connect, ConnectError(\"tcp connect error\", Os { code: 111, kind: ConnectionRefused, message: \"Connection refused\" })) })" Failed during the export process"#,
+        ];
+
+        tracing::subscriber::with_default(subscriber, || {
+            for message in messages {
+                tracing::info!(message);
+            }
+        });
+        tracing::info!("not test message");
+
+        // Get captured output
+        let output = WRITER_OUTPUT.with(|s| s.borrow().clone());
+
+        let json_strings: Vec<&str> = output.lines().collect();
+        assert_eq!(
+            json_strings.len(),
+            messages.len(),
+            "Log didn't have the expected number of json strings."
+        );
+
+        let json_string_shape_regex = regex::Regex::new(
+            r#"\{"timestamp": "\d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2}\.\d{3} GMT", "error_severity": "INFO", "file_name": ".+", "backend_type": "compute_ctl_self", "pg_instance_id": "000", "pg_compute_id": "111", "message": ".+"\}"#
+        ).unwrap();
+
+        for (i, expected_message) in messages.iter().enumerate() {
+            let json_string = json_strings[i];
+            assert!(
+                json_string_shape_regex.is_match(json_string),
+                "Json log didn't match expected pattern:\n{json_string}",
+            );
+            let parsed_json: serde_json::Value = serde_json::from_str(json_string).unwrap();
+            let actual_message = parsed_json["message"].as_str().unwrap();
+            assert_eq!(*expected_message, actual_message);
+        }
+    }
+}

compute_tools/src/migration.rs

@@ -9,15 +9,20 @@ use crate::metrics::DB_MIGRATION_FAILED;
 pub(crate) struct MigrationRunner<'m> {
     client: &'m mut Client,
     migrations: &'m [&'m str],
+    lakebase_mode: bool,
 }
 
 impl<'m> MigrationRunner<'m> {
     /// Create a new migration runner
-    pub fn new(client: &'m mut Client, migrations: &'m [&'m str]) -> Self {
+    pub fn new(client: &'m mut Client, migrations: &'m [&'m str], lakebase_mode: bool) -> Self {
         // The neon_migration.migration_id::id column is a bigint, which is equivalent to an i64
         assert!(migrations.len() + 1 < i64::MAX as usize);
 
-        Self { client, migrations }
+        Self {
+            client,
+            migrations,
+            lakebase_mode,
+        }
     }
 
     /// Get the current value neon_migration.migration_id
@@ -130,8 +135,13 @@ impl<'m> MigrationRunner<'m> {
             // ID is also the next index
             let migration_id = (current_migration + 1) as i64;
             let migration = self.migrations[current_migration];
+            let migration = if self.lakebase_mode {
+                migration.replace("neon_superuser", "databricks_superuser")
+            } else {
+                migration.to_string()
+            };
 
-            match Self::run_migration(self.client, migration_id, migration).await {
+            match Self::run_migration(self.client, migration_id, &migration).await {
                 Ok(_) => {
                     info!("Finished migration id={}", migration_id);
                 }

compute_tools/src/migrations/0001-add_bypass_rls_to_privileged_role.sql

@@ -0,0 +1 @@
+ALTER ROLE {privileged_role_name} BYPASSRLS;

compute_tools/src/migrations/0001-neon_superuser_bypass_rls.sql

@@ -1 +0,0 @@
-ALTER ROLE neon_superuser BYPASSRLS;

compute_tools/src/migrations/0002-alter_roles.sql

@@ -15,7 +15,7 @@ DO $$
 DECLARE
     role_name text;
 BEGIN
-    FOR role_name IN SELECT rolname FROM pg_roles WHERE pg_has_role(rolname, 'neon_superuser', 'member')
+    FOR role_name IN SELECT rolname FROM pg_roles WHERE pg_has_role(rolname, '{privileged_role_name}', 'member')
     LOOP
         RAISE NOTICE 'EXECUTING ALTER ROLE % INHERIT', quote_ident(role_name);
         EXECUTE 'ALTER ROLE ' || quote_ident(role_name) || ' INHERIT';
@@ -23,7 +23,7 @@ BEGIN
 
     FOR role_name IN SELECT rolname FROM pg_roles
         WHERE
-            NOT pg_has_role(rolname, 'neon_superuser', 'member') AND NOT starts_with(rolname, 'pg_')
+            NOT pg_has_role(rolname, '{privileged_role_name}', 'member') AND NOT starts_with(rolname, 'pg_')
     LOOP
         RAISE NOTICE 'EXECUTING ALTER ROLE % NOBYPASSRLS', quote_ident(role_name);
         EXECUTE 'ALTER ROLE ' || quote_ident(role_name) || ' NOBYPASSRLS';

compute_tools/src/migrations/0003-grant_pg_create_subscription_to_privileged_role.sql

@@ -1,6 +1,6 @@
 DO $$
 BEGIN
     IF (SELECT setting::numeric >= 160000 FROM pg_settings WHERE name = 'server_version_num') THEN
-        EXECUTE 'GRANT pg_create_subscription TO neon_superuser';
+        EXECUTE 'GRANT pg_create_subscription TO {privileged_role_name}';
     END IF;
 END $$;

compute_tools/src/migrations/0004-grant_pg_monitor_to_neon_superuser.sql

@@ -1 +0,0 @@
-GRANT pg_monitor TO neon_superuser WITH ADMIN OPTION;

compute_tools/src/migrations/0004-grant_pg_monitor_to_privileged_role.sql

@@ -0,0 +1 @@
+GRANT pg_monitor TO {privileged_role_name} WITH ADMIN OPTION;

compute_tools/src/migrations/0005-grant_all_on_tables_to_privileged_role.sql

@@ -1,4 +1,4 @@
 -- SKIP: Deemed insufficient for allowing relations created by extensions to be
---       interacted with by neon_superuser without permission issues.
+--       interacted with by {privileged_role_name} without permission issues.
 
-ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO neon_superuser;
+ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO {privileged_role_name};

compute_tools/src/migrations/0006-grant_all_on_sequences_to_privileged_role.sql

@@ -1,4 +1,4 @@
 -- SKIP: Deemed insufficient for allowing relations created by extensions to be
---       interacted with by neon_superuser without permission issues.
+--       interacted with by {privileged_role_name} without permission issues.
 
-ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO neon_superuser;
+ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO {privileged_role_name};

compute_tools/src/migrations/0007-grant_all_on_tables_with_grant_option_to_privileged_role.sql

@@ -1,3 +1,3 @@
 -- SKIP: Moved inline to the handle_grants() functions.
 
-ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO neon_superuser WITH GRANT OPTION;
+ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO {privileged_role_name} WITH GRANT OPTION;

compute_tools/src/migrations/0008-grant_all_on_sequences_with_grant_option_to_privileged_role.sql

@@ -1,3 +1,3 @@
 -- SKIP: Moved inline to the handle_grants() functions.
 
-ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO neon_superuser WITH GRANT OPTION;
+ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO {privileged_role_name} WITH GRANT OPTION;

compute_tools/src/migrations/0010-grant_snapshot_synchronization_funcs_to_privileged_role.sql

@@ -1,7 +1,7 @@
 DO $$
 BEGIN
     IF (SELECT setting::numeric >= 160000 FROM pg_settings WHERE name = 'server_version_num') THEN
-       EXECUTE 'GRANT EXECUTE ON FUNCTION pg_export_snapshot TO neon_superuser';
-       EXECUTE 'GRANT EXECUTE ON FUNCTION pg_log_standby_snapshot TO neon_superuser';
+       EXECUTE 'GRANT EXECUTE ON FUNCTION pg_export_snapshot TO {privileged_role_name}';
+       EXECUTE 'GRANT EXECUTE ON FUNCTION pg_log_standby_snapshot TO {privileged_role_name}';
     END IF;
 END $$;

compute_tools/src/migrations/0011-grant_pg_show_replication_origin_status_to_neon_superuser.sql

@@ -1 +0,0 @@
-GRANT EXECUTE ON FUNCTION pg_show_replication_origin_status TO neon_superuser;

compute_tools/src/migrations/0011-grant_pg_show_replication_origin_status_to_privileged_role.sql

@@ -0,0 +1 @@
+GRANT EXECUTE ON FUNCTION pg_show_replication_origin_status TO {privileged_role_name};

compute_tools/src/migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql

@@ -1 +0,0 @@
-GRANT pg_signal_backend TO neon_superuser WITH ADMIN OPTION;

compute_tools/src/migrations/0012-grant_pg_signal_backend_to_privileged_role.sql

@@ -0,0 +1 @@
+GRANT pg_signal_backend TO {privileged_role_name} WITH ADMIN OPTION;

compute_tools/src/monitor.rs

@@ -11,6 +11,7 @@ use tracing::{Level, error, info, instrument, span};
 use crate::compute::ComputeNode;
 use crate::metrics::{PG_CURR_DOWNTIME_MS, PG_TOTAL_DOWNTIME_MS};
 
+const PG_DEFAULT_INIT_TIMEOUIT: Duration = Duration::from_secs(60);
 const MONITOR_CHECK_INTERVAL: Duration = Duration::from_millis(500);
 
 /// Struct to store runtime state of the compute monitor thread.
@@ -352,13 +353,47 @@ impl ComputeMonitor {
 // Hang on condition variable waiting until the compute status is `Running`.
 fn wait_for_postgres_start(compute: &ComputeNode) {
     let mut state = compute.state.lock().unwrap();
+    let pg_init_timeout = compute
+        .params
+        .pg_init_timeout
+        .unwrap_or(PG_DEFAULT_INIT_TIMEOUIT);
+
     while state.status != ComputeStatus::Running {
         info!("compute is not running, waiting before monitoring activity");
-        state = compute.state_changed.wait(state).unwrap();
+        if !compute.params.lakebase_mode {
+            state = compute.state_changed.wait(state).unwrap();
 
-        if state.status == ComputeStatus::Running {
-            break;
+            if state.status == ComputeStatus::Running {
+                break;
+            }
+            continue;
         }
+
+        if state.pg_start_time.is_some()
+            && Utc::now()
+                .signed_duration_since(state.pg_start_time.unwrap())
+                .to_std()
+                .unwrap_or_default()
+                > pg_init_timeout
+        {
+            // If Postgres isn't up and running with working PS/SK connections within POSTGRES_STARTUP_TIMEOUT, it is
+            // possible that we started Postgres with a wrong spec (so it is talking to the wrong PS/SK nodes). To prevent
+            // deadends we simply exit (panic) the compute node so it can restart with the latest spec.
+            //
+            // NB: We skip this check if we have not attempted to start PG yet (indicated by state.pg_start_up == None).
+            // This is to make sure the more appropriate errors are surfaced if we encounter issues before we even attempt
+            // to start PG (e.g., if we can't pull the spec, can't sync safekeepers, or can't get the basebackup).
+            error!(
+                "compute did not enter Running state in {} seconds, exiting",
+                pg_init_timeout.as_secs()
+            );
+            std::process::exit(1);
+        }
+        state = compute
+            .state_changed
+            .wait_timeout(state, Duration::from_secs(5))
+            .unwrap()
+            .0;
     }
 }
 

compute_tools/src/pg_helpers.rs

@@ -11,7 +11,9 @@ use std::time::{Duration, Instant};
 
 use anyhow::{Result, bail};
 use compute_api::responses::TlsConfig;
-use compute_api::spec::{Database, GenericOption, GenericOptions, PgIdent, Role};
+use compute_api::spec::{
+    Database, DatabricksSettings, GenericOption, GenericOptions, PgIdent, Role,
+};
 use futures::StreamExt;
 use indexmap::IndexMap;
 use ini::Ini;
@@ -184,6 +186,42 @@ impl DatabaseExt for Database {
     }
 }
 
+pub trait DatabricksSettingsExt {
+    fn as_pg_settings(&self) -> String;
+}
+
+impl DatabricksSettingsExt for DatabricksSettings {
+    fn as_pg_settings(&self) -> String {
+        // Postgres GUCs rendered from DatabricksSettings
+        vec![
+            // ssl_ca_file
+            Some(format!(
+                "ssl_ca_file = '{}'",
+                self.pg_compute_tls_settings.ca_file
+            )),
+            // [Optional] databricks.workspace_url
+            Some(format!(
+                "databricks.workspace_url = '{}'",
+                &self.databricks_workspace_host
+            )),
+            // todo(vikas.jain): these are not required anymore as they are moved to static
+            // conf but keeping these to avoid image mismatch between hcc and pg.
+            // Once hcc and pg are in sync, we can remove these.
+            //
+            // databricks.enable_databricks_identity_login
+            Some("databricks.enable_databricks_identity_login = true".to_string()),
+            // databricks.enable_sql_restrictions
+            Some("databricks.enable_sql_restrictions = true".to_string()),
+        ]
+        .into_iter()
+        // Removes `None`s
+        .flatten()
+        .collect::<Vec<String>>()
+        .join("\n")
+            + "\n"
+    }
+}
+
 /// Generic trait used to provide quoting / encoding for strings used in the
 /// Postgres SQL queries and DATABASE_URL.
 pub trait Escaping {

compute_tools/src/pg_isready.rs

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

compute_tools/src/spec.rs

@@ -1,4 +1,6 @@
 use std::fs::File;
+use std::fs::{self, Permissions};
+use std::os::unix::fs::PermissionsExt;
 use std::path::Path;
 
 use anyhow::{Result, anyhow, bail};
@@ -9,6 +11,7 @@ use reqwest::StatusCode;
 use tokio_postgres::Client;
 use tracing::{error, info, instrument};
 
+use crate::compute::ComputeNodeParams;
 use crate::config;
 use crate::metrics::{CPLANE_REQUESTS_TOTAL, CPlaneRequestRPC, UNKNOWN_HTTP_STATUS};
 use crate::migration::MigrationRunner;
@@ -132,10 +135,25 @@ pub fn get_config_from_control_plane(base_uri: &str, compute_id: &str) -> Result
 }
 
 /// Check `pg_hba.conf` and update if needed to allow external connections.
-pub fn update_pg_hba(pgdata_path: &Path) -> Result<()> {
+pub fn update_pg_hba(pgdata_path: &Path, databricks_pg_hba: Option<&String>) -> Result<()> {
     // XXX: consider making it a part of config.json
     let pghba_path = pgdata_path.join("pg_hba.conf");
 
+    // Update pg_hba to contains databricks specfic settings before adding neon settings
+    // PG uses the first record that matches to perform authentication, so we need to have
+    // our rules before the default ones from neon.
+    // See https://www.postgresql.org/docs/16/auth-pg-hba-conf.html
+    if let Some(databricks_pg_hba) = databricks_pg_hba {
+        if config::line_in_file(
+            &pghba_path,
+            &format!("include_if_exists {}\n", *databricks_pg_hba),
+        )? {
+            info!("updated pg_hba.conf to include databricks_pg_hba.conf");
+        } else {
+            info!("pg_hba.conf already included databricks_pg_hba.conf");
+        }
+    }
+
     if config::line_in_file(&pghba_path, PG_HBA_ALL_MD5)? {
         info!("updated pg_hba.conf to allow external connections");
     } else {
@@ -145,6 +163,59 @@ pub fn update_pg_hba(pgdata_path: &Path) -> Result<()> {
     Ok(())
 }
 
+/// Check `pg_ident.conf` and update if needed to allow databricks config.
+pub fn update_pg_ident(pgdata_path: &Path, databricks_pg_ident: Option<&String>) -> Result<()> {
+    info!("checking pg_ident.conf");
+    let pghba_path = pgdata_path.join("pg_ident.conf");
+
+    // Update pg_ident to contains databricks specfic settings
+    if let Some(databricks_pg_ident) = databricks_pg_ident {
+        if config::line_in_file(
+            &pghba_path,
+            &format!("include_if_exists {}\n", *databricks_pg_ident),
+        )? {
+            info!("updated pg_ident.conf to include databricks_pg_ident.conf");
+        } else {
+            info!("pg_ident.conf already included databricks_pg_ident.conf");
+        }
+    }
+
+    Ok(())
+}
+
+/// Copy tls key_file and cert_file from k8s secret mount directory
+/// to pgdata and set private key file permissions as expected by Postgres.
+/// See this doc for expected permission <https://www.postgresql.org/docs/current/ssl-tcp.html>
+/// K8s secrets mount on dblet does not honor permission and ownership
+/// specified in the Volume or VolumeMount. So we need to explicitly copy the file and set the permissions.
+pub fn copy_tls_certificates(
+    key_file: &String,
+    cert_file: &String,
+    pgdata_path: &Path,
+) -> Result<()> {
+    let files = [cert_file, key_file];
+    for file in files.iter() {
+        let source = Path::new(file);
+        let dest = pgdata_path.join(source.file_name().unwrap());
+        if !dest.exists() {
+            std::fs::copy(source, &dest)?;
+            info!(
+                "Copying tls file: {} to {}",
+                &source.display(),
+                &dest.display()
+            );
+        }
+        if *file == key_file {
+            // Postgres requires private key to be readable only by the owner by having
+            // chmod 600 permissions.
+            let permissions = Permissions::from_mode(0o600);
+            fs::set_permissions(&dest, permissions)?;
+            info!("Setting permission on {}.", &dest.display());
+        }
+    }
+    Ok(())
+}
+
 /// Create a standby.signal file
 pub fn add_standby_signal(pgdata_path: &Path) -> Result<()> {
     // XXX: consider making it a part of config.json
@@ -169,7 +240,11 @@ pub async fn handle_neon_extension_upgrade(client: &mut Client) -> Result<()> {
 }
 
 #[instrument(skip_all)]
-pub async fn handle_migrations(client: &mut Client) -> Result<()> {
+pub async fn handle_migrations(
+    params: ComputeNodeParams,
+    client: &mut Client,
+    lakebase_mode: bool,
+) -> Result<()> {
     info!("handle migrations");
 
     // !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
@@ -178,29 +253,62 @@ pub async fn handle_migrations(client: &mut Client) -> Result<()> {
 
     // Add new migrations in numerical order.
     let migrations = [
-        include_str!("./migrations/0001-neon_superuser_bypass_rls.sql"),
-        include_str!("./migrations/0002-alter_roles.sql"),
-        include_str!("./migrations/0003-grant_pg_create_subscription_to_neon_superuser.sql"),
-        include_str!("./migrations/0004-grant_pg_monitor_to_neon_superuser.sql"),
-        include_str!("./migrations/0005-grant_all_on_tables_to_neon_superuser.sql"),
-        include_str!("./migrations/0006-grant_all_on_sequences_to_neon_superuser.sql"),
-        include_str!(
-            "./migrations/0007-grant_all_on_tables_to_neon_superuser_with_grant_option.sql"
+        &format!(
+            include_str!("./migrations/0001-add_bypass_rls_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
         ),
-        include_str!(
-            "./migrations/0008-grant_all_on_sequences_to_neon_superuser_with_grant_option.sql"
+        &format!(
+            include_str!("./migrations/0002-alter_roles.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0003-grant_pg_create_subscription_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0004-grant_pg_monitor_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0005-grant_all_on_tables_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0006-grant_all_on_sequences_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!(
+                "./migrations/0007-grant_all_on_tables_with_grant_option_to_privileged_role.sql"
+            ),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!(
+                "./migrations/0008-grant_all_on_sequences_with_grant_option_to_privileged_role.sql"
+            ),
+            privileged_role_name = params.privileged_role_name
         ),
         include_str!("./migrations/0009-revoke_replication_for_previously_allowed_roles.sql"),
-        include_str!(
-            "./migrations/0010-grant_snapshot_synchronization_funcs_to_neon_superuser.sql"
+        &format!(
+            include_str!(
+                "./migrations/0010-grant_snapshot_synchronization_funcs_to_privileged_role.sql"
+            ),
+            privileged_role_name = params.privileged_role_name
         ),
-        include_str!(
-            "./migrations/0011-grant_pg_show_replication_origin_status_to_neon_superuser.sql"
+        &format!(
+            include_str!(
+                "./migrations/0011-grant_pg_show_replication_origin_status_to_privileged_role.sql"
+            ),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0012-grant_pg_signal_backend_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
         ),
-        include_str!("./migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql"),
     ];
 
-    MigrationRunner::new(client, &migrations)
+    MigrationRunner::new(client, &migrations, lakebase_mode)
         .run_migrations()
         .await?;
 

compute_tools/src/spec_apply.rs

@@ -13,14 +13,14 @@ use tokio_postgres::Client;
 use tokio_postgres::error::SqlState;
 use tracing::{Instrument, debug, error, info, info_span, instrument, warn};
 
-use crate::compute::{ComputeNode, ComputeState};
+use crate::compute::{ComputeNode, ComputeNodeParams, ComputeState};
 use crate::pg_helpers::{
     DatabaseExt, Escaping, GenericOptionsSearch, RoleExt, get_existing_dbs_async,
     get_existing_roles_async,
 };
 use crate::spec_apply::ApplySpecPhase::{
-    CreateAndAlterDatabases, CreateAndAlterRoles, CreateAvailabilityCheck, CreateNeonSuperuser,
-    CreatePgauditExtension, CreatePgauditlogtofileExtension, CreateSchemaNeon,
+    CreateAndAlterDatabases, CreateAndAlterRoles, CreateAvailabilityCheck, CreatePgauditExtension,
+    CreatePgauditlogtofileExtension, CreatePrivilegedRole, CreateSchemaNeon,
     DisablePostgresDBPgAudit, DropInvalidDatabases, DropRoles, FinalizeDropLogicalSubscriptions,
     HandleNeonExtension, HandleOtherExtensions, RenameAndDeleteDatabases, RenameRoles,
     RunInEachDatabase,
@@ -49,6 +49,7 @@ impl ComputeNode {
             // Proceed with post-startup configuration. Note, that order of operations is important.
             let client = Self::get_maintenance_client(&conf).await?;
             let spec = spec.clone();
+            let params = Arc::new(self.params.clone());
 
             let databases = get_existing_dbs_async(&client).await?;
             let roles = get_existing_roles_async(&client)
@@ -157,6 +158,7 @@ impl ComputeNode {
 
                     let conf = Arc::new(conf);
                     let fut = Self::apply_spec_sql_db(
+                        params.clone(),
                         spec.clone(),
                         conf,
                         ctx.clone(),
@@ -185,7 +187,7 @@ impl ComputeNode {
             }
 
             for phase in [
-                CreateNeonSuperuser,
+                CreatePrivilegedRole,
                 DropInvalidDatabases,
                 RenameRoles,
                 CreateAndAlterRoles,
@@ -195,6 +197,7 @@ impl ComputeNode {
             ] {
                 info!("Applying phase {:?}", &phase);
                 apply_operations(
+                    params.clone(),
                     spec.clone(),
                     ctx.clone(),
                     jwks_roles.clone(),
@@ -243,6 +246,7 @@ impl ComputeNode {
                     }
 
                     let fut = Self::apply_spec_sql_db(
+                        params.clone(),
                         spec.clone(),
                         conf,
                         ctx.clone(),
@@ -293,6 +297,7 @@ impl ComputeNode {
             for phase in phases {
                 debug!("Applying phase {:?}", &phase);
                 apply_operations(
+                    params.clone(),
                     spec.clone(),
                     ctx.clone(),
                     jwks_roles.clone(),
@@ -313,7 +318,9 @@ impl ComputeNode {
     /// May opt to not connect to databases that don't have any scheduled
     /// operations.  The function is concurrency-controlled with the provided
     /// semaphore.  The caller has to make sure the semaphore isn't exhausted.
+    #[allow(clippy::too_many_arguments)] // TODO: needs bigger refactoring
     async fn apply_spec_sql_db(
+        params: Arc<ComputeNodeParams>,
         spec: Arc<ComputeSpec>,
         conf: Arc<tokio_postgres::Config>,
         ctx: Arc<tokio::sync::RwLock<MutableApplyContext>>,
@@ -328,6 +335,7 @@ impl ComputeNode {
 
         for subphase in subphases {
             apply_operations(
+                params.clone(),
                 spec.clone(),
                 ctx.clone(),
                 jwks_roles.clone(),
@@ -403,7 +411,8 @@ impl ComputeNode {
             .map(|limit| match limit {
                 0..10 => limit,
                 10..30 => 10,
-                30.. => limit / 3,
+                30..300 => limit / 3,
+                300.. => 100,
             })
             // If we didn't find max_connections, default to 10 concurrent connections.
             .unwrap_or(10)
@@ -467,7 +476,7 @@ pub enum PerDatabasePhase {
 
 #[derive(Clone, Debug)]
 pub enum ApplySpecPhase {
-    CreateNeonSuperuser,
+    CreatePrivilegedRole,
     DropInvalidDatabases,
     RenameRoles,
     CreateAndAlterRoles,
@@ -510,6 +519,7 @@ pub struct MutableApplyContext {
 /// - No timeouts have (yet) been implemented.
 /// - The caller is responsible for limiting and/or applying concurrency.
 pub async fn apply_operations<'a, Fut, F>(
+    params: Arc<ComputeNodeParams>,
     spec: Arc<ComputeSpec>,
     ctx: Arc<RwLock<MutableApplyContext>>,
     jwks_roles: Arc<HashSet<String>>,
@@ -527,7 +537,7 @@ where
         debug!("Processing phase {:?}", &apply_spec_phase);
         let ctx = ctx;
 
-        let mut ops = get_operations(&spec, &ctx, &jwks_roles, &apply_spec_phase)
+        let mut ops = get_operations(&params, &spec, &ctx, &jwks_roles, &apply_spec_phase)
             .await?
             .peekable();
 
@@ -588,14 +598,18 @@ where
 /// sort/merge/batch execution, but for now this is a nice way to improve
 /// batching behavior of the commands.
 async fn get_operations<'a>(
+    params: &'a ComputeNodeParams,
     spec: &'a ComputeSpec,
     ctx: &'a RwLock<MutableApplyContext>,
     jwks_roles: &'a HashSet<String>,
     apply_spec_phase: &'a ApplySpecPhase,
 ) -> Result<Box<dyn Iterator<Item = Operation> + 'a + Send>> {
     match apply_spec_phase {
-        ApplySpecPhase::CreateNeonSuperuser => Ok(Box::new(once(Operation {
-            query: include_str!("sql/create_neon_superuser.sql").to_string(),
+        ApplySpecPhase::CreatePrivilegedRole => Ok(Box::new(once(Operation {
+            query: format!(
+                include_str!("sql/create_privileged_role.sql"),
+                privileged_role_name = params.privileged_role_name
+            ),
             comment: None,
         }))),
         ApplySpecPhase::DropInvalidDatabases => {
@@ -697,8 +711,9 @@ async fn get_operations<'a>(
                         None => {
                             let query = if !jwks_roles.contains(role.name.as_str()) {
                                 format!(
-                                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser {}",
+                                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE {} {}",
                                     role.name.pg_quote(),
+                                    params.privileged_role_name,
                                     role.to_pg_options(),
                                 )
                             } else {
@@ -849,8 +864,9 @@ async fn get_operations<'a>(
                                 // ALL PRIVILEGES grants CREATE, CONNECT, and TEMPORARY on the database
                                 // (see https://www.postgresql.org/docs/current/ddl-priv.html)
                                 query: format!(
-                                    "GRANT ALL PRIVILEGES ON DATABASE {} TO neon_superuser",
-                                    db.name.pg_quote()
+                                    "GRANT ALL PRIVILEGES ON DATABASE {} TO {}",
+                                    db.name.pg_quote(),
+                                    params.privileged_role_name
                                 ),
                                 comment: None,
                             },

compute_tools/src/sql/create_neon_superuser.sql

@@ -1,8 +0,0 @@
-DO $$
-    BEGIN
-        IF NOT EXISTS (SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
-        THEN
-            CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
-        END IF;
-    END
-$$;

compute_tools/src/sql/create_privileged_role.sql

@@ -0,0 +1,8 @@
+DO $$
+    BEGIN
+        IF NOT EXISTS (SELECT FROM pg_catalog.pg_roles WHERE rolname = '{privileged_role_name}')
+        THEN
+            CREATE ROLE {privileged_role_name} CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
+        END IF;
+    END
+$$;

control_plane/README.md

@@ -8,10 +8,10 @@ code changes locally, but not suitable for running production systems.
 
 ## Example: Start with Postgres 16
 
-To create and start a local development environment with Postgres 16, you will need to provide `--pg-version` flag to 3 of the start-up commands.
+To create and start a local development environment with Postgres 16, you will need to provide `--pg-version` flag to 2 of the start-up commands.
 
 ```shell
-cargo neon init --pg-version 16
+cargo neon init
 cargo neon start
 cargo neon tenant create --set-default --pg-version 16
 cargo neon endpoint create main --pg-version 16

control_plane/src/bin/neon_local.rs

@@ -407,6 +407,12 @@ struct StorageControllerStartCmdArgs {
         help = "Base port for the storage controller instance idenfified by instance-id (defaults to pageserver cplane api)"
     )]
     base_port: Option<u16>,
+
+    #[clap(
+        long,
+        help = "Whether the storage controller should handle pageserver-reported local disk loss events."
+    )]
+    handle_ps_local_disk_loss: Option<bool>,
 }
 
 #[derive(clap::Args)]
@@ -631,6 +637,10 @@ struct EndpointCreateCmdArgs {
         help = "Allow multiple primary endpoints running on the same branch. Shouldn't be used normally, but useful for tests."
     )]
     allow_multiple: bool,
+
+    /// Only allow changing it on creation
+    #[clap(long, help = "Name of the privileged role for the endpoint")]
+    privileged_role_name: Option<String>,
 }
 
 #[derive(clap::Args)]
@@ -1480,6 +1490,7 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 args.grpc,
                 !args.update_catalog,
                 false,
+                args.privileged_role_name.clone(),
             )?;
         }
         EndpointCmd::Start(args) => {
@@ -1506,7 +1517,7 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
             let endpoint = cplane
                 .endpoints
                 .get(endpoint_id.as_str())
-                .ok_or_else(|| anyhow::anyhow!("endpoint {endpoint_id} not found"))?;
+                .ok_or_else(|| anyhow!("endpoint {endpoint_id} not found"))?;
 
             if !args.allow_multiple {
                 cplane.check_conflicting_endpoints(
@@ -1804,6 +1815,7 @@ async fn handle_storage_controller(
                 instance_id: args.instance_id,
                 base_port: args.base_port,
                 start_timeout: args.start_timeout,
+                handle_ps_local_disk_loss: args.handle_ps_local_disk_loss,
             };
 
             if let Err(e) = svc.start(start_args).await {

control_plane/src/endpoint.rs

@@ -65,7 +65,6 @@ use jsonwebtoken::jwk::{
     OctetKeyPairParameters, OctetKeyPairType, PublicKeyUse,
 };
 use nix::sys::signal::{Signal, kill};
-use pageserver_api::shard::ShardStripeSize;
 use pem::Pem;
 use reqwest::header::CONTENT_TYPE;
 use safekeeper_api::PgMajorVersion;
@@ -77,6 +76,7 @@ use spki::{SubjectPublicKeyInfo, SubjectPublicKeyInfoRef};
 use tracing::debug;
 use url::Host;
 use utils::id::{NodeId, TenantId, TimelineId};
+use utils::shard::ShardStripeSize;
 
 use crate::local_env::LocalEnv;
 use crate::postgresql_conf::PostgresConf;
@@ -99,6 +99,7 @@ pub struct EndpointConf {
     features: Vec<ComputeFeature>,
     cluster: Option<Cluster>,
     compute_ctl_config: ComputeCtlConfig,
+    privileged_role_name: Option<String>,
 }
 
 //
@@ -199,6 +200,7 @@ impl ComputeControlPlane {
         grpc: bool,
         skip_pg_catalog_updates: bool,
         drop_subscriptions_before_start: bool,
+        privileged_role_name: Option<String>,
     ) -> Result<Arc<Endpoint>> {
         let pg_port = pg_port.unwrap_or_else(|| self.get_port());
         let external_http_port = external_http_port.unwrap_or_else(|| self.get_port() + 1);
@@ -236,6 +238,7 @@ impl ComputeControlPlane {
             features: vec![],
             cluster: None,
             compute_ctl_config: compute_ctl_config.clone(),
+            privileged_role_name: privileged_role_name.clone(),
         });
 
         ep.create_endpoint_dir()?;
@@ -257,6 +260,7 @@ impl ComputeControlPlane {
                 features: vec![],
                 cluster: None,
                 compute_ctl_config,
+                privileged_role_name,
             })?,
         )?;
         std::fs::write(
@@ -332,6 +336,9 @@ pub struct Endpoint {
 
     /// The compute_ctl config for the endpoint's compute.
     compute_ctl_config: ComputeCtlConfig,
+
+    /// The name of the privileged role for the endpoint.
+    privileged_role_name: Option<String>,
 }
 
 #[derive(PartialEq, Eq)]
@@ -432,6 +439,7 @@ impl Endpoint {
             features: conf.features,
             cluster: conf.cluster,
             compute_ctl_config: conf.compute_ctl_config,
+            privileged_role_name: conf.privileged_role_name,
         })
     }
 
@@ -464,7 +472,7 @@ impl Endpoint {
         conf.append("max_connections", "100");
         conf.append("wal_level", "logical");
         // wal_sender_timeout is the maximum time to wait for WAL replication.
-        // It also defines how often the walreciever will send a feedback message to the wal sender.
+        // It also defines how often the walreceiver will send a feedback message to the wal sender.
         conf.append("wal_sender_timeout", "5s");
         conf.append("listen_addresses", &self.pg_address.ip().to_string());
         conf.append("port", &self.pg_address.port().to_string());
@@ -870,6 +878,10 @@ impl Endpoint {
             cmd.arg("--dev");
         }
 
+        if let Some(privileged_role_name) = self.privileged_role_name.clone() {
+            cmd.args(["--privileged-role-name", &privileged_role_name]);
+        }
+
         let child = cmd.spawn()?;
         // set up a scopeguard to kill & wait for the child in case we panic or bail below
         let child = scopeguard::guard(child, |mut child| {

control_plane/src/storage_controller.rs

@@ -56,6 +56,7 @@ pub struct NeonStorageControllerStartArgs {
     pub instance_id: u8,
     pub base_port: Option<u16>,
     pub start_timeout: humantime::Duration,
+    pub handle_ps_local_disk_loss: Option<bool>,
 }
 
 impl NeonStorageControllerStartArgs {
@@ -64,6 +65,7 @@ impl NeonStorageControllerStartArgs {
             instance_id: 1,
             base_port: None,
             start_timeout,
+            handle_ps_local_disk_loss: None,
         }
     }
 }
@@ -669,6 +671,10 @@ impl StorageController {
 
         println!("Starting storage controller at {scheme}://{host}:{listen_port}");
 
+        if start_args.handle_ps_local_disk_loss.unwrap_or_default() {
+            args.push("--handle-ps-local-disk-loss".to_string());
+        }
+
         background_process::start_process(
             COMMAND,
             &instance_dir,

control_plane/storcon_cli/src/main.rs

@@ -76,6 +76,12 @@ enum Command {
     NodeStartDelete {
         #[arg(long)]
         node_id: NodeId,
+        /// When `force` is true, skip waiting for shards to prewarm during migration.
+        /// This can significantly speed up node deletion since prewarming all shards
+        /// can take considerable time, but may result in slower initial access to
+        /// migrated shards until they warm up naturally.
+        #[arg(long)]
+        force: bool,
     },
     /// Cancel deletion of the specified pageserver and wait for `timeout`
     /// for the operation to be canceled. May be retried.
@@ -952,13 +958,14 @@ async fn main() -> anyhow::Result<()> {
                 .dispatch::<(), ()>(Method::DELETE, format!("control/v1/node/{node_id}"), None)
                 .await?;
         }
-        Command::NodeStartDelete { node_id } => {
+        Command::NodeStartDelete { node_id, force } => {
+            let query = if force {
+                format!("control/v1/node/{node_id}/delete?force=true")
+            } else {
+                format!("control/v1/node/{node_id}/delete")
+            };
             storcon_client
-                .dispatch::<(), ()>(
-                    Method::PUT,
-                    format!("control/v1/node/{node_id}/delete"),
-                    None,
-                )
+                .dispatch::<(), ()>(Method::PUT, query, None)
                 .await?;
             println!("Delete started for {node_id}");
         }

deny.toml

@@ -35,6 +35,7 @@ reason = "The paste crate is a build-only dependency with no runtime components.
 # More documentation for the licenses section can be found here:
 # https://embarkstudios.github.io/cargo-deny/checks/licenses/cfg.html
 [licenses]
+version = 2
 allow = [
     "0BSD",
     "Apache-2.0",

docs/pageserver-services.md

@@ -75,7 +75,7 @@ CLI examples:
 * AWS S3  : `env AWS_ACCESS_KEY_ID='SOMEKEYAAAAASADSAH*#' AWS_SECRET_ACCESS_KEY='SOMEsEcReTsd292v' ${PAGESERVER_BIN} -c "remote_storage={bucket_name='some-sample-bucket',bucket_region='eu-north-1', prefix_in_bucket='/test_prefix/'}"`
 
 For Amazon AWS S3, a key id and secret access key could be located in `~/.aws/credentials` if awscli was ever configured to work with the desired bucket, on the AWS Settings page for a certain user. Also note, that the bucket names does not contain any protocols when used on AWS.
-For local S3 installations, refer to the their documentation for name format and credentials.
+For local S3 installations, refer to their documentation for name format and credentials.
 
 Similar to other pageserver settings, toml config file can be used to configure either of the storages as backup targets.
 Required sections are:

endpoint_storage/src/app.rs

@@ -233,7 +233,7 @@ mod tests {
                 .unwrap()
                 .as_millis();
             use rand::Rng;
-            let random = rand::thread_rng().r#gen::<u32>();
+            let random = rand::rng().random::<u32>();
 
             let s3_config = remote_storage::S3Config {
                 bucket_name: var(REAL_S3_BUCKET).unwrap(),

libs/compute_api/src/responses.rs

@@ -46,16 +46,33 @@ pub struct ExtensionInstallResponse {
     pub version: ExtVersion,
 }
 
+/// Status of the LFC prewarm process. The same state machine is reused for
+/// both autoprewarm (prewarm after compute/Postgres start using the previously
+/// stored LFC state) and explicit prewarming via API.
 #[derive(Serialize, Default, Debug, Clone, PartialEq)]
 #[serde(tag = "status", rename_all = "snake_case")]
 pub enum LfcPrewarmState {
+    /// Default value when compute boots up.
     #[default]
     NotPrewarmed,
+    /// Prewarming thread is active and loading pages into LFC.
     Prewarming,
+    /// We found requested LFC state in the endpoint storage and
+    /// completed prewarming successfully.
     Completed,
-    Failed {
-        error: String,
-    },
+    /// Unexpected error happened during prewarming. Note, `Not Found 404`
+    /// response from the endpoint storage is explicitly excluded here
+    /// because it can normally happen on the first compute start,
+    /// since LFC state is not available yet.
+    Failed { error: String },
+    /// We tried to fetch the corresponding LFC state from the endpoint storage,
+    /// but received `Not Found 404`. This should normally happen only during the
+    /// first endpoint start after creation with `autoprewarm: true`.
+    ///
+    /// During the orchestrated prewarm via API, when a caller explicitly
+    /// provides the LFC state key to prewarm from, it's the caller responsibility
+    /// to handle this status as an error state in this case.
+    Skipped,
 }
 
 impl Display for LfcPrewarmState {
@@ -64,6 +81,7 @@ impl Display for LfcPrewarmState {
             LfcPrewarmState::NotPrewarmed => f.write_str("NotPrewarmed"),
             LfcPrewarmState::Prewarming => f.write_str("Prewarming"),
             LfcPrewarmState::Completed => f.write_str("Completed"),
+            LfcPrewarmState::Skipped => f.write_str("Skipped"),
             LfcPrewarmState::Failed { error } => write!(f, "Error({error})"),
         }
     }
@@ -90,11 +108,10 @@ pub enum PromoteState {
     Failed { error: String },
 }
 
-#[derive(Deserialize, Serialize, Default, Debug, Clone)]
+#[derive(Deserialize, Default, Debug)]
 #[serde(rename_all = "snake_case")]
-/// Result of /safekeepers_lsn
-pub struct SafekeepersLsn {
-    pub safekeepers: String,
+pub struct PromoteConfig {
+    pub spec: ComputeSpec,
     pub wal_flush_lsn: utils::lsn::Lsn,
 }
 

libs/compute_api/src/spec.rs

@@ -416,6 +416,32 @@ pub struct GenericOption {
     pub vartype: String,
 }
 
+/// Postgres compute TLS settings.
+#[derive(Clone, Debug, Deserialize, Serialize, PartialEq)]
+pub struct PgComputeTlsSettings {
+    // Absolute path to the certificate file for server-side TLS.
+    pub cert_file: String,
+    // Absolute path to the private key file for server-side TLS.
+    pub key_file: String,
+    // Absolute path to the certificate authority file for verifying client certificates.
+    pub ca_file: String,
+}
+
+/// Databricks specific options for compute instance.
+/// This is used to store any other settings that needs to be propagate to Compute
+/// but should not be persisted to ComputeSpec in the database.
+#[derive(Clone, Debug, Deserialize, Serialize, PartialEq)]
+pub struct DatabricksSettings {
+    pub pg_compute_tls_settings: PgComputeTlsSettings,
+    // Absolute file path to databricks_pg_hba.conf file.
+    pub databricks_pg_hba: String,
+    // Absolute file path to databricks_pg_ident.conf file.
+    pub databricks_pg_ident: String,
+    // Hostname portion of the Databricks workspace URL of the endpoint, or empty string if not known.
+    // A valid hostname is required for the compute instance to support PAT logins.
+    pub databricks_workspace_host: String,
+}
+
 /// Optional collection of `GenericOption`'s. Type alias allows us to
 /// declare a `trait` on it.
 pub type GenericOptions = Option<Vec<GenericOption>>;

libs/consumption_metrics/src/lib.rs

@@ -90,7 +90,7 @@ impl<'a> IdempotencyKey<'a> {
         IdempotencyKey {
             now: Utc::now(),
             node_id,
-            nonce: rand::thread_rng().gen_range(0..=9999),
+            nonce: rand::rng().random_range(0..=9999),
         }
     }
 

libs/desim/src/node_os.rs

@@ -41,7 +41,7 @@ impl NodeOs {
 
     /// Generate a random number in range [0, max).
     pub fn random(&self, max: u64) -> u64 {
-        self.internal.rng.lock().gen_range(0..max)
+        self.internal.rng.lock().random_range(0..max)
     }
 
     /// Append a new event to the world event log.

libs/desim/src/options.rs

@@ -32,10 +32,10 @@ impl Delay {
     /// Generate a random delay in range [min, max]. Return None if the
     /// message should be dropped.
     pub fn delay(&self, rng: &mut StdRng) -> Option<u64> {
-        if rng.gen_bool(self.fail_prob) {
+        if rng.random_bool(self.fail_prob) {
             return None;
         }
-        Some(rng.gen_range(self.min..=self.max))
+        Some(rng.random_range(self.min..=self.max))
     }
 }
 

libs/desim/src/world.rs

@@ -69,7 +69,7 @@ impl World {
     /// Create a new random number generator.
     pub fn new_rng(&self) -> StdRng {
         let mut rng = self.rng.lock();
-        StdRng::from_rng(rng.deref_mut()).unwrap()
+        StdRng::from_rng(rng.deref_mut())
     }
 
     /// Create a new node.

libs/metrics/Cargo.toml

@@ -17,5 +17,5 @@ procfs.workspace = true
 measured-process.workspace = true
 
 [dev-dependencies]
-rand = "0.8"
-rand_distr = "0.4.3"
+rand.workspace = true
+rand_distr = "0.5"

libs/metrics/src/hll.rs

@@ -260,7 +260,7 @@ mod tests {
 
     #[test]
     fn test_cardinality_small() {
-        let (actual, estimate) = test_cardinality(100, Zipf::new(100, 1.2f64).unwrap());
+        let (actual, estimate) = test_cardinality(100, Zipf::new(100.0, 1.2f64).unwrap());
 
         assert_eq!(actual, [46, 30, 32]);
         assert!(51.3 < estimate[0] && estimate[0] < 51.4);
@@ -270,7 +270,7 @@ mod tests {
 
     #[test]
     fn test_cardinality_medium() {
-        let (actual, estimate) = test_cardinality(10000, Zipf::new(10000, 1.2f64).unwrap());
+        let (actual, estimate) = test_cardinality(10000, Zipf::new(10000.0, 1.2f64).unwrap());
 
         assert_eq!(actual, [2529, 1618, 1629]);
         assert!(2309.1 < estimate[0] && estimate[0] < 2309.2);
@@ -280,7 +280,8 @@ mod tests {
 
     #[test]
     fn test_cardinality_large() {
-        let (actual, estimate) = test_cardinality(1_000_000, Zipf::new(1_000_000, 1.2f64).unwrap());
+        let (actual, estimate) =
+            test_cardinality(1_000_000, Zipf::new(1_000_000.0, 1.2f64).unwrap());
 
         assert_eq!(actual, [129077, 79579, 79630]);
         assert!(126067.2 < estimate[0] && estimate[0] < 126067.3);
@@ -290,7 +291,7 @@ mod tests {
 
     #[test]
     fn test_cardinality_small2() {
-        let (actual, estimate) = test_cardinality(100, Zipf::new(200, 0.8f64).unwrap());
+        let (actual, estimate) = test_cardinality(100, Zipf::new(200.0, 0.8f64).unwrap());
 
         assert_eq!(actual, [92, 58, 60]);
         assert!(116.1 < estimate[0] && estimate[0] < 116.2);
@@ -300,7 +301,7 @@ mod tests {
 
     #[test]
     fn test_cardinality_medium2() {
-        let (actual, estimate) = test_cardinality(10000, Zipf::new(20000, 0.8f64).unwrap());
+        let (actual, estimate) = test_cardinality(10000, Zipf::new(20000.0, 0.8f64).unwrap());
 
         assert_eq!(actual, [8201, 5131, 5051]);
         assert!(6846.4 < estimate[0] && estimate[0] < 6846.5);
@@ -310,7 +311,8 @@ mod tests {
 
     #[test]
     fn test_cardinality_large2() {
-        let (actual, estimate) = test_cardinality(1_000_000, Zipf::new(2_000_000, 0.8f64).unwrap());
+        let (actual, estimate) =
+            test_cardinality(1_000_000, Zipf::new(2_000_000.0, 0.8f64).unwrap());
 
         assert_eq!(actual, [777847, 482069, 482246]);
         assert!(699437.4 < estimate[0] && estimate[0] < 699437.5);

libs/metrics/src/lib.rs

@@ -4,12 +4,14 @@
 //! a default registry.
 #![deny(clippy::undocumented_unsafe_blocks)]
 
+use std::sync::RwLock;
+
 use measured::label::{LabelGroupSet, LabelGroupVisitor, LabelName, NoLabels};
 use measured::metric::counter::CounterState;
 use measured::metric::gauge::GaugeState;
 use measured::metric::group::Encoding;
 use measured::metric::name::{MetricName, MetricNameEncoder};
-use measured::metric::{MetricEncoding, MetricFamilyEncoding};
+use measured::metric::{MetricEncoding, MetricFamilyEncoding, MetricType};
 use measured::{FixedCardinalityLabel, LabelGroup, MetricGroup};
 use once_cell::sync::Lazy;
 use prometheus::Registry;
@@ -116,12 +118,52 @@ pub fn pow2_buckets(start: usize, end: usize) -> Vec<f64> {
         .collect()
 }
 
+pub struct InfoMetric<L: LabelGroup, M: MetricType = GaugeState> {
+    label: RwLock<L>,
+    metric: M,
+}
+
+impl<L: LabelGroup> InfoMetric<L> {
+    pub fn new(label: L) -> Self {
+        Self::with_metric(label, GaugeState::new(1))
+    }
+}
+
+impl<L: LabelGroup, M: MetricType<Metadata = ()>> InfoMetric<L, M> {
+    pub fn with_metric(label: L, metric: M) -> Self {
+        Self {
+            label: RwLock::new(label),
+            metric,
+        }
+    }
+
+    pub fn set_label(&self, label: L) {
+        *self.label.write().unwrap() = label;
+    }
+}
+
+impl<L, M, E> MetricFamilyEncoding<E> for InfoMetric<L, M>
+where
+    L: LabelGroup,
+    M: MetricEncoding<E, Metadata = ()>,
+    E: Encoding,
+{
+    fn collect_family_into(
+        &self,
+        name: impl measured::metric::name::MetricNameEncoder,
+        enc: &mut E,
+    ) -> Result<(), E::Err> {
+        M::write_type(&name, enc)?;
+        self.metric
+            .collect_into(&(), &*self.label.read().unwrap(), name, enc)
+    }
+}
+
 pub struct BuildInfo {
     pub revision: &'static str,
     pub build_tag: &'static str,
 }
 
-// todo: allow label group without the set
 impl LabelGroup for BuildInfo {
     fn visit_values(&self, v: &mut impl LabelGroupVisitor) {
         const REVISION: &LabelName = LabelName::from_str("revision");
@@ -131,24 +173,6 @@ impl LabelGroup for BuildInfo {
     }
 }
 
-impl<T: Encoding> MetricFamilyEncoding<T> for BuildInfo
-where
-    GaugeState: MetricEncoding<T>,
-{
-    fn collect_family_into(
-        &self,
-        name: impl measured::metric::name::MetricNameEncoder,
-        enc: &mut T,
-    ) -> Result<(), T::Err> {
-        enc.write_help(&name, "Build/version information")?;
-        GaugeState::write_type(&name, enc)?;
-        GaugeState {
-            count: std::sync::atomic::AtomicI64::new(1),
-        }
-        .collect_into(&(), self, name, enc)
-    }
-}
-
 #[derive(MetricGroup)]
 #[metric(new(build_info: BuildInfo))]
 pub struct NeonMetrics {
@@ -165,8 +189,8 @@ pub struct NeonMetrics {
 #[derive(MetricGroup)]
 #[metric(new(build_info: BuildInfo))]
 pub struct LibMetrics {
-    #[metric(init = build_info)]
-    build_info: BuildInfo,
+    #[metric(init = InfoMetric::new(build_info))]
+    build_info: InfoMetric<BuildInfo>,
 
     #[metric(flatten)]
     rusage: Rusage,

libs/neon-shmem/Cargo.toml

@@ -8,6 +8,13 @@ license.workspace = true
 thiserror.workspace = true
 nix.workspace=true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+libc.workspace = true
+lock_api.workspace = true
+rustc-hash.workspace = true
 
 [target.'cfg(target_os = "macos")'.dependencies]
 tempfile = "3.14.0"
+
+[dev-dependencies]
+rand.workspace = true
+rand_distr = "0.5.1"

libs/neon-shmem/src/hash.rs

@@ -0,0 +1,583 @@
+//! Resizable hash table implementation on top of byte-level storage (either a [`ShmemHandle`] or a fixed byte array).
+//!
+//! This hash table has two major components: the bucket array and the dictionary. Each bucket within the
+//! bucket array contains a `Option<(K, V)>` and an index of another bucket. In this way there is both an
+//! implicit freelist within the bucket array (`None` buckets point to other `None` entries) and various hash
+//! chains within the bucket array (a Some bucket will point to other Some buckets that had the same hash).
+//!
+//! Buckets are never moved unless they are within a region that is being shrunk, and so the actual hash-
+//! dependent component is done with the dictionary. When a new key is inserted into the map, a position
+//! within the dictionary is decided based on its hash, the data is inserted into an empty bucket based
+//! off of the freelist, and then the index of said bucket is placed in the dictionary.
+//!
+//! This map is resizable (if initialized on top of a [`ShmemHandle`]). Both growing and shrinking happen
+//! in-place and are at a high level achieved by expanding/reducing the bucket array and rebuilding the
+//! dictionary by rehashing all keys.
+//!
+//! Concurrency is managed very simply: the entire map is guarded by one shared-memory RwLock.
+
+use std::hash::{BuildHasher, Hash};
+use std::mem::MaybeUninit;
+
+use crate::shmem::ShmemHandle;
+use crate::{shmem, sync::*};
+
+mod core;
+pub mod entry;
+
+#[cfg(test)]
+mod tests;
+
+use core::{Bucket, CoreHashMap, INVALID_POS};
+use entry::{Entry, OccupiedEntry, PrevPos, VacantEntry};
+
+use thiserror::Error;
+
+/// Error type for a hashmap shrink operation.
+#[derive(Error, Debug)]
+pub enum HashMapShrinkError {
+    /// There was an error encountered while resizing the memory area.
+    #[error("shmem resize failed: {0}")]
+    ResizeError(shmem::Error),
+    /// Occupied entries in to-be-shrunk space were encountered beginning at the given index.
+    #[error("occupied entry in deallocated space found at {0}")]
+    RemainingEntries(usize),
+}
+
+/// This represents a hash table that (possibly) lives in shared memory.
+/// If a new process is launched with fork(), the child process inherits
+/// this struct.
+#[must_use]
+pub struct HashMapInit<'a, K, V, S = rustc_hash::FxBuildHasher> {
+    shmem_handle: Option<ShmemHandle>,
+    shared_ptr: *mut HashMapShared<'a, K, V>,
+    shared_size: usize,
+    hasher: S,
+    num_buckets: u32,
+}
+
+/// This is a per-process handle to a hash table that (possibly) lives in shared memory.
+/// If a child process is launched with fork(), the child process should
+/// get its own HashMapAccess by calling HashMapInit::attach_writer/reader().
+///
+/// XXX: We're not making use of it at the moment, but this struct could
+/// hold process-local information in the future.
+pub struct HashMapAccess<'a, K, V, S = rustc_hash::FxBuildHasher> {
+    shmem_handle: Option<ShmemHandle>,
+    shared_ptr: *mut HashMapShared<'a, K, V>,
+    hasher: S,
+}
+
+unsafe impl<K: Sync, V: Sync, S> Sync for HashMapAccess<'_, K, V, S> {}
+unsafe impl<K: Send, V: Send, S> Send for HashMapAccess<'_, K, V, S> {}
+
+impl<'a, K: Clone + Hash + Eq, V, S> HashMapInit<'a, K, V, S> {
+    /// Change the 'hasher' used by the hash table.
+    ///
+    /// NOTE: This must be called right after creating the hash table,
+    /// before inserting any entries and before calling attach_writer/reader.
+    /// Otherwise different accessors could be using different hash function,
+    /// with confusing results.
+    pub fn with_hasher<T: BuildHasher>(self, hasher: T) -> HashMapInit<'a, K, V, T> {
+        HashMapInit {
+            hasher,
+            shmem_handle: self.shmem_handle,
+            shared_ptr: self.shared_ptr,
+            shared_size: self.shared_size,
+            num_buckets: self.num_buckets,
+        }
+    }
+
+    /// Loosely (over)estimate the size needed to store a hash table with `num_buckets` buckets.
+    pub fn estimate_size(num_buckets: u32) -> usize {
+        // add some margin to cover alignment etc.
+        CoreHashMap::<K, V>::estimate_size(num_buckets) + size_of::<HashMapShared<K, V>>() + 1000
+    }
+
+    fn new(
+        num_buckets: u32,
+        shmem_handle: Option<ShmemHandle>,
+        area_ptr: *mut u8,
+        area_size: usize,
+        hasher: S,
+    ) -> Self {
+        let mut ptr: *mut u8 = area_ptr;
+        let end_ptr: *mut u8 = unsafe { ptr.add(area_size) };
+
+        // carve out area for the One Big Lock (TM) and the HashMapShared.
+        ptr = unsafe { ptr.add(ptr.align_offset(align_of::<libc::pthread_rwlock_t>())) };
+        let raw_lock_ptr = ptr;
+        ptr = unsafe { ptr.add(size_of::<libc::pthread_rwlock_t>()) };
+        ptr = unsafe { ptr.add(ptr.align_offset(align_of::<HashMapShared<K, V>>())) };
+        let shared_ptr: *mut HashMapShared<K, V> = ptr.cast();
+        ptr = unsafe { ptr.add(size_of::<HashMapShared<K, V>>()) };
+
+        // carve out the buckets
+        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<core::Bucket<K, V>>())) };
+        let buckets_ptr = ptr;
+        ptr = unsafe { ptr.add(size_of::<core::Bucket<K, V>>() * num_buckets as usize) };
+
+        // use remaining space for the dictionary
+        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<u32>())) };
+        assert!(ptr.addr() < end_ptr.addr());
+        let dictionary_ptr = ptr;
+        let dictionary_size = unsafe { end_ptr.byte_offset_from(ptr) / size_of::<u32>() as isize };
+        assert!(dictionary_size > 0);
+
+        let buckets =
+            unsafe { std::slice::from_raw_parts_mut(buckets_ptr.cast(), num_buckets as usize) };
+        let dictionary = unsafe {
+            std::slice::from_raw_parts_mut(dictionary_ptr.cast(), dictionary_size as usize)
+        };
+
+        let hashmap = CoreHashMap::new(buckets, dictionary);
+        unsafe {
+            let lock = RwLock::from_raw(PthreadRwLock::new(raw_lock_ptr.cast()), hashmap);
+            std::ptr::write(shared_ptr, lock);
+        }
+
+        Self {
+            num_buckets,
+            shmem_handle,
+            shared_ptr,
+            shared_size: area_size,
+            hasher,
+        }
+    }
+
+    /// Attach to a hash table for writing.
+    pub fn attach_writer(self) -> HashMapAccess<'a, K, V, S> {
+        HashMapAccess {
+            shmem_handle: self.shmem_handle,
+            shared_ptr: self.shared_ptr,
+            hasher: self.hasher,
+        }
+    }
+
+    /// Initialize a table for reading. Currently identical to [`HashMapInit::attach_writer`].
+    ///
+    /// This is a holdover from a previous implementation and is being kept around for
+    /// backwards compatibility reasons.
+    pub fn attach_reader(self) -> HashMapAccess<'a, K, V, S> {
+        self.attach_writer()
+    }
+}
+
+/// Hash table data that is actually stored in the shared memory area.
+///
+/// NOTE: We carve out the parts from a contiguous chunk. Growing and shrinking the hash table
+/// relies on the memory layout! The data structures are laid out in the contiguous shared memory
+/// area as follows:
+///
+/// [`libc::pthread_rwlock_t`]
+/// [`HashMapShared`]
+/// buckets
+/// dictionary
+///
+/// In between the above parts, there can be padding bytes to align the parts correctly.
+type HashMapShared<'a, K, V> = RwLock<CoreHashMap<'a, K, V>>;
+
+impl<'a, K, V> HashMapInit<'a, K, V, rustc_hash::FxBuildHasher>
+where
+    K: Clone + Hash + Eq,
+{
+    /// Place the hash table within a user-supplied fixed memory area.
+    pub fn with_fixed(num_buckets: u32, area: &'a mut [MaybeUninit<u8>]) -> Self {
+        Self::new(
+            num_buckets,
+            None,
+            area.as_mut_ptr().cast(),
+            area.len(),
+            rustc_hash::FxBuildHasher,
+        )
+    }
+
+    /// Place a new hash map in the given shared memory area
+    ///
+    /// # Panics
+    /// Will panic on failure to resize area to expected map size.
+    pub fn with_shmem(num_buckets: u32, shmem: ShmemHandle) -> Self {
+        let size = Self::estimate_size(num_buckets);
+        shmem
+            .set_size(size)
+            .expect("could not resize shared memory area");
+        let ptr = shmem.data_ptr.as_ptr().cast();
+        Self::new(
+            num_buckets,
+            Some(shmem),
+            ptr,
+            size,
+            rustc_hash::FxBuildHasher,
+        )
+    }
+
+    /// Make a resizable hash map within a new shared memory area with the given name.
+    pub fn new_resizeable_named(num_buckets: u32, max_buckets: u32, name: &str) -> Self {
+        let size = Self::estimate_size(num_buckets);
+        let max_size = Self::estimate_size(max_buckets);
+        let shmem =
+            ShmemHandle::new(name, size, max_size).expect("failed to make shared memory area");
+        let ptr = shmem.data_ptr.as_ptr().cast();
+
+        Self::new(
+            num_buckets,
+            Some(shmem),
+            ptr,
+            size,
+            rustc_hash::FxBuildHasher,
+        )
+    }
+
+    /// Make a resizable hash map within a new anonymous shared memory area.
+    pub fn new_resizeable(num_buckets: u32, max_buckets: u32) -> Self {
+        use std::sync::atomic::{AtomicUsize, Ordering};
+        static COUNTER: AtomicUsize = AtomicUsize::new(0);
+        let val = COUNTER.fetch_add(1, Ordering::Relaxed);
+        let name = format!("neon_shmem_hmap{val}");
+        Self::new_resizeable_named(num_buckets, max_buckets, &name)
+    }
+}
+
+impl<'a, K, V, S: BuildHasher> HashMapAccess<'a, K, V, S>
+where
+    K: Clone + Hash + Eq,
+{
+    /// Hash a key using the map's hasher.
+    #[inline]
+    fn get_hash_value(&self, key: &K) -> u64 {
+        self.hasher.hash_one(key)
+    }
+
+    fn entry_with_hash(&self, key: K, hash: u64) -> Entry<'a, '_, K, V> {
+        let mut map = unsafe { self.shared_ptr.as_ref() }.unwrap().write();
+        let dict_pos = hash as usize % map.dictionary.len();
+        let first = map.dictionary[dict_pos];
+        if first == INVALID_POS {
+            // no existing entry
+            return Entry::Vacant(VacantEntry {
+                map,
+                key,
+                dict_pos: dict_pos as u32,
+            });
+        }
+
+        let mut prev_pos = PrevPos::First(dict_pos as u32);
+        let mut next = first;
+        loop {
+            let bucket = &mut map.buckets[next as usize];
+            let (bucket_key, _bucket_value) = bucket.inner.as_mut().expect("entry is in use");
+            if *bucket_key == key {
+                // found existing entry
+                return Entry::Occupied(OccupiedEntry {
+                    map,
+                    _key: key,
+                    prev_pos,
+                    bucket_pos: next,
+                });
+            }
+
+            if bucket.next == INVALID_POS {
+                // No existing entry
+                return Entry::Vacant(VacantEntry {
+                    map,
+                    key,
+                    dict_pos: dict_pos as u32,
+                });
+            }
+            prev_pos = PrevPos::Chained(next);
+            next = bucket.next;
+        }
+    }
+
+    /// Get a reference to the corresponding value for a key.
+    pub fn get<'e>(&'e self, key: &K) -> Option<ValueReadGuard<'e, V>> {
+        let hash = self.get_hash_value(key);
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+        RwLockReadGuard::try_map(map, |m| m.get_with_hash(key, hash)).ok()
+    }
+
+    /// Get a reference to the entry containing a key.
+    ///
+    /// NB: THis takes a write lock as there's no way to distinguish whether the intention
+    /// is to use the entry for reading or for writing in advance.
+    pub fn entry(&self, key: K) -> Entry<'a, '_, K, V> {
+        let hash = self.get_hash_value(&key);
+        self.entry_with_hash(key, hash)
+    }
+
+    /// Remove a key given its hash. Returns the associated value if it existed.
+    pub fn remove(&self, key: &K) -> Option<V> {
+        let hash = self.get_hash_value(key);
+        match self.entry_with_hash(key.clone(), hash) {
+            Entry::Occupied(e) => Some(e.remove()),
+            Entry::Vacant(_) => None,
+        }
+    }
+
+    /// Insert/update a key. Returns the previous associated value if it existed.
+    ///
+    /// # Errors
+    /// Will return [`core::FullError`] if there is no more space left in the map.
+    pub fn insert(&self, key: K, value: V) -> Result<Option<V>, core::FullError> {
+        let hash = self.get_hash_value(&key);
+        match self.entry_with_hash(key.clone(), hash) {
+            Entry::Occupied(mut e) => Ok(Some(e.insert(value))),
+            Entry::Vacant(e) => {
+                _ = e.insert(value)?;
+                Ok(None)
+            }
+        }
+    }
+
+    /// Optionally return the entry for a bucket at a given index if it exists.
+    ///
+    /// Has more overhead than one would intuitively expect: performs both a clone of the key
+    /// due to the [`OccupiedEntry`] type owning the key and also a hash of the key in order
+    /// to enable repairing the hash chain if the entry is removed.
+    pub fn entry_at_bucket(&self, pos: usize) -> Option<OccupiedEntry<'a, '_, K, V>> {
+        let map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        if pos >= map.buckets.len() {
+            return None;
+        }
+
+        let entry = map.buckets[pos].inner.as_ref();
+        match entry {
+            Some((key, _)) => Some(OccupiedEntry {
+                _key: key.clone(),
+                bucket_pos: pos as u32,
+                prev_pos: entry::PrevPos::Unknown(self.get_hash_value(key)),
+                map,
+            }),
+            _ => None,
+        }
+    }
+
+    /// Returns the number of buckets in the table.
+    pub fn get_num_buckets(&self) -> usize {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+        map.get_num_buckets()
+    }
+
+    /// Return the key and value stored in bucket with given index. This can be used to
+    /// iterate through the hash map.
+    // TODO: An Iterator might be nicer. The communicator's clock algorithm needs to
+    // _slowly_ iterate through all buckets with its clock hand,  without holding a lock.
+    // If we switch to an Iterator, it must not hold the lock.
+    pub fn get_at_bucket(&self, pos: usize) -> Option<ValueReadGuard<(K, V)>> {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+        if pos >= map.buckets.len() {
+            return None;
+        }
+        RwLockReadGuard::try_map(map, |m| m.buckets[pos].inner.as_ref()).ok()
+    }
+
+    /// Returns the index of the bucket a given value corresponds to.
+    pub fn get_bucket_for_value(&self, val_ptr: *const V) -> usize {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+
+        let origin = map.buckets.as_ptr();
+        let idx = (val_ptr as usize - origin as usize) / size_of::<Bucket<K, V>>();
+        assert!(idx < map.buckets.len());
+
+        idx
+    }
+
+    /// Returns the number of occupied buckets in the table.
+    pub fn get_num_buckets_in_use(&self) -> usize {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+        map.buckets_in_use as usize
+    }
+
+    /// Clears all entries in a table. Does not reset any shrinking operations.
+    pub fn clear(&self) {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        map.clear();
+    }
+
+    /// Perform an in-place rehash of some region (0..`rehash_buckets`) of the table and reset
+    /// the `buckets` and `dictionary` slices to be as long as `num_buckets`. Resets the freelist
+    /// in the process.
+    fn rehash_dict(
+        &self,
+        inner: &mut CoreHashMap<'a, K, V>,
+        buckets_ptr: *mut core::Bucket<K, V>,
+        end_ptr: *mut u8,
+        num_buckets: u32,
+        rehash_buckets: u32,
+    ) {
+        inner.free_head = INVALID_POS;
+
+        let buckets;
+        let dictionary;
+        unsafe {
+            let buckets_end_ptr = buckets_ptr.add(num_buckets as usize);
+            let dictionary_ptr: *mut u32 = buckets_end_ptr
+                .byte_add(buckets_end_ptr.align_offset(align_of::<u32>()))
+                .cast();
+            let dictionary_size: usize =
+                end_ptr.byte_offset_from(buckets_end_ptr) as usize / size_of::<u32>();
+
+            buckets = std::slice::from_raw_parts_mut(buckets_ptr, num_buckets as usize);
+            dictionary = std::slice::from_raw_parts_mut(dictionary_ptr, dictionary_size);
+        }
+        for e in dictionary.iter_mut() {
+            *e = INVALID_POS;
+        }
+
+        for (i, bucket) in buckets.iter_mut().enumerate().take(rehash_buckets as usize) {
+            if bucket.inner.is_none() {
+                bucket.next = inner.free_head;
+                inner.free_head = i as u32;
+                continue;
+            }
+
+            let hash = self.hasher.hash_one(&bucket.inner.as_ref().unwrap().0);
+            let pos: usize = (hash % dictionary.len() as u64) as usize;
+            bucket.next = dictionary[pos];
+            dictionary[pos] = i as u32;
+        }
+
+        inner.dictionary = dictionary;
+        inner.buckets = buckets;
+    }
+
+    /// Rehash the map without growing or shrinking.
+    pub fn shuffle(&self) {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        let num_buckets = map.get_num_buckets() as u32;
+        let size_bytes = HashMapInit::<K, V, S>::estimate_size(num_buckets);
+        let end_ptr: *mut u8 = unsafe { self.shared_ptr.byte_add(size_bytes).cast() };
+        let buckets_ptr = map.buckets.as_mut_ptr();
+        self.rehash_dict(&mut map, buckets_ptr, end_ptr, num_buckets, num_buckets);
+    }
+
+    /// Grow the number of buckets within the table.
+    ///
+    /// 1. Grows the underlying shared memory area
+    /// 2. Initializes new buckets and overwrites the current dictionary
+    /// 3. Rehashes the dictionary
+    ///
+    /// # Panics
+    /// Panics if called on a map initialized with [`HashMapInit::with_fixed`].
+    ///
+    /// # Errors
+    /// Returns an [`shmem::Error`] if any errors occur resizing the memory region.
+    pub fn grow(&self, num_buckets: u32) -> Result<(), shmem::Error> {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        let old_num_buckets = map.buckets.len() as u32;
+
+        assert!(
+            num_buckets >= old_num_buckets,
+            "grow called with a smaller number of buckets"
+        );
+        if num_buckets == old_num_buckets {
+            return Ok(());
+        }
+        let shmem_handle = self
+            .shmem_handle
+            .as_ref()
+            .expect("grow called on a fixed-size hash table");
+
+        let size_bytes = HashMapInit::<K, V, S>::estimate_size(num_buckets);
+        shmem_handle.set_size(size_bytes)?;
+        let end_ptr: *mut u8 = unsafe { shmem_handle.data_ptr.as_ptr().add(size_bytes) };
+
+        // Initialize new buckets. The new buckets are linked to the free list.
+        // NB: This overwrites the dictionary!
+        let buckets_ptr = map.buckets.as_mut_ptr();
+        unsafe {
+            for i in old_num_buckets..num_buckets {
+                let bucket = buckets_ptr.add(i as usize);
+                bucket.write(core::Bucket {
+                    next: if i < num_buckets - 1 {
+                        i + 1
+                    } else {
+                        map.free_head
+                    },
+                    inner: None,
+                });
+            }
+        }
+
+        self.rehash_dict(&mut map, buckets_ptr, end_ptr, num_buckets, old_num_buckets);
+        map.free_head = old_num_buckets;
+
+        Ok(())
+    }
+
+    /// Begin a shrink, limiting all new allocations to be in buckets with index below `num_buckets`.
+    ///
+    /// # Panics
+    /// Panics if called on a map initialized with [`HashMapInit::with_fixed`] or if `num_buckets` is
+    /// greater than the number of buckets in the map.
+    pub fn begin_shrink(&mut self, num_buckets: u32) {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        assert!(
+            num_buckets <= map.get_num_buckets() as u32,
+            "shrink called with a larger number of buckets"
+        );
+        _ = self
+            .shmem_handle
+            .as_ref()
+            .expect("shrink called on a fixed-size hash table");
+        map.alloc_limit = num_buckets;
+    }
+
+    /// If a shrink operation is underway, returns the target size of the map. Otherwise, returns None.
+    pub fn shrink_goal(&self) -> Option<usize> {
+        let map = unsafe { self.shared_ptr.as_mut() }.unwrap().read();
+        let goal = map.alloc_limit;
+        if goal == INVALID_POS {
+            None
+        } else {
+            Some(goal as usize)
+        }
+    }
+
+    /// Complete a shrink after caller has evicted entries, removing the unused buckets and rehashing.
+    ///
+    /// # Panics
+    /// The following cases result in a panic:
+    /// - Calling this function on a map initialized with [`HashMapInit::with_fixed`].
+    /// - Calling this function on a map when no shrink operation is in progress.
+    pub fn finish_shrink(&self) -> Result<(), HashMapShrinkError> {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        assert!(
+            map.alloc_limit != INVALID_POS,
+            "called finish_shrink when no shrink is in progress"
+        );
+
+        let num_buckets = map.alloc_limit;
+
+        if map.get_num_buckets() == num_buckets as usize {
+            return Ok(());
+        }
+
+        assert!(
+            map.buckets_in_use <= num_buckets,
+            "called finish_shrink before enough entries were removed"
+        );
+
+        for i in (num_buckets as usize)..map.buckets.len() {
+            if map.buckets[i].inner.is_some() {
+                return Err(HashMapShrinkError::RemainingEntries(i));
+            }
+        }
+
+        let shmem_handle = self
+            .shmem_handle
+            .as_ref()
+            .expect("shrink called on a fixed-size hash table");
+
+        let size_bytes = HashMapInit::<K, V, S>::estimate_size(num_buckets);
+        if let Err(e) = shmem_handle.set_size(size_bytes) {
+            return Err(HashMapShrinkError::ResizeError(e));
+        }
+        let end_ptr: *mut u8 = unsafe { shmem_handle.data_ptr.as_ptr().add(size_bytes) };
+        let buckets_ptr = map.buckets.as_mut_ptr();
+        self.rehash_dict(&mut map, buckets_ptr, end_ptr, num_buckets, num_buckets);
+        map.alloc_limit = INVALID_POS;
+
+        Ok(())
+    }
+}

libs/neon-shmem/src/hash/core.rs

@@ -0,0 +1,174 @@
+//! Simple hash table with chaining.
+
+use std::hash::Hash;
+use std::mem::MaybeUninit;
+
+use crate::hash::entry::*;
+
+/// Invalid position within the map (either within the dictionary or bucket array).
+pub(crate) const INVALID_POS: u32 = u32::MAX;
+
+/// Fundamental storage unit within the hash table. Either empty or contains a key-value pair.
+/// Always part of a chain of some kind (either a freelist if empty or a hash chain if full).
+pub(crate) struct Bucket<K, V> {
+    /// Index of next bucket in the chain.
+    pub(crate) next: u32,
+    /// Key-value pair contained within bucket.
+    pub(crate) inner: Option<(K, V)>,
+}
+
+/// Core hash table implementation.
+pub(crate) struct CoreHashMap<'a, K, V> {
+    /// Dictionary used to map hashes to bucket indices.
+    pub(crate) dictionary: &'a mut [u32],
+    /// Buckets containing key-value pairs.
+    pub(crate) buckets: &'a mut [Bucket<K, V>],
+    /// Head of the freelist.
+    pub(crate) free_head: u32,
+    /// Maximum index of a bucket allowed to be allocated. [`INVALID_POS`] if no limit.
+    pub(crate) alloc_limit: u32,
+    /// The number of currently occupied buckets.
+    pub(crate) buckets_in_use: u32,
+}
+
+/// Error for when there are no empty buckets left but one is needed.
+#[derive(Debug, PartialEq)]
+pub struct FullError;
+
+impl<'a, K: Clone + Hash + Eq, V> CoreHashMap<'a, K, V> {
+    const FILL_FACTOR: f32 = 0.60;
+
+    /// Estimate the size of data contained within the the hash map.
+    pub fn estimate_size(num_buckets: u32) -> usize {
+        let mut size = 0;
+
+        // buckets
+        size += size_of::<Bucket<K, V>>() * num_buckets as usize;
+
+        // dictionary
+        size += (f32::ceil((size_of::<u32>() * num_buckets as usize) as f32 / Self::FILL_FACTOR))
+            as usize;
+
+        size
+    }
+
+    pub fn new(
+        buckets: &'a mut [MaybeUninit<Bucket<K, V>>],
+        dictionary: &'a mut [MaybeUninit<u32>],
+    ) -> Self {
+        // Initialize the buckets
+        for i in 0..buckets.len() {
+            buckets[i].write(Bucket {
+                next: if i < buckets.len() - 1 {
+                    i as u32 + 1
+                } else {
+                    INVALID_POS
+                },
+                inner: None,
+            });
+        }
+
+        // Initialize the dictionary
+        for e in dictionary.iter_mut() {
+            e.write(INVALID_POS);
+        }
+
+        // TODO: use std::slice::assume_init_mut() once it stabilizes
+        let buckets =
+            unsafe { std::slice::from_raw_parts_mut(buckets.as_mut_ptr().cast(), buckets.len()) };
+        let dictionary = unsafe {
+            std::slice::from_raw_parts_mut(dictionary.as_mut_ptr().cast(), dictionary.len())
+        };
+
+        Self {
+            dictionary,
+            buckets,
+            free_head: 0,
+            buckets_in_use: 0,
+            alloc_limit: INVALID_POS,
+        }
+    }
+
+    /// Get the value associated with a key (if it exists) given its hash.
+    pub fn get_with_hash(&self, key: &K, hash: u64) -> Option<&V> {
+        let mut next = self.dictionary[hash as usize % self.dictionary.len()];
+        loop {
+            if next == INVALID_POS {
+                return None;
+            }
+
+            let bucket = &self.buckets[next as usize];
+            let (bucket_key, bucket_value) = bucket.inner.as_ref().expect("entry is in use");
+            if bucket_key == key {
+                return Some(bucket_value);
+            }
+            next = bucket.next;
+        }
+    }
+
+    /// Get number of buckets in map.
+    pub fn get_num_buckets(&self) -> usize {
+        self.buckets.len()
+    }
+
+    /// Clears all entries from the hashmap.
+    ///
+    /// Does not reset any allocation limits, but does clear any entries beyond them.
+    pub fn clear(&mut self) {
+        for i in 0..self.buckets.len() {
+            self.buckets[i] = Bucket {
+                next: if i < self.buckets.len() - 1 {
+                    i as u32 + 1
+                } else {
+                    INVALID_POS
+                },
+                inner: None,
+            }
+        }
+        for i in 0..self.dictionary.len() {
+            self.dictionary[i] = INVALID_POS;
+        }
+
+        self.free_head = 0;
+        self.buckets_in_use = 0;
+    }
+
+    /// Find the position of an unused bucket via the freelist and initialize it.
+    pub(crate) fn alloc_bucket(&mut self, key: K, value: V) -> Result<u32, FullError> {
+        let mut pos = self.free_head;
+
+        // Find the first bucket we're *allowed* to use.
+        let mut prev = PrevPos::First(self.free_head);
+        while pos != INVALID_POS && pos >= self.alloc_limit {
+            let bucket = &mut self.buckets[pos as usize];
+            prev = PrevPos::Chained(pos);
+            pos = bucket.next;
+        }
+        if pos == INVALID_POS {
+            return Err(FullError);
+        }
+
+        // Repair the freelist.
+        match prev {
+            PrevPos::First(_) => {
+                let next_pos = self.buckets[pos as usize].next;
+                self.free_head = next_pos;
+            }
+            PrevPos::Chained(p) => {
+                if p != INVALID_POS {
+                    let next_pos = self.buckets[pos as usize].next;
+                    self.buckets[p as usize].next = next_pos;
+                }
+            }
+            _ => unreachable!(),
+        }
+
+        // Initialize the bucket.
+        let bucket = &mut self.buckets[pos as usize];
+        self.buckets_in_use += 1;
+        bucket.next = INVALID_POS;
+        bucket.inner = Some((key, value));
+
+        Ok(pos)
+    }
+}

libs/neon-shmem/src/hash/entry.rs

@@ -0,0 +1,130 @@
+//! Equivalent of [`std::collections::hash_map::Entry`] for this hashmap.
+
+use crate::hash::core::{CoreHashMap, FullError, INVALID_POS};
+use crate::sync::{RwLockWriteGuard, ValueWriteGuard};
+
+use std::hash::Hash;
+use std::mem;
+
+pub enum Entry<'a, 'b, K, V> {
+    Occupied(OccupiedEntry<'a, 'b, K, V>),
+    Vacant(VacantEntry<'a, 'b, K, V>),
+}
+
+/// Enum representing the previous position within a chain.
+#[derive(Clone, Copy)]
+pub(crate) enum PrevPos {
+    /// Starting index within the dictionary.  
+    First(u32),
+    /// Regular index within the buckets.
+    Chained(u32),
+    /// Unknown - e.g. the associated entry was retrieved by index instead of chain.
+    Unknown(u64),
+}
+
+pub struct OccupiedEntry<'a, 'b, K, V> {
+    /// Mutable reference to the map containing this entry.
+    pub(crate) map: RwLockWriteGuard<'b, CoreHashMap<'a, K, V>>,
+    /// The key of the occupied entry
+    pub(crate) _key: K,
+    /// The index of the previous entry in the chain.
+    pub(crate) prev_pos: PrevPos,
+    /// The position of the bucket in the [`CoreHashMap`] bucket array.
+    pub(crate) bucket_pos: u32,
+}
+
+impl<K, V> OccupiedEntry<'_, '_, K, V> {
+    pub fn get(&self) -> &V {
+        &self.map.buckets[self.bucket_pos as usize]
+            .inner
+            .as_ref()
+            .unwrap()
+            .1
+    }
+
+    pub fn get_mut(&mut self) -> &mut V {
+        &mut self.map.buckets[self.bucket_pos as usize]
+            .inner
+            .as_mut()
+            .unwrap()
+            .1
+    }
+
+    /// Inserts a value into the entry, replacing (and returning) the existing value.
+    pub fn insert(&mut self, value: V) -> V {
+        let bucket = &mut self.map.buckets[self.bucket_pos as usize];
+        // This assumes inner is Some, which it must be for an OccupiedEntry
+        mem::replace(&mut bucket.inner.as_mut().unwrap().1, value)
+    }
+
+    /// Removes the entry from the hash map, returning the value originally stored within it.
+    ///
+    /// This may result in multiple bucket accesses if the entry was obtained by index as the
+    /// previous chain entry needs to be discovered in this case.
+    pub fn remove(mut self) -> V {
+        // If this bucket was queried by index, go ahead and follow its chain from the start.
+        let prev = if let PrevPos::Unknown(hash) = self.prev_pos {
+            let dict_idx = hash as usize % self.map.dictionary.len();
+            let mut prev = PrevPos::First(dict_idx as u32);
+            let mut curr = self.map.dictionary[dict_idx];
+            while curr != self.bucket_pos {
+                assert!(curr != INVALID_POS);
+                prev = PrevPos::Chained(curr);
+                curr = self.map.buckets[curr as usize].next;
+            }
+            prev
+        } else {
+            self.prev_pos
+        };
+
+        // CoreHashMap::remove returns Option<(K, V)>. We know it's Some for an OccupiedEntry.
+        let bucket = &mut self.map.buckets[self.bucket_pos as usize];
+
+        // unlink it from the chain
+        match prev {
+            PrevPos::First(dict_pos) => {
+                self.map.dictionary[dict_pos as usize] = bucket.next;
+            }
+            PrevPos::Chained(bucket_pos) => {
+                self.map.buckets[bucket_pos as usize].next = bucket.next;
+            }
+            _ => unreachable!(),
+        }
+
+        // and add it to the freelist
+        let free = self.map.free_head;
+        let bucket = &mut self.map.buckets[self.bucket_pos as usize];
+        let old_value = bucket.inner.take();
+        bucket.next = free;
+        self.map.free_head = self.bucket_pos;
+        self.map.buckets_in_use -= 1;
+
+        old_value.unwrap().1
+    }
+}
+
+/// An abstract view into a vacant entry within the map.
+pub struct VacantEntry<'a, 'b, K, V> {
+    /// Mutable reference to the map containing this entry.
+    pub(crate) map: RwLockWriteGuard<'b, CoreHashMap<'a, K, V>>,
+    /// The key to be inserted into this entry.
+    pub(crate) key: K,
+    /// The position within the dictionary corresponding to the key's hash.
+    pub(crate) dict_pos: u32,
+}
+
+impl<'b, K: Clone + Hash + Eq, V> VacantEntry<'_, 'b, K, V> {
+    /// Insert a value into the vacant entry, finding and populating an empty bucket in the process.
+    ///
+    /// # Errors
+    /// Will return [`FullError`] if there are no unoccupied buckets in the map.
+    pub fn insert(mut self, value: V) -> Result<ValueWriteGuard<'b, V>, FullError> {
+        let pos = self.map.alloc_bucket(self.key, value)?;
+        self.map.buckets[pos as usize].next = self.map.dictionary[self.dict_pos as usize];
+        self.map.dictionary[self.dict_pos as usize] = pos;
+
+        Ok(RwLockWriteGuard::map(self.map, |m| {
+            &mut m.buckets[pos as usize].inner.as_mut().unwrap().1
+        }))
+    }
+}

libs/neon-shmem/src/hash/tests.rs

@@ -0,0 +1,428 @@
+use std::collections::BTreeMap;
+use std::collections::HashSet;
+use std::fmt::Debug;
+use std::mem::MaybeUninit;
+
+use crate::hash::Entry;
+use crate::hash::HashMapAccess;
+use crate::hash::HashMapInit;
+use crate::hash::core::FullError;
+
+use rand::seq::SliceRandom;
+use rand::{Rng, RngCore};
+use rand_distr::Zipf;
+
+const TEST_KEY_LEN: usize = 16;
+
+#[derive(Clone, Copy, Debug, Hash, PartialEq, Eq, PartialOrd, Ord)]
+struct TestKey([u8; TEST_KEY_LEN]);
+
+impl From<&TestKey> for u128 {
+    fn from(val: &TestKey) -> u128 {
+        u128::from_be_bytes(val.0)
+    }
+}
+
+impl From<u128> for TestKey {
+    fn from(val: u128) -> TestKey {
+        TestKey(val.to_be_bytes())
+    }
+}
+
+impl<'a> From<&'a [u8]> for TestKey {
+    fn from(bytes: &'a [u8]) -> TestKey {
+        TestKey(bytes.try_into().unwrap())
+    }
+}
+
+fn test_inserts<K: Into<TestKey> + Copy>(keys: &[K]) {
+    let w = HashMapInit::<TestKey, usize>::new_resizeable_named(100000, 120000, "test_inserts")
+        .attach_writer();
+
+    for (idx, k) in keys.iter().enumerate() {
+        let res = w.entry((*k).into());
+        match res {
+            Entry::Occupied(mut e) => {
+                e.insert(idx);
+            }
+            Entry::Vacant(e) => {
+                let res = e.insert(idx);
+                assert!(res.is_ok());
+            }
+        };
+    }
+
+    for (idx, k) in keys.iter().enumerate() {
+        let x = w.get(&(*k).into());
+        let value = x.as_deref().copied();
+        assert_eq!(value, Some(idx));
+    }
+}
+
+#[test]
+fn dense() {
+    // This exercises splitting a node with prefix
+    let keys: &[u128] = &[0, 1, 2, 3, 256];
+    test_inserts(keys);
+
+    // Dense keys
+    let mut keys: Vec<u128> = (0..10000).collect();
+    test_inserts(&keys);
+
+    // Do the same in random orders
+    for _ in 1..10 {
+        keys.shuffle(&mut rand::rng());
+        test_inserts(&keys);
+    }
+}
+
+#[test]
+fn sparse() {
+    // sparse keys
+    let mut keys: Vec<TestKey> = Vec::new();
+    let mut used_keys = HashSet::new();
+    for _ in 0..10000 {
+        loop {
+            let key = rand::random::<u128>();
+            if used_keys.contains(&key) {
+                continue;
+            }
+            used_keys.insert(key);
+            keys.push(key.into());
+            break;
+        }
+    }
+    test_inserts(&keys);
+}
+
+#[derive(Clone, Debug)]
+struct TestOp(TestKey, Option<usize>);
+
+fn apply_op(
+    op: &TestOp,
+    map: &mut HashMapAccess<TestKey, usize>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+) {
+    // apply the change to the shadow tree first
+    let shadow_existing = if let Some(v) = op.1 {
+        shadow.insert(op.0, v)
+    } else {
+        shadow.remove(&op.0)
+    };
+
+    let entry = map.entry(op.0);
+    let hash_existing = match op.1 {
+        Some(new) => match entry {
+            Entry::Occupied(mut e) => Some(e.insert(new)),
+            Entry::Vacant(e) => {
+                _ = e.insert(new).unwrap();
+                None
+            }
+        },
+        None => match entry {
+            Entry::Occupied(e) => Some(e.remove()),
+            Entry::Vacant(_) => None,
+        },
+    };
+
+    assert_eq!(shadow_existing, hash_existing);
+}
+
+fn do_random_ops(
+    num_ops: usize,
+    size: u32,
+    del_prob: f64,
+    writer: &mut HashMapAccess<TestKey, usize>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+    rng: &mut rand::rngs::ThreadRng,
+) {
+    for i in 0..num_ops {
+        let key: TestKey = ((rng.next_u32() % size) as u128).into();
+        let op = TestOp(
+            key,
+            if rng.random_bool(del_prob) {
+                Some(i)
+            } else {
+                None
+            },
+        );
+        apply_op(&op, writer, shadow);
+    }
+}
+
+fn do_deletes(
+    num_ops: usize,
+    writer: &mut HashMapAccess<TestKey, usize>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+) {
+    for _ in 0..num_ops {
+        let (k, _) = shadow.pop_first().unwrap();
+        writer.remove(&k);
+    }
+}
+
+fn do_shrink(
+    writer: &mut HashMapAccess<TestKey, usize>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+    from: u32,
+    to: u32,
+) {
+    assert!(writer.shrink_goal().is_none());
+    writer.begin_shrink(to);
+    assert_eq!(writer.shrink_goal(), Some(to as usize));
+    for i in to..from {
+        if let Some(entry) = writer.entry_at_bucket(i as usize) {
+            shadow.remove(&entry._key);
+            entry.remove();
+        }
+    }
+    let old_usage = writer.get_num_buckets_in_use();
+    writer.finish_shrink().unwrap();
+    assert!(writer.shrink_goal().is_none());
+    assert_eq!(writer.get_num_buckets_in_use(), old_usage);
+}
+
+#[test]
+fn random_ops() {
+    let mut writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(100000, 120000, "test_random")
+            .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+
+    let distribution = Zipf::new(u128::MAX as f64, 1.1).unwrap();
+    let mut rng = rand::rng();
+    for i in 0..100000 {
+        let key: TestKey = (rng.sample(distribution) as u128).into();
+
+        let op = TestOp(key, if rng.random_bool(0.75) { Some(i) } else { None });
+
+        apply_op(&op, &mut writer, &mut shadow);
+    }
+}
+
+#[test]
+fn test_shuffle() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1000, 1200, "test_shuf")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+
+    do_random_ops(10000, 1000, 0.75, &mut writer, &mut shadow, &mut rng);
+    writer.shuffle();
+    do_random_ops(10000, 1000, 0.75, &mut writer, &mut shadow, &mut rng);
+}
+
+#[test]
+fn test_grow() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1000, 2000, "test_grow")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+
+    do_random_ops(10000, 1000, 0.75, &mut writer, &mut shadow, &mut rng);
+    let old_usage = writer.get_num_buckets_in_use();
+    writer.grow(1500).unwrap();
+    assert_eq!(writer.get_num_buckets_in_use(), old_usage);
+    assert_eq!(writer.get_num_buckets(), 1500);
+    do_random_ops(10000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
+}
+
+#[test]
+fn test_clear() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_clear")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+    do_random_ops(2000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
+    writer.clear();
+    assert_eq!(writer.get_num_buckets_in_use(), 0);
+    assert_eq!(writer.get_num_buckets(), 1500);
+    while let Some((key, _)) = shadow.pop_first() {
+        assert!(writer.get(&key).is_none());
+    }
+    do_random_ops(2000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
+    for i in 0..(1500 - writer.get_num_buckets_in_use()) {
+        writer.insert((1500 + i as u128).into(), 0).unwrap();
+    }
+    assert_eq!(writer.insert(5000.into(), 0), Err(FullError {}));
+    writer.clear();
+    assert!(writer.insert(5000.into(), 0).is_ok());
+}
+
+#[test]
+fn test_idx_remove() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_clear")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+    do_random_ops(2000, 1500, 0.25, &mut writer, &mut shadow, &mut rng);
+    for _ in 0..100 {
+        let idx = (rng.next_u32() % 1500) as usize;
+        if let Some(e) = writer.entry_at_bucket(idx) {
+            shadow.remove(&e._key);
+            e.remove();
+        }
+    }
+    while let Some((key, val)) = shadow.pop_first() {
+        assert_eq!(*writer.get(&key).unwrap(), val);
+    }
+}
+
+#[test]
+fn test_idx_get() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_clear")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+    do_random_ops(2000, 1500, 0.25, &mut writer, &mut shadow, &mut rng);
+    for _ in 0..100 {
+        let idx = (rng.next_u32() % 1500) as usize;
+        if let Some(pair) = writer.get_at_bucket(idx) {
+            {
+                let v: *const usize = &pair.1;
+                assert_eq!(writer.get_bucket_for_value(v), idx);
+            }
+            {
+                let v: *const usize = &pair.1;
+                assert_eq!(writer.get_bucket_for_value(v), idx);
+            }
+        }
+    }
+}
+
+#[test]
+fn test_shrink() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_shrink")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+
+    do_random_ops(10000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
+    do_shrink(&mut writer, &mut shadow, 1500, 1000);
+    assert_eq!(writer.get_num_buckets(), 1000);
+    do_deletes(500, &mut writer, &mut shadow);
+    do_random_ops(10000, 500, 0.75, &mut writer, &mut shadow, &mut rng);
+    assert!(writer.get_num_buckets_in_use() <= 1000);
+}
+
+#[test]
+fn test_shrink_grow_seq() {
+    let mut writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(1000, 20000, "test_grow_seq")
+            .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+
+    do_random_ops(500, 1000, 0.1, &mut writer, &mut shadow, &mut rng);
+    eprintln!("Shrinking to 750");
+    do_shrink(&mut writer, &mut shadow, 1000, 750);
+    do_random_ops(200, 1000, 0.5, &mut writer, &mut shadow, &mut rng);
+    eprintln!("Growing to 1500");
+    writer.grow(1500).unwrap();
+    do_random_ops(600, 1500, 0.1, &mut writer, &mut shadow, &mut rng);
+    eprintln!("Shrinking to 200");
+    while shadow.len() > 100 {
+        do_deletes(1, &mut writer, &mut shadow);
+    }
+    do_shrink(&mut writer, &mut shadow, 1500, 200);
+    do_random_ops(50, 1500, 0.25, &mut writer, &mut shadow, &mut rng);
+    eprintln!("Growing to 10k");
+    writer.grow(10000).unwrap();
+    do_random_ops(10000, 5000, 0.25, &mut writer, &mut shadow, &mut rng);
+}
+
+#[test]
+fn test_bucket_ops() {
+    let writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1000, 1200, "test_bucket_ops")
+        .attach_writer();
+    match writer.entry(1.into()) {
+        Entry::Occupied(mut e) => {
+            e.insert(2);
+        }
+        Entry::Vacant(e) => {
+            _ = e.insert(2).unwrap();
+        }
+    }
+    assert_eq!(writer.get_num_buckets_in_use(), 1);
+    assert_eq!(writer.get_num_buckets(), 1000);
+    assert_eq!(*writer.get(&1.into()).unwrap(), 2);
+    let pos = match writer.entry(1.into()) {
+        Entry::Occupied(e) => {
+            assert_eq!(e._key, 1.into());
+            e.bucket_pos as usize
+        }
+        Entry::Vacant(_) => {
+            panic!("Insert didn't affect entry");
+        }
+    };
+    assert_eq!(writer.entry_at_bucket(pos).unwrap()._key, 1.into());
+    assert_eq!(*writer.get_at_bucket(pos).unwrap(), (1.into(), 2));
+    {
+        let ptr: *const usize = &*writer.get(&1.into()).unwrap();
+        assert_eq!(writer.get_bucket_for_value(ptr), pos);
+    }
+    writer.remove(&1.into());
+    assert!(writer.get(&1.into()).is_none());
+}
+
+#[test]
+fn test_shrink_zero() {
+    let mut writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_shrink_zero")
+            .attach_writer();
+    writer.begin_shrink(0);
+    for i in 0..1500 {
+        writer.entry_at_bucket(i).map(|x| x.remove());
+    }
+    writer.finish_shrink().unwrap();
+    assert_eq!(writer.get_num_buckets_in_use(), 0);
+    let entry = writer.entry(1.into());
+    if let Entry::Vacant(v) = entry {
+        assert!(v.insert(2).is_err());
+    } else {
+        panic!("Somehow got non-vacant entry in empty map.")
+    }
+    writer.grow(50).unwrap();
+    let entry = writer.entry(1.into());
+    if let Entry::Vacant(v) = entry {
+        assert!(v.insert(2).is_ok());
+    } else {
+        panic!("Somehow got non-vacant entry in empty map.")
+    }
+    assert_eq!(writer.get_num_buckets_in_use(), 1);
+}
+
+#[test]
+#[should_panic]
+fn test_grow_oom() {
+    let writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_grow_oom")
+        .attach_writer();
+    writer.grow(20000).unwrap();
+}
+
+#[test]
+#[should_panic]
+fn test_shrink_bigger() {
+    let mut writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2500, "test_shrink_bigger")
+            .attach_writer();
+    writer.begin_shrink(2000);
+}
+
+#[test]
+#[should_panic]
+fn test_shrink_early_finish() {
+    let writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2500, "test_shrink_early_finish")
+            .attach_writer();
+    writer.finish_shrink().unwrap();
+}
+
+#[test]
+#[should_panic]
+fn test_shrink_fixed_size() {
+    let mut area = [MaybeUninit::uninit(); 10000];
+    let init_struct = HashMapInit::<TestKey, usize>::with_fixed(3, &mut area);
+    let mut writer = init_struct.attach_writer();
+    writer.begin_shrink(1);
+}

libs/neon-shmem/src/lib.rs

@@ -1,418 +1,3 @@
-//! Shared memory utilities for neon communicator
-
-use std::num::NonZeroUsize;
-use std::os::fd::{AsFd, BorrowedFd, OwnedFd};
-use std::ptr::NonNull;
-use std::sync::atomic::{AtomicUsize, Ordering};
-
-use nix::errno::Errno;
-use nix::sys::mman::MapFlags;
-use nix::sys::mman::ProtFlags;
-use nix::sys::mman::mmap as nix_mmap;
-use nix::sys::mman::munmap as nix_munmap;
-use nix::unistd::ftruncate as nix_ftruncate;
-
-/// ShmemHandle represents a shared memory area that can be shared by processes over fork().
-/// Unlike shared memory allocated by Postgres, this area is resizable, up to 'max_size' that's
-/// specified at creation.
-///
-/// The area is backed by an anonymous file created with memfd_create(). The full address space for
-/// 'max_size' is reserved up-front with mmap(), but whenever you call [`ShmemHandle::set_size`],
-/// the underlying file is resized. Do not access the area beyond the current size. Currently, that
-/// will cause the file to be expanded, but we might use mprotect() etc. to enforce that in the
-/// future.
-pub struct ShmemHandle {
-    /// memfd file descriptor
-    fd: OwnedFd,
-
-    max_size: usize,
-
-    // Pointer to the beginning of the shared memory area. The header is stored there.
-    shared_ptr: NonNull<SharedStruct>,
-
-    // Pointer to the beginning of the user data
-    pub data_ptr: NonNull<u8>,
-}
-
-/// This is stored at the beginning in the shared memory area.
-struct SharedStruct {
-    max_size: usize,
-
-    /// Current size of the backing file. The high-order bit is used for the RESIZE_IN_PROGRESS flag
-    current_size: AtomicUsize,
-}
-
-const RESIZE_IN_PROGRESS: usize = 1 << 63;
-
-const HEADER_SIZE: usize = std::mem::size_of::<SharedStruct>();
-
-/// Error type returned by the ShmemHandle functions.
-#[derive(thiserror::Error, Debug)]
-#[error("{msg}: {errno}")]
-pub struct Error {
-    pub msg: String,
-    pub errno: Errno,
-}
-
-impl Error {
-    fn new(msg: &str, errno: Errno) -> Error {
-        Error {
-            msg: msg.to_string(),
-            errno,
-        }
-    }
-}
-
-impl ShmemHandle {
-    /// Create a new shared memory area. To communicate between processes, the processes need to be
-    /// fork()'d after calling this, so that the ShmemHandle is inherited by all processes.
-    ///
-    /// If the ShmemHandle is dropped, the memory is unmapped from the current process. Other
-    /// processes can continue using it, however.
-    pub fn new(name: &str, initial_size: usize, max_size: usize) -> Result<ShmemHandle, Error> {
-        // create the backing anonymous file.
-        let fd = create_backing_file(name)?;
-
-        Self::new_with_fd(fd, initial_size, max_size)
-    }
-
-    fn new_with_fd(
-        fd: OwnedFd,
-        initial_size: usize,
-        max_size: usize,
-    ) -> Result<ShmemHandle, Error> {
-        // We reserve the high-order bit for the RESIZE_IN_PROGRESS flag, and the actual size
-        // is a little larger than this because of the SharedStruct header. Make the upper limit
-        // somewhat smaller than that, because with anything close to that, you'll run out of
-        // memory anyway.
-        if max_size >= 1 << 48 {
-            panic!("max size {max_size} too large");
-        }
-        if initial_size > max_size {
-            panic!("initial size {initial_size} larger than max size {max_size}");
-        }
-
-        // The actual initial / max size is the one given by the caller, plus the size of
-        // 'SharedStruct'.
-        let initial_size = HEADER_SIZE + initial_size;
-        let max_size = NonZeroUsize::new(HEADER_SIZE + max_size).unwrap();
-
-        // Reserve address space for it with mmap
-        //
-        // TODO: Use MAP_HUGETLB if possible
-        let start_ptr = unsafe {
-            nix_mmap(
-                None,
-                max_size,
-                ProtFlags::PROT_READ | ProtFlags::PROT_WRITE,
-                MapFlags::MAP_SHARED,
-                &fd,
-                0,
-            )
-        }
-        .map_err(|e| Error::new("mmap failed: {e}", e))?;
-
-        // Reserve space for the initial size
-        enlarge_file(fd.as_fd(), initial_size as u64)?;
-
-        // Initialize the header
-        let shared: NonNull<SharedStruct> = start_ptr.cast();
-        unsafe {
-            shared.write(SharedStruct {
-                max_size: max_size.into(),
-                current_size: AtomicUsize::new(initial_size),
-            })
-        };
-
-        // The user data begins after the header
-        let data_ptr = unsafe { start_ptr.cast().add(HEADER_SIZE) };
-
-        Ok(ShmemHandle {
-            fd,
-            max_size: max_size.into(),
-            shared_ptr: shared,
-            data_ptr,
-        })
-    }
-
-    // return reference to the header
-    fn shared(&self) -> &SharedStruct {
-        unsafe { self.shared_ptr.as_ref() }
-    }
-
-    /// Resize the shared memory area. 'new_size' must not be larger than the 'max_size' specified
-    /// when creating the area.
-    ///
-    /// This may only be called from one process/thread concurrently. We detect that case
-    /// and return an Error.
-    pub fn set_size(&self, new_size: usize) -> Result<(), Error> {
-        let new_size = new_size + HEADER_SIZE;
-        let shared = self.shared();
-
-        if new_size > self.max_size {
-            panic!(
-                "new size ({} is greater than max size ({})",
-                new_size, self.max_size
-            );
-        }
-        assert_eq!(self.max_size, shared.max_size);
-
-        // Lock the area by setting the bit in 'current_size'
-        //
-        // Ordering::Relaxed would probably be sufficient here, as we don't access any other memory
-        // and the posix_fallocate/ftruncate call is surely a synchronization point anyway. But
-        // since this is not performance-critical, better safe than sorry .
-        let mut old_size = shared.current_size.load(Ordering::Acquire);
-        loop {
-            if (old_size & RESIZE_IN_PROGRESS) != 0 {
-                return Err(Error::new(
-                    "concurrent resize detected",
-                    Errno::UnknownErrno,
-                ));
-            }
-            match shared.current_size.compare_exchange(
-                old_size,
-                new_size,
-                Ordering::Acquire,
-                Ordering::Relaxed,
-            ) {
-                Ok(_) => break,
-                Err(x) => old_size = x,
-            }
-        }
-
-        // Ok, we got the lock.
-        //
-        // NB: If anything goes wrong, we *must* clear the bit!
-        let result = {
-            use std::cmp::Ordering::{Equal, Greater, Less};
-            match new_size.cmp(&old_size) {
-                Less => nix_ftruncate(&self.fd, new_size as i64).map_err(|e| {
-                    Error::new("could not shrink shmem segment, ftruncate failed: {e}", e)
-                }),
-                Equal => Ok(()),
-                Greater => enlarge_file(self.fd.as_fd(), new_size as u64),
-            }
-        };
-
-        // Unlock
-        shared.current_size.store(
-            if result.is_ok() { new_size } else { old_size },
-            Ordering::Release,
-        );
-
-        result
-    }
-
-    /// Returns the current user-visible size of the shared memory segment.
-    ///
-    /// NOTE: a concurrent set_size() call can change the size at any time. It is the caller's
-    /// responsibility not to access the area beyond the current size.
-    pub fn current_size(&self) -> usize {
-        let total_current_size =
-            self.shared().current_size.load(Ordering::Relaxed) & !RESIZE_IN_PROGRESS;
-        total_current_size - HEADER_SIZE
-    }
-}
-
-impl Drop for ShmemHandle {
-    fn drop(&mut self) {
-        // SAFETY: The pointer was obtained from mmap() with the given size.
-        // We unmap the entire region.
-        let _ = unsafe { nix_munmap(self.shared_ptr.cast(), self.max_size) };
-        // The fd is dropped automatically by OwnedFd.
-    }
-}
-
-/// Create a "backing file" for the shared memory area. On Linux, use memfd_create(), to create an
-/// anonymous in-memory file. One macos, fall back to a regular file. That's good enough for
-/// development and testing, but in production we want the file to stay in memory.
-///
-/// disable 'unused_variables' warnings, because in the macos path, 'name' is unused.
-#[allow(unused_variables)]
-fn create_backing_file(name: &str) -> Result<OwnedFd, Error> {
-    #[cfg(not(target_os = "macos"))]
-    {
-        nix::sys::memfd::memfd_create(name, nix::sys::memfd::MFdFlags::empty())
-            .map_err(|e| Error::new("memfd_create failed: {e}", e))
-    }
-    #[cfg(target_os = "macos")]
-    {
-        let file = tempfile::tempfile().map_err(|e| {
-            Error::new(
-                "could not create temporary file to back shmem area: {e}",
-                nix::errno::Errno::from_raw(e.raw_os_error().unwrap_or(0)),
-            )
-        })?;
-        Ok(OwnedFd::from(file))
-    }
-}
-
-fn enlarge_file(fd: BorrowedFd, size: u64) -> Result<(), Error> {
-    // Use posix_fallocate() to enlarge the file. It reserves the space correctly, so that
-    // we don't get a segfault later when trying to actually use it.
-    #[cfg(not(target_os = "macos"))]
-    {
-        nix::fcntl::posix_fallocate(fd, 0, size as i64).map_err(|e| {
-            Error::new(
-                "could not grow shmem segment, posix_fallocate failed: {e}",
-                e,
-            )
-        })
-    }
-    // As a fallback on macos, which doesn't have posix_fallocate, use plain 'fallocate'
-    #[cfg(target_os = "macos")]
-    {
-        nix::unistd::ftruncate(fd, size as i64)
-            .map_err(|e| Error::new("could not grow shmem segment, ftruncate failed: {e}", e))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    use nix::unistd::ForkResult;
-    use std::ops::Range;
-
-    /// check that all bytes in given range have the expected value.
-    fn assert_range(ptr: *const u8, expected: u8, range: Range<usize>) {
-        for i in range {
-            let b = unsafe { *(ptr.add(i)) };
-            assert_eq!(expected, b, "unexpected byte at offset {i}");
-        }
-    }
-
-    /// Write 'b' to all bytes in the given range
-    fn write_range(ptr: *mut u8, b: u8, range: Range<usize>) {
-        unsafe { std::ptr::write_bytes(ptr.add(range.start), b, range.end - range.start) };
-    }
-
-    // simple single-process test of growing and shrinking
-    #[test]
-    fn test_shmem_resize() -> Result<(), Error> {
-        let max_size = 1024 * 1024;
-        let init_struct = ShmemHandle::new("test_shmem_resize", 0, max_size)?;
-
-        assert_eq!(init_struct.current_size(), 0);
-
-        // Initial grow
-        let size1 = 10000;
-        init_struct.set_size(size1).unwrap();
-        assert_eq!(init_struct.current_size(), size1);
-
-        // Write some data
-        let data_ptr = init_struct.data_ptr.as_ptr();
-        write_range(data_ptr, 0xAA, 0..size1);
-        assert_range(data_ptr, 0xAA, 0..size1);
-
-        // Shrink
-        let size2 = 5000;
-        init_struct.set_size(size2).unwrap();
-        assert_eq!(init_struct.current_size(), size2);
-
-        // Grow again
-        let size3 = 20000;
-        init_struct.set_size(size3).unwrap();
-        assert_eq!(init_struct.current_size(), size3);
-
-        // Try to read it. The area that was shrunk and grown again should read as all zeros now
-        assert_range(data_ptr, 0xAA, 0..5000);
-        assert_range(data_ptr, 0, 5000..size1);
-
-        // Try to grow beyond max_size
-        //let size4 = max_size + 1;
-        //assert!(init_struct.set_size(size4).is_err());
-
-        // Dropping init_struct should unmap the memory
-        drop(init_struct);
-
-        Ok(())
-    }
-
-    /// This is used in tests to coordinate between test processes. It's like std::sync::Barrier,
-    /// but is stored in the shared memory area and works across processes. It's implemented by
-    /// polling, because e.g. standard rust mutexes are not guaranteed to work across processes.
-    struct SimpleBarrier {
-        num_procs: usize,
-        count: AtomicUsize,
-    }
-
-    impl SimpleBarrier {
-        unsafe fn init(ptr: *mut SimpleBarrier, num_procs: usize) {
-            unsafe {
-                *ptr = SimpleBarrier {
-                    num_procs,
-                    count: AtomicUsize::new(0),
-                }
-            }
-        }
-
-        pub fn wait(&self) {
-            let old = self.count.fetch_add(1, Ordering::Relaxed);
-
-            let generation = old / self.num_procs;
-
-            let mut current = old + 1;
-            while current < (generation + 1) * self.num_procs {
-                std::thread::sleep(std::time::Duration::from_millis(10));
-                current = self.count.load(Ordering::Relaxed);
-            }
-        }
-    }
-
-    #[test]
-    fn test_multi_process() {
-        // Initialize
-        let max_size = 1_000_000_000_000;
-        let init_struct = ShmemHandle::new("test_multi_process", 0, max_size).unwrap();
-        let ptr = init_struct.data_ptr.as_ptr();
-
-        // Store the SimpleBarrier in the first 1k of the area.
-        init_struct.set_size(10000).unwrap();
-        let barrier_ptr: *mut SimpleBarrier = unsafe {
-            ptr.add(ptr.align_offset(std::mem::align_of::<SimpleBarrier>()))
-                .cast()
-        };
-        unsafe { SimpleBarrier::init(barrier_ptr, 2) };
-        let barrier = unsafe { barrier_ptr.as_ref().unwrap() };
-
-        // Fork another test process. The code after this runs in both processes concurrently.
-        let fork_result = unsafe { nix::unistd::fork().unwrap() };
-
-        // In the parent, fill bytes between 1000..2000. In the child, between 2000..3000
-        if fork_result.is_parent() {
-            write_range(ptr, 0xAA, 1000..2000);
-        } else {
-            write_range(ptr, 0xBB, 2000..3000);
-        }
-        barrier.wait();
-        // Verify the contents. (in both processes)
-        assert_range(ptr, 0xAA, 1000..2000);
-        assert_range(ptr, 0xBB, 2000..3000);
-
-        // Grow, from the child this time
-        let size = 10_000_000;
-        if !fork_result.is_parent() {
-            init_struct.set_size(size).unwrap();
-        }
-        barrier.wait();
-
-        // make some writes at the end
-        if fork_result.is_parent() {
-            write_range(ptr, 0xAA, (size - 10)..size);
-        } else {
-            write_range(ptr, 0xBB, (size - 20)..(size - 10));
-        }
-        barrier.wait();
-
-        // Verify the contents. (This runs in both processes)
-        assert_range(ptr, 0, (size - 1000)..(size - 20));
-        assert_range(ptr, 0xBB, (size - 20)..(size - 10));
-        assert_range(ptr, 0xAA, (size - 10)..size);
-
-        if let ForkResult::Parent { child } = fork_result {
-            nix::sys::wait::waitpid(child, None).unwrap();
-        }
-    }
-}
+pub mod hash;
+pub mod shmem;
+pub mod sync;

libs/neon-shmem/src/shmem.rs

@@ -0,0 +1,409 @@
+//! Dynamically resizable contiguous chunk of shared memory
+
+use std::num::NonZeroUsize;
+use std::os::fd::{AsFd, BorrowedFd, OwnedFd};
+use std::ptr::NonNull;
+use std::sync::atomic::{AtomicUsize, Ordering};
+
+use nix::errno::Errno;
+use nix::sys::mman::MapFlags;
+use nix::sys::mman::ProtFlags;
+use nix::sys::mman::mmap as nix_mmap;
+use nix::sys::mman::munmap as nix_munmap;
+use nix::unistd::ftruncate as nix_ftruncate;
+
+/// `ShmemHandle` represents a shared memory area that can be shared by processes over `fork()`.
+/// Unlike shared memory allocated by Postgres, this area is resizable, up to `max_size` that's
+/// specified at creation.
+///
+/// The area is backed by an anonymous file created with `memfd_create()`. The full address space for
+/// `max_size` is reserved up-front with `mmap()`, but whenever you call [`ShmemHandle::set_size`],
+/// the underlying file is resized. Do not access the area beyond the current size. Currently, that
+/// will cause the file to be expanded, but we might use `mprotect()` etc. to enforce that in the
+/// future.
+pub struct ShmemHandle {
+    /// memfd file descriptor
+    fd: OwnedFd,
+
+    max_size: usize,
+
+    // Pointer to the beginning of the shared memory area. The header is stored there.
+    shared_ptr: NonNull<SharedStruct>,
+
+    // Pointer to the beginning of the user data
+    pub data_ptr: NonNull<u8>,
+}
+
+/// This is stored at the beginning in the shared memory area.
+struct SharedStruct {
+    max_size: usize,
+
+    /// Current size of the backing file. The high-order bit is used for the [`RESIZE_IN_PROGRESS`] flag.
+    current_size: AtomicUsize,
+}
+
+const RESIZE_IN_PROGRESS: usize = 1 << 63;
+
+const HEADER_SIZE: usize = std::mem::size_of::<SharedStruct>();
+
+/// Error type returned by the [`ShmemHandle`] functions.
+#[derive(thiserror::Error, Debug)]
+#[error("{msg}: {errno}")]
+pub struct Error {
+    pub msg: String,
+    pub errno: Errno,
+}
+
+impl Error {
+    fn new(msg: &str, errno: Errno) -> Self {
+        Self {
+            msg: msg.to_string(),
+            errno,
+        }
+    }
+}
+
+impl ShmemHandle {
+    /// Create a new shared memory area. To communicate between processes, the processes need to be
+    /// `fork()`'d after calling this, so that the `ShmemHandle` is inherited by all processes.
+    ///
+    /// If the `ShmemHandle` is dropped, the memory is unmapped from the current process. Other
+    /// processes can continue using it, however.
+    pub fn new(name: &str, initial_size: usize, max_size: usize) -> Result<Self, Error> {
+        // create the backing anonymous file.
+        let fd = create_backing_file(name)?;
+
+        Self::new_with_fd(fd, initial_size, max_size)
+    }
+
+    fn new_with_fd(fd: OwnedFd, initial_size: usize, max_size: usize) -> Result<Self, Error> {
+        // We reserve the high-order bit for the `RESIZE_IN_PROGRESS` flag, and the actual size
+        // is a little larger than this because of the SharedStruct header. Make the upper limit
+        // somewhat smaller than that, because with anything close to that, you'll run out of
+        // memory anyway.
+        assert!(max_size < 1 << 48, "max size {max_size} too large");
+
+        assert!(
+            initial_size <= max_size,
+            "initial size {initial_size} larger than max size {max_size}"
+        );
+
+        // The actual initial / max size is the one given by the caller, plus the size of
+        // 'SharedStruct'.
+        let initial_size = HEADER_SIZE + initial_size;
+        let max_size = NonZeroUsize::new(HEADER_SIZE + max_size).unwrap();
+
+        // Reserve address space for it with mmap
+        //
+        // TODO: Use MAP_HUGETLB if possible
+        let start_ptr = unsafe {
+            nix_mmap(
+                None,
+                max_size,
+                ProtFlags::PROT_READ | ProtFlags::PROT_WRITE,
+                MapFlags::MAP_SHARED,
+                &fd,
+                0,
+            )
+        }
+        .map_err(|e| Error::new("mmap failed", e))?;
+
+        // Reserve space for the initial size
+        enlarge_file(fd.as_fd(), initial_size as u64)?;
+
+        // Initialize the header
+        let shared: NonNull<SharedStruct> = start_ptr.cast();
+        unsafe {
+            shared.write(SharedStruct {
+                max_size: max_size.into(),
+                current_size: AtomicUsize::new(initial_size),
+            });
+        }
+
+        // The user data begins after the header
+        let data_ptr = unsafe { start_ptr.cast().add(HEADER_SIZE) };
+
+        Ok(Self {
+            fd,
+            max_size: max_size.into(),
+            shared_ptr: shared,
+            data_ptr,
+        })
+    }
+
+    // return reference to the header
+    fn shared(&self) -> &SharedStruct {
+        unsafe { self.shared_ptr.as_ref() }
+    }
+
+    /// Resize the shared memory area. `new_size` must not be larger than the `max_size` specified
+    /// when creating the area.
+    ///
+    /// This may only be called from one process/thread concurrently. We detect that case
+    /// and return an [`shmem::Error`](Error).
+    pub fn set_size(&self, new_size: usize) -> Result<(), Error> {
+        let new_size = new_size + HEADER_SIZE;
+        let shared = self.shared();
+
+        assert!(
+            new_size <= self.max_size,
+            "new size ({new_size}) is greater than max size ({})",
+            self.max_size
+        );
+
+        assert_eq!(self.max_size, shared.max_size);
+
+        // Lock the area by setting the bit in `current_size`
+        //
+        // Ordering::Relaxed would probably be sufficient here, as we don't access any other memory
+        // and the `posix_fallocate`/`ftruncate` call is surely a synchronization point anyway. But
+        // since this is not performance-critical, better safe than sorry.
+        let mut old_size = shared.current_size.load(Ordering::Acquire);
+        loop {
+            if (old_size & RESIZE_IN_PROGRESS) != 0 {
+                return Err(Error::new(
+                    "concurrent resize detected",
+                    Errno::UnknownErrno,
+                ));
+            }
+            match shared.current_size.compare_exchange(
+                old_size,
+                new_size,
+                Ordering::Acquire,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => break,
+                Err(x) => old_size = x,
+            }
+        }
+
+        // Ok, we got the lock.
+        //
+        // NB: If anything goes wrong, we *must* clear the bit!
+        let result = {
+            use std::cmp::Ordering::{Equal, Greater, Less};
+            match new_size.cmp(&old_size) {
+                Less => nix_ftruncate(&self.fd, new_size as i64)
+                    .map_err(|e| Error::new("could not shrink shmem segment, ftruncate failed", e)),
+                Equal => Ok(()),
+                Greater => enlarge_file(self.fd.as_fd(), new_size as u64),
+            }
+        };
+
+        // Unlock
+        shared.current_size.store(
+            if result.is_ok() { new_size } else { old_size },
+            Ordering::Release,
+        );
+
+        result
+    }
+
+    /// Returns the current user-visible size of the shared memory segment.
+    ///
+    /// NOTE: a concurrent [`ShmemHandle::set_size()`] call can change the size at any time.
+    /// It is the caller's responsibility not to access the area beyond the current size.
+    pub fn current_size(&self) -> usize {
+        let total_current_size =
+            self.shared().current_size.load(Ordering::Relaxed) & !RESIZE_IN_PROGRESS;
+        total_current_size - HEADER_SIZE
+    }
+}
+
+impl Drop for ShmemHandle {
+    fn drop(&mut self) {
+        // SAFETY: The pointer was obtained from mmap() with the given size.
+        // We unmap the entire region.
+        let _ = unsafe { nix_munmap(self.shared_ptr.cast(), self.max_size) };
+        // The fd is dropped automatically by OwnedFd.
+    }
+}
+
+/// Create a "backing file" for the shared memory area. On Linux, use `memfd_create()`, to create an
+/// anonymous in-memory file. One macos, fall back to a regular file. That's good enough for
+/// development and testing, but in production we want the file to stay in memory.
+///
+/// Disable unused variables warnings because `name` is unused in the macos path.
+#[allow(unused_variables)]
+fn create_backing_file(name: &str) -> Result<OwnedFd, Error> {
+    #[cfg(not(target_os = "macos"))]
+    {
+        nix::sys::memfd::memfd_create(name, nix::sys::memfd::MFdFlags::empty())
+            .map_err(|e| Error::new("memfd_create failed", e))
+    }
+    #[cfg(target_os = "macos")]
+    {
+        let file = tempfile::tempfile().map_err(|e| {
+            Error::new(
+                "could not create temporary file to back shmem area",
+                nix::errno::Errno::from_raw(e.raw_os_error().unwrap_or(0)),
+            )
+        })?;
+        Ok(OwnedFd::from(file))
+    }
+}
+
+fn enlarge_file(fd: BorrowedFd, size: u64) -> Result<(), Error> {
+    // Use posix_fallocate() to enlarge the file. It reserves the space correctly, so that
+    // we don't get a segfault later when trying to actually use it.
+    #[cfg(not(target_os = "macos"))]
+    {
+        nix::fcntl::posix_fallocate(fd, 0, size as i64)
+            .map_err(|e| Error::new("could not grow shmem segment, posix_fallocate failed", e))
+    }
+    // As a fallback on macos, which doesn't have posix_fallocate, use plain 'fallocate'
+    #[cfg(target_os = "macos")]
+    {
+        nix::unistd::ftruncate(fd, size as i64)
+            .map_err(|e| Error::new("could not grow shmem segment, ftruncate failed", e))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    use nix::unistd::ForkResult;
+    use std::ops::Range;
+
+    /// check that all bytes in given range have the expected value.
+    fn assert_range(ptr: *const u8, expected: u8, range: Range<usize>) {
+        for i in range {
+            let b = unsafe { *(ptr.add(i)) };
+            assert_eq!(expected, b, "unexpected byte at offset {i}");
+        }
+    }
+
+    /// Write 'b' to all bytes in the given range
+    fn write_range(ptr: *mut u8, b: u8, range: Range<usize>) {
+        unsafe { std::ptr::write_bytes(ptr.add(range.start), b, range.end - range.start) };
+    }
+
+    // simple single-process test of growing and shrinking
+    #[test]
+    fn test_shmem_resize() -> Result<(), Error> {
+        let max_size = 1024 * 1024;
+        let init_struct = ShmemHandle::new("test_shmem_resize", 0, max_size)?;
+
+        assert_eq!(init_struct.current_size(), 0);
+
+        // Initial grow
+        let size1 = 10000;
+        init_struct.set_size(size1).unwrap();
+        assert_eq!(init_struct.current_size(), size1);
+
+        // Write some data
+        let data_ptr = init_struct.data_ptr.as_ptr();
+        write_range(data_ptr, 0xAA, 0..size1);
+        assert_range(data_ptr, 0xAA, 0..size1);
+
+        // Shrink
+        let size2 = 5000;
+        init_struct.set_size(size2).unwrap();
+        assert_eq!(init_struct.current_size(), size2);
+
+        // Grow again
+        let size3 = 20000;
+        init_struct.set_size(size3).unwrap();
+        assert_eq!(init_struct.current_size(), size3);
+
+        // Try to read it. The area that was shrunk and grown again should read as all zeros now
+        assert_range(data_ptr, 0xAA, 0..5000);
+        assert_range(data_ptr, 0, 5000..size1);
+
+        // Try to grow beyond max_size
+        //let size4 = max_size + 1;
+        //assert!(init_struct.set_size(size4).is_err());
+
+        // Dropping init_struct should unmap the memory
+        drop(init_struct);
+
+        Ok(())
+    }
+
+    /// This is used in tests to coordinate between test processes. It's like `std::sync::Barrier`,
+    /// but is stored in the shared memory area and works across processes. It's implemented by
+    /// polling, because e.g. standard rust mutexes are not guaranteed to work across processes.
+    struct SimpleBarrier {
+        num_procs: usize,
+        count: AtomicUsize,
+    }
+
+    impl SimpleBarrier {
+        unsafe fn init(ptr: *mut SimpleBarrier, num_procs: usize) {
+            unsafe {
+                *ptr = SimpleBarrier {
+                    num_procs,
+                    count: AtomicUsize::new(0),
+                }
+            }
+        }
+
+        pub fn wait(&self) {
+            let old = self.count.fetch_add(1, Ordering::Relaxed);
+
+            let generation = old / self.num_procs;
+
+            let mut current = old + 1;
+            while current < (generation + 1) * self.num_procs {
+                std::thread::sleep(std::time::Duration::from_millis(10));
+                current = self.count.load(Ordering::Relaxed);
+            }
+        }
+    }
+
+    #[test]
+    fn test_multi_process() {
+        // Initialize
+        let max_size = 1_000_000_000_000;
+        let init_struct = ShmemHandle::new("test_multi_process", 0, max_size).unwrap();
+        let ptr = init_struct.data_ptr.as_ptr();
+
+        // Store the SimpleBarrier in the first 1k of the area.
+        init_struct.set_size(10000).unwrap();
+        let barrier_ptr: *mut SimpleBarrier = unsafe {
+            ptr.add(ptr.align_offset(std::mem::align_of::<SimpleBarrier>()))
+                .cast()
+        };
+        unsafe { SimpleBarrier::init(barrier_ptr, 2) };
+        let barrier = unsafe { barrier_ptr.as_ref().unwrap() };
+
+        // Fork another test process. The code after this runs in both processes concurrently.
+        let fork_result = unsafe { nix::unistd::fork().unwrap() };
+
+        // In the parent, fill bytes between 1000..2000. In the child, between 2000..3000
+        if fork_result.is_parent() {
+            write_range(ptr, 0xAA, 1000..2000);
+        } else {
+            write_range(ptr, 0xBB, 2000..3000);
+        }
+        barrier.wait();
+        // Verify the contents. (in both processes)
+        assert_range(ptr, 0xAA, 1000..2000);
+        assert_range(ptr, 0xBB, 2000..3000);
+
+        // Grow, from the child this time
+        let size = 10_000_000;
+        if !fork_result.is_parent() {
+            init_struct.set_size(size).unwrap();
+        }
+        barrier.wait();
+
+        // make some writes at the end
+        if fork_result.is_parent() {
+            write_range(ptr, 0xAA, (size - 10)..size);
+        } else {
+            write_range(ptr, 0xBB, (size - 20)..(size - 10));
+        }
+        barrier.wait();
+
+        // Verify the contents. (This runs in both processes)
+        assert_range(ptr, 0, (size - 1000)..(size - 20));
+        assert_range(ptr, 0xBB, (size - 20)..(size - 10));
+        assert_range(ptr, 0xAA, (size - 10)..size);
+
+        if let ForkResult::Parent { child } = fork_result {
+            nix::sys::wait::waitpid(child, None).unwrap();
+        }
+    }
+}