proxy: add benchmark for custom json logging vs official fmt logger

Don't depend on the postgres_ffi just for one type (#12610 )
We don't want to depend on postgres_ffi in an API crate. If there is no such dependency, we can compile stuff like `storcon_cli` without needing a full working postgres build. Fixes regression of #12548 (before we could compile it).
2026-05-31 12:00:42 +00:00 · 2025-07-15 19:44:41 +02:00 · 2025-07-15 17:28:08 +00:00 · 2025-07-15 14:36:28 +00:00 · 2025-07-15 11:16:29 +00:00 · 2025-07-15 11:06:58 +00:00
187 changed files with 4730 additions and 14456 deletions
--- a/.dockerignore
+++ b/.dockerignore
@@ -27,4 +27,4 @@
 !storage_controller/
 !vendor/postgres-*/
 !workspace_hack/
-!build_tools/patches
+!build-tools/patches
--- a/.github/actionlint.yml
+++ b/.github/actionlint.yml
@@ -31,6 +31,7 @@ config-variables:
  - NEON_PROD_AWS_ACCOUNT_ID
  - PGREGRESS_PG16_PROJECT_ID
  - PGREGRESS_PG17_PROJECT_ID
+  - PREWARM_PGBENCH_SIZE
  - REMOTE_STORAGE_AZURE_CONTAINER
  - REMOTE_STORAGE_AZURE_REGION
  - SLACK_CICD_CHANNEL_ID
--- a/.github/actions/run-python-test-set/action.yml
+++ b/.github/actions/run-python-test-set/action.yml
@@ -176,7 +176,11 @@ runs:
        fi

        if [[ $BUILD_TYPE == "debug" && $RUNNER_ARCH == 'X64' ]]; then
-          cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
+          # We don't use code coverage for regression tests (the step is disabled),
+          # so there's no need to collect it.
+          # Ref https://github.com/neondatabase/neon/issues/4540
+          # cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
+          cov_prefix=()
        else
          cov_prefix=()
        fi
--- a/.github/workflows/_build-and-test-locally.yml
+++ b/.github/workflows/_build-and-test-locally.yml
@@ -150,7 +150,7 @@ jobs:
          secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
          use-fallback: false
          path: pg_install/v14
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}

      - name: Cache postgres v15 build
        id: cache_pg_15
@@ -162,7 +162,7 @@ jobs:
          secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
          use-fallback: false
          path: pg_install/v15
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}

      - name: Cache postgres v16 build
        id: cache_pg_16
@@ -174,7 +174,7 @@ jobs:
          secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
          use-fallback: false
          path: pg_install/v16
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}

      - name: Cache postgres v17 build
        id: cache_pg_17
@@ -186,7 +186,7 @@ jobs:
          secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
          use-fallback: false
          path: pg_install/v17
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v17_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v17_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}

      - name: Build all
        # Note: the Makefile picks up BUILD_TYPE and CARGO_PROFILE from the env variables
--- a/.github/workflows/benchmarking.yml
+++ b/.github/workflows/benchmarking.yml
@@ -219,6 +219,7 @@ jobs:
          --ignore test_runner/performance/test_cumulative_statistics_persistence.py
          --ignore test_runner/performance/test_perf_many_relations.py
          --ignore test_runner/performance/test_perf_oltp_large_tenant.py
+          --ignore test_runner/performance/test_lfc_prewarm.py
      env:
        BENCHMARK_CONNSTR: ${{ steps.create-neon-project.outputs.dsn }}
        VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
@@ -410,6 +411,77 @@ jobs:
      env:
        SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}

+  prewarm-test:
+    if: ${{ github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null }}
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
+    env:
+      PGBENCH_SIZE: ${{ vars.PREWARM_PGBENCH_SIZE }}
+      POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
+      DEFAULT_PG_VERSION: 17
+      TEST_OUTPUT: /tmp/test_output
+      BUILD_TYPE: remote
+      SAVE_PERF_REPORT: ${{ github.event.inputs.save_perf_report || ( github.ref_name == 'main' ) }}
+      PLATFORM: "neon-staging"
+
+    runs-on: [ self-hosted, us-east-2, x64 ]
+    container:
+      image: ghcr.io/neondatabase/build-tools:pinned-bookworm
+      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
+
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@e3dd6a429d7300a6a4c196c26e071d42e0343502 # v4.0.2
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours
+
+    - name: Download Neon artifact
+      uses: ./.github/actions/download
+      with:
+        name: neon-${{ runner.os }}-${{ runner.arch }}-release-artifact
+        path: /tmp/neon/
+        prefix: latest
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+
+    - name: Run prewarm benchmark
+      uses: ./.github/actions/run-python-test-set
+      with:
+        build_type: ${{ env.BUILD_TYPE }}
+        test_selection: performance/test_lfc_prewarm.py
+        run_in_parallel: false
+        save_perf_report: ${{ env.SAVE_PERF_REPORT }}
+        extra_params: -m remote_cluster --timeout 5400
+        pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+      env:
+        VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
+        PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
+        NEON_API_KEY: ${{ secrets.NEON_STAGING_API_KEY }}
+
+    - name: Create Allure report
+      id: create-allure-report
+      if: ${{ !cancelled() }}
+      uses: ./.github/actions/allure-report-generate
+      with:
+        store-test-results-into-db: true
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+      env:
+        REGRESS_TEST_RESULT_CONNSTR_NEW: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
+
  generate-matrices:
    if: ${{ github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null }}
    # Create matrices for the benchmarking jobs, so we run benchmarks on rds only once a week (on Saturday)
--- a/.github/workflows/build-build-tools-image.yml
+++ b/.github/workflows/build-build-tools-image.yml
@@ -72,7 +72,7 @@ jobs:
          ARCHS: ${{ inputs.archs || '["x64","arm64"]' }}
          DEBIANS: ${{ inputs.debians || '["bullseye","bookworm"]' }}
          IMAGE_TAG: |
-            ${{ hashFiles('build-tools.Dockerfile',
+            ${{ hashFiles('build-tools/Dockerfile',
                          '.github/workflows/build-build-tools-image.yml') }}
        run: |
          echo "archs=${ARCHS}"           | tee -a ${GITHUB_OUTPUT}
@@ -144,7 +144,7 @@ jobs:

      - uses: docker/build-push-action@471d1dc4e07e5cdedd4c2171150001c434f0b7a4 # v6.15.0
        with:
-          file: build-tools.Dockerfile
+          file: build-tools/Dockerfile
          context: .
          provenance: false
          push: true
--- a/.gitignore
+++ b/.gitignore
@@ -15,7 +15,6 @@ neon.iml
 /.neon
 /integration_tests/.neon
 compaction-suite-results.*
-pgxn/neon/communicator/communicator_bindings.h
 docker-compose/docker-compose-parallel.yml

 # Coverage
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,16 +1,16 @@
 [submodule "vendor/postgres-v14"]
 	path = vendor/postgres-v14
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_14_STABLE_neon
 [submodule "vendor/postgres-v15"]
 	path = vendor/postgres-v15
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_15_STABLE_neon
 [submodule "vendor/postgres-v16"]
 	path = vendor/postgres-v16
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_16_STABLE_neon
 [submodule "vendor/postgres-v17"]
 	path = vendor/postgres-v17
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_17_STABLE_neon
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -253,17 +253,6 @@ version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a8ab6b55fe97976e46f91ddbed8d147d966475dc29b2032757ba47e02376fbc3"

-[[package]]
-name = "atomic_enum"
-version = "0.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "99e1aca718ea7b89985790c94aad72d77533063fe00bc497bb79a7c2dae6a661"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn 2.0.100",
-]
-
 [[package]]
 name = "autocfg"
 version = "1.1.0"
@@ -698,40 +687,13 @@ dependencies = [
 "tracing",
 ]

-[[package]]
-name = "axum"
-version = "0.7.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "edca88bc138befd0323b20752846e6587272d3b03b0343c8ea28a6f819e6e71f"
-dependencies = [
- "async-trait",
- "axum-core 0.4.5",
- "bytes",
- "futures-util",
- "http 1.1.0",
- "http-body 1.0.0",
- "http-body-util",
- "itoa",
- "matchit 0.7.3",
- "memchr",
- "mime",
- "percent-encoding",
- "pin-project-lite",
- "rustversion",
- "serde",
- "sync_wrapper 1.0.1",
- "tower 0.5.2",
- "tower-layer",
- "tower-service",
-]
-
 [[package]]
 name = "axum"
 version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6d6fd624c75e18b3b4c6b9caf42b1afe24437daaee904069137d8bab077be8b8"
 dependencies = [
- "axum-core 0.5.0",
+ "axum-core",
 "base64 0.22.1",
 "bytes",
 "form_urlencoded",
@@ -739,10 +701,10 @@ dependencies = [
 "http 1.1.0",
 "http-body 1.0.0",
 "http-body-util",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "hyper-util",
 "itoa",
- "matchit 0.8.4",
+ "matchit",
 "memchr",
 "mime",
 "percent-encoding",
@@ -762,26 +724,6 @@ dependencies = [
 "tracing",
 ]

-[[package]]
-name = "axum-core"
-version = "0.4.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09f2bd6146b97ae3359fa0cc6d6b376d9539582c7b4220f041a33ec24c226199"
-dependencies = [
- "async-trait",
- "bytes",
- "futures-util",
- "http 1.1.0",
- "http-body 1.0.0",
- "http-body-util",
- "mime",
- "pin-project-lite",
- "rustversion",
- "sync_wrapper 1.0.1",
- "tower-layer",
- "tower-service",
-]
-
 [[package]]
 name = "axum-core"
 version = "0.5.0"
@@ -808,8 +750,8 @@ version = "0.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "460fc6f625a1f7705c6cf62d0d070794e94668988b1c38111baeec177c715f7b"
 dependencies = [
- "axum 0.8.1",
- "axum-core 0.5.0",
+ "axum",
+ "axum-core",
 "bytes",
 "form_urlencoded",
 "futures-util",
@@ -1346,31 +1288,10 @@ dependencies = [

 [[package]]
 name = "communicator"
-version = "0.0.0"
+version = "0.1.0"
 dependencies = [
- "atomic_enum",
- "axum 0.8.1",
- "bytes",
 "cbindgen",
- "clashmap",
- "http 1.1.0",
- "libc",
- "metrics",
 "neon-shmem",
- "nix 0.30.1",
- "pageserver_api",
- "pageserver_client_grpc",
- "pageserver_page_api",
- "prometheus",
- "prost 0.13.5",
- "thiserror 1.0.69",
- "tokio",
- "tokio-pipe",
- "tonic 0.12.3",
- "tracing",
- "tracing-subscriber",
- "uring-common",
- "utils",
 "workspace_hack",
 ]

@@ -1400,7 +1321,7 @@ dependencies = [
 "aws-sdk-kms",
 "aws-sdk-s3",
 "aws-smithy-types",
- "axum 0.8.1",
+ "axum",
 "axum-extra",
 "base64 0.22.1",
 "bytes",
@@ -1705,9 +1626,9 @@ dependencies = [

 [[package]]
 name = "crossbeam-utils"
-version = "0.8.21"
+version = "0.8.19"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d0a5c400df2834b80a4c3327b3aad3a4c4cd4de0629063962b03235697506a28"
+checksum = "248e3bacc7dc6baa3b21e405ee045c3047101a49145e7e9eca583ab4c2ca5345"

 [[package]]
 name = "crossterm"
@@ -2161,7 +2082,7 @@ name = "endpoint_storage"
 version = "0.0.1"
 dependencies = [
 "anyhow",
- "axum 0.8.1",
+ "axum",
 "axum-extra",
 "camino",
 "camino-tempfile",
@@ -2422,12 +2343,6 @@ version = "1.0.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1"

-[[package]]
-name = "foldhash"
-version = "0.1.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d9c4f5dac5e15c24eb999c26181a6ca40b39fe946cbe4c263c7209467bc83af2"
-
 [[package]]
 name = "form_urlencoded"
 version = "1.2.1"
@@ -2448,7 +2363,7 @@ dependencies = [
 "futures-core",
 "futures-sink",
 "http-body-util",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "hyper-util",
 "pin-project",
 "rand 0.8.5",
@@ -2618,18 +2533,6 @@ dependencies = [
 "wasm-bindgen",
 ]

-[[package]]
-name = "getrandom"
-version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "26145e563e54f2cadc477553f1ec5ee650b00862f0a58bcd12cbdc5f0ea2d2f4"
-dependencies = [
- "cfg-if",
- "libc",
- "r-efi",
- "wasi 0.14.2+wasi-0.2.4",
-]
-
 [[package]]
 name = "gettid"
 version = "0.1.3"
@@ -2795,16 +2698,6 @@ version = "0.15.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bf151400ff0baff5465007dd2f3e717f3fe502074ca563069ce3a6629d07b289"

-[[package]]
-name = "hashbrown"
-version = "0.15.4"
-source = "git+https://github.com/quantumish/hashbrown.git?rev=6610e6d#6610e6d2b1f288ef7b0709a3efefbc846395dc5e"
-dependencies = [
- "allocator-api2",
- "equivalent",
- "foldhash",
-]
-
 [[package]]
 name = "hashlink"
 version = "0.9.1"
@@ -3029,9 +2922,9 @@ dependencies = [

 [[package]]
 name = "httparse"
-version = "1.10.1"
+version = "1.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
+checksum = "d897f394bad6a705d5f4104762e116a75639e470d80901eed05a860a95cb1904"

 [[package]]
 name = "httpdate"
@@ -3081,9 +2974,9 @@ dependencies = [

 [[package]]
 name = "hyper"
-version = "1.6.0"
+version = "1.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cc2b571658e38e0c01b1fdca3bbbe93c00d3d71693ff2770043f8c29bc7d6f80"
+checksum = "50dfd22e0e76d0f662d429a5f80fcaf3855009297eab6a0a9f8543834744ba05"
 dependencies = [
 "bytes",
 "futures-channel",
@@ -3123,7 +3016,7 @@ checksum = "a0bea761b46ae2b24eb4aef630d8d1c398157b6fc29e6350ecf090a0b70c952c"
 dependencies = [
 "futures-util",
 "http 1.1.0",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "hyper-util",
 "rustls 0.22.4",
 "rustls-pki-types",
@@ -3138,7 +3031,7 @@ version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3203a961e5c83b6f5498933e78b6b263e208c197b63e9c6c53cc82ffd3f63793"
 dependencies = [
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "hyper-util",
 "pin-project-lite",
 "tokio",
@@ -3147,21 +3040,20 @@ dependencies = [

 [[package]]
 name = "hyper-util"
-version = "0.1.14"
+version = "0.1.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc2fdfdbff08affe55bb779f33b053aa1fe5dd5b54c257343c17edfa55711bdb"
+checksum = "cde7055719c54e36e95e8719f95883f22072a48ede39db7fc17a4e1d5281e9b9"
 dependencies = [
 "bytes",
 "futures-channel",
- "futures-core",
 "futures-util",
 "http 1.1.0",
 "http-body 1.0.0",
- "hyper 1.6.0",
- "libc",
+ "hyper 1.4.1",
 "pin-project-lite",
 "socket2",
 "tokio",
+ "tower 0.4.13",
 "tower-service",
 "tracing",
 ]
@@ -3714,9 +3606,9 @@ checksum = "4ee93343901ab17bd981295f2cf0026d4ad018c7c31ba84549a4ddbb47a45104"

 [[package]]
 name = "lock_api"
-version = "0.4.13"
+version = "0.4.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "96936507f153605bddfcda068dd804796c84324ed2510809e5b2a624c81da765"
+checksum = "c1cc9717a20b1bb222f333e6a92fd32f7d8a18ddc5a3191a11af45dcbf4dcd16"
 dependencies = [
 "autocfg",
 "scopeguard",
@@ -3759,12 +3651,6 @@ dependencies = [
 "regex-automata 0.1.10",
 ]

-[[package]]
-name = "matchit"
-version = "0.7.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0e7465ac9959cc2b1404e8e2367b43684a6d13790fe23056cc8c6c5a6b7bcb94"
-
 [[package]]
 name = "matchit"
 version = "0.8.4"
@@ -3872,8 +3758,8 @@ dependencies = [
 "procfs",
 "prometheus",
 "rand 0.8.5",
- "rand_distr 0.4.3",
- "twox-hash 1.6.3",
+ "rand_distr",
+ "twox-hash",
 ]

 [[package]]
@@ -3960,33 +3846,10 @@ checksum = "e5ce46fe64a9d73be07dcbe690a38ce1b293be448fd8ce1e6c1b8062c9f72c6a"
 name = "neon-shmem"
 version = "0.1.0"
 dependencies = [
- "ahash",
- "criterion",
- "foldhash",
- "hashbrown 0.15.4",
- "libc",
- "lock_api",
 "nix 0.30.1",
- "rand 0.9.1",
- "rand_distr 0.5.1",
- "rustc-hash 2.1.1",
- "seahash",
 "tempfile",
 "thiserror 1.0.69",
- "twox-hash 2.1.1",
 "workspace_hack",
- "xxhash-rust",
-]
-
-[[package]]
-name = "neonart"
-version = "0.1.0"
-dependencies = [
- "crossbeam-utils",
- "rand 0.9.1",
- "rand_distr 0.5.1",
- "spin",
- "tracing",
 ]

 [[package]]
@@ -4422,20 +4285,18 @@ version = "0.1.0"
 dependencies = [
 "anyhow",
 "async-trait",
- "axum 0.8.1",
 "bytes",
 "camino",
 "clap",
 "futures",
 "hdrhistogram",
- "http 1.1.0",
 "humantime",
 "humantime-serde",
- "metrics",
 "pageserver_api",
 "pageserver_client",
 "pageserver_client_grpc",
 "pageserver_page_api",
+ "pprof",
 "rand 0.8.5",
 "reqwest",
 "serde",
@@ -4520,7 +4381,6 @@ dependencies = [
 "pageserver_client",
 "pageserver_compaction",
 "pageserver_page_api",
- "peekable",
 "pem",
 "pin-project-lite",
 "postgres-protocol",
@@ -4534,7 +4394,6 @@ dependencies = [
 "pprof",
 "pq_proto",
 "procfs",
- "prost 0.13.5",
 "rand 0.8.5",
 "range-set-blaze",
 "regex",
@@ -4571,7 +4430,7 @@ dependencies = [
 "tower 0.5.2",
 "tracing",
 "tracing-utils",
- "twox-hash 1.6.3",
+ "twox-hash",
 "url",
 "utils",
 "uuid",
@@ -4783,7 +4642,7 @@ dependencies = [
 "paste",
 "seq-macro",
 "thrift",
- "twox-hash 1.6.3",
+ "twox-hash",
 "zstd",
 "zstd-sys",
 ]
@@ -4829,15 +4688,6 @@ dependencies = [
 "sha2",
 ]

-[[package]]
-name = "peekable"
-version = "0.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "225f9651e475709164f871dc2f5724956be59cb9edb055372ffeeab01ec2d20b"
-dependencies = [
- "smallvec",
-]
-
 [[package]]
 name = "pem"
 version = "3.0.3"
@@ -5440,6 +5290,7 @@ dependencies = [
 "async-trait",
 "atomic-take",
 "aws-config",
+ "aws-credential-types",
 "aws-sdk-iam",
 "aws-sigv4",
 "base64 0.22.1",
@@ -5452,6 +5303,7 @@ dependencies = [
 "clashmap",
 "compute_api",
 "consumption_metrics",
+ "criterion",
 "ecdsa 0.16.9",
 "ed25519-dalek",
 "env_logger",
@@ -5471,7 +5323,7 @@ dependencies = [
 "humantime",
 "humantime-serde",
 "hyper 0.14.30",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "hyper-util",
 "indexmap 2.9.0",
 "ipnet",
@@ -5479,6 +5331,7 @@ dependencies = [
 "itoa",
 "jose-jwa",
 "jose-jwk",
+ "json",
 "lasso",
 "measured",
 "metrics",
@@ -5495,7 +5348,7 @@ dependencies = [
 "postgres_backend",
 "pq_proto",
 "rand 0.8.5",
- "rand_distr 0.4.3",
+ "rand_distr",
 "rcgen",
 "redis",
 "regex",
@@ -5599,12 +5452,6 @@ dependencies = [
 "proc-macro2",
 ]

-[[package]]
-name = "r-efi"
-version = "5.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "69cdb34c158ceb288df11e18b4bd39de994f6657d83847bdffdbd7f346754b0f"
-
 [[package]]
 name = "rand"
 version = "0.7.3"
@@ -5629,16 +5476,6 @@ dependencies = [
 "rand_core 0.6.4",
 ]

-[[package]]
-name = "rand"
-version = "0.9.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9fbfd9d094a40bf3ae768db9361049ace4c0e04a4fd6b359518bd7b73a73dd97"
-dependencies = [
- "rand_chacha 0.9.0",
- "rand_core 0.9.3",
-]
-
 [[package]]
 name = "rand_chacha"
 version = "0.2.2"
@@ -5659,16 +5496,6 @@ dependencies = [
 "rand_core 0.6.4",
 ]

-[[package]]
-name = "rand_chacha"
-version = "0.9.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d3022b5f1df60f26e1ffddd6c66e8aa15de382ae63b3a0c1bfc0e4d3e3f325cb"
-dependencies = [
- "ppv-lite86",
- "rand_core 0.9.3",
-]
-
 [[package]]
 name = "rand_core"
 version = "0.5.1"
@@ -5687,15 +5514,6 @@ dependencies = [
 "getrandom 0.2.11",
 ]

-[[package]]
-name = "rand_core"
-version = "0.9.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "99d9a13982dcf210057a8a78572b2217b667c3beacbf3a0d8b454f6f82837d38"
-dependencies = [
- "getrandom 0.3.3",
-]
-
 [[package]]
 name = "rand_distr"
 version = "0.4.3"
@@ -5706,16 +5524,6 @@ dependencies = [
 "rand 0.8.5",
 ]

-[[package]]
-name = "rand_distr"
-version = "0.5.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6a8615d50dcf34fa31f7ab52692afec947c4dd0ab803cc87cb3b0b4570ff7463"
-dependencies = [
- "num-traits",
- "rand 0.9.1",
-]
-
 [[package]]
 name = "rand_hc"
 version = "0.2.0"
@@ -5914,7 +5722,7 @@ dependencies = [
 "http-body-util",
 "http-types",
 "humantime-serde",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "itertools 0.10.5",
 "metrics",
 "once_cell",
@@ -5954,7 +5762,7 @@ dependencies = [
 "http 1.1.0",
 "http-body 1.0.0",
 "http-body-util",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "hyper-rustls 0.26.0",
 "hyper-util",
 "ipnet",
@@ -6011,7 +5819,7 @@ dependencies = [
 "futures",
 "getrandom 0.2.11",
 "http 1.1.0",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "parking_lot 0.11.2",
 "reqwest",
 "reqwest-middleware",
@@ -6032,7 +5840,7 @@ dependencies = [
 "async-trait",
 "getrandom 0.2.11",
 "http 1.1.0",
- "matchit 0.8.4",
+ "matchit",
 "opentelemetry",
 "reqwest",
 "reqwest-middleware",
@@ -6404,6 +6212,7 @@ dependencies = [
 "postgres-protocol",
 "postgres_backend",
 "postgres_ffi",
+ "postgres_ffi_types",
 "postgres_versioninfo",
 "pprof",
 "pq_proto",
@@ -6448,7 +6257,7 @@ dependencies = [
 "anyhow",
 "const_format",
 "pageserver_api",
- "postgres_ffi",
+ "postgres_ffi_types",
 "postgres_versioninfo",
 "pq_proto",
 "serde",
@@ -6519,12 +6328,6 @@ version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "621e3680f3e07db4c9c2c3fb07c6223ab2fab2e54bd3c04c3ae037990f428c32"

-[[package]]
-name = "seahash"
-version = "4.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1c107b6f4780854c8b126e228ea8869f4d7b71260f962fefb57b996b8959ba6b"
-
 [[package]]
 name = "sec1"
 version = "0.3.0"
@@ -6986,12 +6789,12 @@ dependencies = [

 [[package]]
 name = "socket2"
-version = "0.5.10"
+version = "0.5.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e22376abed350d73dd1cd119b57ffccad95b4e585a7cda43e286245ce23c0678"
+checksum = "7b5fac59a5cb5dd637972e5fca70daf0523c9067fcdc4842f053dae04a18f8e9"
 dependencies = [
 "libc",
- "windows-sys 0.52.0",
+ "windows-sys 0.48.0",
 ]

 [[package]]
@@ -6999,9 +6802,6 @@ name = "spin"
 version = "0.9.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6980e8d7511241f8acf4aebddbb1ff938df5eebe98691418c4468d0b72a96a67"
-dependencies = [
- "lock_api",
-]

 [[package]]
 name = "spinning_top"
@@ -7060,7 +6860,7 @@ dependencies = [
 "http-body-util",
 "http-utils",
 "humantime",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "hyper-util",
 "metrics",
 "once_cell",
@@ -7196,6 +6996,7 @@ dependencies = [
 "pageserver_api",
 "pageserver_client",
 "reqwest",
+ "safekeeper_api",
 "serde_json",
 "storage_controller_client",
 "tokio",
@@ -7669,16 +7470,6 @@ dependencies = [
 "syn 2.0.100",
 ]

-[[package]]
-name = "tokio-pipe"
-version = "0.2.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f213a84bffbd61b8fa0ba8a044b4bbe35d471d0b518867181e82bd5c15542784"
-dependencies = [
- "libc",
- "tokio",
-]
-
 [[package]]
 name = "tokio-postgres"
 version = "0.7.10"
@@ -7775,6 +7566,7 @@ dependencies = [
 "futures-core",
 "pin-project-lite",
 "tokio",
+ "tokio-util",
 ]

 [[package]]
@@ -7873,25 +7665,16 @@ version = "0.12.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "877c5b330756d856ffcc4553ab34a5684481ade925ecc54bcd1bf02b1d0d4d52"
 dependencies = [
- "async-stream",
 "async-trait",
- "axum 0.7.9",
 "base64 0.22.1",
 "bytes",
- "h2 0.4.4",
 "http 1.1.0",
 "http-body 1.0.0",
 "http-body-util",
- "hyper 1.6.0",
- "hyper-timeout",
- "hyper-util",
 "percent-encoding",
 "pin-project",
 "prost 0.13.5",
- "socket2",
- "tokio",
 "tokio-stream",
- "tower 0.4.13",
 "tower-layer",
 "tower-service",
 "tracing",
@@ -7904,7 +7687,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7e581ba15a835f4d9ea06c55ab1bd4dce26fc53752c69a04aac00703bfb49ba9"
 dependencies = [
 "async-trait",
- "axum 0.8.1",
+ "axum",
 "base64 0.22.1",
 "bytes",
 "flate2",
@@ -7912,7 +7695,7 @@ dependencies = [
 "http 1.1.0",
 "http-body 1.0.0",
 "http-body-util",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "hyper-timeout",
 "hyper-util",
 "percent-encoding",
@@ -7965,16 +7748,11 @@ checksum = "b8fa9be0de6cf49e536ce1851f987bd21a43b771b09473c3549a6c853db37c1c"
 dependencies = [
 "futures-core",
 "futures-util",
- "indexmap 1.9.3",
 "pin-project",
 "pin-project-lite",
- "rand 0.8.5",
- "slab",
 "tokio",
- "tokio-util",
 "tower-layer",
 "tower-service",
- "tracing",
 ]

 [[package]]
@@ -8245,15 +8023,6 @@ dependencies = [
 "static_assertions",
 ]

-[[package]]
-name = "twox-hash"
-version = "2.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b907da542cbced5261bd3256de1b3a1bf340a3d37f93425a07362a1d687de56"
-dependencies = [
- "rand 0.9.1",
-]
-
 [[package]]
 name = "typed-json"
 version = "0.1.1"
@@ -8467,7 +8236,7 @@ name = "vm_monitor"
 version = "0.1.0"
 dependencies = [
 "anyhow",
- "axum 0.8.1",
+ "axum",
 "cgroups-rs",
 "clap",
 "futures",
@@ -8579,15 +8348,6 @@ version = "0.11.0+wasi-snapshot-preview1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9c8d87e72b64a3b4db28d11ce29237c246188f4f51057d65a7eab63b7987e423"

-[[package]]
-name = "wasi"
-version = "0.14.2+wasi-0.2.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9683f9a5a998d873c0d21fcbe3c083009670149a8fab228644b8bd36b2c48cb3"
-dependencies = [
- "wit-bindgen-rt",
-]
-
 [[package]]
 name = "wasite"
 version = "0.1.0"
@@ -8945,15 +8705,6 @@ dependencies = [
 "windows-sys 0.48.0",
 ]

-[[package]]
-name = "wit-bindgen-rt"
-version = "0.39.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6f42320e61fe2cfd34354ecb597f86f413484a798ba44a8ca1165c58d42da6c1"
-dependencies = [
- "bitflags 2.8.0",
-]
-
 [[package]]
 name = "workspace_hack"
 version = "0.1.0"
@@ -8961,8 +8712,8 @@ dependencies = [
 "ahash",
 "anstream",
 "anyhow",
- "axum 0.8.1",
- "axum-core 0.5.0",
+ "axum",
+ "axum-core",
 "base64 0.21.7",
 "base64ct",
 "bytes",
@@ -8996,7 +8747,7 @@ dependencies = [
 "hex",
 "hmac",
 "hyper 0.14.30",
- "hyper 1.6.0",
+ "hyper 1.4.1",
 "hyper-util",
 "indexmap 2.9.0",
 "itertools 0.12.1",
@@ -9121,12 +8872,6 @@ version = "0.13.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4d25c75bf9ea12c4040a97f829154768bbbce366287e2dc044af160cd79a13fd"

-[[package]]
-name = "xxhash-rust"
-version = "0.8.15"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fdd20c5420375476fbd4394763288da7eb0cc0b8c11deed431a91562af7335d3"
-
 [[package]]
 name = "yasna"
 version = "0.5.2"
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -35,7 +35,6 @@ members = [
    "libs/pq_proto",
    "libs/tenant_size_model",
    "libs/metrics",
-    "libs/neonart",
    "libs/postgres_connection",
    "libs/remote_storage",
    "libs/tracing-utils",
@@ -93,7 +92,6 @@ clap = { version = "4.0", features = ["derive", "env"] }
 clashmap = { version = "1.0", features = ["raw-api"] }
 comfy-table = "7.1"
 const_format = "0.2"
-crossbeam-utils = "0.8.21"
 crc32c = "0.6"
 diatomic-waker = { version = "0.2.3" }
 either = "1.8"
@@ -152,7 +150,6 @@ parquet = { version = "53", default-features = false, features = ["zstd"] }
 parquet_derive = "53"
 pbkdf2 = { version = "0.12.1", features = ["simple", "std"] }
 pem = "3.0.3"
-peekable = "0.3.0"
 pin-project-lite = "0.2"
 pprof = { version = "0.14", features = ["criterion", "flamegraph", "frame-pointer", "prost-codec"] }
 procfs = "0.16"
@@ -189,7 +186,6 @@ smallvec = "1.11"
 smol_str = { version = "0.2.0", features = ["serde"] }
 socket2 = "0.5"
 spki = "0.7.3"
-spin = "0.9.8"
 strum = "0.26"
 strum_macros = "0.26"
 "subtle"  = "2.5.0"
@@ -201,10 +197,11 @@ thiserror = "1.0"
 tikv-jemallocator = { version = "0.6", features = ["profiling", "stats", "unprefixed_malloc_on_supported_platforms"] }
 tikv-jemalloc-ctl = { version = "0.6", features = ["stats"] }
 tokio = { version = "1.43.1", features = ["macros"] }
+tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
 tokio-io-timeout = "1.2.0"
 tokio-postgres-rustls = "0.12.0"
 tokio-rustls = { version = "0.26.0", default-features = false, features = ["tls12", "ring"]}
-tokio-stream = "0.1"
+tokio-stream = { version = "0.1", features = ["sync"] }
 tokio-tar = "0.3"
 tokio-util = { version = "0.7.10", features = ["io", "io-util", "rt"] }
 toml = "0.8"
@@ -242,9 +239,6 @@ x509-cert = { version = "0.2.5" }
 env_logger = "0.11"
 log = "0.4"

-tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
-uring-common = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
-
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
 postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
 postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
--- a/build-tools/Dockerfile
+++ b/build-tools/Dockerfile
@@ -35,7 +35,7 @@ RUN echo 'Acquire::Retries "5";' > /etc/apt/apt.conf.d/80-retries && \
    echo -e "retry_connrefused=on\ntimeout=15\ntries=5\nretry-on-host-error=on\n" > /root/.wgetrc && \
    echo -e "--retry-connrefused\n--connect-timeout 15\n--retry 5\n--max-time 300\n" > /root/.curlrc

-COPY build_tools/patches/pgcopydbv017.patch /pgcopydbv017.patch
+COPY build-tools/patches/pgcopydbv017.patch /pgcopydbv017.patch

 RUN if [ "${DEBIAN_VERSION}" = "bookworm" ]; then \
        set -e && \
--- a/build-tools/patches/pgcopydbv017.patch
+++ b/build-tools/patches/pgcopydbv017.patch
--- a/compute/compute-node.Dockerfile
+++ b/compute/compute-node.Dockerfile
@@ -9,7 +9,7 @@
 #
 # build-tools:   This contains Rust compiler toolchain and other tools needed at compile
 #                time. This is also used for the storage builds. This image is defined in
-#                build-tools.Dockerfile.
+#                build-tools/Dockerfile.
 #
 # build-deps:    Contains C compiler, other build tools, and compile-time dependencies
 #                needed to compile PostgreSQL and most extensions. (Some extensions need
@@ -115,7 +115,7 @@ ARG EXTENSIONS=all
 FROM $BASE_IMAGE_SHA AS build-deps
 ARG DEBIAN_VERSION

-# Keep in sync with build-tools.Dockerfile
+# Keep in sync with build-tools/Dockerfile
 ENV PROTOC_VERSION=25.1

 # Use strict mode for bash to catch errors early
@@ -1790,7 +1790,7 @@ RUN set -e \
 #########################################################################################
 FROM build-deps AS exporters
 ARG TARGETARCH
-# Keep sql_exporter version same as in build-tools.Dockerfile and
+# Keep sql_exporter version same as in build-tools/Dockerfile and
 # test_runner/regress/test_compute_metrics.py
 # See comment on the top of the file regading `echo`, `-e` and `\n`
 RUN if [ "$TARGETARCH" = "amd64" ]; then\
--- a/compute_tools/src/compute.rs
+++ b/compute_tools/src/compute.rs
@@ -1,4 +1,4 @@
-use anyhow::{Context, Result, anyhow};
+use anyhow::{Context, Result};
 use chrono::{DateTime, Utc};
 use compute_api::privilege::Privilege;
 use compute_api::responses::{
@@ -6,8 +6,7 @@ use compute_api::responses::{
    LfcPrewarmState, PromoteState, TlsConfig,
 };
 use compute_api::spec::{
-    ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PageserverConnectionInfo,
-    PageserverShardConnectionInfo, PgIdent,
+    ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PageserverProtocol, PgIdent,
 };
 use futures::StreamExt;
 use futures::future::join_all;
@@ -226,7 +225,7 @@ pub struct ParsedSpec {
    pub spec: ComputeSpec,
    pub tenant_id: TenantId,
    pub timeline_id: TimelineId,
-    pub pageserver_conninfo: PageserverConnectionInfo,
+    pub pageserver_connstr: String,
    pub safekeeper_connstrings: Vec<String>,
    pub storage_auth_token: Option<String>,
    /// k8s dns name and port
@@ -273,27 +272,6 @@ impl ParsedSpec {
    }
 }

-fn extract_pageserver_conninfo_from_guc(
-    pageserver_connstring_guc: &str,
-) -> PageserverConnectionInfo {
-    PageserverConnectionInfo {
-        shards: pageserver_connstring_guc
-            .split(',')
-            .enumerate()
-            .map(|(i, connstr)| {
-                (
-                    i as u32,
-                    PageserverShardConnectionInfo {
-                        libpq_url: Some(connstr.to_string()),
-                        grpc_url: None,
-                    },
-                )
-            })
-            .collect(),
-        prefer_grpc: false,
-    }
-}
-
 impl TryFrom<ComputeSpec> for ParsedSpec {
    type Error = String;
    fn try_from(spec: ComputeSpec) -> Result<Self, String> {
@@ -303,17 +281,11 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
        // For backwards-compatibility, the top-level fields in the spec file
        // may be empty. In that case, we need to dig them from the GUCs in the
        // cluster.settings field.
-        let pageserver_conninfo = match &spec.pageserver_connection_info {
-            Some(x) => x.clone(),
-            None => {
-                if let Some(guc) = spec.cluster.settings.find("neon.pageserver_connstring") {
-                    extract_pageserver_conninfo_from_guc(&guc)
-                } else {
-                    return Err("pageserver connstr should be provided".to_string());
-                }
-            }
-        };
-
+        let pageserver_connstr = spec
+            .pageserver_connstring
+            .clone()
+            .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
+            .ok_or("pageserver connstr should be provided")?;
        let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
            if matches!(spec.mode, ComputeMode::Primary) {
                spec.cluster
@@ -363,7 +335,7 @@ impl TryFrom<ComputeSpec> for ParsedSpec {

        let res = ParsedSpec {
            spec,
-            pageserver_conninfo,
+            pageserver_connstr,
            safekeeper_connstrings,
            storage_auth_token,
            tenant_id,
@@ -453,7 +425,7 @@ impl ComputeNode {

        let mut new_state = ComputeState::new();
        if let Some(spec) = config.spec {
-            let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow!(msg))?;
+            let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
            new_state.pspec = Some(pspec);
        }

@@ -1060,13 +1032,16 @@ impl ComputeNode {
    fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
        let spec = compute_state.pspec.as_ref().expect("spec must be set");

+        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
        let started = Instant::now();
-        let (connected, size) = if spec.pageserver_conninfo.prefer_grpc {
-            self.try_get_basebackup_grpc(spec, lsn)?
-        } else {
-            self.try_get_basebackup_libpq(spec, lsn)?
+
+        let (connected, size) = match PageserverProtocol::from_connstring(shard0_connstr)? {
+            PageserverProtocol::Libpq => self.try_get_basebackup_libpq(spec, lsn)?,
+            PageserverProtocol::Grpc => self.try_get_basebackup_grpc(spec, lsn)?,
        };

+        self.fix_zenith_signal_neon_signal()?;
+
        let mut state = self.state.lock().unwrap();
        state.metrics.pageserver_connect_micros =
            connected.duration_since(started).as_micros() as u64;
@@ -1076,24 +1051,44 @@ impl ComputeNode {
        Ok(())
    }

+    /// Move the Zenith signal file to Neon signal file location.
+    /// This makes Compute compatible with older PageServers that don't yet
+    /// know about the Zenith->Neon rename.
+    fn fix_zenith_signal_neon_signal(&self) -> Result<()> {
+        let datadir = Path::new(&self.params.pgdata);
+
+        let neonsig = datadir.join("neon.signal");
+
+        if neonsig.is_file() {
+            return Ok(());
+        }
+
+        let zenithsig = datadir.join("zenith.signal");
+
+        if zenithsig.is_file() {
+            fs::copy(zenithsig, neonsig)?;
+        }
+
+        Ok(())
+    }
+
    /// Fetches a basebackup via gRPC. The connstring must use grpc://. Returns the timestamp when
    /// the connection was established, and the (compressed) size of the basebackup.
    fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
-        let shard0 = spec
-            .pageserver_conninfo
-            .shards
-            .get(&0)
-            .expect("shard 0 connection info missing");
-        let shard0_url = shard0.grpc_url.clone().expect("no grpc_url for shard 0");
-
-        let shard_index = match spec.pageserver_conninfo.shards.len() as u8 {
+        let shard0_connstr = spec
+            .pageserver_connstr
+            .split(',')
+            .next()
+            .unwrap()
+            .to_string();
+        let shard_index = match spec.pageserver_connstr.split(',').count() as u8 {
            0 | 1 => ShardIndex::unsharded(),
            count => ShardIndex::new(ShardNumber(0), ShardCount(count)),
        };

        let (reader, connected) = tokio::runtime::Handle::current().block_on(async move {
            let mut client = page_api::Client::connect(
-                shard0_url,
+                shard0_connstr,
                spec.tenant_id,
                spec.timeline_id,
                shard_index,
@@ -1128,13 +1123,8 @@ impl ComputeNode {
    /// Fetches a basebackup via libpq. The connstring must use postgresql://. Returns the timestamp
    /// when the connection was established, and the (compressed) size of the basebackup.
    fn try_get_basebackup_libpq(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
-        let shard0 = spec
-            .pageserver_conninfo
-            .shards
-            .get(&0)
-            .expect("shard 0 connection info missing");
-        let shard0_connstr = shard0.libpq_url.clone().expect("no libpq_url for shard 0");
-        let mut config = postgres::Config::from_str(&shard0_connstr)?;
+        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
+        let mut config = postgres::Config::from_str(shard0_connstr)?;

        // Use the storage auth token from the config file, if given.
        // Note: this overrides any password set in the connection string.
@@ -1220,7 +1210,10 @@ impl ComputeNode {
                    return result;
                }
                Err(ref e) if attempts < max_attempts => {
-                    warn!("Failed to get basebackup: {e:?} (attempt {attempts}/{max_attempts})");
+                    warn!(
+                        "Failed to get basebackup: {} (attempt {}/{})",
+                        e, attempts, max_attempts
+                    );
                    std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
                    retry_period_ms *= 1.5;
                }
@@ -1293,9 +1286,7 @@ 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 {:?} {:?} {:?}",
@@ -1429,8 +1420,16 @@ impl ComputeNode {
            }
        };

-        self.get_basebackup(compute_state, lsn)
-            .with_context(|| format!("failed to get basebackup@{lsn}"))?;
+        info!(
+            "getting basebackup@{} from pageserver {}",
+            lsn, &pspec.pageserver_connstr
+        );
+        self.get_basebackup(compute_state, lsn).with_context(|| {
+            format!(
+                "failed to get basebackup@{} from pageserver {}",
+                lsn, &pspec.pageserver_connstr
+            )
+        })?;

        // Update pg_hba.conf received with basebackup.
        update_pg_hba(pgdata_path)?;
@@ -2096,7 +2095,7 @@ LIMIT 100",
            self.params
                .remote_ext_base_url
                .as_ref()
-                .ok_or(DownloadError::BadInput(anyhow!(
+                .ok_or(DownloadError::BadInput(anyhow::anyhow!(
                    "Remote extensions storage is not configured",
                )))?;

@@ -2292,7 +2291,7 @@ LIMIT 100",
        let remote_extensions = spec
            .remote_extensions
            .as_ref()
-            .ok_or(anyhow!("Remote extensions are not configured"))?;
+            .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;

        info!("parse shared_preload_libraries from spec.cluster.settings");
        let mut libs_vec = Vec::new();
@@ -2371,22 +2370,22 @@ LIMIT 100",
    /// The operation will time out after a specified duration.
    pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
        let state = self.state.lock().unwrap();
-        let old_pageserver_conninfo = state
+        let old_pageserver_connstr = state
            .pspec
            .as_ref()
            .expect("spec must be set")
-            .pageserver_conninfo
+            .pageserver_connstr
            .clone();
        let mut unchanged = true;
        let _ = self
            .state_changed
            .wait_timeout_while(state, duration, |s| {
-                let pageserver_conninfo = &s
+                let pageserver_connstr = &s
                    .pspec
                    .as_ref()
                    .expect("spec must be set")
-                    .pageserver_conninfo;
-                unchanged = pageserver_conninfo == &old_pageserver_conninfo;
+                    .pageserver_connstr;
+                unchanged = pageserver_connstr == &old_pageserver_connstr;
                unchanged
            })
            .unwrap();
@@ -2486,7 +2485,7 @@ pub async fn installed_extensions(conf: tokio_postgres::Config) -> Result<()> {
                serde_json::to_string(&extensions).expect("failed to serialize extensions list")
            );
        }
-        Err(err) => error!("could not get installed extensions: {err:?}"),
+        Err(err) => error!("could not get installed extensions: {err}"),
    }
    Ok(())
 }
--- a/compute_tools/src/config.rs
+++ b/compute_tools/src/config.rs
@@ -56,51 +56,9 @@ pub fn write_postgres_conf(

    // Add options for connecting to storage
    writeln!(file, "# Neon storage settings")?;
-
-    if let Some(conninfo) = &spec.pageserver_connection_info {
-        let mut libpq_urls: Option<Vec<String>> = Some(Vec::new());
-        let mut grpc_urls: Option<Vec<String>> = Some(Vec::new());
-
-        for shardno in 0..conninfo.shards.len() {
-            let info = conninfo.shards.get(&(shardno as u32)).ok_or_else(|| {
-                anyhow::anyhow!("shard {shardno} missing from pageserver_connection_info shard map")
-            })?;
-
-            if let Some(url) = &info.libpq_url {
-                if let Some(ref mut urls) = libpq_urls {
-                    urls.push(url.clone());
-                }
-            } else {
-                libpq_urls = None
-            }
-            if let Some(url) = &info.grpc_url {
-                if let Some(ref mut urls) = grpc_urls {
-                    urls.push(url.clone());
-                }
-            } else {
-                grpc_urls = None
-            }
-        }
-        if let Some(libpq_urls) = libpq_urls {
-            writeln!(
-                file,
-                "neon.pageserver_connstring={}",
-                escape_conf_value(&libpq_urls.join(","))
-            )?;
-        } else {
-            writeln!(file, "# no neon.pageserver_connstring")?;
-        }
-        if let Some(grpc_urls) = grpc_urls {
-            writeln!(
-                file,
-                "neon.pageserver_grpc_urls={}",
-                escape_conf_value(&grpc_urls.join(","))
-            )?;
-        } else {
-            writeln!(file, "# no neon.pageserver_grpc_urls")?;
-        }
+    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}")?;
    }
--- a/compute_tools/src/installed_extensions.rs
+++ b/compute_tools/src/installed_extensions.rs
@@ -2,6 +2,7 @@ use std::collections::HashMap;

 use anyhow::Result;
 use compute_api::responses::{InstalledExtension, InstalledExtensions};
+use tokio_postgres::error::Error as PostgresError;
 use tokio_postgres::{Client, Config, NoTls};

 use crate::metrics::INSTALLED_EXTENSIONS;
@@ -10,7 +11,7 @@ use crate::metrics::INSTALLED_EXTENSIONS;
 /// and to make database listing query here more explicit.
 ///
 /// Limit the number of databases to 500 to avoid excessive load.
-async fn list_dbs(client: &mut Client) -> Result<Vec<String>> {
+async fn list_dbs(client: &mut Client) -> Result<Vec<String>, PostgresError> {
    // `pg_database.datconnlimit = -2` means that the database is in the
    // invalid state
    let databases = client
@@ -37,7 +38,9 @@ async fn list_dbs(client: &mut Client) -> Result<Vec<String>> {
 /// Same extension can be installed in multiple databases with different versions,
 /// so we report a separate metric (number of databases where it is installed)
 /// for each extension version.
-pub async fn get_installed_extensions(mut conf: Config) -> Result<InstalledExtensions> {
+pub async fn get_installed_extensions(
+    mut conf: Config,
+) -> Result<InstalledExtensions, PostgresError> {
    conf.application_name("compute_ctl:get_installed_extensions");
    let databases: Vec<String> = {
        let (mut client, connection) = conf.connect(NoTls).await?;
--- a/compute_tools/src/lsn_lease.rs
+++ b/compute_tools/src/lsn_lease.rs
@@ -4,7 +4,8 @@ use std::thread;
 use std::time::{Duration, SystemTime};

 use anyhow::{Result, bail};
-use compute_api::spec::{ComputeMode, PageserverConnectionInfo};
+use compute_api::spec::{ComputeMode, PageserverProtocol};
+use itertools::Itertools as _;
 use pageserver_page_api as page_api;
 use postgres::{NoTls, SimpleQueryMessage};
 use tracing::{info, warn};
@@ -77,16 +78,17 @@ fn acquire_lsn_lease_with_retry(

    loop {
        // Note: List of pageservers is dynamic, need to re-read configs before each attempt.
-        let (conninfo, auth) = {
+        let (connstrings, auth) = {
            let state = compute.state.lock().unwrap();
            let spec = state.pspec.as_ref().expect("spec must be set");
            (
-                spec.pageserver_conninfo.clone(),
+                spec.pageserver_connstr.clone(),
                spec.storage_auth_token.clone(),
            )
        };

-        let result = try_acquire_lsn_lease(conninfo, auth.as_deref(), tenant_id, timeline_id, lsn);
+        let result =
+            try_acquire_lsn_lease(&connstrings, auth.as_deref(), tenant_id, timeline_id, lsn);
        match result {
            Ok(Some(res)) => {
                return Ok(res);
@@ -110,16 +112,17 @@ fn acquire_lsn_lease_with_retry(

 /// Tries to acquire LSN leases on all Pageserver shards.
 fn try_acquire_lsn_lease(
-    conninfo: PageserverConnectionInfo,
+    connstrings: &str,
    auth: Option<&str>,
    tenant_id: TenantId,
    timeline_id: TimelineId,
    lsn: Lsn,
 ) -> Result<Option<SystemTime>> {
-    let shard_count = conninfo.shards.len();
+    let connstrings = connstrings.split(',').collect_vec();
+    let shard_count = connstrings.len();
    let mut leases = Vec::new();

-    for (shard_number, shard) in conninfo.shards.into_iter() {
+    for (shard_number, &connstring) in connstrings.iter().enumerate() {
        let tenant_shard_id = match shard_count {
            0 | 1 => TenantShardId::unsharded(tenant_id),
            shard_count => TenantShardId {
@@ -129,22 +132,13 @@ fn try_acquire_lsn_lease(
            },
        };

-        let lease = if conninfo.prefer_grpc {
-            acquire_lsn_lease_grpc(
-                &shard.grpc_url.unwrap(),
-                auth,
-                tenant_shard_id,
-                timeline_id,
-                lsn,
-            )?
-        } else {
-            acquire_lsn_lease_libpq(
-                &shard.libpq_url.unwrap(),
-                auth,
-                tenant_shard_id,
-                timeline_id,
-                lsn,
-            )?
+        let lease = match PageserverProtocol::from_connstring(connstring)? {
+            PageserverProtocol::Libpq => {
+                acquire_lsn_lease_libpq(connstring, auth, tenant_shard_id, timeline_id, lsn)?
+            }
+            PageserverProtocol::Grpc => {
+                acquire_lsn_lease_grpc(connstring, auth, tenant_shard_id, timeline_id, lsn)?
+            }
        };
        leases.push(lease);
    }
--- a/control_plane/src/bin/neon_local.rs
+++ b/control_plane/src/bin/neon_local.rs
@@ -16,7 +16,7 @@ use std::time::Duration;
 use anyhow::{Context, Result, anyhow, bail};
 use clap::Parser;
 use compute_api::requests::ComputeClaimsScope;
-use compute_api::spec::{ComputeMode, PageserverConnectionInfo, PageserverShardConnectionInfo};
+use compute_api::spec::{ComputeMode, PageserverProtocol};
 use control_plane::broker::StorageBroker;
 use control_plane::endpoint::{ComputeControlPlane, EndpointTerminateMode};
 use control_plane::endpoint_storage::{ENDPOINT_STORAGE_DEFAULT_ADDR, EndpointStorage};
@@ -1516,35 +1516,29 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                )?;
            }

-            let (shards, stripe_size) = if let Some(ps_id) = pageserver_id {
-                let conf = env.get_pageserver_conf(ps_id).unwrap();
-                let libpq_url = Some({
-                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
-                    let port = port.unwrap_or(5432);
-                    format!("postgres://no_user@{host}:{port}")
-                });
-                let grpc_url = if let Some(grpc_addr) = &conf.listen_grpc_addr {
+            let (pageservers, stripe_size) = if let Some(pageserver_id) = pageserver_id {
+                let conf = env.get_pageserver_conf(pageserver_id).unwrap();
+                // Use gRPC if requested.
+                let pageserver = if endpoint.grpc {
+                    let grpc_addr = conf.listen_grpc_addr.as_ref().expect("bad config");
                    let (host, port) = parse_host_port(grpc_addr)?;
                    let port = port.unwrap_or(DEFAULT_PAGESERVER_GRPC_PORT);
-                    Some(format!("grpc://no_user@{host}:{port}"))
+                    (PageserverProtocol::Grpc, host, port)
                } else {
-                    None
+                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
+                    let port = port.unwrap_or(5432);
+                    (PageserverProtocol::Libpq, host, port)
                };
-                let pageserver = PageserverShardConnectionInfo {
-                    libpq_url,
-                    grpc_url,
-                };
-
                // If caller is telling us what pageserver to use, this is not a tenant which is
                // fully managed by storage controller, therefore not sharded.
-                (vec![(0, pageserver)], DEFAULT_STRIPE_SIZE)
+                (vec![pageserver], DEFAULT_STRIPE_SIZE)
            } else {
                // Look up the currently attached location of the tenant, and its striping metadata,
                // to pass these on to postgres.
                let storage_controller = StorageController::from_env(env);
                let locate_result = storage_controller.tenant_locate(endpoint.tenant_id).await?;
-                let shards = futures::future::try_join_all(locate_result.shards.into_iter().map(
-                    |shard| async move {
+                let pageservers = futures::future::try_join_all(
+                    locate_result.shards.into_iter().map(|shard| async move {
                        if let ComputeMode::Static(lsn) = endpoint.mode {
                            // Initialize LSN leases for static computes.
                            let conf = env.get_pageserver_conf(shard.node_id).unwrap();
@@ -1556,34 +1550,28 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                                .await?;
                        }

-                        let libpq_host = Host::parse(&shard.listen_pg_addr)?;
-                        let libpq_port = shard.listen_pg_port;
-                        let libpq_url =
-                            Some(format!("postgres://no_user@{libpq_host}:{libpq_port}"));
-
-                        let grpc_url = if let Some(grpc_host) = shard.listen_grpc_addr {
-                            let grpc_port = shard.listen_grpc_port.expect("no gRPC port");
-                            Some(format!("grpc://no_user@{grpc_host}:{grpc_port}"))
+                        let pageserver = if endpoint.grpc {
+                            (
+                                PageserverProtocol::Grpc,
+                                Host::parse(&shard.listen_grpc_addr.expect("no gRPC address"))?,
+                                shard.listen_grpc_port.expect("no gRPC port"),
+                            )
                        } else {
-                            None
+                            (
+                                PageserverProtocol::Libpq,
+                                Host::parse(&shard.listen_pg_addr)?,
+                                shard.listen_pg_port,
+                            )
                        };
-                        let pageserver = PageserverShardConnectionInfo {
-                            libpq_url,
-                            grpc_url,
-                        };
-                        anyhow::Ok((shard.shard_id.shard_number.0 as u32, pageserver))
-                    },
-                ))
+                        anyhow::Ok(pageserver)
+                    }),
+                )
                .await?;
                let stripe_size = locate_result.shard_params.stripe_size;

-                (shards, stripe_size)
-            };
-            assert!(!shards.is_empty());
-            let pageserver_conninfo = PageserverConnectionInfo {
-                shards: shards.into_iter().collect(),
-                prefer_grpc: endpoint.grpc,
+                (pageservers, stripe_size)
            };
+            assert!(!pageservers.is_empty());

            let ps_conf = env.get_pageserver_conf(DEFAULT_PAGESERVER_ID)?;
            let auth_token = if matches!(ps_conf.pg_auth_type, AuthType::NeonJWT) {
@@ -1613,7 +1601,7 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                endpoint_storage_addr,
                safekeepers_generation,
                safekeepers,
-                pageserver_conninfo,
+                pageservers,
                remote_ext_base_url: remote_ext_base_url.clone(),
                shard_stripe_size: stripe_size.0 as usize,
                create_test_user: args.create_test_user,
@@ -1632,27 +1620,20 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                .endpoints
                .get(endpoint_id.as_str())
                .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
-            let shards = if let Some(ps_id) = args.endpoint_pageserver_id {
+            let pageservers = if let Some(ps_id) = args.endpoint_pageserver_id {
                let conf = env.get_pageserver_conf(ps_id)?;
-                let libpq_url = Some({
-                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
-                    let port = port.unwrap_or(5432);
-                    format!("postgres://no_user@{host}:{port}")
-                });
-                let grpc_url = if let Some(grpc_addr) = &conf.listen_grpc_addr {
+                // Use gRPC if requested.
+                let pageserver = if endpoint.grpc {
+                    let grpc_addr = conf.listen_grpc_addr.as_ref().expect("bad config");
                    let (host, port) = parse_host_port(grpc_addr)?;
                    let port = port.unwrap_or(DEFAULT_PAGESERVER_GRPC_PORT);
-                    Some(format!("grpc://no_user@{host}:{port}"))
+                    (PageserverProtocol::Grpc, host, port)
                } else {
-                    None
+                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
+                    let port = port.unwrap_or(5432);
+                    (PageserverProtocol::Libpq, host, port)
                };
-                let pageserver = PageserverShardConnectionInfo {
-                    libpq_url,
-                    grpc_url,
-                };
-                // If caller is telling us what pageserver to use, this is not a tenant which is
-                // fully managed by storage controller, therefore not sharded.
-                vec![(0, pageserver)]
+                vec![pageserver]
            } else {
                let storage_controller = StorageController::from_env(env);
                storage_controller
@@ -1662,36 +1643,28 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                    .into_iter()
                    .map(|shard| {
                        // Use gRPC if requested.
-                        let libpq_host = Host::parse(&shard.listen_pg_addr).expect("bad hostname");
-                        let libpq_port = shard.listen_pg_port;
-                        let libpq_url =
-                            Some(format!("postgres://no_user@{libpq_host}:{libpq_port}"));
-
-                        let grpc_url = if let Some(grpc_host) = shard.listen_grpc_addr {
-                            let grpc_port = shard.listen_grpc_port.expect("no gRPC port");
-                            Some(format!("grpc://no_user@{grpc_host}:{grpc_port}"))
+                        if endpoint.grpc {
+                            (
+                                PageserverProtocol::Grpc,
+                                Host::parse(&shard.listen_grpc_addr.expect("no gRPC address"))
+                                    .expect("bad hostname"),
+                                shard.listen_grpc_port.expect("no gRPC port"),
+                            )
                        } else {
-                            None
-                        };
-                        (
-                            shard.shard_id.shard_number.0 as u32,
-                            PageserverShardConnectionInfo {
-                                libpq_url,
-                                grpc_url,
-                            },
-                        )
+                            (
+                                PageserverProtocol::Libpq,
+                                Host::parse(&shard.listen_pg_addr).expect("bad hostname"),
+                                shard.listen_pg_port,
+                            )
+                        }
                    })
                    .collect::<Vec<_>>()
            };
-            let pageserver_conninfo = PageserverConnectionInfo {
-                shards: shards.into_iter().collect(),
-                prefer_grpc: endpoint.grpc,
-            };
            // If --safekeepers argument is given, use only the listed
            // safekeeper nodes; otherwise all from the env.
            let safekeepers = parse_safekeepers(&args.safekeepers)?;
            endpoint
-                .reconfigure(Some(pageserver_conninfo), None, safekeepers, None)
+                .reconfigure(Some(pageservers), None, safekeepers, None)
                .await?;
        }
        EndpointCmd::Stop(args) => {
--- a/control_plane/src/broker.rs
+++ b/control_plane/src/broker.rs
@@ -36,7 +36,7 @@ impl StorageBroker {
    pub async fn start(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
        let broker = &self.env.broker;

-        print!("Starting neon broker at {}", broker.client_url());
+        println!("Starting neon broker at {}", broker.client_url());

        let mut args = Vec::new();

--- a/control_plane/src/endpoint.rs
+++ b/control_plane/src/endpoint.rs
@@ -32,7 +32,8 @@
 //!     config.json                 - passed to `compute_ctl`
 //!     pgdata/
 //!         postgresql.conf       - copy of postgresql.conf created by `compute_ctl`
-//!         zenith.signal
+//!         neon.signal
+//!         zenith.signal         - copy of neon.signal, for backward compatibility
 //!         <other PostgreSQL files>
 //! ```
 //!
@@ -56,13 +57,9 @@ use compute_api::responses::{
    TlsConfig,
 };
 use compute_api::spec::{
-    Cluster, ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, Database, PgIdent,
-    RemoteExtSpec, Role,
+    Cluster, ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, Database, PageserverProtocol,
+    PgIdent, RemoteExtSpec, Role,
 };
-
-// re-export these, because they're used in the reconfigure() function
-pub use compute_api::spec::{PageserverConnectionInfo, PageserverShardConnectionInfo};
-
 use jsonwebtoken::jwk::{
    AlgorithmParameters, CommonParameters, EllipticCurve, Jwk, JwkSet, KeyAlgorithm, KeyOperations,
    OctetKeyPairParameters, OctetKeyPairType, PublicKeyUse,
@@ -78,6 +75,7 @@ use sha2::{Digest, Sha256};
 use spki::der::Decode;
 use spki::{SubjectPublicKeyInfo, SubjectPublicKeyInfoRef};
 use tracing::debug;
+use url::Host;
 use utils::id::{NodeId, TenantId, TimelineId};

 use crate::local_env::LocalEnv;
@@ -382,7 +380,7 @@ pub struct EndpointStartArgs {
    pub endpoint_storage_addr: String,
    pub safekeepers_generation: Option<SafekeeperGeneration>,
    pub safekeepers: Vec<NodeId>,
-    pub pageserver_conninfo: PageserverConnectionInfo,
+    pub pageservers: Vec<(PageserverProtocol, Host, u16)>,
    pub remote_ext_base_url: Option<String>,
    pub shard_stripe_size: usize,
    pub create_test_user: bool,
@@ -466,7 +464,7 @@ 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());
@@ -656,6 +654,14 @@ impl Endpoint {
        }
    }

+    fn build_pageserver_connstr(pageservers: &[(PageserverProtocol, Host, u16)]) -> String {
+        pageservers
+            .iter()
+            .map(|(scheme, host, port)| format!("{scheme}://no_user@{host}:{port}"))
+            .collect::<Vec<_>>()
+            .join(",")
+    }
+
    /// Map safekeepers ids to the actual connection strings.
    fn build_safekeepers_connstrs(&self, sk_ids: Vec<NodeId>) -> Result<Vec<String>> {
        let mut safekeeper_connstrings = Vec::new();
@@ -701,6 +707,9 @@ impl Endpoint {
            std::fs::remove_dir_all(self.pgdata())?;
        }

+        let pageserver_connstring = Self::build_pageserver_connstr(&args.pageservers);
+        assert!(!pageserver_connstring.is_empty());
+
        let safekeeper_connstrings = self.build_safekeepers_connstrs(args.safekeepers)?;

        // check for file remote_extensions_spec.json
@@ -759,7 +768,7 @@ impl Endpoint {
                branch_id: None,
                endpoint_id: Some(self.endpoint_id.clone()),
                mode: self.mode,
-                pageserver_connection_info: Some(args.pageserver_conninfo),
+                pageserver_connstring: Some(pageserver_connstring),
                safekeepers_generation: args.safekeepers_generation.map(|g| g.into_inner()),
                safekeeper_connstrings,
                storage_auth_token: args.auth_token.clone(),
@@ -973,7 +982,7 @@ impl Endpoint {

    pub async fn reconfigure(
        &self,
-        pageserver_conninfo: Option<PageserverConnectionInfo>,
+        pageservers: Option<Vec<(PageserverProtocol, Host, u16)>>,
        stripe_size: Option<ShardStripeSize>,
        safekeepers: Option<Vec<NodeId>>,
        safekeeper_generation: Option<SafekeeperGeneration>,
@@ -989,17 +998,15 @@ impl Endpoint {
        let postgresql_conf = self.read_postgresql_conf()?;
        spec.cluster.postgresql_conf = Some(postgresql_conf);

-        if let Some(pageserver_conninfo) = pageserver_conninfo {
-            // If pageservers are provided, we need to ensure that they are not empty.
-            // This is a requirement for the compute_ctl configuration.
-            anyhow::ensure!(
-                !pageserver_conninfo.shards.is_empty(),
-                "no pageservers provided"
-            );
-            spec.pageserver_connection_info = Some(pageserver_conninfo);
-        }
-        if stripe_size.is_some() {
-            spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
+        // If pageservers are not specified, don't change them.
+        if let Some(pageservers) = pageservers {
+            anyhow::ensure!(!pageservers.is_empty(), "no pageservers provided");
+
+            let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
+            spec.pageserver_connstring = Some(pageserver_connstr);
+            if stripe_size.is_some() {
+                spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
+            }
        }

        // If safekeepers are not specified, don't change them.
@@ -1048,7 +1055,7 @@ impl Endpoint {

    pub async fn reconfigure_pageservers(
        &self,
-        pageservers: PageserverConnectionInfo,
+        pageservers: Vec<(PageserverProtocol, Host, u16)>,
        stripe_size: Option<ShardStripeSize>,
    ) -> Result<()> {
        self.reconfigure(Some(pageservers), stripe_size, None, None)
--- a/control_plane/src/local_env.rs
+++ b/control_plane/src/local_env.rs
@@ -217,6 +217,9 @@ pub struct NeonStorageControllerConf {
    pub posthog_config: Option<PostHogConfig>,

    pub kick_secondary_downloads: Option<bool>,
+
+    #[serde(with = "humantime_serde")]
+    pub shard_split_request_timeout: Option<Duration>,
 }

 impl NeonStorageControllerConf {
@@ -250,6 +253,7 @@ impl Default for NeonStorageControllerConf {
            timeline_safekeeper_count: None,
            posthog_config: None,
            kick_secondary_downloads: None,
+            shard_split_request_timeout: None,
        }
    }
 }
--- a/control_plane/src/pageserver.rs
+++ b/control_plane/src/pageserver.rs
@@ -303,7 +303,7 @@ impl PageServerNode {
    async fn start_node(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
        // TODO: using a thread here because start_process() is not async but we need to call check_status()
        let datadir = self.repo_path();
-        print!(
+        println!(
            "Starting pageserver node {} at '{}' in {:?}, retrying for {:?}",
            self.conf.id,
            self.pg_connection_config.raw_address(),
--- a/control_plane/src/safekeeper.rs
+++ b/control_plane/src/safekeeper.rs
@@ -127,7 +127,7 @@ impl SafekeeperNode {
        extra_opts: &[String],
        retry_timeout: &Duration,
    ) -> anyhow::Result<()> {
-        print!(
+        println!(
            "Starting safekeeper at '{}' in '{}', retrying for {:?}",
            self.pg_connection_config.raw_address(),
            self.datadir_path().display(),
--- a/control_plane/src/storage_controller.rs
+++ b/control_plane/src/storage_controller.rs
@@ -648,6 +648,13 @@ impl StorageController {
            args.push(format!("--timeline-safekeeper-count={sk_cnt}"));
        }

+        if let Some(duration) = self.config.shard_split_request_timeout {
+            args.push(format!(
+                "--shard-split-request-timeout={}",
+                humantime::Duration::from(duration)
+            ));
+        }
+
        let mut envs = vec![
            ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
            ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
@@ -660,7 +667,7 @@ impl StorageController {
            ));
        }

-        println!("Starting storage controller");
+        println!("Starting storage controller at {scheme}://{host}:{listen_port}");

        background_process::start_process(
            COMMAND,
--- a/control_plane/storcon_cli/Cargo.toml
+++ b/control_plane/storcon_cli/Cargo.toml
@@ -14,6 +14,7 @@ humantime.workspace = true
 pageserver_api.workspace = true
 pageserver_client.workspace = true
 reqwest.workspace = true
+safekeeper_api.workspace=true
 serde_json = { workspace = true, features = ["raw_value"] }
 storage_controller_client.workspace = true
 tokio.workspace = true
--- a/control_plane/storcon_cli/src/main.rs
+++ b/control_plane/storcon_cli/src/main.rs
@@ -11,7 +11,7 @@ use pageserver_api::controller_api::{
    PlacementPolicy, SafekeeperDescribeResponse, SafekeeperSchedulingPolicyRequest,
    ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
    SkSchedulingPolicy, TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
-    TenantShardMigrateRequest, TenantShardMigrateResponse,
+    TenantShardMigrateRequest, TenantShardMigrateResponse, TimelineSafekeeperMigrateRequest,
 };
 use pageserver_api::models::{
    EvictionPolicy, EvictionPolicyLayerAccessThreshold, ShardParameters, TenantConfig,
@@ -21,6 +21,7 @@ use pageserver_api::models::{
 use pageserver_api::shard::{ShardStripeSize, TenantShardId};
 use pageserver_client::mgmt_api::{self};
 use reqwest::{Certificate, Method, StatusCode, Url};
+use safekeeper_api::models::TimelineLocateResponse;
 use storage_controller_client::control_api::Client;
 use utils::id::{NodeId, TenantId, TimelineId};

@@ -279,6 +280,23 @@ enum Command {
        #[arg(long)]
        concurrency: Option<usize>,
    },
+    /// Locate safekeepers for a timeline from the storcon DB.
+    TimelineLocate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+    },
+    /// Migrate a timeline to a new set of safekeepers
+    TimelineSafekeeperMigrate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+        /// Example: --new-sk-set 1,2,3
+        #[arg(long, required = true, value_delimiter = ',')]
+        new_sk_set: Vec<NodeId>,
+    },
 }

 #[derive(Parser)]
@@ -458,6 +476,7 @@ async fn main() -> anyhow::Result<()> {
                        listen_http_port,
                        listen_https_port,
                        availability_zone_id: AvailabilityZone(availability_zone_id),
+                        node_ip_addr: None,
                    }),
                )
                .await?;
@@ -1324,7 +1343,7 @@ async fn main() -> anyhow::Result<()> {
            concurrency,
        } => {
            let mut path = format!(
-                "/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
+                "v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
            );

            if let Some(c) = concurrency {
@@ -1335,6 +1354,41 @@ async fn main() -> anyhow::Result<()> {
                .dispatch::<(), ()>(Method::POST, path, None)
                .await?;
        }
+        Command::TimelineLocate {
+            tenant_id,
+            timeline_id,
+        } => {
+            let path = format!("debug/v1/tenant/{tenant_id}/timeline/{timeline_id}/locate");
+
+            let resp = storcon_client
+                .dispatch::<(), TimelineLocateResponse>(Method::GET, path, None)
+                .await?;
+
+            let sk_set = resp.sk_set.iter().map(|id| id.0 as i64).collect::<Vec<_>>();
+            let new_sk_set = resp
+                .new_sk_set
+                .as_ref()
+                .map(|ids| ids.iter().map(|id| id.0 as i64).collect::<Vec<_>>());
+
+            println!("generation = {}", resp.generation);
+            println!("sk_set = {sk_set:?}");
+            println!("new_sk_set = {new_sk_set:?}");
+        }
+        Command::TimelineSafekeeperMigrate {
+            tenant_id,
+            timeline_id,
+            new_sk_set,
+        } => {
+            let path = format!("v1/tenant/{tenant_id}/timeline/{timeline_id}/safekeeper_migrate");
+
+            storcon_client
+                .dispatch::<_, ()>(
+                    Method::POST,
+                    path,
+                    Some(TimelineSafekeeperMigrateRequest { new_sk_set }),
+                )
+                .await?;
+        }
    }

    Ok(())
--- a/docs/core_changes.md
+++ b/docs/core_changes.md
@@ -129,9 +129,10 @@ segment to bootstrap the WAL writing, but it doesn't contain the checkpoint reco
 changes in xlog.c, to allow starting the compute node without reading the last checkpoint record
 from WAL.

-This includes code to read the `zenith.signal` file, which tells the startup code the LSN to start
-at. When the `zenith.signal` file is present, the startup uses that LSN instead of the last
-checkpoint's LSN. The system is known to be consistent at that LSN, without any WAL redo.
+This includes code to read the `neon.signal` (also `zenith.signal`) file, which tells the startup 
+code the LSN to start at. When the `neon.signal` file is present, the startup uses that LSN
+instead of the last checkpoint's LSN. The system is known to be consistent at that LSN, without 
+any WAL redo.


 ### How to get rid of the patch
--- a/docs/pageserver-services.md
+++ b/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:
--- a/libs/compute_api/src/spec.rs
+++ b/libs/compute_api/src/spec.rs
@@ -105,11 +105,7 @@ pub struct ComputeSpec {
    // updated to fill these fields, we can make these non optional.
    pub tenant_id: Option<TenantId>,
    pub timeline_id: Option<TimelineId>,
-
-    // Pageserver information can be passed in two different ways:
-    // 1. Here
-    // 2. in cluster.settings. This is legacy, we are switching to method 1.
-    pub pageserver_connection_info: Option<PageserverConnectionInfo>,
+    pub pageserver_connstring: Option<String>,

    // More neon ids that we expose to the compute_ctl
    // and to postgres as neon extension GUCs.
@@ -218,20 +214,6 @@ pub enum ComputeFeature {
    UnknownFeature,
 }

-/// Feature flag to signal `compute_ctl` to enable certain experimental functionality.
-#[derive(Clone, Debug, Default, Deserialize, Serialize, Eq, PartialEq)]
-pub struct PageserverConnectionInfo {
-    pub shards: HashMap<u32, PageserverShardConnectionInfo>,
-
-    pub prefer_grpc: bool,
-}
-
-#[derive(Clone, Debug, Default, Deserialize, Serialize, Eq, PartialEq)]
-pub struct PageserverShardConnectionInfo {
-    pub libpq_url: Option<String>,
-    pub grpc_url: Option<String>,
-}
-
 #[derive(Clone, Debug, Default, Deserialize, Serialize)]
 pub struct RemoteExtSpec {
    pub public_extensions: Option<Vec<String>>,
@@ -349,12 +331,6 @@ impl ComputeMode {
    }
 }

-impl Display for ComputeMode {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.write_str(self.to_type_str())
-    }
-}
-
 /// Log level for audit logging
 #[derive(Clone, Debug, Default, Eq, PartialEq, Deserialize, Serialize)]
 pub enum ComputeAudit {
--- a/libs/neon-shmem/Cargo.toml
+++ b/libs/neon-shmem/Cargo.toml
@@ -6,27 +6,8 @@ license.workspace = true

 [dependencies]
 thiserror.workspace = true
-nix.workspace = true
+nix.workspace=true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
-rustc-hash = { version = "2.1.1" }
-rand = "0.9.1"
-libc.workspace = true
-lock_api = "0.4.13"
-
-[dev-dependencies]
-criterion = { workspace = true, features = ["html_reports"] }
-rand_distr = "0.5.1"
-xxhash-rust = { version = "0.8.15", features = ["xxh3"] }
-ahash.workspace = true
-twox-hash = { version = "2.1.1" }
-seahash = "4.1.0"
-hashbrown = { git = "https://github.com/quantumish/hashbrown.git", rev = "6610e6d" }
-foldhash = "0.1.5"
-

 [target.'cfg(target_os = "macos")'.dependencies]
 tempfile = "3.14.0"
-
-[[bench]]
-name = "hmap_resize"
-harness = false
--- a/libs/neon-shmem/benches/hmap_resize.rs
+++ b/libs/neon-shmem/benches/hmap_resize.rs
@@ -1,330 +0,0 @@
-use criterion::{BatchSize, BenchmarkId, Criterion, criterion_group, criterion_main};
-use neon_shmem::hash::HashMapAccess;
-use neon_shmem::hash::HashMapInit;
-use neon_shmem::hash::entry::Entry;
-use rand::distr::{Distribution, StandardUniform};
-use rand::prelude::*;
-use std::default::Default;
-use std::hash::BuildHasher;
-
-// Taken from bindings to C code
-
-#[derive(Clone, Debug, Hash, Eq, PartialEq)]
-#[repr(C)]
-pub struct FileCacheKey {
-    pub _spc_id: u32,
-    pub _db_id: u32,
-    pub _rel_number: u32,
-    pub _fork_num: u32,
-    pub _block_num: u32,
-}
-
-impl Distribution<FileCacheKey> for StandardUniform {
-    // questionable, but doesn't need to be good randomness
-    fn sample<R: Rng + ?Sized>(&self, rng: &mut R) -> FileCacheKey {
-        FileCacheKey {
-            _spc_id: rng.random(),
-            _db_id: rng.random(),
-            _rel_number: rng.random(),
-            _fork_num: rng.random(),
-            _block_num: rng.random(),
-        }
-    }
-}
-
-#[derive(Clone, Debug)]
-#[repr(C)]
-pub struct FileCacheEntry {
-    pub _offset: u32,
-    pub _access_count: u32,
-    pub _prev: *mut FileCacheEntry,
-    pub _next: *mut FileCacheEntry,
-    pub _state: [u32; 8],
-}
-
-impl FileCacheEntry {
-    fn dummy() -> Self {
-        Self {
-            _offset: 0,
-            _access_count: 0,
-            _prev: std::ptr::null_mut(),
-            _next: std::ptr::null_mut(),
-            _state: [0; 8],
-        }
-    }
-}
-
-// Utilities for applying operations.
-
-#[derive(Clone, Debug)]
-struct TestOp<K, V>(K, Option<V>);
-
-fn apply_op<K: Clone + std::hash::Hash + Eq, V, S: std::hash::BuildHasher>(
-    op: TestOp<K, V>,
-    map: &mut HashMapAccess<K, V, S>,
-) {
-    let entry = map.entry(op.0);
-
-    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,
-        },
-    };
-}
-
-// Hash utilities
-
-struct SeaRandomState {
-    k1: u64,
-    k2: u64,
-    k3: u64,
-    k4: u64,
-}
-
-impl std::hash::BuildHasher for SeaRandomState {
-    type Hasher = seahash::SeaHasher;
-
-    fn build_hasher(&self) -> Self::Hasher {
-        seahash::SeaHasher::with_seeds(self.k1, self.k2, self.k3, self.k4)
-    }
-}
-
-impl SeaRandomState {
-    fn new() -> Self {
-        let mut rng = rand::rng();
-        Self {
-            k1: rng.random(),
-            k2: rng.random(),
-            k3: rng.random(),
-            k4: rng.random(),
-        }
-    }
-}
-
-fn small_benchs(c: &mut Criterion) {
-    let mut group = c.benchmark_group("Small maps");
-    group.sample_size(10);
-
-    group.bench_function("small_rehash", |b| {
-        let ideal_filled = 4_000_000;
-        let size = 5_000_000;
-        let mut writer = HashMapInit::new_resizeable(size, size * 2).attach_writer();
-        let mut rng = rand::rng();
-        while writer.get_num_buckets_in_use() < ideal_filled as usize {
-            let key: FileCacheKey = rng.random();
-            let val = FileCacheEntry::dummy();
-            apply_op(TestOp(key, Some(val)), &mut writer);
-        }
-        b.iter(|| writer.shuffle());
-    });
-
-    group.bench_function("small_rehash_xxhash", |b| {
-        let ideal_filled = 4_000_000;
-        let size = 5_000_000;
-        let mut writer = HashMapInit::new_resizeable(size, size * 2)
-            .with_hasher(twox_hash::xxhash64::RandomState::default())
-            .attach_writer();
-        let mut rng = rand::rng();
-        while writer.get_num_buckets_in_use() < ideal_filled as usize {
-            let key: FileCacheKey = rng.random();
-            let val = FileCacheEntry::dummy();
-            apply_op(TestOp(key, Some(val)), &mut writer);
-        }
-        b.iter(|| writer.shuffle());
-    });
-
-    group.bench_function("small_rehash_ahash", |b| {
-        let ideal_filled = 4_000_000;
-        let size = 5_000_000;
-        let mut writer = HashMapInit::new_resizeable(size, size * 2)
-            .with_hasher(ahash::RandomState::default())
-            .attach_writer();
-        let mut rng = rand::rng();
-        while writer.get_num_buckets_in_use() < ideal_filled as usize {
-            let key: FileCacheKey = rng.random();
-            let val = FileCacheEntry::dummy();
-            apply_op(TestOp(key, Some(val)), &mut writer);
-        }
-        b.iter(|| writer.shuffle());
-    });
-
-    group.bench_function("small_rehash_seahash", |b| {
-        let ideal_filled = 4_000_000;
-        let size = 5_000_000;
-        let mut writer = HashMapInit::new_resizeable(size, size * 2)
-            .with_hasher(SeaRandomState::new())
-            .attach_writer();
-        let mut rng = rand::rng();
-        while writer.get_num_buckets_in_use() < ideal_filled as usize {
-            let key: FileCacheKey = rng.random();
-            let val = FileCacheEntry::dummy();
-            apply_op(TestOp(key, Some(val)), &mut writer);
-        }
-        b.iter(|| writer.shuffle());
-    });
-
-    group.finish();
-}
-
-fn real_benchs(c: &mut Criterion) {
-    let mut group = c.benchmark_group("Realistic workloads");
-    group.sample_size(10);
-    group.bench_function("real_bulk_insert", |b| {
-        let size = 125_000_000;
-        let ideal_filled = 100_000_000;
-        let mut rng = rand::rng();
-        b.iter_batched(
-            || HashMapInit::new_resizeable(size, size * 2).attach_writer(),
-            |writer| {
-                for _ in 0..ideal_filled {
-                    let key: FileCacheKey = rng.random();
-                    let val = FileCacheEntry::dummy();
-                    let entry = writer.entry(key);
-                    std::hint::black_box(match entry {
-                        Entry::Occupied(mut e) => {
-                            e.insert(val);
-                        }
-                        Entry::Vacant(e) => {
-                            _ = e.insert(val).unwrap();
-                        }
-                    })
-                }
-            },
-            BatchSize::SmallInput,
-        )
-    });
-
-    group.bench_function("real_rehash", |b| {
-        let size = 125_000_000;
-        let ideal_filled = 100_000_000;
-        let mut writer = HashMapInit::new_resizeable(size, size).attach_writer();
-        let mut rng = rand::rng();
-        while writer.get_num_buckets_in_use() < ideal_filled {
-            let key: FileCacheKey = rng.random();
-            let val = FileCacheEntry::dummy();
-            apply_op(TestOp(key, Some(val)), &mut writer);
-        }
-        b.iter(|| writer.shuffle());
-    });
-
-    group.bench_function("real_rehash_hashbrown", |b| {
-        let size = 125_000_000;
-        let ideal_filled = 100_000_000;
-        let mut writer = hashbrown::raw::RawTable::new();
-        let mut rng = rand::rng();
-        let hasher = rustc_hash::FxBuildHasher::default();
-        unsafe {
-            writer
-                .resize(
-                    size,
-                    |(k, _)| hasher.hash_one(&k),
-                    hashbrown::raw::Fallibility::Infallible,
-                )
-                .unwrap();
-        }
-        while writer.len() < ideal_filled as usize {
-            let key: FileCacheKey = rng.random();
-            let val = FileCacheEntry::dummy();
-            writer.insert(hasher.hash_one(&key), (key, val), |(k, _)| {
-                hasher.hash_one(&k)
-            });
-        }
-        b.iter(|| unsafe {
-            writer.table.rehash_in_place(
-                &|table, index| {
-                    hasher.hash_one(
-                        &table
-                            .bucket::<(FileCacheKey, FileCacheEntry)>(index)
-                            .as_ref()
-                            .0,
-                    )
-                },
-                std::mem::size_of::<(FileCacheKey, FileCacheEntry)>(),
-                if std::mem::needs_drop::<(FileCacheKey, FileCacheEntry)>() {
-                    Some(|ptr| std::ptr::drop_in_place(ptr as *mut (FileCacheKey, FileCacheEntry)))
-                } else {
-                    None
-                },
-            )
-        });
-    });
-
-    for elems in [2, 4, 8, 16, 32, 64, 96, 112] {
-        group.bench_with_input(
-            BenchmarkId::new("real_rehash_varied", elems),
-            &elems,
-            |b, &size| {
-                let ideal_filled = size * 1_000_000;
-                let size = 125_000_000;
-                let mut writer = HashMapInit::new_resizeable(size, size).attach_writer();
-                let mut rng = rand::rng();
-                while writer.get_num_buckets_in_use() < ideal_filled as usize {
-                    let key: FileCacheKey = rng.random();
-                    let val = FileCacheEntry::dummy();
-                    apply_op(TestOp(key, Some(val)), &mut writer);
-                }
-                b.iter(|| writer.shuffle());
-            },
-        );
-        group.bench_with_input(
-            BenchmarkId::new("real_rehash_varied_hashbrown", elems),
-            &elems,
-            |b, &size| {
-                let ideal_filled = size * 1_000_000;
-                let size = 125_000_000;
-                let mut writer = hashbrown::raw::RawTable::new();
-                let mut rng = rand::rng();
-                let hasher = rustc_hash::FxBuildHasher::default();
-                unsafe {
-                    writer
-                        .resize(
-                            size,
-                            |(k, _)| hasher.hash_one(&k),
-                            hashbrown::raw::Fallibility::Infallible,
-                        )
-                        .unwrap();
-                }
-                while writer.len() < ideal_filled as usize {
-                    let key: FileCacheKey = rng.random();
-                    let val = FileCacheEntry::dummy();
-                    writer.insert(hasher.hash_one(&key), (key, val), |(k, _)| {
-                        hasher.hash_one(&k)
-                    });
-                }
-                b.iter(|| unsafe {
-                    writer.table.rehash_in_place(
-                        &|table, index| {
-                            hasher.hash_one(
-                                &table
-                                    .bucket::<(FileCacheKey, FileCacheEntry)>(index)
-                                    .as_ref()
-                                    .0,
-                            )
-                        },
-                        std::mem::size_of::<(FileCacheKey, FileCacheEntry)>(),
-                        if std::mem::needs_drop::<(FileCacheKey, FileCacheEntry)>() {
-                            Some(|ptr| {
-                                std::ptr::drop_in_place(ptr as *mut (FileCacheKey, FileCacheEntry))
-                            })
-                        } else {
-                            None
-                        },
-                    )
-                });
-            },
-        );
-    }
-
-    group.finish();
-}
-
-criterion_group!(benches, small_benchs, real_benchs);
-criterion_main!(benches);
--- a/libs/neon-shmem/src/hash.rs
+++ b/libs/neon-shmem/src/hash.rs
@@ -1,598 +0,0 @@
-//! 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.
-
-use std::fmt::Debug;
-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};
-
-/// 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,
-}
-
-impl<'a, K, V, S> Debug for HashMapInit<'a, K, V, S>
-where
-    K: Debug,
-    V: Debug,
-{
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.debug_struct("HashMapInit")
-            .field("shmem_handle", &self.shmem_handle)
-            .field("shared_ptr", &self.shared_ptr)
-            .field("shared_size", &self.shared_size)
-            // .field("hasher", &self.hasher)
-            .field("num_buckets", &self.num_buckets)
-            .finish()
-    }
-}
-
-/// 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, V, S> Debug for HashMapAccess<'a, K, V, S>
-where
-    K: Debug,
-    V: Debug,
-{
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.debug_struct("HashMapAccess")
-            .field("shmem_handle", &self.shmem_handle)
-            .field("shared_ptr", &self.shared_ptr)
-            // .field("hasher", &self.hasher)
-            .finish()
-    }
-}
-
-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);
-        let lock = RwLock::from_raw(PthreadRwLock::new(raw_lock_ptr.cast()), hashmap);
-        unsafe {
-            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`].
-    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.
-    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.
-    /// - Calling this function on a map with `shrink_mode` set to [`HashMapShrinkMode::Remap`] and
-    ///   there are more buckets in use than the value returned by [`HashMapAccess::shrink_goal`].
-    ///
-    /// # Errors
-    /// Returns an [`shmem::Error`] if any errors occur resizing the memory region.
-    pub fn finish_shrink(&self) -> Result<(), shmem::Error> {
-        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 let Some((k, v)) = map.buckets[i].inner.take() {
-                // alloc_bucket increases count, so need to decrease since we're just moving
-                map.buckets_in_use -= 1;
-                map.alloc_bucket(k, v).unwrap();
-            }
-        }
-
-        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);
-        shmem_handle.set_size(size_bytes)?;
-        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(())
-    }
-}
--- a/libs/neon-shmem/src/hash/core.rs
+++ b/libs/neon-shmem/src/hash/core.rs
@@ -1,208 +0,0 @@
-//! Simple hash table with chaining.
-
-use std::fmt::Debug;
-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)>,
-}
-
-impl<K, V> Debug for Bucket<K, V>
-where
-    K: Debug,
-    V: Debug,
-{
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.debug_struct("Bucket")
-            .field("next", &self.next)
-            .field("inner", &self.inner)
-            .finish()
-    }
-}
-
-/// 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,
-    // pub(crate) lock: libc::pthread_mutex_t,
-    // Unclear what the purpose of this is.
-    pub(crate) _user_list_head: u32,
-}
-
-impl<'a, K, V> Debug for CoreHashMap<'a, K, V>
-where
-    K: Debug,
-    V: Debug,
-{
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.debug_struct("CoreHashMap")
-            .field("dictionary", &self.dictionary)
-            .field("buckets", &self.buckets)
-            .field("free_head", &self.free_head)
-            .field("alloc_limit", &self.alloc_limit)
-            .field("buckets_in_use", &self.buckets_in_use)
-            .finish()
-    }
-}
-
-/// 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,
-            _user_list_head: INVALID_POS,
-            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)
-    }
-}
--- a/libs/neon-shmem/src/hash/entry.rs
+++ b/libs/neon-shmem/src/hash/entry.rs
@@ -1,138 +0,0 @@
-//! 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.
-    ///
-    /// # Panics
-    /// Panics if the `prev_pos` field is equal to [`PrevPos::Unknown`]. In practice, this means
-    /// the entry was obtained via calling something like [`CoreHashMap::entry_at_bucket`].
-    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) => {
-                // println!("we think prev of {} is {bucket_pos}", self.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)?;
-        if pos == INVALID_POS {
-            return Err(FullError());
-        }
-        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
-        }))
-    }
-}
--- a/libs/neon-shmem/src/hash/tests.rs
+++ b/libs/neon-shmem/src/hash/tests.rs
@@ -1,429 +0,0 @@
-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>,
-    to: u32,
-) {
-    assert!(writer.shrink_goal().is_none());
-    writer.begin_shrink(to);
-    assert_eq!(writer.shrink_goal(), Some(to as usize));
-    while writer.get_num_buckets_in_use() > to as usize {
-        let (k, _) = shadow.pop_first().unwrap();
-        let entry = writer.entry(k);
-        if let Entry::Occupied(e) = entry {
-            e.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, 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, 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, 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());
-            let pos = e.bucket_pos as usize;
-            pos
-        }
-        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);
-}
--- a/libs/neon-shmem/src/lib.rs
+++ b/libs/neon-shmem/src/lib.rs
@@ -1,5 +1,418 @@
 //! Shared memory utilities for neon communicator

-pub mod hash;
-pub mod shmem;
-pub mod sync;
+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();
+        }
+    }
+}
--- a/libs/neon-shmem/src/shmem.rs
+++ b/libs/neon-shmem/src/shmem.rs
@@ -1,411 +0,0 @@
-//! 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.
-#[derive(Debug)]
-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.
-#[derive(Debug)]
-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();
-        }
-    }
-}
--- a/libs/neon-shmem/src/sync.rs
+++ b/libs/neon-shmem/src/sync.rs
@@ -1,104 +0,0 @@
-//! Simple utilities akin to what's in [`std::sync`] but designed to work with shared memory.
-
-use std::mem::MaybeUninit;
-use std::ptr::NonNull;
-
-use nix::errno::Errno;
-
-pub type RwLock<T> = lock_api::RwLock<PthreadRwLock, T>;
-pub(crate) type RwLockReadGuard<'a, T> = lock_api::RwLockReadGuard<'a, PthreadRwLock, T>;
-pub type RwLockWriteGuard<'a, T> = lock_api::RwLockWriteGuard<'a, PthreadRwLock, T>;
-pub type ValueReadGuard<'a, T> = lock_api::MappedRwLockReadGuard<'a, PthreadRwLock, T>;
-pub type ValueWriteGuard<'a, T> = lock_api::MappedRwLockWriteGuard<'a, PthreadRwLock, T>;
-
-/// Shared memory read-write lock.
-pub struct PthreadRwLock(Option<NonNull<libc::pthread_rwlock_t>>);
-
-impl PthreadRwLock {
-    pub fn new(lock: *mut libc::pthread_rwlock_t) -> Self {
-        unsafe {
-            let mut attrs = MaybeUninit::uninit();
-            // Ignoring return value here - only possible error is OOM.
-            libc::pthread_rwlockattr_init(attrs.as_mut_ptr());
-            libc::pthread_rwlockattr_setpshared(attrs.as_mut_ptr(), libc::PTHREAD_PROCESS_SHARED);
-            // TODO(quantumish): worth making this function return Result?
-            libc::pthread_rwlock_init(lock, attrs.as_mut_ptr());
-            // Safety: POSIX specifies that "any function affecting the attributes
-            // object (including destruction) shall not affect any previously
-            // initialized read-write locks".
-            libc::pthread_rwlockattr_destroy(attrs.as_mut_ptr());
-            Self(Some(NonNull::new_unchecked(lock)))
-        }
-    }
-
-    fn inner(&self) -> NonNull<libc::pthread_rwlock_t> {
-        match self.0 {
-            None => {
-                panic!("PthreadRwLock constructed badly - something likely used RawMutex::INIT")
-            }
-            Some(x) => x,
-        }
-    }
-}
-
-unsafe impl lock_api::RawRwLock for PthreadRwLock {
-    type GuardMarker = lock_api::GuardSend;
-    const INIT: Self = Self(None);
-
-    fn lock_shared(&self) {
-        unsafe {
-            let res = libc::pthread_rwlock_rdlock(self.inner().as_ptr());
-            if res != 0 {
-                panic!("rdlock failed with {}", Errno::from_raw(res));
-            }
-        }
-    }
-
-    fn try_lock_shared(&self) -> bool {
-        unsafe {
-            let res = libc::pthread_rwlock_tryrdlock(self.inner().as_ptr());
-            match res {
-                0 => true,
-                libc::EAGAIN => false,
-                _ => panic!("try_rdlock failed with {}", Errno::from_raw(res)),
-            }
-        }
-    }
-
-    fn lock_exclusive(&self) {
-        unsafe {
-            let res = libc::pthread_rwlock_wrlock(self.inner().as_ptr());
-            if res != 0 {
-                panic!("wrlock failed with {}", Errno::from_raw(res));
-            }
-        }
-    }
-
-    fn try_lock_exclusive(&self) -> bool {
-        unsafe {
-            let res = libc::pthread_rwlock_trywrlock(self.inner().as_ptr());
-            match res {
-                0 => true,
-                libc::EAGAIN => false,
-                _ => panic!("try_wrlock failed with {}", Errno::from_raw(res)),
-            }
-        }
-    }
-
-    unsafe fn unlock_exclusive(&self) {
-        unsafe {
-            let res = libc::pthread_rwlock_unlock(self.inner().as_ptr());
-            if res != 0 {
-                panic!("unlock failed with {}", Errno::from_raw(res));
-            }
-        }
-    }
-    unsafe fn unlock_shared(&self) {
-        unsafe {
-            let res = libc::pthread_rwlock_unlock(self.inner().as_ptr());
-            if res != 0 {
-                panic!("unlock failed with {}", Errno::from_raw(res));
-            }
-        }
-    }
-}
--- a/libs/neonart/Cargo.toml
+++ b/libs/neonart/Cargo.toml
@@ -1,14 +0,0 @@
-[package]
-name = "neonart"
-version = "0.1.0"
-edition.workspace = true
-license.workspace = true
-
-[dependencies]
-crossbeam-utils.workspace = true
-spin.workspace = true
-tracing.workspace = true
-
-[dev-dependencies]
-rand = "0.9.1"
-rand_distr = "0.5.1"
--- a/libs/neonart/src/algorithm.rs
+++ b/libs/neonart/src/algorithm.rs
@@ -1,599 +0,0 @@
-mod lock_and_version;
-pub(crate) mod node_ptr;
-mod node_ref;
-
-use std::vec::Vec;
-
-use crate::algorithm::lock_and_version::ConcurrentUpdateError;
-use crate::algorithm::node_ptr::MAX_PREFIX_LEN;
-use crate::algorithm::node_ref::{NewNodeRef, NodeRef, ReadLockedNodeRef, WriteLockedNodeRef};
-use crate::allocator::OutOfMemoryError;
-
-use crate::TreeWriteGuard;
-use crate::UpdateAction;
-use crate::allocator::ArtAllocator;
-use crate::epoch::EpochPin;
-use crate::{Key, Value};
-
-pub(crate) type RootPtr<V> = node_ptr::NodePtr<V>;
-
-#[derive(Debug)]
-pub enum ArtError {
-    ConcurrentUpdate, // need to retry
-    OutOfMemory,
-}
-
-impl From<ConcurrentUpdateError> for ArtError {
-    fn from(_: ConcurrentUpdateError) -> ArtError {
-        ArtError::ConcurrentUpdate
-    }
-}
-
-impl From<OutOfMemoryError> for ArtError {
-    fn from(_: OutOfMemoryError) -> ArtError {
-        ArtError::OutOfMemory
-    }
-}
-
-pub fn new_root<V: Value>(
-    allocator: &impl ArtAllocator<V>,
-) -> Result<RootPtr<V>, OutOfMemoryError> {
-    node_ptr::new_root(allocator)
-}
-
-pub(crate) fn search<'e, K: Key, V: Value>(
-    key: &K,
-    root: RootPtr<V>,
-    epoch_pin: &'e EpochPin,
-) -> Option<&'e V> {
-    loop {
-        let root_ref = NodeRef::from_root_ptr(root);
-        if let Ok(result) = lookup_recurse(key.as_bytes(), root_ref, None, epoch_pin) {
-            break result;
-        }
-        // retry
-    }
-}
-
-pub(crate) fn iter_next<'e, V: Value>(
-    key: &[u8],
-    root: RootPtr<V>,
-    epoch_pin: &'e EpochPin,
-) -> Option<(Vec<u8>, &'e V)> {
-    loop {
-        let mut path = Vec::new();
-        let root_ref = NodeRef::from_root_ptr(root);
-
-        match next_recurse(key, &mut path, root_ref, epoch_pin) {
-            Ok(Some(v)) => {
-                assert_eq!(path.len(), key.len());
-                break Some((path, v));
-            }
-            Ok(None) => break None,
-            Err(ConcurrentUpdateError()) => {
-                // retry
-                continue;
-            }
-        }
-    }
-}
-
-pub(crate) fn update_fn<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>, F>(
-    key: &K,
-    value_fn: F,
-    root: RootPtr<V>,
-    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
-) -> Result<(), OutOfMemoryError>
-where
-    F: FnOnce(Option<&V>) -> UpdateAction<V>,
-{
-    let value_fn_cell = std::cell::Cell::new(Some(value_fn));
-    loop {
-        let root_ref = NodeRef::from_root_ptr(root);
-        let this_value_fn = |arg: Option<&V>| value_fn_cell.take().unwrap()(arg);
-        let key_bytes = key.as_bytes();
-
-        match update_recurse(
-            key_bytes,
-            this_value_fn,
-            root_ref,
-            None,
-            None,
-            guard,
-            0,
-            key_bytes,
-        ) {
-            Ok(()) => break Ok(()),
-            Err(ArtError::ConcurrentUpdate) => {
-                continue; // retry
-            }
-            Err(ArtError::OutOfMemory) => break Err(OutOfMemoryError()),
-        }
-    }
-}
-
-// Error means you must retry.
-//
-// This corresponds to the 'lookupOpt' function in the paper
-#[allow(clippy::only_used_in_recursion)]
-fn lookup_recurse<'e, V: Value>(
-    key: &[u8],
-    node: NodeRef<'e, V>,
-    parent: Option<ReadLockedNodeRef<V>>,
-    epoch_pin: &'e EpochPin,
-) -> Result<Option<&'e V>, ConcurrentUpdateError> {
-    let rnode = node.read_lock_or_restart()?;
-    if let Some(parent) = parent {
-        parent.read_unlock_or_restart()?;
-    }
-
-    // check if the prefix matches, may increment level
-    let prefix_len = if let Some(prefix_len) = rnode.prefix_matches(key) {
-        prefix_len
-    } else {
-        rnode.read_unlock_or_restart()?;
-        return Ok(None);
-    };
-
-    if rnode.is_leaf() {
-        assert_eq!(key.len(), prefix_len);
-        let vptr = rnode.get_leaf_value_ptr()?;
-        // safety: It's OK to return a ref of the pointer because we checked the version
-        // and the lifetime of 'epoch_pin' enforces that the reference is only accessible
-        // as long as the epoch is pinned.
-        let v = unsafe { vptr.as_ref().unwrap() };
-        return Ok(Some(v));
-    }
-
-    let key = &key[prefix_len..];
-
-    // find child (or leaf value)
-    let next_node = rnode.find_child_or_restart(key[0])?;
-
-    match next_node {
-        None => Ok(None), // key not found
-        Some(child) => lookup_recurse(&key[1..], child, Some(rnode), epoch_pin),
-    }
-}
-
-#[allow(clippy::only_used_in_recursion)]
-fn next_recurse<'e, V: Value>(
-    min_key: &[u8],
-    path: &mut Vec<u8>,
-    node: NodeRef<'e, V>,
-    epoch_pin: &'e EpochPin,
-) -> Result<Option<&'e V>, ConcurrentUpdateError> {
-    let rnode = node.read_lock_or_restart()?;
-    let prefix = rnode.get_prefix();
-    if !prefix.is_empty() {
-        path.extend_from_slice(prefix);
-    }
-
-    use std::cmp::Ordering;
-    let comparison = path.as_slice().cmp(&min_key[0..path.len()]);
-    if comparison == Ordering::Less {
-        rnode.read_unlock_or_restart()?;
-        return Ok(None);
-    }
-
-    if rnode.is_leaf() {
-        assert_eq!(path.len(), min_key.len());
-        let vptr = rnode.get_leaf_value_ptr()?;
-        // safety: It's OK to return a ref of the pointer because we checked the version
-        // and the lifetime of 'epoch_pin' enforces that the reference is only accessible
-        // as long as the epoch is pinned.
-        let v = unsafe { vptr.as_ref().unwrap() };
-        return Ok(Some(v));
-    }
-
-    let mut min_key_byte = match comparison {
-        Ordering::Less => unreachable!(), // checked this above already
-        Ordering::Equal => min_key[path.len()],
-        Ordering::Greater => 0,
-    };
-
-    loop {
-        match rnode.find_next_child_or_restart(min_key_byte)? {
-            None => {
-                return Ok(None);
-            }
-            Some((key_byte, child_ref)) => {
-                let path_len = path.len();
-                path.push(key_byte);
-                let result = next_recurse(min_key, path, child_ref, epoch_pin)?;
-                if result.is_some() {
-                    return Ok(result);
-                }
-                if key_byte == u8::MAX {
-                    return Ok(None);
-                }
-                path.truncate(path_len);
-                min_key_byte = key_byte + 1;
-            }
-        }
-    }
-}
-
-// This corresponds to the 'insertOpt' function in the paper
-#[allow(clippy::only_used_in_recursion)]
-#[allow(clippy::too_many_arguments)]
-pub(crate) fn update_recurse<'e, K: Key, V: Value, A: ArtAllocator<V>, F>(
-    key: &[u8],
-    value_fn: F,
-    node: NodeRef<'e, V>,
-    rparent: Option<(ReadLockedNodeRef<V>, u8)>,
-    rgrandparent: Option<(ReadLockedNodeRef<V>, u8)>,
-    guard: &'_ mut TreeWriteGuard<'e, K, V, A>,
-    level: usize,
-    orig_key: &[u8],
-) -> Result<(), ArtError>
-where
-    F: FnOnce(Option<&V>) -> UpdateAction<V>,
-{
-    let rnode = node.read_lock_or_restart()?;
-
-    let prefix_match_len = rnode.prefix_matches(key);
-    if prefix_match_len.is_none() {
-        let (rparent, parent_key) = rparent.expect("direct children of the root have no prefix");
-        let mut wparent = rparent.upgrade_to_write_lock_or_restart()?;
-        let mut wnode = rnode.upgrade_to_write_lock_or_restart()?;
-
-        match value_fn(None) {
-            UpdateAction::Nothing => {}
-            UpdateAction::Insert(new_value) => {
-                insert_split_prefix(key, new_value, &mut wnode, &mut wparent, parent_key, guard)?;
-            }
-            UpdateAction::Remove => {
-                panic!("unexpected Remove action on insertion");
-            }
-        }
-        wnode.write_unlock();
-        wparent.write_unlock();
-        return Ok(());
-    }
-    let prefix_match_len = prefix_match_len.unwrap();
-    let key = &key[prefix_match_len..];
-    let level = level + prefix_match_len;
-
-    if rnode.is_leaf() {
-        assert_eq!(key.len(), 0);
-        let (rparent, parent_key) = rparent.expect("root cannot be leaf");
-        let mut wparent = rparent.upgrade_to_write_lock_or_restart()?;
-        let mut wnode = rnode.upgrade_to_write_lock_or_restart()?;
-
-        // safety: Now that we have acquired the write lock, we have exclusive access to the
-        // value. XXX: There might be concurrent reads though?
-        let value_mut = wnode.get_leaf_value_mut();
-
-        match value_fn(Some(value_mut)) {
-            UpdateAction::Nothing => {
-                wparent.write_unlock();
-                wnode.write_unlock();
-            }
-            UpdateAction::Insert(_) => panic!("cannot insert over existing value"),
-            UpdateAction::Remove => {
-                guard.remember_obsolete_node(wnode.as_ptr());
-                wparent.delete_child(parent_key);
-                wnode.write_unlock_obsolete();
-
-                if let Some(rgrandparent) = rgrandparent {
-                    // FIXME: Ignore concurrency error. It doesn't lead to
-                    // corruption, but it means we might leak something. Until
-                    // another update cleans it up.
-                    let _ = cleanup_parent(wparent, rgrandparent, guard);
-                }
-            }
-        }
-
-        return Ok(());
-    }
-
-    let next_node = rnode.find_child_or_restart(key[0])?;
-
-    if next_node.is_none() {
-        if rnode.is_full() {
-            let (rparent, parent_key) = rparent.expect("root node cannot become full");
-            let mut wparent = rparent.upgrade_to_write_lock_or_restart()?;
-            let wnode = rnode.upgrade_to_write_lock_or_restart()?;
-
-            match value_fn(None) {
-                UpdateAction::Nothing => {
-                    wnode.write_unlock();
-                    wparent.write_unlock();
-                }
-                UpdateAction::Insert(new_value) => {
-                    insert_and_grow(key, new_value, wnode, &mut wparent, parent_key, guard)?;
-                    wparent.write_unlock();
-                }
-                UpdateAction::Remove => {
-                    panic!("unexpected Remove action on insertion");
-                }
-            };
-        } else {
-            let mut wnode = rnode.upgrade_to_write_lock_or_restart()?;
-            if let Some((rparent, _)) = rparent {
-                rparent.read_unlock_or_restart()?;
-            }
-            match value_fn(None) {
-                UpdateAction::Nothing => {}
-                UpdateAction::Insert(new_value) => {
-                    insert_to_node(&mut wnode, key, new_value, guard)?;
-                }
-                UpdateAction::Remove => {
-                    panic!("unexpected Remove action on insertion");
-                }
-            };
-            wnode.write_unlock();
-        }
-        Ok(())
-    } else {
-        let next_child = next_node.unwrap(); // checked above it's not None
-        if let Some((ref rparent, _)) = rparent {
-            rparent.check_or_restart()?;
-        }
-
-        // recurse to next level
-        update_recurse(
-            &key[1..],
-            value_fn,
-            next_child,
-            Some((rnode, key[0])),
-            rparent,
-            guard,
-            level + 1,
-            orig_key,
-        )
-    }
-}
-
-#[derive(Clone)]
-enum PathElement {
-    Prefix(Vec<u8>),
-    KeyByte(u8),
-}
-
-impl std::fmt::Debug for PathElement {
-    fn fmt(&self, fmt: &mut std::fmt::Formatter<'_>) -> Result<(), std::fmt::Error> {
-        match self {
-            PathElement::Prefix(prefix) => write!(fmt, "{prefix:?}"),
-            PathElement::KeyByte(key_byte) => write!(fmt, "{key_byte}"),
-        }
-    }
-}
-
-pub(crate) fn dump_tree<V: Value + std::fmt::Debug>(
-    root: RootPtr<V>,
-    epoch_pin: &'_ EpochPin,
-    dst: &mut dyn std::io::Write,
-) {
-    let root_ref = NodeRef::from_root_ptr(root);
-
-    let _ = dump_recurse(&[], root_ref, epoch_pin, 0, dst);
-}
-
-// TODO: return an Err if writeln!() returns error, instead of unwrapping
-#[allow(clippy::only_used_in_recursion)]
-fn dump_recurse<'e, V: Value + std::fmt::Debug>(
-    path: &[PathElement],
-    node: NodeRef<'e, V>,
-    epoch_pin: &'e EpochPin,
-    level: usize,
-    dst: &mut dyn std::io::Write,
-) -> Result<(), ConcurrentUpdateError> {
-    let indent = str::repeat(" ", level);
-
-    let rnode = node.read_lock_or_restart()?;
-    let mut path = Vec::from(path);
-    let prefix = rnode.get_prefix();
-    if !prefix.is_empty() {
-        path.push(PathElement::Prefix(Vec::from(prefix)));
-    }
-
-    if rnode.is_leaf() {
-        let vptr = rnode.get_leaf_value_ptr()?;
-        // safety: It's OK to return a ref of the pointer because we checked the version
-        // and the lifetime of 'epoch_pin' enforces that the reference is only accessible
-        // as long as the epoch is pinned.
-        let val = unsafe { vptr.as_ref().unwrap() };
-        writeln!(dst, "{indent} {path:?}: {val:?}").unwrap();
-        return Ok(());
-    }
-
-    for key_byte in 0..=u8::MAX {
-        match rnode.find_child_or_restart(key_byte)? {
-            None => continue,
-            Some(child_ref) => {
-                let rchild = child_ref.read_lock_or_restart()?;
-                writeln!(
-                    dst,
-                    "{} {:?}, {}: prefix {:?}",
-                    indent,
-                    &path,
-                    key_byte,
-                    rchild.get_prefix()
-                )
-                .unwrap();
-
-                let mut child_path = path.clone();
-                child_path.push(PathElement::KeyByte(key_byte));
-
-                dump_recurse(&child_path, child_ref, epoch_pin, level + 1, dst)?;
-            }
-        }
-    }
-
-    Ok(())
-}
-
-///```text
-///        [fooba]r -> value
-///
-/// [foo]b -> [a]r  -> value
-///      e -> [ls]e -> value
-///```
-fn insert_split_prefix<K: Key, V: Value, A: ArtAllocator<V>>(
-    key: &[u8],
-    value: V,
-    node: &mut WriteLockedNodeRef<V>,
-    parent: &mut WriteLockedNodeRef<V>,
-    parent_key: u8,
-    guard: &'_ TreeWriteGuard<K, V, A>,
-) -> Result<(), OutOfMemoryError> {
-    let old_node = node;
-    let old_prefix = old_node.get_prefix();
-    let common_prefix_len = common_prefix(key, old_prefix);
-
-    // Allocate a node for the new value.
-    let new_value_node = allocate_node_for_value(
-        &key[common_prefix_len + 1..],
-        value,
-        guard.tree_writer.allocator,
-    )?;
-
-    // Allocate a new internal node with the common prefix
-    // FIXME: deallocate 'new_value_node' on OOM
-    let mut prefix_node =
-        node_ref::new_internal(&key[..common_prefix_len], guard.tree_writer.allocator)?;
-
-    // Add the old node and the new nodes to the new internal node
-    prefix_node.insert_old_child(old_prefix[common_prefix_len], old_node);
-    prefix_node.insert_new_child(key[common_prefix_len], new_value_node);
-
-    // Modify the prefix of the old child in place
-    old_node.truncate_prefix(old_prefix.len() - common_prefix_len - 1);
-
-    // replace the pointer in the parent
-    parent.replace_child(parent_key, prefix_node.into_ptr());
-
-    Ok(())
-}
-
-fn insert_to_node<K: Key, V: Value, A: ArtAllocator<V>>(
-    wnode: &mut WriteLockedNodeRef<V>,
-    key: &[u8],
-    value: V,
-    guard: &'_ TreeWriteGuard<K, V, A>,
-) -> Result<(), OutOfMemoryError> {
-    let value_child = allocate_node_for_value(&key[1..], value, guard.tree_writer.allocator)?;
-    wnode.insert_child(key[0], value_child.into_ptr());
-    Ok(())
-}
-
-// On entry: 'parent' and 'node' are locked
-fn insert_and_grow<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>>(
-    key: &[u8],
-    value: V,
-    wnode: WriteLockedNodeRef<V>,
-    parent: &mut WriteLockedNodeRef<V>,
-    parent_key_byte: u8,
-    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
-) -> Result<(), ArtError> {
-    let mut bigger_node = wnode.grow(guard.tree_writer.allocator)?;
-
-    // FIXME: deallocate 'bigger_node' on OOM
-    let value_child = allocate_node_for_value(&key[1..], value, guard.tree_writer.allocator)?;
-    bigger_node.insert_new_child(key[0], value_child);
-
-    // Replace the pointer in the parent
-    parent.replace_child(parent_key_byte, bigger_node.into_ptr());
-
-    guard.remember_obsolete_node(wnode.as_ptr());
-    wnode.write_unlock_obsolete();
-
-    Ok(())
-}
-
-fn cleanup_parent<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>>(
-    wparent: WriteLockedNodeRef<V>,
-    rgrandparent: (ReadLockedNodeRef<V>, u8),
-    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
-) -> Result<(), ArtError> {
-    let (rgrandparent, grandparent_key_byte) = rgrandparent;
-
-    // If the parent becomes completely empty after the deletion, remove the parent from the
-    // grandparent. (This case is possible because we reserve only 8 bytes for the prefix.)
-    // TODO: not implemented.
-
-    // If the parent has only one child, replace the parent with the remaining child. (This is not
-    // possible if the child's prefix field cannot absorb the parent's)
-    if wparent.num_children() == 1 {
-        // Try to lock the remaining child. This can fail if the child is updated
-        // concurrently.
-        let (key_byte, remaining_child) = wparent.find_remaining_child();
-
-        let mut wremaining_child = remaining_child.write_lock_or_restart()?;
-
-        if 1 + wremaining_child.get_prefix().len() + wparent.get_prefix().len() <= MAX_PREFIX_LEN {
-            let mut wgrandparent = rgrandparent.upgrade_to_write_lock_or_restart()?;
-
-            // Ok, we have locked the leaf, the parent, the grandparent, and the parent's only
-            // remaining leaf. Proceed with the updates.
-
-            // Update the prefix on the remaining leaf
-            wremaining_child.prepend_prefix(wparent.get_prefix(), key_byte);
-
-            // Replace the pointer in the grandparent to point directly to the remaining leaf
-            wgrandparent.replace_child(grandparent_key_byte, wremaining_child.as_ptr());
-
-            // Mark the parent as deleted.
-            guard.remember_obsolete_node(wparent.as_ptr());
-            wparent.write_unlock_obsolete();
-            return Ok(());
-        }
-    }
-
-    // If the parent's children would fit on a smaller node type after the deletion, replace it with
-    // a smaller node.
-    if wparent.can_shrink() {
-        let mut wgrandparent = rgrandparent.upgrade_to_write_lock_or_restart()?;
-        let smaller_node = wparent.shrink(guard.tree_writer.allocator)?;
-
-        // Replace the pointer in the grandparent
-        wgrandparent.replace_child(grandparent_key_byte, smaller_node.into_ptr());
-
-        guard.remember_obsolete_node(wparent.as_ptr());
-        wparent.write_unlock_obsolete();
-        return Ok(());
-    }
-
-    // nothing to do
-    wparent.write_unlock();
-    Ok(())
-}
-
-// Allocate a new leaf node to hold 'value'. If the key is long, we
-// may need to allocate new internal nodes to hold it too
-fn allocate_node_for_value<'a, V: Value, A: ArtAllocator<V>>(
-    key: &[u8],
-    value: V,
-    allocator: &'a A,
-) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError> {
-    let mut prefix_off = key.len().saturating_sub(MAX_PREFIX_LEN);
-
-    let leaf_node = node_ref::new_leaf(&key[prefix_off..key.len()], value, allocator)?;
-
-    let mut node = leaf_node;
-    while prefix_off > 0 {
-        // Need another internal node
-        let remain_prefix = &key[0..prefix_off];
-
-        prefix_off = remain_prefix.len().saturating_sub(MAX_PREFIX_LEN + 1);
-        let mut internal_node = node_ref::new_internal(
-            &remain_prefix[prefix_off..remain_prefix.len() - 1],
-            allocator,
-        )?;
-        internal_node.insert_new_child(*remain_prefix.last().unwrap(), node);
-        node = internal_node;
-    }
-
-    Ok(node)
-}
-
-fn common_prefix(a: &[u8], b: &[u8]) -> usize {
-    for i in 0..MAX_PREFIX_LEN {
-        if a[i] != b[i] {
-            return i;
-        }
-    }
-    panic!("prefixes are equal");
-}
--- a/libs/neonart/src/algorithm/lock_and_version.rs
+++ b/libs/neonart/src/algorithm/lock_and_version.rs
@@ -1,117 +0,0 @@
-//! Each node in the tree has contains one atomic word that stores three things:
-//!
-//! Bit 0: set if the node is "obsolete". An obsolete node has been removed from the tree,
-//!        but might still be accessed by concurrent readers until the epoch expires.
-//! Bit 1: set if the node is currently write-locked. Used as a spinlock.
-//! Bits 2-63: Version number, incremented every time the node is modified.
-//!
-//! AtomicLockAndVersion represents that.
-
-use std::sync::atomic::{AtomicU64, Ordering};
-
-pub(crate) struct ConcurrentUpdateError();
-
-pub(crate) struct AtomicLockAndVersion {
-    inner: AtomicU64,
-}
-
-impl AtomicLockAndVersion {
-    pub(crate) fn new() -> AtomicLockAndVersion {
-        AtomicLockAndVersion {
-            inner: AtomicU64::new(0),
-        }
-    }
-}
-
-impl AtomicLockAndVersion {
-    pub(crate) fn read_lock_or_restart(&self) -> Result<u64, ConcurrentUpdateError> {
-        let version = self.await_node_unlocked();
-        if is_obsolete(version) {
-            return Err(ConcurrentUpdateError());
-        }
-        Ok(version)
-    }
-
-    pub(crate) fn check_or_restart(&self, version: u64) -> Result<(), ConcurrentUpdateError> {
-        self.read_unlock_or_restart(version)
-    }
-
-    pub(crate) fn read_unlock_or_restart(&self, version: u64) -> Result<(), ConcurrentUpdateError> {
-        if self.inner.load(Ordering::Acquire) != version {
-            return Err(ConcurrentUpdateError());
-        }
-        Ok(())
-    }
-
-    pub(crate) fn upgrade_to_write_lock_or_restart(
-        &self,
-        version: u64,
-    ) -> Result<(), ConcurrentUpdateError> {
-        if self
-            .inner
-            .compare_exchange(
-                version,
-                set_locked_bit(version),
-                Ordering::Acquire,
-                Ordering::Relaxed,
-            )
-            .is_err()
-        {
-            return Err(ConcurrentUpdateError());
-        }
-        Ok(())
-    }
-
-    pub(crate) fn write_lock_or_restart(&self) -> Result<(), ConcurrentUpdateError> {
-        let old = self.inner.load(Ordering::Relaxed);
-        if is_obsolete(old) || is_locked(old) {
-            return Err(ConcurrentUpdateError());
-        }
-        if self
-            .inner
-            .compare_exchange(
-                old,
-                set_locked_bit(old),
-                Ordering::Acquire,
-                Ordering::Relaxed,
-            )
-            .is_err()
-        {
-            return Err(ConcurrentUpdateError());
-        }
-        Ok(())
-    }
-
-    pub(crate) fn write_unlock(&self) {
-        // reset locked bit and overflow into version
-        self.inner.fetch_add(2, Ordering::Release);
-    }
-
-    pub(crate) fn write_unlock_obsolete(&self) {
-        // set obsolete, reset locked, overflow into version
-        self.inner.fetch_add(3, Ordering::Release);
-    }
-
-    // Helper functions
-    fn await_node_unlocked(&self) -> u64 {
-        let mut version = self.inner.load(Ordering::Acquire);
-        while is_locked(version) {
-            // spinlock
-            std::thread::yield_now();
-            version = self.inner.load(Ordering::Acquire)
-        }
-        version
-    }
-}
-
-fn set_locked_bit(version: u64) -> u64 {
-    version + 2
-}
-
-fn is_obsolete(version: u64) -> bool {
-    (version & 1) == 1
-}
-
-fn is_locked(version: u64) -> bool {
-    (version & 2) == 2
-}
--- a/libs/neonart/src/algorithm/node_ptr.rs
+++ b/libs/neonart/src/algorithm/node_ptr.rs
--- a/libs/neonart/src/algorithm/node_ref.rs
+++ b/libs/neonart/src/algorithm/node_ref.rs
@@ -1,349 +0,0 @@
-use std::fmt::Debug;
-use std::marker::PhantomData;
-
-use super::node_ptr;
-use super::node_ptr::NodePtr;
-use crate::EpochPin;
-use crate::Value;
-use crate::algorithm::lock_and_version::AtomicLockAndVersion;
-use crate::algorithm::lock_and_version::ConcurrentUpdateError;
-use crate::allocator::ArtAllocator;
-use crate::allocator::OutOfMemoryError;
-
-pub struct NodeRef<'e, V> {
-    ptr: NodePtr<V>,
-
-    phantom: PhantomData<&'e EpochPin<'e>>,
-}
-
-impl<'e, V> Debug for NodeRef<'e, V> {
-    fn fmt(&self, fmt: &mut std::fmt::Formatter<'_>) -> Result<(), std::fmt::Error> {
-        write!(fmt, "{:?}", self.ptr)
-    }
-}
-
-impl<'e, V: Value> NodeRef<'e, V> {
-    pub(crate) fn from_root_ptr(root_ptr: NodePtr<V>) -> NodeRef<'e, V> {
-        NodeRef {
-            ptr: root_ptr,
-            phantom: PhantomData,
-        }
-    }
-
-    pub(crate) fn read_lock_or_restart(
-        &self,
-    ) -> Result<ReadLockedNodeRef<'e, V>, ConcurrentUpdateError> {
-        let version = self.lockword().read_lock_or_restart()?;
-        Ok(ReadLockedNodeRef {
-            ptr: self.ptr,
-            version,
-            phantom: self.phantom,
-        })
-    }
-
-    pub(crate) fn write_lock_or_restart(
-        &self,
-    ) -> Result<WriteLockedNodeRef<'e, V>, ConcurrentUpdateError> {
-        self.lockword().write_lock_or_restart()?;
-        Ok(WriteLockedNodeRef {
-            ptr: self.ptr,
-            phantom: self.phantom,
-        })
-    }
-
-    fn lockword(&self) -> &AtomicLockAndVersion {
-        self.ptr.lockword()
-    }
-}
-
-/// A reference to a node that has been optimistically read-locked. The functions re-check
-/// the version after each read.
-pub struct ReadLockedNodeRef<'e, V> {
-    ptr: NodePtr<V>,
-    version: u64,
-
-    phantom: PhantomData<&'e EpochPin<'e>>,
-}
-
-impl<'e, V: Value> ReadLockedNodeRef<'e, V> {
-    pub(crate) fn is_leaf(&self) -> bool {
-        self.ptr.is_leaf()
-    }
-
-    pub(crate) fn is_full(&self) -> bool {
-        self.ptr.is_full()
-    }
-
-    pub(crate) fn get_prefix(&self) -> &[u8] {
-        self.ptr.get_prefix()
-    }
-
-    /// Note: because we're only holding a read lock, the prefix can change concurrently.
-    /// You must be prepared to restart, if read_unlock() returns error later.
-    ///
-    /// Returns the length of the prefix, or None if it's not a match
-    pub(crate) fn prefix_matches(&self, key: &[u8]) -> Option<usize> {
-        self.ptr.prefix_matches(key)
-    }
-
-    pub(crate) fn find_child_or_restart(
-        &self,
-        key_byte: u8,
-    ) -> Result<Option<NodeRef<'e, V>>, ConcurrentUpdateError> {
-        let child_or_value = self.ptr.find_child(key_byte);
-        self.ptr.lockword().check_or_restart(self.version)?;
-
-        match child_or_value {
-            None => Ok(None),
-            Some(child_ptr) => Ok(Some(NodeRef {
-                ptr: child_ptr,
-                phantom: self.phantom,
-            })),
-        }
-    }
-
-    pub(crate) fn find_next_child_or_restart(
-        &self,
-        min_key_byte: u8,
-    ) -> Result<Option<(u8, NodeRef<'e, V>)>, ConcurrentUpdateError> {
-        let child_or_value = self.ptr.find_next_child(min_key_byte);
-        self.ptr.lockword().check_or_restart(self.version)?;
-
-        match child_or_value {
-            None => Ok(None),
-            Some((k, child_ptr)) => Ok(Some((
-                k,
-                NodeRef {
-                    ptr: child_ptr,
-                    phantom: self.phantom,
-                },
-            ))),
-        }
-    }
-
-    pub(crate) fn get_leaf_value_ptr(&self) -> Result<*const V, ConcurrentUpdateError> {
-        let result = self.ptr.get_leaf_value();
-        self.ptr.lockword().check_or_restart(self.version)?;
-
-        // Extend the lifetime.
-        let result = std::ptr::from_ref(result);
-
-        Ok(result)
-    }
-
-    pub(crate) fn upgrade_to_write_lock_or_restart(
-        self,
-    ) -> Result<WriteLockedNodeRef<'e, V>, ConcurrentUpdateError> {
-        self.ptr
-            .lockword()
-            .upgrade_to_write_lock_or_restart(self.version)?;
-
-        Ok(WriteLockedNodeRef {
-            ptr: self.ptr,
-            phantom: self.phantom,
-        })
-    }
-
-    pub(crate) fn read_unlock_or_restart(self) -> Result<(), ConcurrentUpdateError> {
-        self.ptr.lockword().check_or_restart(self.version)?;
-        Ok(())
-    }
-
-    pub(crate) fn check_or_restart(&self) -> Result<(), ConcurrentUpdateError> {
-        self.ptr.lockword().check_or_restart(self.version)?;
-        Ok(())
-    }
-}
-
-/// A reference to a node that has been optimistically read-locked. The functions re-check
-/// the version after each read.
-pub struct WriteLockedNodeRef<'e, V> {
-    ptr: NodePtr<V>,
-    phantom: PhantomData<&'e EpochPin<'e>>,
-}
-
-impl<'e, V: Value> WriteLockedNodeRef<'e, V> {
-    pub(crate) fn can_shrink(&self) -> bool {
-        self.ptr.can_shrink()
-    }
-
-    pub(crate) fn num_children(&self) -> usize {
-        self.ptr.num_children()
-    }
-
-    pub(crate) fn write_unlock(mut self) {
-        self.ptr.lockword().write_unlock();
-        self.ptr = NodePtr::null();
-    }
-
-    pub(crate) fn write_unlock_obsolete(mut self) {
-        self.ptr.lockword().write_unlock_obsolete();
-        self.ptr = NodePtr::null();
-    }
-
-    pub(crate) fn get_prefix(&self) -> &[u8] {
-        self.ptr.get_prefix()
-    }
-
-    pub(crate) fn truncate_prefix(&mut self, new_prefix_len: usize) {
-        self.ptr.truncate_prefix(new_prefix_len)
-    }
-
-    pub(crate) fn prepend_prefix(&mut self, prefix: &[u8], prefix_byte: u8) {
-        self.ptr.prepend_prefix(prefix, prefix_byte)
-    }
-
-    pub(crate) fn insert_child(&mut self, key_byte: u8, child: NodePtr<V>) {
-        self.ptr.insert_child(key_byte, child)
-    }
-
-    pub(crate) fn get_leaf_value_mut(&mut self) -> &mut V {
-        self.ptr.get_leaf_value_mut()
-    }
-
-    pub(crate) fn grow<'a, A>(
-        &self,
-        allocator: &'a A,
-    ) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
-    where
-        A: ArtAllocator<V>,
-    {
-        let new_node = self.ptr.grow(allocator)?;
-        Ok(NewNodeRef {
-            ptr: new_node,
-            allocator,
-            extra_nodes: Vec::new(),
-        })
-    }
-
-    pub(crate) fn shrink<'a, A>(
-        &self,
-        allocator: &'a A,
-    ) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
-    where
-        A: ArtAllocator<V>,
-    {
-        let new_node = self.ptr.shrink(allocator)?;
-        Ok(NewNodeRef {
-            ptr: new_node,
-            allocator,
-            extra_nodes: Vec::new(),
-        })
-    }
-
-    pub(crate) fn as_ptr(&self) -> NodePtr<V> {
-        self.ptr
-    }
-
-    pub(crate) fn replace_child(&mut self, key_byte: u8, replacement: NodePtr<V>) {
-        self.ptr.replace_child(key_byte, replacement);
-    }
-
-    pub(crate) fn delete_child(&mut self, key_byte: u8) {
-        self.ptr.delete_child(key_byte);
-    }
-
-    pub(crate) fn find_remaining_child(&self) -> (u8, NodeRef<'e, V>) {
-        assert_eq!(self.num_children(), 1);
-        let child_or_value = self.ptr.find_next_child(0);
-
-        match child_or_value {
-            None => panic!("could not find only child in node"),
-            Some((k, child_ptr)) => (
-                k,
-                NodeRef {
-                    ptr: child_ptr,
-                    phantom: self.phantom,
-                },
-            ),
-        }
-    }
-}
-
-impl<'e, V> Drop for WriteLockedNodeRef<'e, V> {
-    fn drop(&mut self) {
-        if !self.ptr.is_null() {
-            self.ptr.lockword().write_unlock();
-        }
-    }
-}
-
-pub(crate) struct NewNodeRef<'a, V, A>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    ptr: NodePtr<V>,
-    allocator: &'a A,
-
-    extra_nodes: Vec<NodePtr<V>>,
-}
-
-impl<'a, V, A> NewNodeRef<'a, V, A>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    pub(crate) fn insert_old_child(&mut self, key_byte: u8, child: &WriteLockedNodeRef<V>) {
-        self.ptr.insert_child(key_byte, child.as_ptr())
-    }
-
-    pub(crate) fn into_ptr(mut self) -> NodePtr<V> {
-        let ptr = self.ptr;
-        self.ptr = NodePtr::null();
-        ptr
-    }
-
-    pub(crate) fn insert_new_child(&mut self, key_byte: u8, child: NewNodeRef<'a, V, A>) {
-        let child_ptr = child.into_ptr();
-        self.ptr.insert_child(key_byte, child_ptr);
-        self.extra_nodes.push(child_ptr);
-    }
-}
-
-impl<'a, V, A> Drop for NewNodeRef<'a, V, A>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    /// This drop implementation deallocates the newly allocated node, if into_ptr() was not called.
-    fn drop(&mut self) {
-        if !self.ptr.is_null() {
-            self.ptr.deallocate(self.allocator);
-            for p in self.extra_nodes.iter() {
-                p.deallocate(self.allocator);
-            }
-        }
-    }
-}
-
-pub(crate) fn new_internal<'a, V, A>(
-    prefix: &[u8],
-    allocator: &'a A,
-) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    Ok(NewNodeRef {
-        ptr: node_ptr::new_internal(prefix, allocator)?,
-        allocator,
-        extra_nodes: Vec::new(),
-    })
-}
-
-pub(crate) fn new_leaf<'a, V, A>(
-    prefix: &[u8],
-    value: V,
-    allocator: &'a A,
-) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    Ok(NewNodeRef {
-        ptr: node_ptr::new_leaf(prefix, value, allocator)?,
-        allocator,
-        extra_nodes: Vec::new(),
-    })
-}
--- a/libs/neonart/src/allocator.rs
+++ b/libs/neonart/src/allocator.rs
@@ -1,156 +0,0 @@
-pub mod block;
-mod multislab;
-mod slab;
-pub mod r#static;
-
-use std::alloc::Layout;
-use std::marker::PhantomData;
-use std::mem::MaybeUninit;
-use std::sync::atomic::Ordering;
-
-use crate::allocator::multislab::MultiSlabAllocator;
-use crate::allocator::r#static::alloc_from_slice;
-
-use spin;
-
-use crate::Tree;
-pub use crate::algorithm::node_ptr::{
-    NodeInternal4, NodeInternal16, NodeInternal48, NodeInternal256, NodeLeaf,
-};
-
-#[derive(Debug)]
-pub struct OutOfMemoryError();
-
-pub trait ArtAllocator<V: crate::Value> {
-    fn alloc_tree(&self) -> *mut Tree<V>;
-
-    fn alloc_node_internal4(&self) -> *mut NodeInternal4<V>;
-    fn alloc_node_internal16(&self) -> *mut NodeInternal16<V>;
-    fn alloc_node_internal48(&self) -> *mut NodeInternal48<V>;
-    fn alloc_node_internal256(&self) -> *mut NodeInternal256<V>;
-    fn alloc_node_leaf(&self) -> *mut NodeLeaf<V>;
-
-    fn dealloc_node_internal4(&self, ptr: *mut NodeInternal4<V>);
-    fn dealloc_node_internal16(&self, ptr: *mut NodeInternal16<V>);
-    fn dealloc_node_internal48(&self, ptr: *mut NodeInternal48<V>);
-    fn dealloc_node_internal256(&self, ptr: *mut NodeInternal256<V>);
-    fn dealloc_node_leaf(&self, ptr: *mut NodeLeaf<V>);
-}
-
-pub struct ArtMultiSlabAllocator<'t, V>
-where
-    V: crate::Value,
-{
-    tree_area: spin::Mutex<Option<&'t mut MaybeUninit<Tree<V>>>>,
-
-    pub(crate) inner: MultiSlabAllocator<'t, 5>,
-
-    phantom_val: PhantomData<V>,
-}
-
-impl<'t, V: crate::Value> ArtMultiSlabAllocator<'t, V> {
-    const LAYOUTS: [Layout; 5] = [
-        Layout::new::<NodeInternal4<V>>(),
-        Layout::new::<NodeInternal16<V>>(),
-        Layout::new::<NodeInternal48<V>>(),
-        Layout::new::<NodeInternal256<V>>(),
-        Layout::new::<NodeLeaf<V>>(),
-    ];
-
-    pub fn new(area: &'t mut [MaybeUninit<u8>]) -> &'t mut ArtMultiSlabAllocator<'t, V> {
-        let (allocator_area, remain) = alloc_from_slice::<ArtMultiSlabAllocator<V>>(area);
-        let (tree_area, remain) = alloc_from_slice::<Tree<V>>(remain);
-
-        allocator_area.write(ArtMultiSlabAllocator {
-            tree_area: spin::Mutex::new(Some(tree_area)),
-            inner: MultiSlabAllocator::new(remain, &Self::LAYOUTS),
-            phantom_val: PhantomData,
-        })
-    }
-}
-
-impl<'t, V: crate::Value> ArtAllocator<V> for ArtMultiSlabAllocator<'t, V> {
-    fn alloc_tree(&self) -> *mut Tree<V> {
-        let mut t = self.tree_area.lock();
-        if let Some(tree_area) = t.take() {
-            return tree_area.as_mut_ptr().cast();
-        }
-        panic!("cannot allocate more than one tree");
-    }
-
-    fn alloc_node_internal4(&self) -> *mut NodeInternal4<V> {
-        self.inner.alloc_slab(0).cast()
-    }
-    fn alloc_node_internal16(&self) -> *mut NodeInternal16<V> {
-        self.inner.alloc_slab(1).cast()
-    }
-    fn alloc_node_internal48(&self) -> *mut NodeInternal48<V> {
-        self.inner.alloc_slab(2).cast()
-    }
-    fn alloc_node_internal256(&self) -> *mut NodeInternal256<V> {
-        self.inner.alloc_slab(3).cast()
-    }
-    fn alloc_node_leaf(&self) -> *mut NodeLeaf<V> {
-        self.inner.alloc_slab(4).cast()
-    }
-
-    fn dealloc_node_internal4(&self, ptr: *mut NodeInternal4<V>) {
-        self.inner.dealloc_slab(0, ptr.cast())
-    }
-
-    fn dealloc_node_internal16(&self, ptr: *mut NodeInternal16<V>) {
-        self.inner.dealloc_slab(1, ptr.cast())
-    }
-    fn dealloc_node_internal48(&self, ptr: *mut NodeInternal48<V>) {
-        self.inner.dealloc_slab(2, ptr.cast())
-    }
-    fn dealloc_node_internal256(&self, ptr: *mut NodeInternal256<V>) {
-        self.inner.dealloc_slab(3, ptr.cast())
-    }
-    fn dealloc_node_leaf(&self, ptr: *mut NodeLeaf<V>) {
-        self.inner.dealloc_slab(4, ptr.cast())
-    }
-}
-
-impl<'t, V: crate::Value> ArtMultiSlabAllocator<'t, V> {
-    pub(crate) fn get_statistics(&self) -> ArtMultiSlabStats {
-        ArtMultiSlabStats {
-            num_internal4: self.inner.slab_descs[0]
-                .num_allocated
-                .load(Ordering::Relaxed),
-            num_internal16: self.inner.slab_descs[1]
-                .num_allocated
-                .load(Ordering::Relaxed),
-            num_internal48: self.inner.slab_descs[2]
-                .num_allocated
-                .load(Ordering::Relaxed),
-            num_internal256: self.inner.slab_descs[3]
-                .num_allocated
-                .load(Ordering::Relaxed),
-            num_leaf: self.inner.slab_descs[4]
-                .num_allocated
-                .load(Ordering::Relaxed),
-
-            num_blocks_internal4: self.inner.slab_descs[0].num_blocks.load(Ordering::Relaxed),
-            num_blocks_internal16: self.inner.slab_descs[1].num_blocks.load(Ordering::Relaxed),
-            num_blocks_internal48: self.inner.slab_descs[2].num_blocks.load(Ordering::Relaxed),
-            num_blocks_internal256: self.inner.slab_descs[3].num_blocks.load(Ordering::Relaxed),
-            num_blocks_leaf: self.inner.slab_descs[4].num_blocks.load(Ordering::Relaxed),
-        }
-    }
-}
-
-#[derive(Clone, Debug)]
-pub struct ArtMultiSlabStats {
-    pub num_internal4: u64,
-    pub num_internal16: u64,
-    pub num_internal48: u64,
-    pub num_internal256: u64,
-    pub num_leaf: u64,
-
-    pub num_blocks_internal4: u64,
-    pub num_blocks_internal16: u64,
-    pub num_blocks_internal48: u64,
-    pub num_blocks_internal256: u64,
-    pub num_blocks_leaf: u64,
-}
--- a/libs/neonart/src/allocator/block.rs
+++ b/libs/neonart/src/allocator/block.rs
@@ -1,191 +0,0 @@
-//! Simple allocator of fixed-size blocks
-
-use std::mem::MaybeUninit;
-use std::sync::atomic::{AtomicU64, Ordering};
-
-use spin;
-
-pub const BLOCK_SIZE: usize = 16 * 1024;
-
-const INVALID_BLOCK: u64 = u64::MAX;
-
-pub(crate) struct BlockAllocator<'t> {
-    blocks_ptr: &'t [MaybeUninit<u8>],
-    num_blocks: u64,
-    num_initialized: AtomicU64,
-
-    freelist_head: spin::Mutex<u64>,
-}
-
-struct FreeListBlock {
-    inner: spin::Mutex<FreeListBlockInner>,
-}
-
-struct FreeListBlockInner {
-    next: u64,
-
-    num_free_blocks: u64,
-    free_blocks: [u64; 100], // FIXME: fill the rest of the block
-}
-
-impl<'t> BlockAllocator<'t> {
-    pub(crate) fn new(area: &'t mut [MaybeUninit<u8>]) -> Self {
-        // Use all the space for the blocks
-        let padding = area.as_ptr().align_offset(BLOCK_SIZE);
-        let remain = &mut area[padding..];
-
-        let num_blocks = (remain.len() / BLOCK_SIZE) as u64;
-
-        BlockAllocator {
-            blocks_ptr: remain,
-            num_blocks,
-            num_initialized: AtomicU64::new(0),
-            freelist_head: spin::Mutex::new(INVALID_BLOCK),
-        }
-    }
-
-    /// safety: you must hold a lock on the pointer to this block, otherwise it might get
-    /// reused for another kind of block
-    fn read_freelist_block(&self, blkno: u64) -> &FreeListBlock {
-        let ptr: *const FreeListBlock = self.get_block_ptr(blkno).cast();
-        unsafe { ptr.as_ref().unwrap() }
-    }
-
-    fn get_block_ptr(&self, blkno: u64) -> *mut u8 {
-        assert!(blkno < self.num_blocks);
-        unsafe {
-            self.blocks_ptr
-                .as_ptr()
-                .byte_offset(blkno as isize * BLOCK_SIZE as isize)
-        }
-        .cast_mut()
-        .cast()
-    }
-
-    #[allow(clippy::mut_from_ref)]
-    pub(crate) fn alloc_block(&self) -> &mut [MaybeUninit<u8>] {
-        // FIXME: handle OOM
-        let blkno = self.alloc_block_internal();
-        if blkno == INVALID_BLOCK {
-            panic!("out of memory");
-        }
-
-        let ptr: *mut MaybeUninit<u8> = self.get_block_ptr(blkno).cast();
-        unsafe { std::slice::from_raw_parts_mut(ptr, BLOCK_SIZE) }
-    }
-
-    fn alloc_block_internal(&self) -> u64 {
-        //  check the free list.
-        {
-            let mut freelist_head = self.freelist_head.lock();
-            if *freelist_head != INVALID_BLOCK {
-                let freelist_block = self.read_freelist_block(*freelist_head);
-
-                // acquire lock on the freelist block before releasing the lock on the parent (i.e. lock coupling)
-                let mut g = freelist_block.inner.lock();
-
-                if g.num_free_blocks > 0 {
-                    g.num_free_blocks -= 1;
-                    let result = g.free_blocks[g.num_free_blocks as usize];
-                    return result;
-                } else {
-                    // consume the freelist block itself
-                    let result = *freelist_head;
-                    *freelist_head = g.next;
-                    // This freelist block is now unlinked and can be repurposed
-                    drop(g);
-                    return result;
-                }
-            }
-        }
-
-        // If there are some blocks left that we've never used, pick next such block
-        let mut next_uninitialized = self.num_initialized.load(Ordering::Relaxed);
-        while next_uninitialized < self.num_blocks {
-            match self.num_initialized.compare_exchange(
-                next_uninitialized,
-                next_uninitialized + 1,
-                Ordering::Relaxed,
-                Ordering::Relaxed,
-            ) {
-                Ok(_) => {
-                    return next_uninitialized;
-                }
-                Err(old) => {
-                    next_uninitialized = old;
-                    continue;
-                }
-            }
-        }
-
-        // out of blocks
-        INVALID_BLOCK
-    }
-
-    // TODO: this is currently unused. The slab allocator never releases blocks
-    #[allow(dead_code)]
-    pub(crate) fn release_block(&self, block_ptr: *mut u8) {
-        let blockno = unsafe { block_ptr.byte_offset_from(self.blocks_ptr) / BLOCK_SIZE as isize };
-        self.release_block_internal(blockno as u64);
-    }
-
-    fn release_block_internal(&self, blockno: u64) {
-        let mut freelist_head = self.freelist_head.lock();
-        if *freelist_head != INVALID_BLOCK {
-            let freelist_block = self.read_freelist_block(*freelist_head);
-
-            // acquire lock on the freelist block before releasing the lock on the parent (i.e. lock coupling)
-            let mut g = freelist_block.inner.lock();
-
-            let num_free_blocks = g.num_free_blocks;
-            if num_free_blocks < g.free_blocks.len() as u64 {
-                g.free_blocks[num_free_blocks as usize] = blockno;
-                g.num_free_blocks += 1;
-                return;
-            }
-        }
-
-        // Convert the block into a new freelist block
-        let block_ptr: *mut FreeListBlock = self.get_block_ptr(blockno).cast();
-        let init = FreeListBlock {
-            inner: spin::Mutex::new(FreeListBlockInner {
-                next: *freelist_head,
-                num_free_blocks: 0,
-                free_blocks: [INVALID_BLOCK; 100],
-            }),
-        };
-        unsafe { (*block_ptr) = init };
-        *freelist_head = blockno;
-    }
-
-    // for debugging
-    pub(crate) fn get_statistics(&self) -> BlockAllocatorStats {
-        let mut num_free_blocks = 0;
-
-        let mut _prev_lock = None;
-        let head_lock = self.freelist_head.lock();
-        let mut next_blk = *head_lock;
-        let mut _head_lock = Some(head_lock);
-        while next_blk != INVALID_BLOCK {
-            let freelist_block = self.read_freelist_block(next_blk);
-            let lock = freelist_block.inner.lock();
-            num_free_blocks += lock.num_free_blocks;
-            next_blk = lock.next;
-            _prev_lock = Some(lock); // hold the lock until we've read the next block
-            _head_lock = None;
-        }
-
-        BlockAllocatorStats {
-            num_blocks: self.num_blocks,
-            num_initialized: self.num_initialized.load(Ordering::Relaxed),
-            num_free_blocks,
-        }
-    }
-}
-
-#[derive(Clone, Debug)]
-pub struct BlockAllocatorStats {
-    pub num_blocks: u64,
-    pub num_initialized: u64,
-    pub num_free_blocks: u64,
-}
--- a/libs/neonart/src/allocator/multislab.rs
+++ b/libs/neonart/src/allocator/multislab.rs
@@ -1,33 +0,0 @@
-use std::alloc::Layout;
-use std::mem::MaybeUninit;
-
-use crate::allocator::block::BlockAllocator;
-use crate::allocator::slab::SlabDesc;
-
-pub struct MultiSlabAllocator<'t, const N: usize> {
-    pub(crate) block_allocator: BlockAllocator<'t>,
-
-    pub(crate) slab_descs: [SlabDesc; N],
-}
-
-impl<'t, const N: usize> MultiSlabAllocator<'t, N> {
-    pub(crate) fn new(
-        area: &'t mut [MaybeUninit<u8>],
-        layouts: &[Layout; N],
-    ) -> MultiSlabAllocator<'t, N> {
-        let block_allocator = BlockAllocator::new(area);
-        MultiSlabAllocator {
-            block_allocator,
-
-            slab_descs: std::array::from_fn(|i| SlabDesc::new(&layouts[i])),
-        }
-    }
-
-    pub(crate) fn alloc_slab(&self, slab_idx: usize) -> *mut u8 {
-        self.slab_descs[slab_idx].alloc_chunk(&self.block_allocator)
-    }
-
-    pub(crate) fn dealloc_slab(&self, slab_idx: usize, ptr: *mut u8) {
-        self.slab_descs[slab_idx].dealloc_chunk(ptr, &self.block_allocator)
-    }
-}
--- a/libs/neonart/src/allocator/slab.rs
+++ b/libs/neonart/src/allocator/slab.rs
@@ -1,433 +0,0 @@
-//! A slab allocator that carves out fixed-size chunks from larger blocks.
-//!
-//!
-
-use std::alloc::Layout;
-use std::mem::MaybeUninit;
-use std::ops::Deref;
-use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
-
-use spin;
-
-use super::alloc_from_slice;
-use super::block::BlockAllocator;
-
-use crate::allocator::block::BLOCK_SIZE;
-
-pub(crate) struct SlabDesc {
-    pub(crate) layout: Layout,
-
-    block_lists: spin::RwLock<BlockLists>,
-
-    pub(crate) num_blocks: AtomicU64,
-    pub(crate) num_allocated: AtomicU64,
-}
-
-// FIXME: Not sure if SlabDesc is really Sync or Send. It probably is when it's empty, but
-// 'block_lists' contains pointers when it's not empty. In the current use as part of the
-// the art tree, SlabDescs are only moved during initialization.
-unsafe impl Sync for SlabDesc {}
-unsafe impl Send for SlabDesc {}
-
-#[derive(Default, Debug)]
-struct BlockLists {
-    full_blocks: BlockList,
-    nonfull_blocks: BlockList,
-}
-
-impl BlockLists {
-    // Unlink a node. It must be in either one of the two lists.
-    unsafe fn unlink(&mut self, elem: *mut SlabBlockHeader) {
-        let list = unsafe {
-            if (*elem).next.is_null() {
-                if self.full_blocks.tail == elem {
-                    Some(&mut self.full_blocks)
-                } else {
-                    Some(&mut self.nonfull_blocks)
-                }
-            } else if (*elem).prev.is_null() {
-                if self.full_blocks.head == elem {
-                    Some(&mut self.full_blocks)
-                } else {
-                    Some(&mut self.nonfull_blocks)
-                }
-            } else {
-                None
-            }
-        };
-        unsafe { unlink_slab_block(list, elem) };
-    }
-}
-
-unsafe fn unlink_slab_block(mut list: Option<&mut BlockList>, elem: *mut SlabBlockHeader) {
-    unsafe {
-        if (*elem).next.is_null() {
-            assert_eq!(list.as_ref().unwrap().tail, elem);
-            list.as_mut().unwrap().tail = (*elem).prev;
-        } else {
-            assert_eq!((*(*elem).next).prev, elem);
-            (*(*elem).next).prev = (*elem).prev;
-        }
-        if (*elem).prev.is_null() {
-            assert_eq!(list.as_ref().unwrap().head, elem);
-            list.as_mut().unwrap().head = (*elem).next;
-        } else {
-            assert_eq!((*(*elem).prev).next, elem);
-            (*(*elem).prev).next = (*elem).next;
-        }
-    }
-}
-
-#[derive(Debug)]
-struct BlockList {
-    head: *mut SlabBlockHeader,
-    tail: *mut SlabBlockHeader,
-}
-
-impl Default for BlockList {
-    fn default() -> Self {
-        BlockList {
-            head: std::ptr::null_mut(),
-            tail: std::ptr::null_mut(),
-        }
-    }
-}
-
-impl BlockList {
-    unsafe fn push_head(&mut self, elem: *mut SlabBlockHeader) {
-        unsafe {
-            if self.is_empty() {
-                self.tail = elem;
-                (*elem).next = std::ptr::null_mut();
-            } else {
-                (*elem).next = self.head;
-                (*self.head).prev = elem;
-            }
-            (*elem).prev = std::ptr::null_mut();
-            self.head = elem;
-        }
-    }
-
-    fn is_empty(&self) -> bool {
-        self.head.is_null()
-    }
-
-    unsafe fn unlink(&mut self, elem: *mut SlabBlockHeader) {
-        unsafe { unlink_slab_block(Some(self), elem) }
-    }
-
-    #[cfg(test)]
-    fn dump(&self) {
-        let mut next = self.head;
-
-        while !next.is_null() {
-            let n = unsafe { next.as_ref() }.unwrap();
-            eprintln!(
-                "  blk {:?} (free {}/{})",
-                next,
-                n.num_free_chunks.load(Ordering::Relaxed),
-                n.num_chunks
-            );
-            next = n.next;
-        }
-    }
-}
-
-impl SlabDesc {
-    pub(crate) fn new(layout: &Layout) -> SlabDesc {
-        SlabDesc {
-            layout: *layout,
-            block_lists: spin::RwLock::new(BlockLists::default()),
-            num_allocated: AtomicU64::new(0),
-            num_blocks: AtomicU64::new(0),
-        }
-    }
-}
-
-#[derive(Debug)]
-struct SlabBlockHeader {
-    free_chunks_head: spin::Mutex<*mut FreeChunk>,
-    num_free_chunks: AtomicU32,
-    num_chunks: u32, // this is really a constant for a given Layout
-
-    // these fields are protected by the lock on the BlockLists
-    prev: *mut SlabBlockHeader,
-    next: *mut SlabBlockHeader,
-}
-
-struct FreeChunk {
-    next: *mut FreeChunk,
-}
-
-enum ReadOrWriteGuard<'a, T> {
-    Read(spin::RwLockReadGuard<'a, T>),
-    Write(spin::RwLockWriteGuard<'a, T>),
-}
-
-impl<'a, T> Deref for ReadOrWriteGuard<'a, T> {
-    type Target = T;
-
-    fn deref(&self) -> &<Self as Deref>::Target {
-        match self {
-            ReadOrWriteGuard::Read(g) => g.deref(),
-            ReadOrWriteGuard::Write(g) => g.deref(),
-        }
-    }
-}
-
-impl SlabDesc {
-    pub fn alloc_chunk(&self, block_allocator: &BlockAllocator) -> *mut u8 {
-        // Are there any free chunks?
-        let mut acquire_write = false;
-        'outer: loop {
-            let mut block_lists_guard = if acquire_write {
-                ReadOrWriteGuard::Write(self.block_lists.write())
-            } else {
-                ReadOrWriteGuard::Read(self.block_lists.read())
-            };
-            'inner: loop {
-                let block_ptr = block_lists_guard.nonfull_blocks.head;
-                if block_ptr.is_null() {
-                    break 'outer;
-                }
-                unsafe {
-                    let mut free_chunks_head = (*block_ptr).free_chunks_head.lock();
-                    if !(*free_chunks_head).is_null() {
-                        let result = *free_chunks_head;
-                        (*free_chunks_head) = (*result).next;
-                        let _old = (*block_ptr).num_free_chunks.fetch_sub(1, Ordering::Relaxed);
-
-                        self.num_allocated.fetch_add(1, Ordering::Relaxed);
-                        return result.cast();
-                    }
-                }
-
-                // The block at the head of the list was full. Grab write lock and retry
-                match block_lists_guard {
-                    ReadOrWriteGuard::Read(_) => {
-                        acquire_write = true;
-                        continue 'outer;
-                    }
-                    ReadOrWriteGuard::Write(ref mut g) => {
-                        // move the node to the list of full blocks
-                        unsafe {
-                            g.nonfull_blocks.unlink(block_ptr);
-                            g.full_blocks.push_head(block_ptr);
-                        };
-                        continue 'inner;
-                    }
-                }
-            }
-        }
-
-        // no free chunks. Allocate a new block (and the chunk from that)
-        let (new_block, new_chunk) = self.alloc_block_and_chunk(block_allocator);
-        self.num_blocks.fetch_add(1, Ordering::Relaxed);
-
-        // Add the block to the list in the SlabDesc
-        unsafe {
-            let mut block_lists_guard = self.block_lists.write();
-            block_lists_guard.nonfull_blocks.push_head(new_block);
-        }
-        self.num_allocated.fetch_add(1, Ordering::Relaxed);
-        new_chunk
-    }
-
-    pub fn dealloc_chunk(&self, chunk_ptr: *mut u8, _block_allocator: &BlockAllocator) {
-        // Find the block it belongs to. You can find the block from the address. (And knowing the
-        // layout, you could calculate the chunk number too.)
-        let block_ptr: *mut SlabBlockHeader = {
-            let block_addr = (chunk_ptr.addr() / BLOCK_SIZE) * BLOCK_SIZE;
-            chunk_ptr.with_addr(block_addr).cast()
-        };
-        let chunk_ptr: *mut FreeChunk = chunk_ptr.cast();
-
-        // Mark the chunk as free in 'freechunks' list
-        let num_chunks;
-        let num_free_chunks;
-        unsafe {
-            let mut free_chunks_head = (*block_ptr).free_chunks_head.lock();
-            (*chunk_ptr).next = *free_chunks_head;
-            *free_chunks_head = chunk_ptr;
-
-            num_free_chunks = (*block_ptr).num_free_chunks.fetch_add(1, Ordering::Relaxed) + 1;
-            num_chunks = (*block_ptr).num_chunks;
-        }
-
-        if num_free_chunks == 1 {
-            // If the block was full previously, add it to the nonfull blocks list. Note that
-            // we're not holding the lock anymore, so it can immediately become full again.
-            // That's harmless, it will be moved back to the full list again when a call
-            // to alloc_chunk() sees it.
-            let mut block_lists = self.block_lists.write();
-            unsafe {
-                block_lists.unlink(block_ptr);
-                block_lists.nonfull_blocks.push_head(block_ptr);
-            };
-        } else if num_free_chunks == num_chunks {
-            // If the block became completely empty, move it to the free list
-            // TODO
-            // FIXME: we're still holding the spinlock. It's not exactly safe to return it to
-            // the free blocks list, is it? Defer it as garbage to wait out concurrent updates?
-            //block_allocator.release_block()
-        }
-
-        // update stats
-        self.num_allocated.fetch_sub(1, Ordering::Relaxed);
-    }
-
-    fn alloc_block_and_chunk(
-        &self,
-        block_allocator: &BlockAllocator,
-    ) -> (*mut SlabBlockHeader, *mut u8) {
-        // fixme: handle OOM
-        let block_slice: &mut [MaybeUninit<u8>] = block_allocator.alloc_block();
-        let (block_header, remain) = alloc_from_slice::<SlabBlockHeader>(block_slice);
-
-        let padding = remain.as_ptr().align_offset(self.layout.align());
-
-        let num_chunks = (remain.len() - padding) / self.layout.size();
-
-        let first_chunk_ptr: *mut FreeChunk = remain[padding..].as_mut_ptr().cast();
-
-        unsafe {
-            let mut chunk_ptr = first_chunk_ptr;
-            for _ in 0..num_chunks - 1 {
-                let next_chunk_ptr = chunk_ptr.byte_add(self.layout.size());
-                (*chunk_ptr).next = next_chunk_ptr;
-                chunk_ptr = next_chunk_ptr;
-            }
-            (*chunk_ptr).next = std::ptr::null_mut();
-
-            let result_chunk = first_chunk_ptr;
-
-            let block_header = block_header.write(SlabBlockHeader {
-                free_chunks_head: spin::Mutex::new((*first_chunk_ptr).next),
-                prev: std::ptr::null_mut(),
-                next: std::ptr::null_mut(),
-                num_chunks: num_chunks as u32,
-                num_free_chunks: AtomicU32::new(num_chunks as u32 - 1),
-            });
-
-            (block_header, result_chunk.cast())
-        }
-    }
-
-    #[cfg(test)]
-    fn dump(&self) {
-        eprintln!(
-            "slab dump ({} blocks, {} allocated chunks)",
-            self.num_blocks.load(Ordering::Relaxed),
-            self.num_allocated.load(Ordering::Relaxed)
-        );
-        let lists = self.block_lists.read();
-
-        eprintln!("nonfull blocks:");
-        lists.nonfull_blocks.dump();
-        eprintln!("full blocks:");
-        lists.full_blocks.dump();
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    use rand::Rng;
-    use rand_distr::Zipf;
-
-    struct TestObject {
-        val: usize,
-        _dummy: [u8; BLOCK_SIZE / 4],
-    }
-
-    struct TestObjectSlab<'a>(SlabDesc, BlockAllocator<'a>);
-    impl<'a> TestObjectSlab<'a> {
-        fn new(block_allocator: BlockAllocator) -> TestObjectSlab {
-            TestObjectSlab(SlabDesc::new(&Layout::new::<TestObject>()), block_allocator)
-        }
-
-        fn alloc(&self, val: usize) -> *mut TestObject {
-            let obj: *mut TestObject = self.0.alloc_chunk(&self.1).cast();
-            unsafe { (*obj).val = val };
-            obj
-        }
-
-        fn dealloc(&self, obj: *mut TestObject) {
-            self.0.dealloc_chunk(obj.cast(), &self.1)
-        }
-    }
-
-    #[test]
-    fn test_slab_alloc() {
-        const MEM_SIZE: usize = 100000000;
-        let mut area = Box::new_uninit_slice(MEM_SIZE);
-        let block_allocator = BlockAllocator::new(&mut area);
-
-        let slab = TestObjectSlab::new(block_allocator);
-
-        let mut all: Vec<*mut TestObject> = Vec::new();
-        for i in 0..11 {
-            all.push(slab.alloc(i));
-        }
-        #[allow(clippy::needless_range_loop)]
-        for i in 0..11 {
-            assert!(unsafe { (*all[i]).val == i });
-        }
-
-        let distribution = Zipf::new(10.0, 1.1).unwrap();
-        let mut rng = rand::rng();
-        for _ in 0..100000 {
-            slab.0.dump();
-            let idx = rng.sample(distribution) as usize;
-            let ptr: *mut TestObject = all[idx];
-            if !ptr.is_null() {
-                assert_eq!(unsafe { (*ptr).val }, idx);
-                slab.dealloc(ptr);
-                all[idx] = std::ptr::null_mut();
-            } else {
-                all[idx] = slab.alloc(idx);
-            }
-        }
-    }
-
-    fn new_test_blk(i: u32) -> *mut SlabBlockHeader {
-        Box::into_raw(Box::new(SlabBlockHeader {
-            free_chunks_head: spin::Mutex::new(std::ptr::null_mut()),
-            num_free_chunks: AtomicU32::new(0),
-            num_chunks: i,
-            prev: std::ptr::null_mut(),
-            next: std::ptr::null_mut(),
-        }))
-    }
-
-    #[test]
-    fn test_block_linked_list() {
-        // note: these are leaked, but that's OK for tests
-        let a = new_test_blk(0);
-        let b = new_test_blk(1);
-
-        let mut list = BlockList::default();
-        assert!(list.is_empty());
-
-        unsafe {
-            list.push_head(a);
-            assert!(!list.is_empty());
-            list.unlink(a);
-        }
-        assert!(list.is_empty());
-
-        unsafe {
-            list.push_head(b);
-            list.push_head(a);
-            assert_eq!(list.head, a);
-            assert_eq!((*a).next, b);
-            assert_eq!((*b).prev, a);
-            assert_eq!(list.tail, b);
-
-            list.unlink(a);
-            list.unlink(b);
-            assert!(list.is_empty());
-        }
-    }
-}
--- a/libs/neonart/src/allocator/static.rs
+++ b/libs/neonart/src/allocator/static.rs
@@ -1,44 +0,0 @@
-use std::mem::MaybeUninit;
-
-pub fn alloc_from_slice<T>(
-    area: &mut [MaybeUninit<u8>],
-) -> (&mut MaybeUninit<T>, &mut [MaybeUninit<u8>]) {
-    let layout = std::alloc::Layout::new::<T>();
-
-    let area_start = area.as_mut_ptr();
-
-    // pad to satisfy alignment requirements
-    let padding = area_start.align_offset(layout.align());
-    if padding + layout.size() > area.len() {
-        panic!("out of memory");
-    }
-    let area = &mut area[padding..];
-    let (result_area, remain) = area.split_at_mut(layout.size());
-
-    let result_ptr: *mut MaybeUninit<T> = result_area.as_mut_ptr().cast();
-    let result = unsafe { result_ptr.as_mut().unwrap() };
-
-    (result, remain)
-}
-
-pub fn alloc_array_from_slice<T>(
-    area: &mut [MaybeUninit<u8>],
-    len: usize,
-) -> (&mut [MaybeUninit<T>], &mut [MaybeUninit<u8>]) {
-    let layout = std::alloc::Layout::new::<T>();
-
-    let area_start = area.as_mut_ptr();
-
-    // pad to satisfy alignment requirements
-    let padding = area_start.align_offset(layout.align());
-    if padding + layout.size() * len > area.len() {
-        panic!("out of memory");
-    }
-    let area = &mut area[padding..];
-    let (result_area, remain) = area.split_at_mut(layout.size() * len);
-
-    let result_ptr: *mut MaybeUninit<T> = result_area.as_mut_ptr().cast();
-    let result = unsafe { std::slice::from_raw_parts_mut(result_ptr.as_mut().unwrap(), len) };
-
-    (result, remain)
-}
--- a/libs/neonart/src/epoch.rs
+++ b/libs/neonart/src/epoch.rs
@@ -1,142 +0,0 @@
-//! This is similar to crossbeam_epoch crate, but works in shared memory
-
-use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
-
-use crossbeam_utils::CachePadded;
-
-const NUM_SLOTS: usize = 1000;
-
-/// This is the struct that is stored in shmem
-///
-/// bit 0: is it pinned or not?
-/// rest of the bits are the epoch counter.
-pub struct EpochShared {
-    global_epoch: AtomicU64,
-    participants: [CachePadded<AtomicU64>; NUM_SLOTS],
-
-    broadcast_lock: spin::Mutex<()>,
-}
-
-impl EpochShared {
-    pub fn new() -> EpochShared {
-        EpochShared {
-            global_epoch: AtomicU64::new(2),
-            participants: [const { CachePadded::new(AtomicU64::new(2)) }; NUM_SLOTS],
-            broadcast_lock: spin::Mutex::new(()),
-        }
-    }
-
-    pub fn register(&self) -> LocalHandle {
-        LocalHandle {
-            global: self,
-            last_slot: AtomicUsize::new(0), // todo: choose more intelligently
-        }
-    }
-
-    fn release_pin(&self, slot: usize, _epoch: u64) {
-        let global_epoch = self.global_epoch.load(Ordering::Relaxed);
-        self.participants[slot].store(global_epoch, Ordering::Relaxed);
-    }
-
-    fn pin_internal(&self, slot_hint: usize) -> (usize, u64) {
-        // pick a slot
-        let mut slot = slot_hint;
-        let epoch = loop {
-            let old = self.participants[slot].fetch_or(1, Ordering::Relaxed);
-            if old & 1 == 0 {
-                // Got this slot
-                break old;
-            }
-
-            // the slot was busy by another thread / process. try a different slot
-            slot += 1;
-            if slot == NUM_SLOTS {
-                slot = 0;
-            }
-            continue;
-        };
-        (slot, epoch)
-    }
-
-    pub(crate) fn advance(&self) -> u64 {
-        // Advance the global epoch
-        let old_epoch = self.global_epoch.fetch_add(2, Ordering::Relaxed);
-        // Anyone that release their pin after this will update their slot.
-        old_epoch + 2
-    }
-
-    pub(crate) fn broadcast(&self) {
-        let Some(_guard) = self.broadcast_lock.try_lock() else {
-            return;
-        };
-
-        let epoch = self.global_epoch.load(Ordering::Relaxed);
-        let old_epoch = epoch.wrapping_sub(2);
-
-        // Update all free slots.
-        for i in 0..NUM_SLOTS {
-            // TODO: check result, as a sanity check. It should either be the old epoch, or pinned
-            let _ = self.participants[i].compare_exchange(
-                old_epoch,
-                epoch,
-                Ordering::Relaxed,
-                Ordering::Relaxed,
-            );
-        }
-
-        // FIXME: memory fence here, since we used Relaxed?
-    }
-
-    pub(crate) fn get_oldest(&self) -> u64 {
-        // Read all slots.
-        let now = self.global_epoch.load(Ordering::Relaxed);
-        let mut oldest = now;
-        for i in 0..NUM_SLOTS {
-            let this_epoch = self.participants[i].load(Ordering::Relaxed);
-            let delta = now.wrapping_sub(this_epoch);
-            if delta > u64::MAX / 2 {
-                // this is very recent
-            } else if delta > now.wrapping_sub(oldest) {
-                oldest = this_epoch;
-            }
-        }
-        oldest
-    }
-
-    pub(crate) fn get_current(&self) -> u64 {
-        self.global_epoch.load(Ordering::Relaxed)
-    }
-}
-
-pub(crate) struct EpochPin<'e> {
-    slot: usize,
-    pub(crate) epoch: u64,
-
-    handle: &'e LocalHandle<'e>,
-}
-
-impl<'e> Drop for EpochPin<'e> {
-    fn drop(&mut self) {
-        self.handle.global.release_pin(self.slot, self.epoch);
-    }
-}
-
-pub struct LocalHandle<'g> {
-    global: &'g EpochShared,
-
-    last_slot: AtomicUsize,
-}
-
-impl<'g> LocalHandle<'g> {
-    pub fn pin(&self) -> EpochPin {
-        let (slot, epoch) = self
-            .global
-            .pin_internal(self.last_slot.load(Ordering::Relaxed));
-        self.last_slot.store(slot, Ordering::Relaxed);
-        EpochPin {
-            handle: self,
-            epoch,
-            slot,
-        }
-    }
-}
--- a/libs/neonart/src/lib.rs
+++ b/libs/neonart/src/lib.rs
@@ -1,583 +0,0 @@
-//! Adaptive Radix Tree (ART) implementation, with Optimistic Lock Coupling.
-//!
-//! The data structure is described in these two papers:
-//!
-//! [1] Leis, V. & Kemper, Alfons & Neumann, Thomas. (2013).
-//!     The adaptive radix tree: ARTful indexing for main-memory databases.
-//!     Proceedings - International Conference on Data Engineering. 38-49. 10.1109/ICDE.2013.6544812.
-//!     https://db.in.tum.de/~leis/papers/ART.pdf
-//!
-//! [2] Leis, Viktor & Scheibner, Florian & Kemper, Alfons & Neumann, Thomas. (2016).
-//!     The ART of practical synchronization.
-//!     1-8. 10.1145/2933349.2933352.
-//!     https://db.in.tum.de/~leis/papers/artsync.pdf
-//!
-//! [1] describes the base data structure, and [2] describes the Optimistic Lock Coupling that we
-//! use.
-//!
-//! The papers mention a few different variants. We have made the following choices in this
-//! implementation:
-//!
-//! - All keys have the same length
-//!
-//! - Single-value leaves.
-//!
-//! - For collapsing inner nodes, we use the Pessimistic approach, where each inner node stores a
-//!   variable length "prefix", which stores the keys of all the one-way nodes which have been
-//!   removed. However, similar to the "hybrid" approach described in the paper, each node only has
-//!   space for a constant-size prefix of 8 bytes. If a node would have a longer prefix, then we
-//!   create create one-way nodes to store them. (There was no particular reason for this choice,
-//!   the "hybrid" approach described in the paper might be better.)
-//!
-//! - For concurrency, we use Optimistic Lock Coupling. The paper [2] also describes another method,
-//!   ROWEX, which generally performs better when there is contention, but that is not important
-//!   for use and Optimisic Lock Coupling is simpler to implement.
-//!
-//! ## Requirements
-//!
-//! This data structure is currently used for the integrated LFC, relsize and last-written LSN cache
-//! in the compute communicator, part of the 'neon' Postgres extension. We have some unique
-//! requirements, which is why we had to write our own. Namely:
-//!
-//! - The data structure has to live in fixed-sized shared memory segment. That rules out any
-//!   built-in Rust collections and most crates. (Except possibly with the 'allocator_api' rust
-//!   feature, which still nightly-only experimental as of this writing).
-//!
-//! - The data structure is accessed from multiple processes. Only one process updates the data
-//!   structure, but other processes perform reads. That rules out using built-in Rust locking
-//!   primitives like Mutex and RwLock, and most crates too.
-//!
-//! - Within the one process with write-access, multiple threads can perform updates concurrently.
-//!   That rules out using PostgreSQL LWLocks for the locking.
-//!
-//! The implementation is generic, and doesn't depend on any PostgreSQL specifics, but it has been
-//! written with that usage and the above constraints in mind. Some noteworthy assumptions:
-//!
-//! - Contention is assumed to be rare. In the integrated cache in PostgreSQL, there's higher level
-//!   locking in the PostgreSQL buffer manager, which ensures that two backends should not try to
-//!   read / write the same page at the same time. (Prefetching can conflict with actual reads,
-//!   however.)
-//!
-//!  - The keys in the integrated cache are 17 bytes long.
-//!
-//! ## Usage
-//!
-//! Because this is designed to be used as a Postgres shared memory data structure, initialization
-//! happens in three stages:
-//!
-//! 0. A fixed area of shared memory is allocated at postmaster startup.
-//!
-//! 1. TreeInitStruct::new() is called to initialize it, still in Postmaster process, before any
-//!    other process or thread is running. It returns a TreeInitStruct, which is inherited by all
-//!    the processes through fork().
-//!
-//! 2. One process may have write-access to the struct, by calling
-//!    [TreeInitStruct::attach_writer]. (That process is the communicator process.)
-//!
-//! 3. Other processes get read-access to the struct, by calling [TreeInitStruct::attach_reader]
-//!
-//! "Write access" means that you can insert / update / delete values in the tree.
-//!
-//! NOTE: The Values stored in the tree are sometimes moved, when a leaf node fills up and a new
-//! larger node needs to be allocated. The versioning and epoch-based allocator ensure that the data
-//! structure stays consistent, but if the Value has interior mutability, like atomic fields,
-//! updates to such fields might be lost if the leaf node is concurrently moved! If that becomes a
-//! problem, the version check could be passed up to the caller, so that the caller could detect the
-//! lost updates and retry the operation.
-//!
-//! ## Implementation
-//!
-//! node_ptr: Provides low-level implementations of the four different node types (eight actually,
-//! since there is an Internal and Leaf variant of each)
-//!
-//! lock_and_version.rs: Provides an abstraction for the combined lock and version counter on each
-//! node.
-//!
-//! node_ref.rs: The code in node_ptr.rs deals with raw pointers. node_ref.rs provides more type-safe
-//!   abstractions on top.
-//!
-//! algorithm.rs: Contains the functions to implement lookups and updates in the tree
-//!
-//! allocator.rs: Provides a facility to allocate memory for the tree nodes. (We must provide our
-//!   own abstraction for that because we need the data structure to live in a pre-allocated shared
-//!   memory segment).
-//!
-//! epoch.rs: The data structure requires that when a node is removed from the tree, it is not
-//!   immediately deallocated, but stays around for as long as concurrent readers might still have
-//!   pointers to them. This is enforced by an epoch system. This is similar to
-//!   e.g. crossbeam_epoch, but we couldn't use that either because it has to work across processes
-//!   communicating over the shared memory segment.
-//!
-//! ## See also
-//!
-//! There are some existing Rust ART implementations out there, but none of them filled all
-//! the requirements:
-//!
-//! - https://github.com/XiangpengHao/congee
-//! - https://github.com/declanvk/blart
-//!
-//! ## TODO
-//!
-//! - Removing values has not been implemented
-
-mod algorithm;
-pub mod allocator;
-mod epoch;
-
-use algorithm::RootPtr;
-use algorithm::node_ptr::NodePtr;
-
-use std::collections::VecDeque;
-use std::fmt::Debug;
-use std::marker::PhantomData;
-use std::ptr::NonNull;
-use std::sync::atomic::{AtomicBool, Ordering};
-
-use crate::epoch::EpochPin;
-
-#[cfg(test)]
-mod tests;
-
-use allocator::ArtAllocator;
-pub use allocator::ArtMultiSlabAllocator;
-pub use allocator::OutOfMemoryError;
-
-/// Fixed-length key type.
-///
-pub trait Key: Debug {
-    const KEY_LEN: usize;
-
-    fn as_bytes(&self) -> &[u8];
-}
-
-/// Values stored in the tree
-///
-/// Values need to be Cloneable, because when a node "grows", the value is copied to a new node and
-/// the old sticks around until all readers that might see the old value are gone.
-// fixme obsolete, no longer needs Clone
-pub trait Value {}
-
-const MAX_GARBAGE: usize = 1024;
-
-/// The root of the tree, plus other tree-wide data. This is stored in the shared memory.
-pub struct Tree<V: Value> {
-    /// For simplicity, so that we never need to grow or shrink the root, the root node is always an
-    /// Internal256 node. Also, it never has a prefix (that's actually a bit wasteful, incurring one
-    /// indirection to every lookup)
-    root: RootPtr<V>,
-
-    writer_attached: AtomicBool,
-
-    epoch: epoch::EpochShared,
-}
-
-unsafe impl<V: Value + Sync> Sync for Tree<V> {}
-unsafe impl<V: Value + Send> Send for Tree<V> {}
-
-struct GarbageQueue<V>(VecDeque<(NodePtr<V>, u64)>);
-
-unsafe impl<V: Value + Sync> Sync for GarbageQueue<V> {}
-unsafe impl<V: Value + Send> Send for GarbageQueue<V> {}
-
-impl<V> GarbageQueue<V> {
-    fn new() -> GarbageQueue<V> {
-        GarbageQueue(VecDeque::with_capacity(MAX_GARBAGE))
-    }
-
-    fn remember_obsolete_node(&mut self, ptr: NodePtr<V>, epoch: u64) {
-        self.0.push_front((ptr, epoch));
-    }
-
-    fn next_obsolete(&mut self, cutoff_epoch: u64) -> Option<NodePtr<V>> {
-        if let Some(back) = self.0.back() {
-            if back.1 < cutoff_epoch {
-                return Some(self.0.pop_back().unwrap().0);
-            }
-        }
-        None
-    }
-}
-
-/// Struct created at postmaster startup
-pub struct TreeInitStruct<'t, K: Key, V: Value, A: ArtAllocator<V>> {
-    tree: &'t Tree<V>,
-
-    allocator: &'t A,
-
-    phantom_key: PhantomData<K>,
-}
-
-/// The worker process has a reference to this. The write operations are only safe
-/// from the worker process
-pub struct TreeWriteAccess<'t, K: Key, V: Value, A: ArtAllocator<V>>
-where
-    K: Key,
-    V: Value,
-{
-    tree: &'t Tree<V>,
-
-    pub allocator: &'t A,
-
-    epoch_handle: epoch::LocalHandle<'t>,
-
-    phantom_key: PhantomData<K>,
-
-    /// Obsolete nodes that cannot be recycled until their epoch expires.
-    garbage: spin::Mutex<GarbageQueue<V>>,
-}
-
-/// The backends have a reference to this. It cannot be used to modify the tree
-pub struct TreeReadAccess<'t, K: Key, V: Value>
-where
-    K: Key,
-    V: Value,
-{
-    tree: &'t Tree<V>,
-
-    epoch_handle: epoch::LocalHandle<'t>,
-
-    phantom_key: PhantomData<K>,
-}
-
-impl<'t, K: Key, V: Value, A: ArtAllocator<V>> TreeInitStruct<'t, K, V, A> {
-    pub fn new(allocator: &'t A) -> TreeInitStruct<'t, K, V, A> {
-        let tree_ptr = allocator.alloc_tree();
-        let tree_ptr = NonNull::new(tree_ptr).expect("out of memory");
-        let init = Tree {
-            root: algorithm::new_root(allocator).expect("out of memory"),
-            writer_attached: AtomicBool::new(false),
-            epoch: epoch::EpochShared::new(),
-        };
-        unsafe { tree_ptr.write(init) };
-
-        TreeInitStruct {
-            tree: unsafe { tree_ptr.as_ref() },
-            allocator,
-            phantom_key: PhantomData,
-        }
-    }
-
-    pub fn attach_writer(self) -> TreeWriteAccess<'t, K, V, A> {
-        let previously_attached = self.tree.writer_attached.swap(true, Ordering::Relaxed);
-        if previously_attached {
-            panic!("writer already attached");
-        }
-        TreeWriteAccess {
-            tree: self.tree,
-            allocator: self.allocator,
-            phantom_key: PhantomData,
-            epoch_handle: self.tree.epoch.register(),
-            garbage: spin::Mutex::new(GarbageQueue::new()),
-        }
-    }
-
-    pub fn attach_reader(self) -> TreeReadAccess<'t, K, V> {
-        TreeReadAccess {
-            tree: self.tree,
-            phantom_key: PhantomData,
-            epoch_handle: self.tree.epoch.register(),
-        }
-    }
-}
-
-impl<'t, K: Key, V: Value, A: ArtAllocator<V>> TreeWriteAccess<'t, K, V, A> {
-    pub fn start_write<'g>(&'t self) -> TreeWriteGuard<'g, K, V, A>
-    where
-        't: 'g,
-    {
-        TreeWriteGuard {
-            tree_writer: self,
-            epoch_pin: self.epoch_handle.pin(),
-            phantom_key: PhantomData,
-            created_garbage: false,
-        }
-    }
-
-    pub fn start_read(&'t self) -> TreeReadGuard<'t, K, V> {
-        TreeReadGuard {
-            tree: self.tree,
-            epoch_pin: self.epoch_handle.pin(),
-            phantom_key: PhantomData,
-        }
-    }
-}
-
-impl<'t, K: Key, V: Value> TreeReadAccess<'t, K, V> {
-    pub fn start_read(&'t self) -> TreeReadGuard<'t, K, V> {
-        TreeReadGuard {
-            tree: self.tree,
-            epoch_pin: self.epoch_handle.pin(),
-            phantom_key: PhantomData,
-        }
-    }
-}
-
-pub struct TreeReadGuard<'e, K, V>
-where
-    K: Key,
-    V: Value,
-{
-    tree: &'e Tree<V>,
-
-    epoch_pin: EpochPin<'e>,
-    phantom_key: PhantomData<K>,
-}
-
-impl<'e, K: Key, V: Value> TreeReadGuard<'e, K, V> {
-    pub fn get(&'e self, key: &K) -> Option<&'e V> {
-        algorithm::search(key, self.tree.root, &self.epoch_pin)
-    }
-}
-
-pub struct TreeWriteGuard<'e, K, V, A>
-where
-    K: Key,
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    tree_writer: &'e TreeWriteAccess<'e, K, V, A>,
-
-    epoch_pin: EpochPin<'e>,
-    phantom_key: PhantomData<K>,
-
-    created_garbage: bool,
-}
-
-pub enum UpdateAction<V> {
-    Nothing,
-    Insert(V),
-    Remove,
-}
-
-impl<'e, K: Key, V: Value, A: ArtAllocator<V>> TreeWriteGuard<'e, K, V, A> {
-    /// Get a value
-    pub fn get(&'e mut self, key: &K) -> Option<&'e V> {
-        algorithm::search(key, self.tree_writer.tree.root, &self.epoch_pin)
-    }
-
-    /// Insert a value
-    pub fn insert(self, key: &K, value: V) -> Result<bool, OutOfMemoryError> {
-        let mut success = None;
-
-        self.update_with_fn(key, |existing| {
-            if existing.is_some() {
-                success = Some(false);
-                UpdateAction::Nothing
-            } else {
-                success = Some(true);
-                UpdateAction::Insert(value)
-            }
-        })?;
-        Ok(success.expect("value_fn not called"))
-    }
-
-    /// Remove value. Returns true if it existed
-    pub fn remove(self, key: &K) -> bool {
-        let mut result = false;
-        // FIXME: It's not clear if OOM is expected while removing. It seems
-        // not nice, but shrinking a node can OOM. Then again, we could opt
-        // to not shrink a node if we cannot allocate, to live a little longer.
-        self.update_with_fn(key, |existing| match existing {
-            Some(_) => {
-                result = true;
-                UpdateAction::Remove
-            }
-            None => UpdateAction::Nothing,
-        })
-        .expect("out of memory while removing");
-        result
-    }
-
-    /// Try to remove value and return the old value.
-    pub fn remove_and_return(self, key: &K) -> Option<V>
-    where
-        V: Clone,
-    {
-        let mut old = None;
-        self.update_with_fn(key, |existing| {
-            old = existing.cloned();
-            UpdateAction::Remove
-        })
-        .expect("out of memory while removing");
-        old
-    }
-
-    /// Update key using the given function. All the other modifying operations are based on this.
-    ///
-    /// The function is passed a reference to the existing value, if any. If the function
-    /// returns None, the value is removed from the tree (or if there was no existing value,
-    /// does nothing). If the function returns Some, the existing value is replaced, of if there
-    /// was no existing value, it is inserted. FIXME: update comment
-    pub fn update_with_fn<F>(mut self, key: &K, value_fn: F) -> Result<(), OutOfMemoryError>
-    where
-        F: FnOnce(Option<&V>) -> UpdateAction<V>,
-    {
-        algorithm::update_fn(key, value_fn, self.tree_writer.tree.root, &mut self)?;
-
-        if self.created_garbage {
-            let _ = self.collect_garbage();
-        }
-        Ok(())
-    }
-
-    fn remember_obsolete_node(&mut self, ptr: NodePtr<V>) {
-        self.tree_writer
-            .garbage
-            .lock()
-            .remember_obsolete_node(ptr, self.epoch_pin.epoch);
-        self.created_garbage = true;
-    }
-
-    // returns number of nodes recycled
-    fn collect_garbage(&self) -> usize {
-        self.tree_writer.tree.epoch.advance();
-        self.tree_writer.tree.epoch.broadcast();
-
-        let cutoff_epoch = self.tree_writer.tree.epoch.get_oldest();
-
-        let mut result = 0;
-        let mut garbage_queue = self.tree_writer.garbage.lock();
-        while let Some(ptr) = garbage_queue.next_obsolete(cutoff_epoch) {
-            ptr.deallocate(self.tree_writer.allocator);
-            result += 1;
-        }
-        result
-    }
-}
-
-pub struct TreeIterator<K>
-where
-    K: Key + for<'a> From<&'a [u8]>,
-{
-    done: bool,
-    pub next_key: Vec<u8>,
-    max_key: Option<Vec<u8>>,
-
-    phantom_key: PhantomData<K>,
-}
-
-impl<K> TreeIterator<K>
-where
-    K: Key + for<'a> From<&'a [u8]>,
-{
-    pub fn new_wrapping() -> TreeIterator<K> {
-        TreeIterator {
-            done: false,
-            next_key: vec![0; K::KEY_LEN],
-            max_key: None,
-            phantom_key: PhantomData,
-        }
-    }
-
-    pub fn new(range: &std::ops::Range<K>) -> TreeIterator<K> {
-        let result = TreeIterator {
-            done: false,
-            next_key: Vec::from(range.start.as_bytes()),
-            max_key: Some(Vec::from(range.end.as_bytes())),
-            phantom_key: PhantomData,
-        };
-        assert_eq!(result.next_key.len(), K::KEY_LEN);
-        assert_eq!(result.max_key.as_ref().unwrap().len(), K::KEY_LEN);
-
-        result
-    }
-
-    pub fn next<'g, V>(&mut self, read_guard: &'g TreeReadGuard<'g, K, V>) -> Option<(K, &'g V)>
-    where
-        V: Value,
-    {
-        if self.done {
-            return None;
-        }
-
-        let mut wrapped_around = false;
-        loop {
-            assert_eq!(self.next_key.len(), K::KEY_LEN);
-            if let Some((k, v)) =
-                algorithm::iter_next(&self.next_key, read_guard.tree.root, &read_guard.epoch_pin)
-            {
-                assert_eq!(k.len(), K::KEY_LEN);
-                assert_eq!(self.next_key.len(), K::KEY_LEN);
-
-                // Check if we reached the end of the range
-                if let Some(max_key) = &self.max_key {
-                    if k.as_slice() >= max_key.as_slice() {
-                        self.done = true;
-                        break None;
-                    }
-                }
-
-                // increment the key
-                self.next_key = k.clone();
-                increment_key(self.next_key.as_mut_slice());
-                let k = k.as_slice().into();
-
-                break Some((k, v));
-            } else {
-                if self.max_key.is_some() {
-                    self.done = true;
-                } else {
-                    // Start from beginning
-                    if !wrapped_around {
-                        for i in 0..K::KEY_LEN {
-                            self.next_key[i] = 0;
-                        }
-                        wrapped_around = true;
-                        continue;
-                    } else {
-                        // The tree is completely empty
-                        // FIXME: perhaps we should remember the starting point instead.
-                        // Currently this will scan some ranges twice.
-                        break None;
-                    }
-                }
-                break None;
-            }
-        }
-    }
-}
-
-fn increment_key(key: &mut [u8]) -> bool {
-    for i in (0..key.len()).rev() {
-        let (byte, overflow) = key[i].overflowing_add(1);
-        key[i] = byte;
-        if !overflow {
-            return false;
-        }
-    }
-    true
-}
-
-// Debugging functions
-impl<'e, K: Key, V: Value + Debug, A: ArtAllocator<V>> TreeWriteGuard<'e, K, V, A> {
-    pub fn dump(&mut self, dst: &mut dyn std::io::Write) {
-        algorithm::dump_tree(self.tree_writer.tree.root, &self.epoch_pin, dst)
-    }
-}
-impl<'e, K: Key, V: Value + Debug> TreeReadGuard<'e, K, V> {
-    pub fn dump(&mut self, dst: &mut dyn std::io::Write) {
-        algorithm::dump_tree(self.tree.root, &self.epoch_pin, dst)
-    }
-}
-impl<'e, K: Key, V: Value> TreeWriteAccess<'e, K, V, ArtMultiSlabAllocator<'e, V>> {
-    pub fn get_statistics(&self) -> ArtTreeStatistics {
-        self.allocator.get_statistics();
-        ArtTreeStatistics {
-            blocks: self.allocator.inner.block_allocator.get_statistics(),
-            slabs: self.allocator.get_statistics(),
-            epoch: self.tree.epoch.get_current(),
-            oldest_epoch: self.tree.epoch.get_oldest(),
-            num_garbage: self.garbage.lock().0.len() as u64,
-        }
-    }
-}
-
-#[derive(Clone, Debug)]
-pub struct ArtTreeStatistics {
-    pub blocks: allocator::block::BlockAllocatorStats,
-    pub slabs: allocator::ArtMultiSlabStats,
-
-    pub epoch: u64,
-    pub oldest_epoch: u64,
-    pub num_garbage: u64,
-}
--- a/libs/neonart/src/tests.rs
+++ b/libs/neonart/src/tests.rs
@@ -1,236 +0,0 @@
-use std::collections::BTreeMap;
-use std::collections::HashSet;
-use std::fmt::{Debug, Formatter};
-use std::sync::atomic::{AtomicUsize, Ordering};
-
-use crate::ArtAllocator;
-use crate::ArtMultiSlabAllocator;
-use crate::TreeInitStruct;
-use crate::TreeIterator;
-use crate::TreeWriteAccess;
-use crate::UpdateAction;
-
-use crate::{Key, Value};
-
-use rand::Rng;
-use rand::seq::SliceRandom;
-use rand_distr::Zipf;
-
-const TEST_KEY_LEN: usize = 16;
-
-#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord)]
-struct TestKey([u8; TEST_KEY_LEN]);
-
-impl TestKey {
-    const MIN: TestKey = TestKey([0; TEST_KEY_LEN]);
-    const MAX: TestKey = TestKey([u8::MAX; TEST_KEY_LEN]);
-}
-
-impl Key for TestKey {
-    const KEY_LEN: usize = TEST_KEY_LEN;
-    fn as_bytes(&self) -> &[u8] {
-        &self.0
-    }
-}
-
-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())
-    }
-}
-
-impl Value for usize {}
-
-fn test_inserts<K: Into<TestKey> + Copy>(keys: &[K]) {
-    const MEM_SIZE: usize = 10000000;
-    let mut area = Box::new_uninit_slice(MEM_SIZE);
-
-    let allocator = ArtMultiSlabAllocator::new(&mut area);
-
-    let init_struct = TreeInitStruct::<TestKey, usize, _>::new(allocator);
-    let tree_writer = init_struct.attach_writer();
-
-    for (idx, k) in keys.iter().enumerate() {
-        let w = tree_writer.start_write();
-        let res = w.insert(&(*k).into(), idx);
-        assert!(res.is_ok());
-    }
-
-    for (idx, k) in keys.iter().enumerate() {
-        let r = tree_writer.start_read();
-        let value = r.get(&(*k).into());
-        assert_eq!(value, Some(idx).as_ref());
-    }
-
-    eprintln!("stats: {:?}", tree_writer.get_statistics());
-}
-
-#[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);
-}
-
-struct TestValue(AtomicUsize);
-
-impl TestValue {
-    fn new(val: usize) -> TestValue {
-        TestValue(AtomicUsize::new(val))
-    }
-
-    fn load(&self) -> usize {
-        self.0.load(Ordering::Relaxed)
-    }
-}
-
-impl Value for TestValue {}
-
-impl Clone for TestValue {
-    fn clone(&self) -> TestValue {
-        TestValue::new(self.load())
-    }
-}
-
-impl Debug for TestValue {
-    fn fmt(&self, fmt: &mut Formatter<'_>) -> Result<(), std::fmt::Error> {
-        write!(fmt, "{:?}", self.load())
-    }
-}
-
-#[derive(Clone, Debug)]
-struct TestOp(TestKey, Option<usize>);
-
-fn apply_op<A: ArtAllocator<TestValue>>(
-    op: &TestOp,
-    tree: &TreeWriteAccess<TestKey, TestValue, A>,
-    shadow: &mut BTreeMap<TestKey, usize>,
-) {
-    eprintln!("applying op: {op:?}");
-
-    // 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)
-    };
-
-    // apply to Art tree
-    let w = tree.start_write();
-    w.update_with_fn(&op.0, |existing| {
-        assert_eq!(existing.map(TestValue::load), shadow_existing);
-
-        match (existing, op.1) {
-            (None, None) => UpdateAction::Nothing,
-            (None, Some(new_val)) => UpdateAction::Insert(TestValue::new(new_val)),
-            (Some(_old_val), None) => UpdateAction::Remove,
-            (Some(old_val), Some(new_val)) => {
-                old_val.0.store(new_val, Ordering::Relaxed);
-                UpdateAction::Nothing
-            }
-        }
-    })
-    .expect("out of memory");
-}
-
-fn test_iter<A: ArtAllocator<TestValue>>(
-    tree: &TreeWriteAccess<TestKey, TestValue, A>,
-    shadow: &BTreeMap<TestKey, usize>,
-) {
-    let mut shadow_iter = shadow.iter();
-    let mut iter = TreeIterator::new(&(TestKey::MIN..TestKey::MAX));
-
-    loop {
-        let shadow_item = shadow_iter.next().map(|(k, v)| (*k, *v));
-        let r = tree.start_read();
-        let item = iter.next(&r);
-
-        if shadow_item != item.map(|(k, v)| (k, v.load())) {
-            eprintln!("FAIL: iterator returned {item:?}, expected {shadow_item:?}");
-            tree.start_read().dump(&mut std::io::stderr());
-
-            eprintln!("SHADOW:");
-            for si in shadow {
-                eprintln!("key: {:?}, val: {}", si.0, si.1);
-            }
-            panic!("FAIL: iterator returned {item:?}, expected {shadow_item:?}");
-        }
-        if item.is_none() {
-            break;
-        }
-    }
-}
-
-#[test]
-fn random_ops() {
-    const MEM_SIZE: usize = 10000000;
-    let mut area = Box::new_uninit_slice(MEM_SIZE);
-
-    let allocator = ArtMultiSlabAllocator::new(&mut area);
-
-    let init_struct = TreeInitStruct::<TestKey, TestValue, _>::new(allocator);
-    let tree_writer = init_struct.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 mut key: TestKey = (rng.sample(distribution) as u128).into();
-
-        if rng.random_bool(0.10) {
-            key = TestKey::from(u128::from(&key) | 0xffffffff);
-        }
-
-        let op = TestOp(key, if rng.random_bool(0.75) { Some(i) } else { None });
-
-        apply_op(&op, &tree_writer, &mut shadow);
-
-        if i % 1000 == 0 {
-            eprintln!("{i} ops processed");
-            eprintln!("stats: {:?}", tree_writer.get_statistics());
-            test_iter(&tree_writer, &shadow);
-        }
-    }
-}
--- a/libs/pageserver_api/src/controller_api.rs
+++ b/libs/pageserver_api/src/controller_api.rs
@@ -1,5 +1,6 @@
 use std::collections::{HashMap, HashSet};
 use std::fmt::Display;
+use std::net::IpAddr;
 use std::str::FromStr;
 use std::time::{Duration, Instant};

@@ -10,7 +11,7 @@ use serde::{Deserialize, Serialize};
 use utils::id::{NodeId, TenantId, TimelineId};
 use utils::lsn::Lsn;

-use crate::models::{PageserverUtilization, ShardParameters, TenantConfig};
+use crate::models::{PageserverUtilization, ShardParameters, TenantConfig, TimelineInfo};
 use crate::shard::{ShardStripeSize, TenantShardId};

 #[derive(Serialize, Deserialize, Debug)]
@@ -60,6 +61,11 @@ pub struct NodeRegisterRequest {
    pub listen_https_port: Option<u16>,

    pub availability_zone_id: AvailabilityZone,
+
+    // Reachable IP address of the PS/SK registering, if known.
+    // Hadron Cluster Coordiantor will update the DNS record of the registering node
+    // with this IP address.
+    pub node_ip_addr: Option<IpAddr>,
 }

 #[derive(Serialize, Deserialize)]
@@ -126,6 +132,13 @@ pub struct TenantDescribeResponse {
    pub config: TenantConfig,
 }

+#[derive(Serialize, Deserialize, Debug)]
+pub struct TenantTimelineDescribeResponse {
+    pub shards: Vec<TimelineInfo>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_consistent_lsn: Option<Lsn>,
+}
+
 #[derive(Serialize, Deserialize, Debug)]
 pub struct NodeShardResponse {
    pub node_id: NodeId,
@@ -538,6 +551,39 @@ pub struct SafekeeperDescribeResponse {
    pub scheduling_policy: SkSchedulingPolicy,
 }

+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct TimelineSafekeeperPeer {
+    pub node_id: NodeId,
+    pub listen_http_addr: String,
+    pub http_port: i32,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SCSafekeeperTimeline {
+    // SC does not know the tenant id.
+    pub timeline_id: TimelineId,
+    pub peers: Vec<NodeId>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SCSafekeeperTimelinesResponse {
+    pub timelines: Vec<SCSafekeeperTimeline>,
+    pub safekeeper_peers: Vec<TimelineSafekeeperPeer>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SafekeeperTimeline {
+    pub tenant_id: TenantId,
+    pub timeline_id: TimelineId,
+    pub peers: Vec<NodeId>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SafekeeperTimelinesResponse {
+    pub timelines: Vec<SafekeeperTimeline>,
+    pub safekeeper_peers: Vec<TimelineSafekeeperPeer>,
+}
+
 #[derive(Serialize, Deserialize, Clone)]
 pub struct SafekeeperSchedulingPolicyRequest {
    pub scheduling_policy: SkSchedulingPolicy,
--- a/libs/pageserver_api/src/models.rs
+++ b/libs/pageserver_api/src/models.rs
@@ -1622,6 +1622,9 @@ pub struct TimelineInfo {

    /// Whether the timeline is invisible in synthetic size calculations.
    pub is_invisible: Option<bool>,
+    // HADRON: the largest LSN below which all page updates have been included in the image layers.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_consistent_lsn: Option<Lsn>,
 }

 #[derive(Debug, Clone, Serialize, Deserialize)]
--- a/libs/postgres_ffi/build.rs
+++ b/libs/postgres_ffi/build.rs
@@ -110,7 +110,6 @@ fn main() -> anyhow::Result<()> {
            .allowlist_type("XLogRecPtr")
            .allowlist_type("XLogSegNo")
            .allowlist_type("TimeLineID")
-            .allowlist_type("TimestampTz")
            .allowlist_type("MultiXactId")
            .allowlist_type("MultiXactOffset")
            .allowlist_type("MultiXactStatus")
--- a/libs/postgres_ffi/src/lib.rs
+++ b/libs/postgres_ffi/src/lib.rs
@@ -227,8 +227,7 @@ pub mod walrecord;
 // Export some widely used datatypes that are unlikely to change across Postgres versions
 pub use v14::bindings::{
    BlockNumber, CheckPoint, ControlFileData, MultiXactId, OffsetNumber, Oid, PageHeaderData,
-    RepOriginId, TimeLineID, TimestampTz, TransactionId, XLogRecPtr, XLogRecord, XLogSegNo, uint32,
-    uint64,
+    RepOriginId, TimeLineID, TransactionId, XLogRecPtr, XLogRecord, XLogSegNo, uint32, uint64,
 };
 // Likewise for these, although the assumption that these don't change is a little more iffy.
 pub use v14::bindings::{MultiXactOffset, MultiXactStatus};
--- a/libs/postgres_ffi/src/walrecord.rs
+++ b/libs/postgres_ffi/src/walrecord.rs
@@ -4,13 +4,14 @@
 //! TODO: Generate separate types for each supported PG version

 use bytes::{Buf, Bytes};
+use postgres_ffi_types::TimestampTz;
 use serde::{Deserialize, Serialize};
 use utils::bin_ser::DeserializeError;
 use utils::lsn::Lsn;

 use crate::{
    BLCKSZ, BlockNumber, MultiXactId, MultiXactOffset, MultiXactStatus, Oid, PgMajorVersion,
-    RepOriginId, TimestampTz, TransactionId, XLOG_SIZE_OF_XLOG_RECORD, XLogRecord, pg_constants,
+    RepOriginId, TransactionId, XLOG_SIZE_OF_XLOG_RECORD, XLogRecord, pg_constants,
 };

 #[repr(C)]
@@ -863,7 +864,8 @@ pub mod v17 {
        XlHeapDelete, XlHeapInsert, XlHeapLock, XlHeapMultiInsert, XlHeapUpdate, XlParameterChange,
        rm_neon,
    };
-    pub use crate::{TimeLineID, TimestampTz};
+    pub use crate::TimeLineID;
+    pub use postgres_ffi_types::TimestampTz;

    #[repr(C)]
    #[derive(Debug)]
--- a/libs/postgres_ffi/src/xlog_utils.rs
+++ b/libs/postgres_ffi/src/xlog_utils.rs
@@ -9,10 +9,11 @@

 use super::super::waldecoder::WalStreamDecoder;
 use super::bindings::{
-    CheckPoint, ControlFileData, DBState_DB_SHUTDOWNED, FullTransactionId, TimeLineID, TimestampTz,
+    CheckPoint, ControlFileData, DBState_DB_SHUTDOWNED, FullTransactionId, TimeLineID,
    XLogLongPageHeaderData, XLogPageHeaderData, XLogRecPtr, XLogRecord, XLogSegNo, XLOG_PAGE_MAGIC,
    MY_PGVERSION
 };
+use postgres_ffi_types::TimestampTz;
 use super::wal_generator::LogicalMessageGenerator;
 use crate::pg_constants;
 use crate::PG_TLI;
--- a/libs/postgres_ffi_types/src/lib.rs
+++ b/libs/postgres_ffi_types/src/lib.rs
@@ -11,3 +11,4 @@ pub mod forknum;

 pub type Oid = u32;
 pub type RepOriginId = u16;
+pub type TimestampTz = i64;
--- a/libs/remote_storage/src/simulate_failures.rs
+++ b/libs/remote_storage/src/simulate_failures.rs
@@ -31,6 +31,7 @@ pub struct UnreliableWrapper {
    /* BEGIN_HADRON */
    // This the probability of failure for each operation, ranged from [0, 100].
    // The probability is default to 100, which means that all operations will fail.
+    // Storage will fail by probability up to attempts_to_fail times.
    attempt_failure_probability: u64,
    /* END_HADRON */
 }
--- a/libs/safekeeper_api/Cargo.toml
+++ b/libs/safekeeper_api/Cargo.toml
@@ -9,7 +9,7 @@ anyhow.workspace = true
 const_format.workspace = true
 serde.workspace = true
 serde_json.workspace = true
-postgres_ffi.workspace = true
+postgres_ffi_types.workspace = true
 postgres_versioninfo.workspace = true
 pq_proto.workspace = true
 tokio.workspace = true
--- a/libs/safekeeper_api/src/models.rs
+++ b/libs/safekeeper_api/src/models.rs
@@ -3,7 +3,7 @@
 use std::net::SocketAddr;

 use pageserver_api::shard::ShardIdentity;
-use postgres_ffi::TimestampTz;
+use postgres_ffi_types::TimestampTz;
 use postgres_versioninfo::PgVersionId;
 use serde::{Deserialize, Serialize};
 use tokio::time::Instant;
@@ -11,7 +11,7 @@ use utils::id::{NodeId, TenantId, TenantTimelineId, TimelineId};
 use utils::lsn::Lsn;
 use utils::pageserver_feedback::PageserverFeedback;

-use crate::membership::Configuration;
+use crate::membership::{Configuration, SafekeeperGeneration};
 use crate::{ServerInfo, Term};

 #[derive(Debug, Serialize, Deserialize)]
@@ -311,3 +311,12 @@ pub struct PullTimelineResponse {
    pub safekeeper_host: Option<String>,
    // TODO: add more fields?
 }
+
+/// Response to a timeline locate request.
+/// Storcon-only API.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct TimelineLocateResponse {
+    pub generation: SafekeeperGeneration,
+    pub sk_set: Vec<NodeId>,
+    pub new_sk_set: Option<Vec<NodeId>>,
+}
--- a/libs/utils/src/env.rs
+++ b/libs/utils/src/env.rs
@@ -47,6 +47,7 @@ where

 /* BEGIN_HADRON */
 pub enum DeploymentMode {
+    Local,
    Dev,
    Staging,
    Prod,
@@ -64,7 +65,7 @@ pub fn get_deployment_mode() -> Option<DeploymentMode> {
            }
        },
        Err(_) => {
-            tracing::error!("DEPLOYMENT_MODE not set");
+            // tracing::error!("DEPLOYMENT_MODE not set");
            None
        }
    }
--- a/libs/utils/src/ip_address.rs
+++ b/libs/utils/src/ip_address.rs
@@ -0,0 +1,73 @@
+use std::env::{VarError, var};
+use std::error::Error;
+use std::net::IpAddr;
+use std::str::FromStr;
+
+/// Name of the environment variable containing the reachable IP address of the node. If set, the IP address contained in this
+/// environment variable is used as the reachable IP address of the pageserver or safekeeper node during node registration.
+/// In a Kubernetes environment, this environment variable should be set by Kubernetes to the Pod IP (specified in the Pod
+/// template).
+pub const HADRON_NODE_IP_ADDRESS: &str = "HADRON_NODE_IP_ADDRESS";
+
+/// Read the reachable IP address of this page server from env var HADRON_NODE_IP_ADDRESS.
+/// In Kubernetes this environment variable is set to the Pod IP (specified in the Pod template).
+pub fn read_node_ip_addr_from_env() -> Result<Option<IpAddr>, Box<dyn Error>> {
+    match var(HADRON_NODE_IP_ADDRESS) {
+        Ok(v) => {
+            if let Ok(addr) = IpAddr::from_str(&v) {
+                Ok(Some(addr))
+            } else {
+                Err(format!("Invalid IP address string: {v}. Cannot be parsed as either an IPv4 or an IPv6 address.").into())
+            }
+        }
+        Err(VarError::NotPresent) => Ok(None),
+        Err(e) => Err(e.into()),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::env;
+    use std::net::{Ipv4Addr, Ipv6Addr};
+
+    #[test]
+    fn test_read_node_ip_addr_from_env() {
+        // SAFETY: test code
+        unsafe {
+            // Test with a valid IPv4 address
+            env::set_var(HADRON_NODE_IP_ADDRESS, "192.168.1.1");
+            let result = read_node_ip_addr_from_env().unwrap();
+            assert_eq!(result, Some(IpAddr::V4(Ipv4Addr::new(192, 168, 1, 1))));
+
+            // Test with a valid IPv6 address
+            env::set_var(
+                HADRON_NODE_IP_ADDRESS,
+                "2001:0db8:85a3:0000:0000:8a2e:0370:7334",
+            );
+        }
+        let result = read_node_ip_addr_from_env().unwrap();
+        assert_eq!(
+            result,
+            Some(IpAddr::V6(
+                Ipv6Addr::from_str("2001:0db8:85a3:0000:0000:8a2e:0370:7334").unwrap()
+            ))
+        );
+
+        // Test with an invalid IP address
+        // SAFETY: test code
+        unsafe {
+            env::set_var(HADRON_NODE_IP_ADDRESS, "invalid_ip");
+        }
+        let result = read_node_ip_addr_from_env();
+        assert!(result.is_err());
+
+        // Test with no environment variable set
+        // SAFETY: test code
+        unsafe {
+            env::remove_var(HADRON_NODE_IP_ADDRESS);
+        }
+        let result = read_node_ip_addr_from_env().unwrap();
+        assert_eq!(result, None);
+    }
+}
--- a/libs/utils/src/lib.rs
+++ b/libs/utils/src/lib.rs
@@ -26,6 +26,9 @@ pub mod auth;
 // utility functions and helper traits for unified unique id generation/serialization etc.
 pub mod id;

+// utility functions to obtain reachable IP addresses in PS/SK nodes.
+pub mod ip_address;
+
 pub mod shard;

 mod hex;
--- a/libs/utils/src/logging.rs
+++ b/libs/utils/src/logging.rs
@@ -1,4 +1,5 @@
 use std::future::Future;
+use std::pin::Pin;
 use std::str::FromStr;
 use std::time::Duration;

@@ -7,7 +8,7 @@ use metrics::{IntCounter, IntCounterVec};
 use once_cell::sync::Lazy;
 use strum_macros::{EnumString, VariantNames};
 use tokio::time::Instant;
-use tracing::info;
+use tracing::{info, warn};

 /// Logs a critical error, similarly to `tracing::error!`. This will:
 ///
@@ -377,10 +378,11 @@ impl std::fmt::Debug for SecretString {
 ///
 /// TODO: consider upgrading this to a warning, but currently it fires too often.
 #[inline]
-pub async fn log_slow<F, O>(name: &str, threshold: Duration, f: std::pin::Pin<&mut F>) -> O
-where
-    F: Future<Output = O>,
-{
+pub async fn log_slow<O>(
+    name: &str,
+    threshold: Duration,
+    f: Pin<&mut impl Future<Output = O>>,
+) -> O {
    monitor_slow_future(
        threshold,
        threshold, // period = threshold
@@ -394,16 +396,42 @@ where
            if !is_slow {
                return;
            }
+            let elapsed = elapsed_total.as_secs_f64();
            if ready {
-                info!(
-                    "slow {name} completed after {:.3}s",
-                    elapsed_total.as_secs_f64()
-                );
+                info!("slow {name} completed after {elapsed:.3}s");
            } else {
-                info!(
-                    "slow {name} still running after {:.3}s",
-                    elapsed_total.as_secs_f64()
-                );
+                info!("slow {name} still running after {elapsed:.3}s");
+            }
+        },
+    )
+    .await
+}
+
+/// Logs a periodic warning if a future is slow to complete.
+#[inline]
+pub async fn warn_slow<O>(
+    name: &str,
+    threshold: Duration,
+    f: Pin<&mut impl Future<Output = O>>,
+) -> O {
+    monitor_slow_future(
+        threshold,
+        threshold, // period = threshold
+        f,
+        |MonitorSlowFutureCallback {
+             ready,
+             is_slow,
+             elapsed_total,
+             elapsed_since_last_callback: _,
+         }| {
+            if !is_slow {
+                return;
+            }
+            let elapsed = elapsed_total.as_secs_f64();
+            if ready {
+                warn!("slow {name} completed after {elapsed:.3}s");
+            } else {
+                warn!("slow {name} still running after {elapsed:.3}s");
            }
        },
    )
@@ -416,7 +444,7 @@ where
 pub async fn monitor_slow_future<F, O>(
    threshold: Duration,
    period: Duration,
-    mut fut: std::pin::Pin<&mut F>,
+    mut fut: Pin<&mut F>,
    mut cb: impl FnMut(MonitorSlowFutureCallback),
 ) -> O
 where
--- a/libs/wal_decoder/src/models/record.rs
+++ b/libs/wal_decoder/src/models/record.rs
@@ -2,7 +2,8 @@

 use bytes::Bytes;
 use postgres_ffi::walrecord::{MultiXactMember, describe_postgres_wal_record};
-use postgres_ffi::{MultiXactId, MultiXactOffset, TimestampTz, TransactionId};
+use postgres_ffi::{MultiXactId, MultiXactOffset, TransactionId};
+use postgres_ffi_types::TimestampTz;
 use serde::{Deserialize, Serialize};
 use utils::bin_ser::DeserializeError;

--- a/pageserver/Cargo.toml
+++ b/pageserver/Cargo.toml
@@ -54,7 +54,6 @@ pageserver_api.workspace = true
 pageserver_client.workspace = true # for ResponseErrorMessageExt TOOD refactor that
 pageserver_compaction.workspace = true
 pageserver_page_api.workspace = true
-peekable.workspace = true
 pem.workspace = true
 pin-project-lite.workspace = true
 postgres_backend.workspace = true
@@ -67,7 +66,6 @@ postgres-types.workspace = true
 posthog_client_lite.workspace = true
 pprof.workspace = true
 pq_proto.workspace = true
-prost.workspace = true
 rand.workspace = true
 range-set-blaze = { version = "0.1.16", features = ["alloc"] }
 regex.workspace = true
--- a/pageserver/client_grpc/src/client.rs
+++ b/pageserver/client_grpc/src/client.rs
@@ -1,13 +1,16 @@
 use std::collections::HashMap;
 use std::num::NonZero;
+use std::pin::pin;
 use std::sync::Arc;
+use std::time::{Duration, Instant};

 use anyhow::anyhow;
 use arc_swap::ArcSwap;
 use futures::stream::FuturesUnordered;
 use futures::{FutureExt as _, StreamExt as _};
 use tonic::codec::CompressionEncoding;
-use tracing::instrument;
+use tracing::{debug, instrument};
+use utils::logging::warn_slow;

 use crate::pool::{ChannelPool, ClientGuard, ClientPool, StreamGuard, StreamPool};
 use crate::retry::Retry;
@@ -21,28 +24,40 @@ use utils::shard::{ShardCount, ShardIndex, ShardNumber};
 /// Max number of concurrent clients per channel (i.e. TCP connection). New channels will be spun up
 /// when full.
 ///
+/// Normal requests are small, and we don't pipeline them, so we can afford a large number of
+/// streams per connection.
+///
 /// TODO: tune all of these constants, and consider making them configurable.
-/// TODO: consider separate limits for unary and streaming clients, so we don't fill up channels
-/// with only streams.
-const MAX_CLIENTS_PER_CHANNEL: NonZero<usize> = NonZero::new(16).unwrap();
+const MAX_CLIENTS_PER_CHANNEL: NonZero<usize> = NonZero::new(64).unwrap();

-/// Max number of concurrent unary request clients per shard.
-const MAX_UNARY_CLIENTS: NonZero<usize> = NonZero::new(64).unwrap();
+/// Max number of concurrent bulk GetPage streams per channel (i.e. TCP connection). These use a
+/// dedicated channel pool with a lower client limit, to avoid TCP-level head-of-line blocking and
+/// transmission delays. This also concentrates large window sizes on a smaller set of
+/// streams/connections, presumably reducing memory use.
+const MAX_BULK_CLIENTS_PER_CHANNEL: NonZero<usize> = NonZero::new(16).unwrap();

-/// Max number of concurrent GetPage streams per shard. The max number of concurrent GetPage
-/// requests is given by `MAX_STREAMS * MAX_STREAM_QUEUE_DEPTH`.
-const MAX_STREAMS: NonZero<usize> = NonZero::new(64).unwrap();
+/// The batch size threshold at which a GetPage request will use the bulk stream pool.
+///
+/// The gRPC initial window size is 64 KB. Each page is 8 KB, so let's avoid increasing the window
+/// size for the normal stream pool, and route requests for >= 5 pages (>32 KB) to the bulk pool.
+const BULK_THRESHOLD_BATCH_SIZE: usize = 5;

-/// Max number of pipelined requests per stream.
-const MAX_STREAM_QUEUE_DEPTH: NonZero<usize> = NonZero::new(2).unwrap();
+/// The overall request call timeout, including retries and pool acquisition.
+/// TODO: should we retry forever? Should the caller decide?
+const CALL_TIMEOUT: Duration = Duration::from_secs(60);

-/// Max number of concurrent bulk GetPage streams per shard, used e.g. for prefetches. Because these
-/// are more throughput-oriented, we have a smaller limit but higher queue depth.
-const MAX_BULK_STREAMS: NonZero<usize> = NonZero::new(16).unwrap();
+/// The per-request (retry attempt) timeout, including any lazy connection establishment.
+const REQUEST_TIMEOUT: Duration = Duration::from_secs(10);

-/// Max number of pipelined requests per bulk stream. These are more throughput-oriented and thus
-/// get a larger queue depth.
-const MAX_BULK_STREAM_QUEUE_DEPTH: NonZero<usize> = NonZero::new(4).unwrap();
+/// The initial request retry backoff duration. The first retry does not back off.
+/// TODO: use a different backoff for ResourceExhausted (rate limiting)? Needs server support.
+const BASE_BACKOFF: Duration = Duration::from_millis(5);
+
+/// The maximum request retry backoff duration.
+const MAX_BACKOFF: Duration = Duration::from_secs(5);
+
+/// Threshold and interval for warning about slow operation.
+const SLOW_THRESHOLD: Duration = Duration::from_secs(3);

 /// A rich Pageserver gRPC client for a single tenant timeline. This client is more capable than the
 /// basic `page_api::Client` gRPC client, and supports:
@@ -50,10 +65,19 @@ const MAX_BULK_STREAM_QUEUE_DEPTH: NonZero<usize> = NonZero::new(4).unwrap();
 /// * Sharded tenants across multiple Pageservers.
 /// * Pooling of connections, clients, and streams for efficient resource use.
 /// * Concurrent use by many callers.
-/// * Internal handling of GetPage bidirectional streams, with pipelining and error handling.
+/// * Internal handling of GetPage bidirectional streams.
 /// * Automatic retries.
 /// * Observability.
 ///
+/// The client has dedicated connection/client/stream pools per shard, for resource reuse. These
+/// pools are unbounded: we allow scaling out as many concurrent streams as needed to serve all
+/// concurrent callers, which mostly eliminates head-of-line blocking. Idle streams are fairly
+/// cheap: the server task currently uses 26 KB of memory, so we can comfortably fit 100,000
+/// concurrent idle streams (2.5 GB memory). The worst case degenerates to the old libpq case with
+/// one stream per backend, but without the TCP connection overhead. In the common case we expect
+/// significantly lower stream counts due to stream sharing, driven e.g. by idle backends, LFC hits,
+/// read coalescing, sharding (backends typically only talk to one shard at a time), etc.
+///
 /// TODO: this client does not support base backups or LSN leases, as these are only used by
 /// compute_ctl. Consider adding this, but LSN leases need concurrent requests on all shards.
 pub struct PageserverClient {
@@ -67,8 +91,6 @@ pub struct PageserverClient {
    compression: Option<CompressionEncoding>,
    /// The shards for this tenant.
    shards: ArcSwap<Shards>,
-    /// The retry configuration.
-    retry: Retry,
 }

 impl PageserverClient {
@@ -94,7 +116,6 @@ impl PageserverClient {
            auth_token,
            compression,
            shards: ArcSwap::new(Arc::new(shards)),
-            retry: Retry,
        })
    }

@@ -142,13 +163,15 @@ impl PageserverClient {
        &self,
        req: page_api::CheckRelExistsRequest,
    ) -> tonic::Result<page_api::CheckRelExistsResponse> {
-        self.retry
-            .with(async |_| {
-                // Relation metadata is only available on shard 0.
-                let mut client = self.shards.load_full().get_zero().client().await?;
-                client.check_rel_exists(req).await
-            })
-            .await
+        debug!("sending request: {req:?}");
+        let resp = Self::with_retries(CALL_TIMEOUT, async |_| {
+            // Relation metadata is only available on shard 0.
+            let mut client = self.shards.load_full().get_zero().client().await?;
+            Self::with_timeout(REQUEST_TIMEOUT, client.check_rel_exists(req)).await
+        })
+        .await?;
+        debug!("received response: {resp:?}");
+        Ok(resp)
    }

    /// Returns the total size of a database, as # of bytes.
@@ -157,13 +180,15 @@ impl PageserverClient {
        &self,
        req: page_api::GetDbSizeRequest,
    ) -> tonic::Result<page_api::GetDbSizeResponse> {
-        self.retry
-            .with(async |_| {
-                // Relation metadata is only available on shard 0.
-                let mut client = self.shards.load_full().get_zero().client().await?;
-                client.get_db_size(req).await
-            })
-            .await
+        debug!("sending request: {req:?}");
+        let resp = Self::with_retries(CALL_TIMEOUT, async |_| {
+            // Relation metadata is only available on shard 0.
+            let mut client = self.shards.load_full().get_zero().client().await?;
+            Self::with_timeout(REQUEST_TIMEOUT, client.get_db_size(req)).await
+        })
+        .await?;
+        debug!("received response: {resp:?}");
+        Ok(resp)
    }

    /// Fetches pages. The `request_id` must be unique across all in-flight requests, and the
@@ -193,6 +218,8 @@ impl PageserverClient {
            return Err(tonic::Status::invalid_argument("request attempt must be 0"));
        }

+        debug!("sending request: {req:?}");
+
        // The shards may change while we're fetching pages. We execute the request using a stable
        // view of the shards (especially important for requests that span shards), but retry the
        // top-level (pre-split) request to pick up shard changes. This can lead to unnecessary
@@ -201,13 +228,16 @@ impl PageserverClient {
        //
        // TODO: the gRPC server and client doesn't yet properly support shard splits. Revisit this
        // once we figure out how to handle these.
-        self.retry
-            .with(async |attempt| {
-                let mut req = req.clone();
-                req.request_id.attempt = attempt as u32;
-                Self::get_page_with_shards(req, &self.shards.load_full()).await
-            })
-            .await
+        let resp = Self::with_retries(CALL_TIMEOUT, async |attempt| {
+            let mut req = req.clone();
+            req.request_id.attempt = attempt as u32;
+            let shards = self.shards.load_full();
+            Self::with_timeout(REQUEST_TIMEOUT, Self::get_page_with_shards(req, &shards)).await
+        })
+        .await?;
+
+        debug!("received response: {resp:?}");
+        Ok(resp)
    }

    /// Fetches pages using the given shards. This uses a stable view of the shards, regardless of
@@ -246,7 +276,7 @@ impl PageserverClient {
        req: page_api::GetPageRequest,
        shard: &Shard,
    ) -> tonic::Result<page_api::GetPageResponse> {
-        let stream = shard.stream(req.request_class.is_bulk()).await;
+        let mut stream = shard.stream(Self::is_bulk(&req)).await?;
        let resp = stream.send(req.clone()).await?;

        // Convert per-request errors into a tonic::Status.
@@ -290,13 +320,15 @@ impl PageserverClient {
        &self,
        req: page_api::GetRelSizeRequest,
    ) -> tonic::Result<page_api::GetRelSizeResponse> {
-        self.retry
-            .with(async |_| {
-                // Relation metadata is only available on shard 0.
-                let mut client = self.shards.load_full().get_zero().client().await?;
-                client.get_rel_size(req).await
-            })
-            .await
+        debug!("sending request: {req:?}");
+        let resp = Self::with_retries(CALL_TIMEOUT, async |_| {
+            // Relation metadata is only available on shard 0.
+            let mut client = self.shards.load_full().get_zero().client().await?;
+            Self::with_timeout(REQUEST_TIMEOUT, client.get_rel_size(req)).await
+        })
+        .await?;
+        debug!("received response: {resp:?}");
+        Ok(resp)
    }

    /// Fetches an SLRU segment.
@@ -305,13 +337,50 @@ impl PageserverClient {
        &self,
        req: page_api::GetSlruSegmentRequest,
    ) -> tonic::Result<page_api::GetSlruSegmentResponse> {
-        self.retry
-            .with(async |_| {
-                // SLRU segments are only available on shard 0.
-                let mut client = self.shards.load_full().get_zero().client().await?;
-                client.get_slru_segment(req).await
-            })
-            .await
+        debug!("sending request: {req:?}");
+        let resp = Self::with_retries(CALL_TIMEOUT, async |_| {
+            // SLRU segments are only available on shard 0.
+            let mut client = self.shards.load_full().get_zero().client().await?;
+            Self::with_timeout(REQUEST_TIMEOUT, client.get_slru_segment(req)).await
+        })
+        .await?;
+        debug!("received response: {resp:?}");
+        Ok(resp)
+    }
+
+    /// Runs the given async closure with retries up to the given timeout. Only certain gRPC status
+    /// codes are retried, see [`Retry::should_retry`]. Returns `DeadlineExceeded` on timeout.
+    async fn with_retries<T, F, O>(timeout: Duration, f: F) -> tonic::Result<T>
+    where
+        F: FnMut(usize) -> O, // pass attempt number, starting at 0
+        O: Future<Output = tonic::Result<T>>,
+    {
+        Retry {
+            timeout: Some(timeout),
+            base_backoff: BASE_BACKOFF,
+            max_backoff: MAX_BACKOFF,
+        }
+        .with(f)
+        .await
+    }
+
+    /// Runs the given future with a timeout. Returns `DeadlineExceeded` on timeout.
+    async fn with_timeout<T>(
+        timeout: Duration,
+        f: impl Future<Output = tonic::Result<T>>,
+    ) -> tonic::Result<T> {
+        let started = Instant::now();
+        tokio::time::timeout(timeout, f).await.map_err(|_| {
+            tonic::Status::deadline_exceeded(format!(
+                "request timed out after {:.3}s",
+                started.elapsed().as_secs_f64()
+            ))
+        })?
+    }
+
+    /// Returns true if the request is considered a bulk request and should use the bulk pool.
+    fn is_bulk(req: &page_api::GetPageRequest) -> bool {
+        req.block_numbers.len() >= BULK_THRESHOLD_BATCH_SIZE
    }
 }

@@ -440,15 +509,23 @@ impl Shards {
    }
 }

-/// A single shard. Uses dedicated resource pools with the following structure:
+/// A single shard. Has dedicated resource pools with the following structure:
 ///
-/// * Channel pool: unbounded.
-///   * Unary client pool: MAX_UNARY_CLIENTS.
-///   * Stream client pool: unbounded.
-///     * Stream pool: MAX_STREAMS and MAX_STREAM_QUEUE_DEPTH.
-/// * Bulk channel pool: unbounded.
+/// * Channel pool: MAX_CLIENTS_PER_CHANNEL.
+///   * Client pool: unbounded.
+///     * Stream pool: unbounded.
+/// * Bulk channel pool: MAX_BULK_CLIENTS_PER_CHANNEL.
 ///   * Bulk client pool: unbounded.
-///     * Bulk stream pool: MAX_BULK_STREAMS and MAX_BULK_STREAM_QUEUE_DEPTH.
+///     * Bulk stream pool: unbounded.
+///
+/// We use a separate bulk channel pool with a lower concurrency limit for large batch requests.
+/// This avoids TCP-level head-of-line blocking, and also concentrates large window sizes on a
+/// smaller set of streams/connections, which presumably reduces memory use. Neither of these pools
+/// are bounded, nor do they pipeline requests, so the latency characteristics should be mostly
+/// similar (except for TCP transmission time).
+///
+/// TODO: since we never use bounded pools, we could consider removing the pool limiters. However,
+/// the code is fairly trivial, so we may as well keep them around for now in case we need them.
 struct Shard {
    /// The shard ID.
    id: ShardIndex,
@@ -456,7 +533,7 @@ struct Shard {
    client_pool: Arc<ClientPool>,
    /// GetPage stream pool.
    stream_pool: Arc<StreamPool>,
-    /// GetPage stream pool for bulk requests, e.g. prefetches.
+    /// GetPage stream pool for bulk requests.
    bulk_stream_pool: Arc<StreamPool>,
 }

@@ -470,50 +547,30 @@ impl Shard {
        auth_token: Option<String>,
        compression: Option<CompressionEncoding>,
    ) -> anyhow::Result<Self> {
-        // Common channel pool for unary and stream requests. Bounded by client/stream pools.
-        let channel_pool = ChannelPool::new(url.clone(), MAX_CLIENTS_PER_CHANNEL)?;
-
-        // Client pool for unary requests.
+        // Shard pools for unary requests and non-bulk GetPage requests.
        let client_pool = ClientPool::new(
-            channel_pool.clone(),
+            ChannelPool::new(url.clone(), MAX_CLIENTS_PER_CHANNEL)?,
            tenant_id,
            timeline_id,
            shard_id,
            auth_token.clone(),
            compression,
-            Some(MAX_UNARY_CLIENTS),
+            None, // unbounded
        );
+        let stream_pool = StreamPool::new(client_pool.clone(), None); // unbounded

-        // GetPage stream pool. Uses a dedicated client pool to avoid starving out unary clients,
-        // but shares a channel pool with it (as it's unbounded).
-        let stream_pool = StreamPool::new(
-            ClientPool::new(
-                channel_pool.clone(),
-                tenant_id,
-                timeline_id,
-                shard_id,
-                auth_token.clone(),
-                compression,
-                None, // unbounded, limited by stream pool
-            ),
-            Some(MAX_STREAMS),
-            MAX_STREAM_QUEUE_DEPTH,
-        );
-
-        // Bulk GetPage stream pool, e.g. for prefetches. Uses dedicated channel/client/stream pools
-        // to avoid head-of-line blocking of latency-sensitive requests.
+        // Bulk GetPage stream pool for large batches (prefetches, sequential scans, vacuum, etc.).
        let bulk_stream_pool = StreamPool::new(
            ClientPool::new(
-                ChannelPool::new(url, MAX_CLIENTS_PER_CHANNEL)?,
+                ChannelPool::new(url, MAX_BULK_CLIENTS_PER_CHANNEL)?,
                tenant_id,
                timeline_id,
                shard_id,
                auth_token,
                compression,
-                None, // unbounded, limited by stream pool
+                None, // unbounded,
            ),
-            Some(MAX_BULK_STREAMS),
-            MAX_BULK_STREAM_QUEUE_DEPTH,
+            None, // unbounded
        );

        Ok(Self {
@@ -525,19 +582,23 @@ impl Shard {
    }

    /// Returns a pooled client for this shard.
+    #[instrument(skip_all)]
    async fn client(&self) -> tonic::Result<ClientGuard> {
-        self.client_pool
-            .get()
-            .await
-            .map_err(|err| tonic::Status::internal(format!("failed to get client: {err}")))
+        warn_slow(
+            "client pool acquisition",
+            SLOW_THRESHOLD,
+            pin!(self.client_pool.get()),
+        )
+        .await
    }

-    /// Returns a pooled stream for this shard. If `bulk` is `true`, uses the dedicated bulk stream
-    /// pool (e.g. for prefetches).
-    async fn stream(&self, bulk: bool) -> StreamGuard {
-        match bulk {
-            false => self.stream_pool.get().await,
-            true => self.bulk_stream_pool.get().await,
-        }
+    /// Returns a pooled stream for this shard. If `bulk` is `true`, uses the dedicated bulk pool.
+    #[instrument(skip_all, fields(bulk))]
+    async fn stream(&self, bulk: bool) -> tonic::Result<StreamGuard> {
+        let pool = match bulk {
+            false => &self.stream_pool,
+            true => &self.bulk_stream_pool,
+        };
+        warn_slow("stream pool acquisition", SLOW_THRESHOLD, pin!(pool.get())).await
    }
 }
--- a/pageserver/client_grpc/src/lib.rs
+++ b/pageserver/client_grpc/src/lib.rs
@@ -4,4 +4,3 @@ mod retry;
 mod split;

 pub use client::{PageserverClient, ShardSpec};
-pub use pageserver_api::shard::ShardStripeSize; // used in ShardSpec
--- a/pageserver/client_grpc/src/pool.rs
+++ b/pageserver/client_grpc/src/pool.rs
@@ -9,19 +9,36 @@
 //!
 //! * ChannelPool: manages gRPC channels (TCP connections) to a single Pageserver. Multiple clients
 //!   can acquire and use the same channel concurrently (via HTTP/2 stream multiplexing), up to a
-//!   per-channel client limit. Channels may be closed when they are no longer used by any clients.
+//!   per-channel client limit. Channels are closed immediately when empty, and indirectly rely on
+//!   client/stream idle timeouts.
 //!
 //! * ClientPool: manages gRPC clients for a single tenant shard. Each client acquires a (shared)
 //!   channel from the ChannelPool for the client's lifetime. A client can only be acquired by a
-//!   single caller at a time, and is returned to the pool when dropped. Idle clients may be removed
-//!   from the pool after some time, to free up the channel.
+//!   single caller at a time, and is returned to the pool when dropped. Idle clients are removed
+//!   from the pool after a while to free up resources.
 //!
 //! * StreamPool: manages bidirectional gRPC GetPage streams. Each stream acquires a client from the
-//!   ClientPool for the stream's lifetime. Internal streams are not exposed to callers; instead, it
-//!   returns a guard that can be used to send a single request, to properly enforce queue depth and
-//!   route responses. Internally, the pool will reuse or spin up a suitable stream for the request,
-//!   possibly pipelining multiple requests from multiple callers on the same stream (up to some
-//!   queue depth). Idle streams may be removed from the pool after a while to free up the client.
+//!   ClientPool for the stream's lifetime. A stream can only be acquired by a single caller at a
+//!   time, and is returned to the pool when dropped. Idle streams are removed from the pool after
+//!   a while to free up resources.
+//!
+//!   The stream only supports sending a single, synchronous request at a time, and does not support
+//!   pipelining multiple requests from different callers onto the same stream -- instead, we scale
+//!   out concurrent streams to improve throughput. There are many reasons for this design choice:
+//!
+//!     * It (mostly) eliminates head-of-line blocking. A single stream is processed sequentially by
+//!       a single server task, which may block e.g. on layer downloads, LSN waits, etc.
+//!
+//!     * Cancellation becomes trivial, by closing the stream. Otherwise, if a caller goes away
+//!       (e.g. because of a timeout), the request would still be processed by the server and block
+//!       requests behind it in the stream. It might even block its own timeout retry.
+//!
+//!     * Stream scheduling becomes significantly simpler and cheaper.
+//!
+//!     * Individual callers can still use client-side batching for pipelining.
+//!
+//!     * Idle streams are cheap. Benchmarks show that an idle GetPage stream takes up about 26 KB
+//!       per stream (2.5 GB for 100,000 streams), so we can afford to scale out.
 //!
 //! Each channel corresponds to one TCP connection. Each client unary request and each stream
 //! corresponds to one HTTP/2 stream and server task.
@@ -29,33 +46,31 @@
 //! TODO: error handling (including custom error types).
 //! TODO: observability.

-use std::collections::{BTreeMap, HashMap};
+use std::collections::BTreeMap;
 use std::num::NonZero;
 use std::ops::{Deref, DerefMut};
+use std::pin::Pin;
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex, Weak};
 use std::time::{Duration, Instant};

-use futures::StreamExt as _;
-use tokio::sync::mpsc::{Receiver, Sender};
-use tokio::sync::{OwnedSemaphorePermit, Semaphore, mpsc, oneshot};
+use futures::{Stream, StreamExt as _};
+use tokio::sync::{OwnedSemaphorePermit, Semaphore, watch};
+use tokio_stream::wrappers::WatchStream;
 use tokio_util::sync::CancellationToken;
 use tonic::codec::CompressionEncoding;
 use tonic::transport::{Channel, Endpoint};
-use tracing::{error, warn};

 use pageserver_page_api as page_api;
 use utils::id::{TenantId, TimelineId};
 use utils::shard::ShardIndex;

-/// Reap channels/clients/streams that have been idle for this long.
+/// Reap clients/streams that have been idle for this long. Channels are reaped immediately when
+/// empty, and indirectly rely on the client/stream idle timeouts.
 ///
-/// TODO: this is per-pool. For nested pools, it can take up to 3x as long for a TCP connection to
-/// be reaped. First, we must wait for an idle stream to be reaped, which marks its client as idle.
-/// Then, we must wait for the idle client to be reaped, which marks its channel as idle. Then, we
-/// must wait for the idle channel to be reaped. Is that a problem? Maybe not, we just have to
-/// account for it when setting the reap threshold. Alternatively, we can immediately reap empty
-/// channels, and/or stream pool clients.
+/// A stream's client will be reaped after 2x the idle threshold (first stream the client), but
+/// that's okay -- if the stream closes abruptly (e.g. due to timeout or cancellation), we want to
+/// keep its client around in the pool for a while.
 const REAP_IDLE_THRESHOLD: Duration = match cfg!(any(test, feature = "testing")) {
    false => Duration::from_secs(180),
    true => Duration::from_secs(1), // exercise reaping in tests
@@ -83,8 +98,6 @@ pub struct ChannelPool {
    max_clients_per_channel: NonZero<usize>,
    /// Open channels.
    channels: Mutex<BTreeMap<ChannelID, ChannelEntry>>,
-    /// Reaps idle channels.
-    idle_reaper: Reaper,
    /// Channel ID generator.
    next_channel_id: AtomicUsize,
 }
@@ -96,9 +109,6 @@ struct ChannelEntry {
    channel: Channel,
    /// Number of clients using this channel.
    clients: usize,
-    /// The channel has been idle (no clients) since this time. None if channel is in use.
-    /// INVARIANT: Some if clients == 0, otherwise None.
-    idle_since: Option<Instant>,
 }

 impl ChannelPool {
@@ -108,15 +118,12 @@ impl ChannelPool {
        E: TryInto<Endpoint> + Send + Sync + 'static,
        <E as TryInto<Endpoint>>::Error: std::error::Error + Send + Sync,
    {
-        let pool = Arc::new(Self {
+        Ok(Arc::new(Self {
            endpoint: endpoint.try_into()?,
            max_clients_per_channel,
            channels: Mutex::default(),
-            idle_reaper: Reaper::new(REAP_IDLE_THRESHOLD, REAP_IDLE_INTERVAL),
            next_channel_id: AtomicUsize::default(),
-        });
-        pool.idle_reaper.spawn(&pool);
-        Ok(pool)
+        }))
    }

    /// Acquires a gRPC channel for a client. Multiple clients may acquire the same channel.
@@ -137,22 +144,17 @@ impl ChannelPool {
        let mut channels = self.channels.lock().unwrap();

        // Try to find an existing channel with available capacity. We check entries in BTreeMap
-        // order, to fill up the lower-ordered channels first. The ClientPool also prefers clients
-        // with lower-ordered channel IDs first. This will cluster clients in lower-ordered
+        // order, to fill up the lower-ordered channels first. The client/stream pools also prefer
+        // clients with lower-ordered channel IDs first. This will cluster clients in lower-ordered
        // channels, and free up higher-ordered channels such that they can be reaped.
        for (&id, entry) in channels.iter_mut() {
            assert!(
                entry.clients <= self.max_clients_per_channel.get(),
                "channel overflow"
            );
-            assert_eq!(
-                entry.idle_since.is_some(),
-                entry.clients == 0,
-                "incorrect channel idle state"
-            );
+            assert_ne!(entry.clients, 0, "empty channel not reaped");
            if entry.clients < self.max_clients_per_channel.get() {
                entry.clients += 1;
-                entry.idle_since = None;
                return ChannelGuard {
                    pool: Arc::downgrade(self),
                    id,
@@ -169,7 +171,6 @@ impl ChannelPool {
        let entry = ChannelEntry {
            channel: channel.clone(),
            clients: 1, // account for the guard below
-            idle_since: None,
        };
        channels.insert(id, entry);

@@ -181,20 +182,6 @@ impl ChannelPool {
    }
 }

-impl Reapable for ChannelPool {
-    /// Reaps channels that have been idle since before the cutoff.
-    fn reap_idle(&self, cutoff: Instant) {
-        self.channels.lock().unwrap().retain(|_, entry| {
-            let Some(idle_since) = entry.idle_since else {
-                assert_ne!(entry.clients, 0, "empty channel not marked idle");
-                return true;
-            };
-            assert_eq!(entry.clients, 0, "idle channel has clients");
-            idle_since >= cutoff
-        })
-    }
-}
-
 /// Tracks a channel acquired from the pool. The owned inner channel can be obtained with `take()`,
 /// since the gRPC client requires an owned `Channel`.
 pub struct ChannelGuard {
@@ -211,7 +198,7 @@ impl ChannelGuard {
    }
 }

-/// Returns the channel to the pool.
+/// Returns the channel to the pool. The channel is closed when empty.
 impl Drop for ChannelGuard {
    fn drop(&mut self) {
        let Some(pool) = self.pool.upgrade() else {
@@ -220,11 +207,12 @@ impl Drop for ChannelGuard {

        let mut channels = pool.channels.lock().unwrap();
        let entry = channels.get_mut(&self.id).expect("unknown channel");
-        assert!(entry.idle_since.is_none(), "active channel marked idle");
        assert!(entry.clients > 0, "channel underflow");
        entry.clients -= 1;
+
+        // Reap empty channels immediately.
        if entry.clients == 0 {
-            entry.idle_since = Some(Instant::now()); // mark channel as idle
+            channels.remove(&self.id);
        }
    }
 }
@@ -253,8 +241,7 @@ pub struct ClientPool {
    ///
    /// The first client in the map will be acquired next. The map is sorted by client ID, which in
    /// turn is sorted by its channel ID, such that we prefer acquiring idle clients from
-    /// lower-ordered channels. This allows us to free up and reap higher-numbered channels as idle
-    /// clients are reaped.
+    /// lower-ordered channels. This allows us to free up and reap higher-ordered channels.
    idle: Mutex<BTreeMap<ClientID, ClientEntry>>,
    /// Reaps idle clients.
    idle_reaper: Reaper,
@@ -310,7 +297,7 @@ impl ClientPool {
    /// This is moderately performance-sensitive. It is called for every unary request, but these
    /// establish a new gRPC stream per request so they're already expensive. GetPage requests use
    /// the `StreamPool` instead.
-    pub async fn get(self: &Arc<Self>) -> anyhow::Result<ClientGuard> {
+    pub async fn get(self: &Arc<Self>) -> tonic::Result<ClientGuard> {
        // Acquire a permit if the pool is bounded.
        let mut permit = None;
        if let Some(limiter) = self.limiter.clone() {
@@ -328,7 +315,7 @@ impl ClientPool {
            });
        }

-        // Slow path: construct a new client.
+        // Construct a new client.
        let mut channel_guard = self.channel_pool.get();
        let client = page_api::Client::new(
            channel_guard.take(),
@@ -337,7 +324,8 @@ impl ClientPool {
            self.shard_id,
            self.auth_token.clone(),
            self.compression,
-        )?;
+        )
+        .map_err(|err| tonic::Status::internal(format!("failed to create client: {err}")))?;

        Ok(ClientGuard {
            pool: Arc::downgrade(self),
@@ -407,287 +395,187 @@ impl Drop for ClientGuard {
 /// A pool of bidirectional gRPC streams. Currently only used for GetPage streams. Each stream
 /// acquires a client from the inner `ClientPool` for the stream's lifetime.
 ///
-/// Individual streams are not exposed to callers -- instead, the returned guard can be used to send
-/// a single request and await the response. Internally, requests are multiplexed across streams and
-/// channels. This allows proper queue depth enforcement and response routing.
+/// Individual streams only send a single request at a time, and do not pipeline multiple callers
+/// onto the same stream. Instead, we scale out the number of concurrent streams. This is primarily
+/// to eliminate head-of-line blocking. See the module documentation for more details.
 ///
 /// TODO: consider making this generic over request and response types; not currently needed.
 pub struct StreamPool {
    /// The client pool to acquire clients from. Must be unbounded.
    client_pool: Arc<ClientPool>,
-    /// All pooled streams.
+    /// Idle pooled streams. Acquired streams are removed from here and returned on drop.
    ///
-    /// Incoming requests will be sent over an existing stream with available capacity. If all
-    /// streams are full, a new one is spun up and added to the pool (up to `max_streams`). Each
-    /// stream has an associated Tokio task that processes requests and responses.
-    streams: Mutex<HashMap<StreamID, StreamEntry>>,
-    /// The max number of concurrent streams, or None if unbounded.
-    max_streams: Option<NonZero<usize>>,
-    /// The max number of concurrent requests per stream.
-    max_queue_depth: NonZero<usize>,
-    /// Limits the max number of concurrent requests, given by `max_streams * max_queue_depth`.
-    /// None if the pool is unbounded.
+    /// The first stream in the map will be acquired next. The map is sorted by stream ID, which is
+    /// equivalent to the client ID and in turn sorted by its channel ID. This way we prefer
+    /// acquiring idle streams from lower-ordered channels, which allows us to free up and reap
+    /// higher-ordered channels.
+    idle: Mutex<BTreeMap<StreamID, StreamEntry>>,
+    /// Limits the max number of concurrent streams. None if the pool is unbounded.
    limiter: Option<Arc<Semaphore>>,
    /// Reaps idle streams.
    idle_reaper: Reaper,
-    /// Stream ID generator.
-    next_stream_id: AtomicUsize,
 }

-type StreamID = usize;
-type RequestSender = Sender<(page_api::GetPageRequest, ResponseSender)>;
-type RequestReceiver = Receiver<(page_api::GetPageRequest, ResponseSender)>;
-type ResponseSender = oneshot::Sender<tonic::Result<page_api::GetPageResponse>>;
+/// The stream ID. Reuses the inner client ID.
+type StreamID = ClientID;

+/// A pooled stream.
 struct StreamEntry {
-    /// Sends caller requests to the stream task. The stream task exits when this is dropped.
-    sender: RequestSender,
-    /// Number of in-flight requests on this stream.
-    queue_depth: usize,
-    /// The time when this stream went idle (queue_depth == 0).
-    /// INVARIANT: Some if queue_depth == 0, otherwise None.
-    idle_since: Option<Instant>,
+    /// The bidirectional stream.
+    stream: BiStream,
+    /// The time when this stream was last used, i.e. when it was put back into `StreamPool::idle`.
+    idle_since: Instant,
+}
+
+/// A bidirectional GetPage stream and its client. Can send requests and receive responses.
+struct BiStream {
+    /// The owning client. Holds onto the channel slot while the stream is alive.
+    client: ClientGuard,
+    /// Stream for sending requests. Uses a watch channel, so it can only send a single request at a
+    /// time, and the caller must await the response before sending another request. This is
+    /// enforced by `StreamGuard::send`.
+    sender: watch::Sender<page_api::GetPageRequest>,
+    /// Stream for receiving responses.
+    receiver: Pin<Box<dyn Stream<Item = tonic::Result<page_api::GetPageResponse>> + Send>>,
 }

 impl StreamPool {
-    /// Creates a new stream pool, using the given client pool. It will send up to `max_queue_depth`
-    /// concurrent requests on each stream, and use up to `max_streams` concurrent streams.
+    /// Creates a new stream pool, using the given client pool. It will use up to `max_streams`
+    /// concurrent streams.
    ///
    /// The client pool must be unbounded. The stream pool will enforce its own limits, and because
    /// streams are long-lived they can cause persistent starvation if they exhaust the client pool.
    /// The stream pool should generally have its own dedicated client pool (but it can share a
    /// channel pool with others since these are always unbounded).
-    pub fn new(
-        client_pool: Arc<ClientPool>,
-        max_streams: Option<NonZero<usize>>,
-        max_queue_depth: NonZero<usize>,
-    ) -> Arc<Self> {
+    pub fn new(client_pool: Arc<ClientPool>, max_streams: Option<NonZero<usize>>) -> Arc<Self> {
        assert!(client_pool.limiter.is_none(), "bounded client pool");
        let pool = Arc::new(Self {
            client_pool,
-            streams: Mutex::default(),
-            limiter: max_streams.map(|max_streams| {
-                Arc::new(Semaphore::new(max_streams.get() * max_queue_depth.get()))
-            }),
-            max_streams,
-            max_queue_depth,
+            idle: Mutex::default(),
+            limiter: max_streams.map(|max_streams| Arc::new(Semaphore::new(max_streams.get()))),
            idle_reaper: Reaper::new(REAP_IDLE_THRESHOLD, REAP_IDLE_INTERVAL),
-            next_stream_id: AtomicUsize::default(),
        });
        pool.idle_reaper.spawn(&pool);
        pool
    }

-    /// Acquires an available stream from the pool, or spins up a new stream async if all streams
-    /// are full. Returns a guard that can be used to send a single request on the stream and await
-    /// the response, with queue depth quota already acquired. Blocks if the pool is at capacity
-    /// (i.e. `CLIENT_LIMIT * STREAM_QUEUE_DEPTH` requests in flight).
+    /// Acquires an available stream from the pool, or spins up a new stream if all streams are
+    /// full. Returns a guard that can be used to send requests and await the responses. Blocks if
+    /// the pool is full.
    ///
    /// This is very performance-sensitive, as it is on the GetPage hot path.
    ///
-    /// TODO: this must do something more sophisticated for performance. We want:
-    ///
-    /// * Cheap, concurrent access in the common case where we can use a pooled stream.
-    /// * Quick acquisition of pooled streams with available capacity.
-    /// * Prefer streams that belong to lower-numbered channels, to reap idle channels.
-    /// * Prefer filling up existing streams' queue depth before spinning up new streams.
-    /// * Don't hold a lock while spinning up new streams.
-    /// * Allow concurrent clients to join onto streams while they're spun up.
-    /// * Allow spinning up multiple streams concurrently, but don't overshoot limits.
-    ///
-    /// For now, we just do something simple but inefficient (linear scan under mutex).
-    pub async fn get(self: &Arc<Self>) -> StreamGuard {
+    /// TODO: is a `Mutex<BTreeMap>` performant enough? Will it become too contended? We can't
+    /// trivially use e.g. DashMap or sharding, because we want to pop lower-ordered streams first
+    /// to free up higher-ordered channels.
+    pub async fn get(self: &Arc<Self>) -> tonic::Result<StreamGuard> {
        // Acquire a permit if the pool is bounded.
        let mut permit = None;
        if let Some(limiter) = self.limiter.clone() {
            permit = Some(limiter.acquire_owned().await.expect("never closed"));
        }
-        let mut streams = self.streams.lock().unwrap();

-        // Look for a pooled stream with available capacity.
-        for (&id, entry) in streams.iter_mut() {
-            assert!(
-                entry.queue_depth <= self.max_queue_depth.get(),
-                "stream queue overflow"
-            );
-            assert_eq!(
-                entry.idle_since.is_some(),
-                entry.queue_depth == 0,
-                "incorrect stream idle state"
-            );
-            if entry.queue_depth < self.max_queue_depth.get() {
-                entry.queue_depth += 1;
-                entry.idle_since = None;
-                return StreamGuard {
-                    pool: Arc::downgrade(self),
-                    id,
-                    sender: entry.sender.clone(),
-                    permit,
-                };
-            }
+        // Fast path: acquire an idle stream from the pool.
+        if let Some((_, entry)) = self.idle.lock().unwrap().pop_first() {
+            return Ok(StreamGuard {
+                pool: Arc::downgrade(self),
+                stream: Some(entry.stream),
+                can_reuse: true,
+                permit,
+            });
        }

-        // No available stream, spin up a new one. We install the stream entry in the pool first and
-        // return the guard, while spinning up the stream task async. This allows other callers to
-        // join onto this stream and also create additional streams concurrently if this fills up.
-        let id = self.next_stream_id.fetch_add(1, Ordering::Relaxed);
-        let (req_tx, req_rx) = mpsc::channel(self.max_queue_depth.get());
-        let entry = StreamEntry {
-            sender: req_tx.clone(),
-            queue_depth: 1, // reserve quota for this caller
-            idle_since: None,
-        };
-        streams.insert(id, entry);
+        // Spin up a new stream. Uses a watch channel to send a single request at a time, since
+        // `StreamGuard::send` enforces this anyway and it avoids unnecessary channel overhead.
+        let mut client = self.client_pool.get().await?;

-        if let Some(max_streams) = self.max_streams {
-            assert!(streams.len() <= max_streams.get(), "stream overflow");
-        };
+        let (req_tx, req_rx) = watch::channel(page_api::GetPageRequest::default());
+        let req_stream = WatchStream::from_changes(req_rx);
+        let resp_stream = client.get_pages(req_stream).await?;

-        let client_pool = self.client_pool.clone();
-        let pool = Arc::downgrade(self);
-
-        tokio::spawn(async move {
-            if let Err(err) = Self::run_stream(client_pool, req_rx).await {
-                error!("stream failed: {err}");
-            }
-            // Remove stream from pool on exit. Weak reference to avoid holding the pool alive.
-            if let Some(pool) = pool.upgrade() {
-                let entry = pool.streams.lock().unwrap().remove(&id);
-                assert!(entry.is_some(), "unknown stream ID: {id}");
-            }
-        });
-
-        StreamGuard {
+        Ok(StreamGuard {
            pool: Arc::downgrade(self),
-            id,
-            sender: req_tx,
+            stream: Some(BiStream {
+                client,
+                sender: req_tx,
+                receiver: Box::pin(resp_stream),
+            }),
+            can_reuse: true,
            permit,
-        }
-    }
-
-    /// Runs a stream task. This acquires a client from the `ClientPool` and establishes a
-    /// bidirectional GetPage stream, then forwards requests and responses between callers and the
-    /// stream. It does not track or enforce queue depths -- that's done by `get()` since it must be
-    /// atomic with pool stream acquisition.
-    ///
-    /// The task exits when the request channel is closed, or on a stream error. The caller is
-    /// responsible for removing the stream from the pool on exit.
-    async fn run_stream(
-        client_pool: Arc<ClientPool>,
-        mut caller_rx: RequestReceiver,
-    ) -> anyhow::Result<()> {
-        // Acquire a client from the pool and create a stream.
-        let mut client = client_pool.get().await?;
-
-        // NB: use an unbounded channel such that the stream send never blocks. Otherwise, we could
-        // theoretically deadlock if both the client and server block on sends (since we're not
-        // reading responses while sending). This is unlikely to happen due to gRPC/TCP buffers and
-        // low queue depths, but it was seen to happen with the libpq protocol so better safe than
-        // sorry. It should never buffer more than the queue depth anyway, but using an unbounded
-        // channel guarantees that it will never block.
-        let (req_tx, req_rx) = mpsc::unbounded_channel();
-        let req_stream = tokio_stream::wrappers::UnboundedReceiverStream::new(req_rx);
-        let mut resp_stream = client.get_pages(req_stream).await?;
-
-        // Track caller response channels by request ID. If the task returns early, these response
-        // channels will be dropped and the waiting callers will receive an error.
-        //
-        // NB: this will leak entries if the server doesn't respond to a request (by request ID).
-        // It shouldn't happen, and if it does it will often hold onto queue depth quota anyway and
-        // block further use. But we could consider reaping closed channels after some time.
-        let mut callers = HashMap::new();
-
-        // Process requests and responses.
-        loop {
-            tokio::select! {
-                // Receive requests from callers and send them to the stream.
-                req = caller_rx.recv() => {
-                    // Shut down if request channel is closed.
-                    let Some((req, resp_tx)) = req else {
-                        return Ok(());
-                    };
-
-                    // Store the response channel by request ID.
-                    if callers.contains_key(&req.request_id) {
-                        // Error on request ID duplicates. Ignore callers that went away.
-                        _ = resp_tx.send(Err(tonic::Status::invalid_argument(
-                            format!("duplicate request ID: {}", req.request_id),
-                        )));
-                        continue;
-                    }
-                    callers.insert(req.request_id, resp_tx);
-
-                    // Send the request on the stream. Bail out if the stream is closed.
-                    req_tx.send(req).map_err(|_| {
-                        tonic::Status::unavailable("stream closed")
-                    })?;
-                }
-
-                // Receive responses from the stream and send them to callers.
-                resp = resp_stream.next() => {
-                    // Shut down if the stream is closed, and bail out on stream errors.
-                    let Some(resp) = resp.transpose()? else {
-                        return Ok(())
-                    };
-
-                    // Send the response to the caller. Ignore errors if the caller went away.
-                    let Some(resp_tx) = callers.remove(&resp.request_id) else {
-                        warn!("received response for unknown request ID: {}", resp.request_id);
-                        continue;
-                    };
-                    _ = resp_tx.send(Ok(resp));
-                }
-            }
-        }
+        })
    }
 }

 impl Reapable for StreamPool {
    /// Reaps streams that have been idle since before the cutoff.
    fn reap_idle(&self, cutoff: Instant) {
-        self.streams.lock().unwrap().retain(|_, entry| {
-            let Some(idle_since) = entry.idle_since else {
-                assert_ne!(entry.queue_depth, 0, "empty stream not marked idle");
-                return true;
-            };
-            assert_eq!(entry.queue_depth, 0, "idle stream has requests");
-            idle_since >= cutoff
-        });
+        self.idle
+            .lock()
+            .unwrap()
+            .retain(|_, entry| entry.idle_since >= cutoff);
    }
 }

-/// A pooled stream reference. Can be used to send a single request, to properly enforce queue
-/// depth. Queue depth is already reserved and will be returned on drop.
+/// A stream acquired from the pool. Returned to the pool when dropped, unless there are still
+/// in-flight requests on the stream, or the stream failed.
 pub struct StreamGuard {
    pool: Weak<StreamPool>,
-    id: StreamID,
-    sender: RequestSender,
+    stream: Option<BiStream>,             // Some until dropped
+    can_reuse: bool,                      // returned to pool if true
    permit: Option<OwnedSemaphorePermit>, // None if pool is unbounded
 }

 impl StreamGuard {
-    /// Sends a request on the stream and awaits the response. Consumes the guard, since it's only
-    /// valid for a single request (to enforce queue depth). This also drops the guard on return and
-    /// returns the queue depth quota to the pool.
+    /// Sends a request on the stream and awaits the response. If the future is dropped before it
+    /// resolves (e.g. due to a timeout or cancellation), the stream will be closed to cancel the
+    /// request and is not returned to the pool. The same is true if the stream errors, in which
+    /// case the caller can't send further requests on the stream.
    ///
-    /// The `GetPageRequest::request_id` must be unique across in-flight requests.
+    /// We only support sending a single request at a time, to eliminate head-of-line blocking. See
+    /// module documentation for details.
    ///
    /// NB: errors are often returned as `GetPageResponse::status_code` instead of `tonic::Status`
    /// to avoid tearing down the stream for per-request errors. Callers must check this.
    pub async fn send(
-        self,
+        &mut self,
        req: page_api::GetPageRequest,
    ) -> tonic::Result<page_api::GetPageResponse> {
-        let (resp_tx, resp_rx) = oneshot::channel();
+        let req_id = req.request_id;
+        let stream = self.stream.as_mut().expect("not dropped");

-        self.sender
-            .send((req, resp_tx))
-            .await
+        // Mark the stream as not reusable while the request is in flight. We can't return the
+        // stream to the pool until we receive the response, to avoid head-of-line blocking and
+        // stale responses. Failed streams can't be reused either.
+        if !self.can_reuse {
+            return Err(tonic::Status::internal("stream can't be reused"));
+        }
+        self.can_reuse = false;
+
+        // Send the request and receive the response.
+        //
+        // NB: this uses a watch channel, so it's unsafe to change this code to pipeline requests.
+        stream
+            .sender
+            .send(req)
            .map_err(|_| tonic::Status::unavailable("stream closed"))?;

-        resp_rx
+        let resp = stream
+            .receiver
+            .next()
            .await
-            .map_err(|_| tonic::Status::unavailable("stream closed"))?
+            .ok_or_else(|| tonic::Status::unavailable("stream closed"))??;
+
+        if resp.request_id != req_id {
+            return Err(tonic::Status::internal(format!(
+                "response ID {} does not match request ID {}",
+                resp.request_id, req_id
+            )));
+        }
+
+        // Success, mark the stream as reusable.
+        self.can_reuse = true;
+
+        Ok(resp)
    }
 }

@@ -697,26 +585,21 @@ impl Drop for StreamGuard {
            return; // pool was dropped
        };

-        // Release the queue depth reservation on drop. This can prematurely decrement it if dropped
-        // before the response is received, but that's okay.
-        //
-        // TODO: actually, it's probably not okay. Queue depth release should be moved into the
-        // stream task, such that it continues to account for the queue depth slot until the server
-        // responds. Otherwise, if a slow request times out and keeps blocking the stream, the
-        // server will keep waiting on it and we can pile on subsequent requests (including the
-        // timeout retry) in the same stream and get blocked. But we may also want to avoid blocking
-        // requests on e.g. LSN waits and layer downloads, instead returning early to free up the
-        // stream. Or just scale out streams with a queue depth of 1 to sidestep all head-of-line
-        // blocking. TBD.
-        let mut streams = pool.streams.lock().unwrap();
-        let entry = streams.get_mut(&self.id).expect("unknown stream");
-        assert!(entry.idle_since.is_none(), "active stream marked idle");
-        assert!(entry.queue_depth > 0, "stream queue underflow");
-        entry.queue_depth -= 1;
-        if entry.queue_depth == 0 {
-            entry.idle_since = Some(Instant::now()); // mark stream as idle
+        // If the stream isn't reusable, it can't be returned to the pool.
+        if !self.can_reuse {
+            return;
        }

+        // Place the idle stream back into the pool.
+        let entry = StreamEntry {
+            stream: self.stream.take().expect("dropped once"),
+            idle_since: Instant::now(),
+        };
+        pool.idle
+            .lock()
+            .unwrap()
+            .insert(entry.stream.client.id, entry);
+
        _ = self.permit; // returned on drop, referenced for visibility
    }
 }
--- a/pageserver/client_grpc/src/retry.rs
+++ b/pageserver/client_grpc/src/retry.rs
@@ -1,5 +1,6 @@
 use std::time::Duration;

+use futures::future::pending;
 use tokio::time::Instant;
 use tracing::{error, info, warn};

@@ -8,60 +9,54 @@ use utils::backoff::exponential_backoff_duration;
 /// A retry handler for Pageserver gRPC requests.
 ///
 /// This is used instead of backoff::retry for better control and observability.
-pub struct Retry;
+pub struct Retry {
+    /// Timeout across all retry attempts. If None, retries forever.
+    pub timeout: Option<Duration>,
+    /// The initial backoff duration. The first retry does not use a backoff.
+    pub base_backoff: Duration,
+    /// The maximum backoff duration.
+    pub max_backoff: Duration,
+}

 impl Retry {
-    /// The per-request timeout.
-    // TODO: tune these, and/or make them configurable. Should we retry forever?
-    const REQUEST_TIMEOUT: Duration = Duration::from_secs(10);
-    /// The total timeout across all attempts
-    const TOTAL_TIMEOUT: Duration = Duration::from_secs(60);
-    /// The initial backoff duration.
-    const BASE_BACKOFF: Duration = Duration::from_millis(10);
-    /// The maximum backoff duration.
-    const MAX_BACKOFF: Duration = Duration::from_secs(10);
-    /// If true, log successful requests. For debugging.
-    const LOG_SUCCESS: bool = false;
-
-    /// Runs the given async closure with timeouts and retries (exponential backoff), passing the
-    /// attempt number starting at 0. Logs errors, using the current tracing span for context.
+    /// Runs the given async closure with timeouts and retries (exponential backoff). Logs errors,
+    /// using the current tracing span for context.
    ///
-    /// Only certain gRPC status codes are retried, see [`Self::should_retry`]. For default
-    /// timeouts, see [`Self::REQUEST_TIMEOUT`] and [`Self::TOTAL_TIMEOUT`].
+    /// Only certain gRPC status codes are retried, see [`Self::should_retry`].
    pub async fn with<T, F, O>(&self, mut f: F) -> tonic::Result<T>
    where
-        F: FnMut(usize) -> O, // takes attempt number, starting at 0
+        F: FnMut(usize) -> O, // pass attempt number, starting at 0
        O: Future<Output = tonic::Result<T>>,
    {
        let started = Instant::now();
-        let deadline = started + Self::TOTAL_TIMEOUT;
+        let deadline = self.timeout.map(|timeout| started + timeout);
        let mut last_error = None;
        let mut retries = 0;
        loop {
-            // Set up a future to wait for the backoff (if any) and run the request with a timeout.
+            // Set up a future to wait for the backoff, if any, and run the closure.
            let backoff_and_try = async {
                // NB: sleep() always sleeps 1ms, even when given a 0 argument. See:
                // https://github.com/tokio-rs/tokio/issues/6866
-                if let Some(backoff) = Self::backoff_duration(retries) {
+                if let Some(backoff) = self.backoff_duration(retries) {
                    tokio::time::sleep(backoff).await;
                }

-                let request_started = Instant::now();
-                tokio::time::timeout(Self::REQUEST_TIMEOUT, f(retries))
-                    .await
-                    .map_err(|_| {
-                        tonic::Status::deadline_exceeded(format!(
-                            "request timed out after {:.3}s",
-                            request_started.elapsed().as_secs_f64()
-                        ))
-                    })?
+                f(retries).await
            };

-            // Wait for the backoff and request, or bail out if the total timeout is exceeded.
+            // Set up a future for the timeout, if any.
+            let timeout = async {
+                match deadline {
+                    Some(deadline) => tokio::time::sleep_until(deadline).await,
+                    None => pending().await,
+                }
+            };
+
+            // Wait for the backoff and request, or bail out if the timeout is exceeded.
            let result = tokio::select! {
                result = backoff_and_try => result,

-                _ = tokio::time::sleep_until(deadline) => {
+                _ = timeout => {
                    let last_error = last_error.unwrap_or_else(|| {
                        tonic::Status::deadline_exceeded(format!(
                            "request timed out after {:.3}s",
@@ -79,7 +74,7 @@ impl Retry {
            match result {
                // Success, return the result.
                Ok(result) => {
-                    if retries > 0 || Self::LOG_SUCCESS {
+                    if retries > 0 {
                        info!(
                            "request succeeded after {retries} retries in {:.3}s",
                            started.elapsed().as_secs_f64(),
@@ -112,12 +107,13 @@ impl Retry {
        }
    }

-    /// Returns the backoff duration for the given retry attempt, or None for no backoff.
-    fn backoff_duration(retry: usize) -> Option<Duration> {
+    /// Returns the backoff duration for the given retry attempt, or None for no backoff. The first
+    /// attempt and first retry never backs off, so this returns None for 0 and 1 retries.
+    fn backoff_duration(&self, retries: usize) -> Option<Duration> {
        let backoff = exponential_backoff_duration(
-            retry as u32,
-            Self::BASE_BACKOFF.as_secs_f64(),
-            Self::MAX_BACKOFF.as_secs_f64(),
+            (retries as u32).saturating_sub(1), // first retry does not back off
+            self.base_backoff.as_secs_f64(),
+            self.max_backoff.as_secs_f64(),
        );
        (!backoff.is_zero()).then_some(backoff)
    }
--- a/pageserver/page_api/src/model.rs
+++ b/pageserver/page_api/src/model.rs
@@ -33,8 +33,6 @@ pub enum ProtocolError {
    Invalid(&'static str, String),
    #[error("required field '{0}' is missing")]
    Missing(&'static str),
-    #[error("invalid combination of not_modified_lsn '{0}' and request_lsn '{1}'")]
-    InvalidLsns(Lsn, Lsn),
 }

 impl ProtocolError {
@@ -51,7 +49,7 @@ impl From<ProtocolError> for tonic::Status {
 }

 /// The LSN a request should read at.
-#[derive(Clone, Copy, Debug)]
+#[derive(Clone, Copy, Debug, Default)]
 pub struct ReadLsn {
    /// The request's read LSN.
    pub request_lsn: Lsn,
@@ -87,9 +85,9 @@ impl TryFrom<proto::ReadLsn> for ReadLsn {
            return Err(ProtocolError::invalid("request_lsn", pb.request_lsn));
        }
        if pb.not_modified_since_lsn > pb.request_lsn {
-            return Err(ProtocolError::InvalidLsns(
-                Lsn(pb.not_modified_since_lsn),
-                Lsn(pb.request_lsn),
+            return Err(ProtocolError::invalid(
+                "not_modified_since_lsn",
+                pb.not_modified_since_lsn,
            ));
        }
        Ok(Self {
@@ -331,7 +329,7 @@ impl From<GetDbSizeResponse> for proto::GetDbSizeResponse {
 }

 /// Requests one or more pages.
-#[derive(Clone, Debug)]
+#[derive(Clone, Debug, Default)]
 pub struct GetPageRequest {
    /// A request ID. Will be included in the response. Should be unique for in-flight requests on
    /// the stream.
@@ -432,12 +430,13 @@ impl From<RequestID> for proto::RequestId {
 }

 /// A GetPage request class.
-#[derive(Clone, Copy, Debug, strum_macros::Display)]
+#[derive(Clone, Copy, Debug, Default, strum_macros::Display)]
 pub enum GetPageClass {
    /// Unknown class. For backwards compatibility: used when an older client version sends a class
    /// that a newer server version has removed.
    Unknown,
    /// A normal request. This is the default.
+    #[default]
    Normal,
    /// A prefetch request. NB: can only be classified on pg < 18.
    Prefetch,
@@ -445,19 +444,6 @@ pub enum GetPageClass {
    Background,
 }

-impl GetPageClass {
-    /// Returns true if this is considered a bulk request (i.e. more throughput-oriented rather than
-    /// latency-sensitive).
-    pub fn is_bulk(&self) -> bool {
-        match self {
-            Self::Unknown => false,
-            Self::Normal => false,
-            Self::Prefetch => true,
-            Self::Background => true,
-        }
-    }
-}
-
 impl From<proto::GetPageClass> for GetPageClass {
    fn from(pb: proto::GetPageClass) -> Self {
        match pb {
--- a/pageserver/pagebench/Cargo.toml
+++ b/pageserver/pagebench/Cargo.toml
@@ -16,6 +16,7 @@ futures.workspace = true
 hdrhistogram.workspace = true
 humantime.workspace = true
 humantime-serde.workspace = true
+pprof.workspace = true
 rand.workspace = true
 reqwest.workspace = true
 serde.workspace = true
@@ -24,9 +25,6 @@ tracing.workspace = true
 tokio.workspace = true
 tokio-stream.workspace = true
 tokio-util.workspace = true
-axum.workspace = true
-http.workspace = true
-metrics.workspace = true
 tonic.workspace = true
 url.workspace = true

--- a/pageserver/pagebench/src/cmd/getpage_latest_lsn.rs
+++ b/pageserver/pagebench/src/cmd/getpage_latest_lsn.rs
@@ -34,10 +34,6 @@ use crate::util::{request_stats, tokio_thread_local_stats};
 /// GetPage@LatestLSN, uniformly distributed across the compute-accessible keyspace.
 #[derive(clap::Parser)]
 pub(crate) struct Args {
-    #[clap(long, default_value = "false")]
-    grpc: bool,
-    #[clap(long, default_value = "false")]
-    grpc_stream: bool,
    #[clap(long, default_value = "http://localhost:9898")]
    mgmt_api_endpoint: String,
    /// Pageserver connection string. Supports postgresql:// and grpc:// protocols.
@@ -82,9 +78,6 @@ pub(crate) struct Args {
    #[clap(long)]
    set_io_mode: Option<pageserver_api::models::virtual_file::IoMode>,

-    #[clap(long)]
-    only_relnode: Option<u32>,
-
    /// Queue depth generated in each client.
    #[clap(long, default_value = "1")]
    queue_depth: NonZeroUsize,
@@ -99,31 +92,10 @@ pub(crate) struct Args {
    #[clap(long, default_value = "1")]
    batch_size: NonZeroUsize,

+    #[clap(long)]
+    only_relnode: Option<u32>,
+
    targets: Option<Vec<TenantTimelineId>>,
-
-    #[clap(long, default_value = "100")]
-    pool_max_consumers: NonZeroUsize,
-
-    #[clap(long, default_value = "5")]
-    pool_error_threshold: NonZeroUsize,
-
-    #[clap(long, default_value = "5000")]
-    pool_connect_timeout: NonZeroUsize,
-
-    #[clap(long, default_value = "1000")]
-    pool_connect_backoff: NonZeroUsize,
-
-    #[clap(long, default_value = "60000")]
-    pool_max_idle_duration: NonZeroUsize,
-
-    #[clap(long, default_value = "0")]
-    max_delay_ms: usize,
-
-    #[clap(long, default_value = "0")]
-    percent_drops: usize,
-
-    #[clap(long, default_value = "0")]
-    percent_hangs: usize,
 }

 /// State shared by all clients
@@ -180,6 +152,7 @@ pub(crate) fn main(args: Args) -> anyhow::Result<()> {
        main_impl(args, thread_local_stats)
    })
 }
+
 async fn main_impl(
    args: Args,
    all_thread_local_stats: AllThreadLocalStats<request_stats::Stats>,
@@ -344,7 +317,6 @@ async fn main_impl(
    let rps_period = args
        .per_client_rate
        .map(|rps_limit| Duration::from_secs_f64(1.0 / (rps_limit as f64)));
-
    let make_worker: &dyn Fn(WorkerId) -> Pin<Box<dyn Send + Future<Output = ()>>> = &|worker_id| {
        let ss = shared_state.clone();
        let cancel = cancel.clone();
--- a/pageserver/pagebench/src/cmd/idle_streams.rs
+++ b/pageserver/pagebench/src/cmd/idle_streams.rs
@@ -0,0 +1,127 @@
+use std::sync::Arc;
+
+use anyhow::anyhow;
+use futures::StreamExt;
+use tonic::transport::Endpoint;
+use tracing::info;
+
+use pageserver_page_api::{GetPageClass, GetPageRequest, GetPageStatusCode, ReadLsn, RelTag};
+use utils::id::TenantTimelineId;
+use utils::lsn::Lsn;
+use utils::shard::ShardIndex;
+
+/// Starts a large number of idle gRPC GetPage streams.
+#[derive(clap::Parser)]
+pub(crate) struct Args {
+    /// The Pageserver to connect to. Must use grpc://.
+    #[clap(long, default_value = "grpc://localhost:51051")]
+    server: String,
+    /// The Pageserver HTTP API.
+    #[clap(long, default_value = "http://localhost:9898")]
+    http_server: String,
+    /// The number of streams to open.
+    #[clap(long, default_value = "100000")]
+    count: usize,
+    /// Number of streams per connection.
+    #[clap(long, default_value = "100")]
+    per_connection: usize,
+    /// Send a single GetPage request on each stream.
+    #[clap(long, default_value_t = false)]
+    send_request: bool,
+}
+
+pub(crate) fn main(args: Args) -> anyhow::Result<()> {
+    let rt = tokio::runtime::Builder::new_multi_thread()
+        .enable_all()
+        .build()?;
+
+    rt.block_on(main_impl(args))
+}
+
+async fn main_impl(args: Args) -> anyhow::Result<()> {
+    // Discover a tenant and timeline to use.
+    let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
+        reqwest::Client::new(),
+        args.http_server.clone(),
+        None,
+    ));
+    let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
+        &mgmt_api_client,
+        crate::util::cli::targets::Spec {
+            limit_to_first_n_targets: Some(1),
+            targets: None,
+        },
+    )
+    .await?;
+    let ttid = timelines
+        .first()
+        .ok_or_else(|| anyhow!("no timelines found"))?;
+
+    // Set up the initial client.
+    let endpoint = Endpoint::from_shared(args.server.clone())?;
+
+    let connect = async || {
+        pageserver_page_api::Client::new(
+            endpoint.connect().await?,
+            ttid.tenant_id,
+            ttid.timeline_id,
+            ShardIndex::unsharded(),
+            None,
+            None,
+        )
+    };
+
+    let mut client = connect().await?;
+    let mut streams = Vec::with_capacity(args.count);
+
+    // Create streams.
+    for i in 0..args.count {
+        if i % 100 == 0 {
+            info!("opened {}/{} streams", i, args.count);
+        }
+        if i % args.per_connection == 0 && i > 0 {
+            client = connect().await?;
+        }
+
+        let (req_tx, req_rx) = tokio::sync::mpsc::unbounded_channel();
+        let req_stream = tokio_stream::wrappers::UnboundedReceiverStream::new(req_rx);
+        let mut resp_stream = client.get_pages(req_stream).await?;
+
+        // Send request if specified.
+        if args.send_request {
+            req_tx.send(GetPageRequest {
+                request_id: 1.into(),
+                request_class: GetPageClass::Normal,
+                read_lsn: ReadLsn {
+                    request_lsn: Lsn::MAX,
+                    not_modified_since_lsn: Some(Lsn(1)),
+                },
+                rel: RelTag {
+                    spcnode: 1664, // pg_global
+                    dbnode: 0,     // shared database
+                    relnode: 1262, // pg_authid
+                    forknum: 0,    // init
+                },
+                block_numbers: vec![0],
+            })?;
+            let resp = resp_stream
+                .next()
+                .await
+                .transpose()?
+                .ok_or_else(|| anyhow!("no response"))?;
+            if resp.status_code != GetPageStatusCode::Ok {
+                return Err(anyhow!("{} response", resp.status_code));
+            }
+        }
+
+        // Hold onto streams to avoid closing them.
+        streams.push((req_tx, resp_stream));
+    }
+
+    info!("opened {} streams, sleeping", args.count);
+
+    // Block forever, to hold the idle streams open for inspection.
+    futures::future::pending::<()>().await;
+
+    Ok(())
+}
--- a/pageserver/pagebench/src/main.rs
+++ b/pageserver/pagebench/src/main.rs
@@ -1,4 +1,7 @@
+use std::fs::File;
+
 use clap::Parser;
+use tracing::info;
 use utils::logging;

 /// Re-usable pieces of code that aren't CLI-specific.
@@ -17,38 +20,73 @@ mod cmd {
    pub(super) mod aux_files;
    pub(super) mod basebackup;
    pub(super) mod getpage_latest_lsn;
+    pub(super) mod idle_streams;
    pub(super) mod ondemand_download_churn;
    pub(super) mod trigger_initial_size_calculation;
 }

 /// Component-level performance test for pageserver.
 #[derive(clap::Parser)]
-enum Args {
+struct Args {
+    /// Takes a client CPU profile into profile.svg. The benchmark must exit cleanly before it's
+    /// written, e.g. via --runtime.
+    #[arg(long)]
+    profile: bool,
+
+    #[command(subcommand)]
+    subcommand: Subcommand,
+}
+
+#[derive(clap::Subcommand)]
+enum Subcommand {
    Basebackup(cmd::basebackup::Args),
    GetPageLatestLsn(cmd::getpage_latest_lsn::Args),
    TriggerInitialSizeCalculation(cmd::trigger_initial_size_calculation::Args),
    OndemandDownloadChurn(cmd::ondemand_download_churn::Args),
    AuxFiles(cmd::aux_files::Args),
+    IdleStreams(cmd::idle_streams::Args),
 }

-fn main() {
+fn main() -> anyhow::Result<()> {
    logging::init(
        logging::LogFormat::Plain,
        logging::TracingErrorLayerEnablement::Disabled,
        logging::Output::Stderr,
-    )
-    .unwrap();
+    )?;
    logging::replace_panic_hook_with_tracing_panic_hook().forget();

    let args = Args::parse();
-    match args {
-        Args::Basebackup(args) => cmd::basebackup::main(args),
-        Args::GetPageLatestLsn(args) => cmd::getpage_latest_lsn::main(args),
-        Args::TriggerInitialSizeCalculation(args) => {
+
+    // Start a CPU profile if requested.
+    let mut profiler = None;
+    if args.profile {
+        profiler = Some(
+            pprof::ProfilerGuardBuilder::default()
+                .frequency(1000)
+                .blocklist(&["libc", "libgcc", "pthread", "vdso"])
+                .build()?,
+        );
+    }
+
+    match args.subcommand {
+        Subcommand::Basebackup(args) => cmd::basebackup::main(args),
+        Subcommand::GetPageLatestLsn(args) => cmd::getpage_latest_lsn::main(args),
+        Subcommand::TriggerInitialSizeCalculation(args) => {
            cmd::trigger_initial_size_calculation::main(args)
        }
-        Args::OndemandDownloadChurn(args) => cmd::ondemand_download_churn::main(args),
-        Args::AuxFiles(args) => cmd::aux_files::main(args),
+        Subcommand::OndemandDownloadChurn(args) => cmd::ondemand_download_churn::main(args),
+        Subcommand::AuxFiles(args) => cmd::aux_files::main(args),
+        Subcommand::IdleStreams(args) => cmd::idle_streams::main(args),
+    }?;
+
+    // Generate a CPU flamegraph if requested.
+    if let Some(profiler) = profiler {
+        let report = profiler.report().build()?;
+        drop(profiler); // stop profiling
+        let file = File::create("profile.svg")?;
+        report.flamegraph(file)?;
+        info!("wrote CPU profile flamegraph to profile.svg")
    }
-    .unwrap()
+
+    Ok(())
 }
--- a/pageserver/src/basebackup.rs
+++ b/pageserver/src/basebackup.rs
@@ -114,7 +114,7 @@ where
    // Compute postgres doesn't have any previous WAL files, but the first
    // record that it's going to write needs to include the LSN of the
    // previous record (xl_prev). We include prev_record_lsn in the
-    // "zenith.signal" file, so that postgres can read it during startup.
+    // "neon.signal" file, so that postgres can read it during startup.
    //
    // We don't keep full history of record boundaries in the page server,
    // however, only the predecessor of the latest record on each
@@ -751,34 +751,39 @@ where

    //
    // Add generated pg_control file and bootstrap WAL segment.
-    // Also send zenith.signal file with extra bootstrap data.
+    // Also send neon.signal and zenith.signal file with extra bootstrap data.
    //
    async fn add_pgcontrol_file(
        &mut self,
        pg_control_bytes: Bytes,
        system_identifier: u64,
    ) -> Result<(), BasebackupError> {
-        // add zenith.signal file
-        let mut zenith_signal = String::new();
+        // add neon.signal file
+        let mut neon_signal = String::new();
        if self.prev_record_lsn == Lsn(0) {
            if self.timeline.is_ancestor_lsn(self.lsn) {
-                write!(zenith_signal, "PREV LSN: none")
+                write!(neon_signal, "PREV LSN: none")
                    .map_err(|e| BasebackupError::Server(e.into()))?;
            } else {
-                write!(zenith_signal, "PREV LSN: invalid")
+                write!(neon_signal, "PREV LSN: invalid")
                    .map_err(|e| BasebackupError::Server(e.into()))?;
            }
        } else {
-            write!(zenith_signal, "PREV LSN: {}", self.prev_record_lsn)
+            write!(neon_signal, "PREV LSN: {}", self.prev_record_lsn)
                .map_err(|e| BasebackupError::Server(e.into()))?;
        }
-        self.ar
-            .append(
-                &new_tar_header("zenith.signal", zenith_signal.len() as u64)?,
-                zenith_signal.as_bytes(),
-            )
-            .await
-            .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,zenith.signal"))?;
+
+        // TODO: Remove zenith.signal once all historical computes have been replaced
+        // ... and thus support the neon.signal file.
+        for signalfilename in ["neon.signal", "zenith.signal"] {
+            self.ar
+                .append(
+                    &new_tar_header(signalfilename, neon_signal.len() as u64)?,
+                    neon_signal.as_bytes(),
+                )
+                .await
+                .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,neon.signal"))?;
+        }

        //send pg_control
        let header = new_tar_header("global/pg_control", pg_control_bytes.len() as u64)?;
--- a/pageserver/src/bin/pageserver.rs
+++ b/pageserver/src/bin/pageserver.rs
@@ -917,11 +917,6 @@ async fn create_remote_storage_client(
    // If `test_remote_failures` is non-zero, wrap the client with a
    // wrapper that simulates failures.
    if conf.test_remote_failures > 0 {
-        if !cfg!(feature = "testing") {
-            anyhow::bail!(
-                "test_remote_failures option is not available because pageserver was compiled without the 'testing' feature"
-            );
-        }
        info!(
            "Simulating remote failures for first {} attempts of each op",
            conf.test_remote_failures
--- a/pageserver/src/controller_upcall_client.rs
+++ b/pageserver/src/controller_upcall_client.rs
@@ -194,6 +194,7 @@ impl StorageControllerUpcallApi for StorageControllerUpcallClient {
                        listen_http_port: m.http_port,
                        listen_https_port: m.https_port,
                        availability_zone_id: az_id.expect("Checked above"),
+                        node_ip_addr: None,
                    })
                }
                Err(e) => {
--- a/pageserver/src/deletion_queue/validator.rs
+++ b/pageserver/src/deletion_queue/validator.rs
@@ -1,5 +1,5 @@
 //! The validator is responsible for validating DeletionLists for execution,
-//! based on whethe the generation in the DeletionList is still the latest
+//! based on whether the generation in the DeletionList is still the latest
 //! generation for a tenant.
 //!
 //! The purpose of validation is to ensure split-brain safety in the cluster
--- a/pageserver/src/http/routes.rs
+++ b/pageserver/src/http/routes.rs
@@ -10,6 +10,7 @@ use std::sync::Arc;
 use std::time::Duration;

 use anyhow::{Context, Result, anyhow};
+use bytes::Bytes;
 use enumset::EnumSet;
 use futures::future::join_all;
 use futures::{StreamExt, TryFutureExt};
@@ -46,6 +47,7 @@ use pageserver_api::shard::{ShardCount, TenantShardId};
 use postgres_ffi::PgMajorVersion;
 use remote_storage::{DownloadError, GenericRemoteStorage, TimeTravelError};
 use scopeguard::defer;
+use serde::{Deserialize, Serialize};
 use serde_json::json;
 use tenant_size_model::svg::SvgBranchKind;
 use tenant_size_model::{SizeResult, StorageModel};
@@ -57,6 +59,7 @@ use utils::auth::SwappableJwtAuth;
 use utils::generation::Generation;
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
+use wal_decoder::models::record::NeonWalRecord;

 use crate::config::PageServerConf;
 use crate::context;
@@ -77,6 +80,7 @@ use crate::tenant::remote_timeline_client::{
 };
 use crate::tenant::secondary::SecondaryController;
 use crate::tenant::size::ModelInputs;
+use crate::tenant::storage_layer::ValuesReconstructState;
 use crate::tenant::storage_layer::{IoConcurrency, LayerAccessStatsReset, LayerName};
 use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 use crate::tenant::timeline::offload::{OffloadError, offload_timeline};
@@ -397,6 +401,7 @@ async fn build_timeline_info(
    timeline: &Arc<Timeline>,
    include_non_incremental_logical_size: bool,
    force_await_initial_logical_size: bool,
+    include_image_consistent_lsn: bool,
    ctx: &RequestContext,
 ) -> anyhow::Result<TimelineInfo> {
    crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id();
@@ -421,6 +426,10 @@ async fn build_timeline_info(
                .await?,
        );
    }
+    // HADRON
+    if include_image_consistent_lsn {
+        info.image_consistent_lsn = Some(timeline.compute_image_consistent_lsn().await?);
+    }
    Ok(info)
 }

@@ -510,6 +519,8 @@ async fn build_timeline_info_common(
        is_invisible: Some(is_invisible),

        walreceiver_status,
+        // HADRON
+        image_consistent_lsn: None,
    };
    Ok(info)
 }
@@ -712,6 +723,8 @@ async fn timeline_list_handler(
        parse_query_param(&request, "include-non-incremental-logical-size")?;
    let force_await_initial_logical_size: Option<bool> =
        parse_query_param(&request, "force-await-initial-logical-size")?;
+    let include_image_consistent_lsn: Option<bool> =
+        parse_query_param(&request, "include-image-consistent-lsn")?;
    check_permission(&request, Some(tenant_shard_id.tenant_id))?;

    let state = get_state(&request);
@@ -732,6 +745,7 @@ async fn timeline_list_handler(
                &timeline,
                include_non_incremental_logical_size.unwrap_or(false),
                force_await_initial_logical_size.unwrap_or(false),
+                include_image_consistent_lsn.unwrap_or(false),
                &ctx,
            )
            .instrument(info_span!("build_timeline_info", timeline_id = %timeline.timeline_id))
@@ -760,6 +774,9 @@ async fn timeline_and_offloaded_list_handler(
        parse_query_param(&request, "include-non-incremental-logical-size")?;
    let force_await_initial_logical_size: Option<bool> =
        parse_query_param(&request, "force-await-initial-logical-size")?;
+    let include_image_consistent_lsn: Option<bool> =
+        parse_query_param(&request, "include-image-consistent-lsn")?;
+
    check_permission(&request, Some(tenant_shard_id.tenant_id))?;

    let state = get_state(&request);
@@ -780,6 +797,7 @@ async fn timeline_and_offloaded_list_handler(
                &timeline,
                include_non_incremental_logical_size.unwrap_or(false),
                force_await_initial_logical_size.unwrap_or(false),
+                include_image_consistent_lsn.unwrap_or(false),
                &ctx,
            )
            .instrument(info_span!("build_timeline_info", timeline_id = %timeline.timeline_id))
@@ -964,6 +982,9 @@ async fn timeline_detail_handler(
        parse_query_param(&request, "include-non-incremental-logical-size")?;
    let force_await_initial_logical_size: Option<bool> =
        parse_query_param(&request, "force-await-initial-logical-size")?;
+    // HADRON
+    let include_image_consistent_lsn: Option<bool> =
+        parse_query_param(&request, "include-image-consistent-lsn")?;
    check_permission(&request, Some(tenant_shard_id.tenant_id))?;

    // Logical size calculation needs downloading.
@@ -984,6 +1005,7 @@ async fn timeline_detail_handler(
            &timeline,
            include_non_incremental_logical_size.unwrap_or(false),
            force_await_initial_logical_size.unwrap_or(false),
+            include_image_consistent_lsn.unwrap_or(false),
            ctx,
        )
        .await
@@ -2690,6 +2712,16 @@ async fn deletion_queue_flush(
    }
 }

+/// Try if `GetPage@Lsn` is successful, useful for manual debugging.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+struct GetPageResponse {
+    pub page: Bytes,
+    pub layers_visited: u32,
+    pub delta_layers_visited: u32,
+    pub records: Vec<(Lsn, NeonWalRecord)>,
+    pub img: Option<(Lsn, Bytes)>,
+}
+
 async fn getpage_at_lsn_handler(
    request: Request<Body>,
    cancel: CancellationToken,
@@ -2740,21 +2772,24 @@ async fn getpage_at_lsn_handler_inner(

        // Use last_record_lsn if no lsn is provided
        let lsn = lsn.unwrap_or_else(|| timeline.get_last_record_lsn());
-        let page = timeline.get(key.0, lsn, &ctx).await?;

        if touch {
            json_response(StatusCode::OK, ())
        } else {
-            Result::<_, ApiError>::Ok(
-                Response::builder()
-                    .status(StatusCode::OK)
-                    .header(header::CONTENT_TYPE, "application/octet-stream")
-                    .body(hyper::Body::from(page))
-                    .unwrap(),
-            )
+            let mut reconstruct_state = ValuesReconstructState::new_with_debug(IoConcurrency::sequential());
+            let page = timeline.debug_get(key.0, lsn, &ctx, &mut reconstruct_state).await?;
+            let response = GetPageResponse {
+                page,
+                layers_visited: reconstruct_state.get_layers_visited(),
+                delta_layers_visited: reconstruct_state.get_delta_layers_visited(),
+                records: reconstruct_state.debug_state.records.clone(),
+                img: reconstruct_state.debug_state.img.clone(),
+            };
+
+            json_response(StatusCode::OK, response)
        }
    }
-    .instrument(info_span!("timeline_get", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
+    .instrument(info_span!("timeline_debug_get", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
    .await
 }

@@ -3643,6 +3678,7 @@ async fn activate_post_import_handler(
        let timeline_info = build_timeline_info(
            &timeline, false, // include_non_incremental_logical_size,
            false, // force_await_initial_logical_size
+            false, // include_image_consistent_lsn
            &ctx,
        )
        .await
@@ -4164,7 +4200,7 @@ pub fn make_router(
        })
        .get(
            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/getpage",
-            |r| testing_api_handler("getpage@lsn", r, getpage_at_lsn_handler),
+            |r|  testing_api_handler("getpage@lsn", r, getpage_at_lsn_handler),
        )
        .get(
            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/touchpage",
--- a/pageserver/src/import_datadir.rs
+++ b/pageserver/src/import_datadir.rs
@@ -610,13 +610,13 @@ async fn import_file(
        debug!("imported twophase file");
    } else if file_path.starts_with("pg_wal") {
        debug!("found wal file in base section. ignore it");
-    } else if file_path.starts_with("zenith.signal") {
+    } else if file_path.starts_with("zenith.signal") || file_path.starts_with("neon.signal") {
        // Parse zenith signal file to set correct previous LSN
        let bytes = read_all_bytes(reader).await?;
-        // zenith.signal format is "PREV LSN: prev_lsn"
+        // neon.signal format is "PREV LSN: prev_lsn"
        // TODO write serialization and deserialization in the same place.
-        let zenith_signal = std::str::from_utf8(&bytes)?.trim();
-        let prev_lsn = match zenith_signal {
+        let neon_signal = std::str::from_utf8(&bytes)?.trim();
+        let prev_lsn = match neon_signal {
            "PREV LSN: none" => Lsn(0),
            "PREV LSN: invalid" => Lsn(0),
            other => {
@@ -624,17 +624,17 @@ async fn import_file(
                split[1]
                    .trim()
                    .parse::<Lsn>()
-                    .context("can't parse zenith.signal")?
+                    .context("can't parse neon.signal")?
            }
        };

-        // zenith.signal is not necessarily the last file, that we handle
+        // neon.signal is not necessarily the last file, that we handle
        // but it is ok to call `finish_write()`, because final `modification.commit()`
        // will update lsn once more to the final one.
        let writer = modification.tline.writer().await;
        writer.finish_write(prev_lsn);

-        debug!("imported zenith signal {}", prev_lsn);
+        debug!("imported neon signal {}", prev_lsn);
    } else if file_path.starts_with("pg_tblspc") {
        // TODO Backups exported from neon won't have pg_tblspc, but we will need
        // this to import arbitrary postgres databases.
--- a/pageserver/src/page_service.rs
+++ b/pageserver/src/page_service.rs
@@ -3218,7 +3218,6 @@ where
 pub struct GrpcPageServiceHandler {
    tenant_manager: Arc<TenantManager>,
    ctx: RequestContext,
-    cancel: CancellationToken,
    gate_guard: GateGuard,
    get_vectored_concurrent_io: GetVectoredConcurrentIo,
 }
@@ -3271,7 +3270,6 @@ impl GrpcPageServiceHandler {
        let page_service_handler = GrpcPageServiceHandler {
            tenant_manager,
            ctx,
-            cancel: cancel.clone(),
            gate_guard: gate.enter().expect("gate was just created"),
            get_vectored_concurrent_io,
        };
@@ -3408,8 +3406,6 @@ impl GrpcPageServiceHandler {
    /// TODO: get_vectored() currently enforces a batch limit of 32. Postgres will typically send
    /// batches up to effective_io_concurrency = 100. Either we have to accept large batches, or
    /// split them up in the client or server.
-    ///
-    /// TODO: verify that the given keys belong to this shard.
    #[instrument(skip_all, fields(req_id, rel, blkno, blks, req_lsn, mod_lsn))]
    async fn get_page(
        ctx: &RequestContext,
@@ -3731,7 +3727,6 @@ impl proto::PageService for GrpcPageServiceHandler {
        // Spawn a task to handle the GetPageRequest stream.
        let span = Span::current();
        let ctx = self.ctx.attached_child();
-        let cancel = self.cancel.clone();
        let mut reqs = req.into_inner();

        let resps = async_stream::try_stream! {
@@ -3739,18 +3734,8 @@ impl proto::PageService for GrpcPageServiceHandler {
                .get(ttid.tenant_id, ttid.timeline_id, shard_selector)
                .await?
                .downgrade();
-
-            loop {
-                let req = tokio::select! {
-                    req = reqs.message() => req,
-                    _ = cancel.cancelled() => {
-                        tracing::info!("closing getpages stream due to shutdown");
-                        break;
-                    },
-                };
-                let Some(req) = req? else { break };
+            while let Some(req) = reqs.message().await? {
                let req_id = req.request_id.map(page_api::RequestID::from).unwrap_or_default();
-
                let result = Self::get_page(&ctx, &timeline, req, io_concurrency.clone())
                    .instrument(span.clone()) // propagate request span
                    .await;
--- a/pageserver/src/pgdatadir_mapping.rs
+++ b/pageserver/src/pgdatadir_mapping.rs
@@ -25,9 +25,9 @@ use pageserver_api::keyspace::{KeySpaceRandomAccum, SparseKeySpace};
 use pageserver_api::models::RelSizeMigration;
 use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
 use pageserver_api::shard::ShardIdentity;
-use postgres_ffi::{BLCKSZ, PgMajorVersion, TimestampTz, TransactionId};
+use postgres_ffi::{BLCKSZ, PgMajorVersion, TransactionId};
 use postgres_ffi_types::forknum::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
-use postgres_ffi_types::{Oid, RepOriginId};
+use postgres_ffi_types::{Oid, RepOriginId, TimestampTz};
 use serde::{Deserialize, Serialize};
 use strum::IntoEnumIterator;
 use tokio_util::sync::CancellationToken;
--- a/pageserver/src/tenant.rs
+++ b/pageserver/src/tenant.rs
@@ -3393,7 +3393,13 @@ impl TenantShard {
                .collect_vec();

            for timeline in timelines {
-                timeline.maybe_freeze_ephemeral_layer().await;
+                // Include a span with the timeline ID. The parent span already has the tenant ID.
+                let span =
+                    info_span!("maybe_freeze_ephemeral_layer", timeline_id = %timeline.timeline_id);
+                timeline
+                    .maybe_freeze_ephemeral_layer()
+                    .instrument(span)
+                    .await;
            }
        }

@@ -12816,6 +12822,40 @@ mod tests {
                },
            ]
        );
+
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_get_force_image_creation_lsn() -> anyhow::Result<()> {
+        let tenant_conf = pageserver_api::models::TenantConfig {
+            pitr_interval: Some(Duration::from_secs(7 * 3600)),
+            image_layer_force_creation_period: Some(Duration::from_secs(3600)),
+            ..Default::default()
+        };
+
+        let tenant_id = TenantId::generate();
+
+        let harness = TenantHarness::create_custom(
+            "test_get_force_image_creation_lsn",
+            tenant_conf,
+            tenant_id,
+            ShardIdentity::unsharded(),
+            Generation::new(1),
+        )
+        .await?;
+        let (tenant, ctx) = harness.load().await;
+        let timeline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
+            .await?;
+        timeline.gc_info.write().unwrap().cutoffs.time = Some(Lsn(100));
+        {
+            let writer = timeline.writer().await;
+            writer.finish_write(Lsn(5000));
+        }
+
+        let image_creation_lsn = timeline.get_force_image_creation_lsn().unwrap();
+        assert_eq!(image_creation_lsn, Lsn(4300));
        Ok(())
    }
 }
--- a/pageserver/src/tenant/layer_map.rs
+++ b/pageserver/src/tenant/layer_map.rs
@@ -46,10 +46,11 @@
 mod historic_layer_coverage;
 mod layer_coverage;

-use std::collections::{HashMap, VecDeque};
+use std::collections::{BTreeMap, HashMap, VecDeque};
 use std::iter::Peekable;
 use std::ops::Range;
 use std::sync::Arc;
+use std::time::Instant;

 use anyhow::Result;
 use historic_layer_coverage::BufferedHistoricLayerCoverage;
@@ -904,6 +905,103 @@ impl LayerMap {
        max_stacked_deltas
    }

+    /* BEGIN_HADRON */
+    /**
+     * Compute the image consistent LSN, the largest LSN below which all pages have been redone successfully.
+     * It works by first finding the latest image layers and store them into a map. Then for each delta layer,
+     * find all overlapping image layers in order to potentially increase the image LSN in case there are gaps
+     * (e.g., if an image is created at LSN 100 but the delta layer spans LSN [150, 200], then we can increase
+     * image LSN to 150 because there is no WAL record in between).
+     * Finally, the image consistent LSN is computed by taking the minimum of all image layers.
+     */
+    pub fn compute_image_consistent_lsn(&self, disk_consistent_lsn: Lsn) -> Lsn {
+        struct ImageLayerInfo {
+            // creation LSN of the image layer
+            image_lsn: Lsn,
+            // the current minimum LSN of newer delta layers with overlapping key ranges
+            min_delta_lsn: Lsn,
+        }
+        let started_at = Instant::now();
+
+        let min_l0_deltas_lsn = {
+            let l0_deltas = self.level0_deltas();
+            l0_deltas
+                .iter()
+                .map(|layer| layer.get_lsn_range().start)
+                .min()
+                .unwrap_or(disk_consistent_lsn)
+        };
+        let global_key_range = Key::MIN..Key::MAX;
+
+        // step 1: collect all most recent image layers into a map
+        // map: end key to image_layer_info
+        let mut image_map: BTreeMap<Key, ImageLayerInfo> = BTreeMap::new();
+        for (img_range, img) in self.image_coverage(&global_key_range, disk_consistent_lsn) {
+            let img_lsn = img.map(|layer| layer.get_lsn_range().end).unwrap_or(Lsn(0));
+            image_map.insert(
+                img_range.end,
+                ImageLayerInfo {
+                    image_lsn: img_lsn,
+                    min_delta_lsn: min_l0_deltas_lsn,
+                },
+            );
+        }
+
+        // step 2: go through all delta layers, and update the image layer info with overlapping
+        // key ranges
+        for layer in self.historic.iter() {
+            if !layer.is_delta {
+                continue;
+            }
+            let delta_key_range = layer.get_key_range();
+            let delta_lsn_range = layer.get_lsn_range();
+            for (img_end_key, img_info) in image_map.range_mut(delta_key_range.start..Key::MAX) {
+                debug_assert!(img_end_key >= &delta_key_range.start);
+                if delta_lsn_range.end > img_info.image_lsn {
+                    // the delta layer includes WAL records after the image
+                    // it's possibel that the delta layer's start LSN < image LSN, which will be simply ignored by step 3
+                    img_info.min_delta_lsn =
+                        std::cmp::min(img_info.min_delta_lsn, delta_lsn_range.start);
+                }
+                if img_end_key >= &delta_key_range.end {
+                    // we have fully processed all overlapping image layers
+                    break;
+                }
+            }
+        }
+
+        // step 3, go through all image layers and find the image consistent LSN
+        let mut img_consistent_lsn = min_l0_deltas_lsn.checked_sub(Lsn(1)).unwrap();
+        let mut prev_key = Key::MIN;
+        for (img_key, img_info) in image_map {
+            tracing::debug!(
+                "Image layer {:?}:{} has min delta lsn {}",
+                Range {
+                    start: prev_key,
+                    end: img_key,
+                },
+                img_info.image_lsn,
+                img_info.min_delta_lsn,
+            );
+            let image_lsn = std::cmp::max(
+                img_info.image_lsn,
+                img_info.min_delta_lsn.checked_sub(Lsn(1)).unwrap_or(Lsn(0)),
+            );
+            img_consistent_lsn = std::cmp::min(img_consistent_lsn, image_lsn);
+            prev_key = img_key;
+        }
+        tracing::info!(
+            "computed image_consistent_lsn {} for disk_consistent_lsn {} in {}ms. Processed {} layrs in total.",
+            img_consistent_lsn,
+            disk_consistent_lsn,
+            started_at.elapsed().as_millis(),
+            self.historic.len()
+        );
+        img_consistent_lsn
+    }
+
+    /* END_HADRON */
+
    /// Return all L0 delta layers
    pub fn level0_deltas(&self) -> &Vec<Arc<PersistentLayerDesc>> {
        &self.l0_delta_layers
@@ -1579,6 +1677,138 @@ mod tests {
            LayerVisibilityHint::Visible
        ));
    }
+
+    /* BEGIN_HADRON */
+    #[test]
+    fn test_compute_image_consistent_lsn() {
+        let mut layer_map = LayerMap::default();
+
+        let disk_consistent_lsn = Lsn(1000);
+        // case 1: empty layer map
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(
+            disk_consistent_lsn.checked_sub(Lsn(1)).unwrap(),
+            image_consistent_lsn
+        );
+
+        // case 2: only L0 delta layer
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(100),
+                Lsn(900)..Lsn(990),
+                true,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(100),
+                Lsn(850)..Lsn(899),
+                true,
+            ));
+        }
+
+        // should use min L0 delta LSN - 1 as image consistent LSN
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(849), image_consistent_lsn);
+
+        // case 3: 3 images, no L1 delta
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(40),
+                Lsn(100)..Lsn(100),
+                false,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(40)..Key::from_i128(70),
+                Lsn(200)..Lsn(200),
+                false,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(70)..Key::from_i128(100),
+                Lsn(150)..Lsn(150),
+                false,
+            ));
+        }
+        // should use min L0 delta LSN - 1 as image consistent LSN
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(849), image_consistent_lsn);
+
+        // case 4: 3 images with 1 L1 delta
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(50),
+                Lsn(300)..Lsn(350),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(299), image_consistent_lsn);
+
+        // case 5: 3 images with 1 more L1 delta with smaller LSN
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(50)..Key::from_i128(72),
+                Lsn(200)..Lsn(300),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(199), image_consistent_lsn);
+
+        // case 6: 3 images with more newer L1 deltas (no impact on final results)
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(30),
+                Lsn(400)..Lsn(500),
+                true,
+            ));
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(35)..Key::from_i128(100),
+                Lsn(450)..Lsn(600),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(199), image_consistent_lsn);
+
+        // case 7: 3 images with more older L1 deltas (no impact on final results)
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(40),
+                Lsn(0)..Lsn(50),
+                true,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(50)..Key::from_i128(100),
+                Lsn(10)..Lsn(60),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(199), image_consistent_lsn);
+
+        // case 8: 3 images with one more L1 delta with overlapping LSN range
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(50),
+                Lsn(50)..Lsn(250),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(100), image_consistent_lsn);
+    }
+
+    /* END_HADRON */
 }

 #[cfg(test)]
--- a/pageserver/src/tenant/mgr.rs
+++ b/pageserver/src/tenant/mgr.rs
@@ -1678,6 +1678,8 @@ impl TenantManager {
        // Phase 6: Release the InProgress on the parent shard
        drop(parent_slot_guard);

+        utils::pausable_failpoint!("shard-split-post-finish-pause");
+
        Ok(child_shards)
    }

--- a/pageserver/src/tenant/storage_layer.rs
+++ b/pageserver/src/tenant/storage_layer.rs
@@ -75,7 +75,7 @@ where
 /// the same ValueReconstructState struct in the next 'get_value_reconstruct_data'
 /// call, to collect more records.
 ///
-#[derive(Debug, Default)]
+#[derive(Debug, Default, Clone)]
 pub(crate) struct ValueReconstructState {
    pub(crate) records: Vec<(Lsn, NeonWalRecord)>,
    pub(crate) img: Option<(Lsn, Bytes)>,
@@ -308,6 +308,9 @@ pub struct ValuesReconstructState {
    layers_visited: u32,
    delta_layers_visited: u32,

+    pub(crate) enable_debug: bool,
+    pub(crate) debug_state: ValueReconstructState,
+
    pub(crate) io_concurrency: IoConcurrency,
    num_active_ios: Arc<AtomicUsize>,

@@ -657,6 +660,23 @@ impl ValuesReconstructState {
            layers_visited: 0,
            delta_layers_visited: 0,
            io_concurrency,
+            enable_debug: false,
+            debug_state: ValueReconstructState::default(),
+            num_active_ios: Arc::new(AtomicUsize::new(0)),
+            read_path: None,
+        }
+    }
+
+    pub(crate) fn new_with_debug(io_concurrency: IoConcurrency) -> Self {
+        Self {
+            keys: HashMap::new(),
+            keys_done: KeySpaceRandomAccum::new(),
+            keys_with_image_coverage: None,
+            layers_visited: 0,
+            delta_layers_visited: 0,
+            io_concurrency,
+            enable_debug: true,
+            debug_state: ValueReconstructState::default(),
            num_active_ios: Arc::new(AtomicUsize::new(0)),
            read_path: None,
        }
@@ -670,6 +690,12 @@ impl ValuesReconstructState {
        self.io_concurrency.spawn_io(fut).await;
    }

+    pub(crate) fn set_debug_state(&mut self, debug_state: &ValueReconstructState) {
+        if self.enable_debug {
+            self.debug_state = debug_state.clone();
+        }
+    }
+
    pub(crate) fn on_layer_visited(&mut self, layer: &ReadableLayer) {
        self.layers_visited += 1;
        if let ReadableLayer::PersistentLayer(layer) = layer {
--- a/pageserver/src/tenant/timeline.rs
+++ b/pageserver/src/tenant/timeline.rs
@@ -351,13 +351,6 @@ pub struct Timeline {
    last_image_layer_creation_check_at: AtomicLsn,
    last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,

-    // HADRON
-    /// If a key range has writes with LSN > force_image_creation_lsn, then we should force image layer creation
-    /// on this key range.
-    force_image_creation_lsn: AtomicLsn,
-    /// The last time instant when force_image_creation_lsn is computed.
-    force_image_creation_lsn_computed_at: std::sync::Mutex<Option<Instant>>,
-
    /// Current logical size of the "datadir", at the last LSN.
    current_logical_size: LogicalSize,

@@ -1260,6 +1253,57 @@ impl Timeline {
        }
    }

+    #[inline(always)]
+    pub(crate) async fn debug_get(
+        &self,
+        key: Key,
+        lsn: Lsn,
+        ctx: &RequestContext,
+        reconstruct_state: &mut ValuesReconstructState,
+    ) -> Result<Bytes, PageReconstructError> {
+        if !lsn.is_valid() {
+            return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
+        }
+
+        // This check is debug-only because of the cost of hashing, and because it's a double-check: we
+        // already checked the key against the shard_identity when looking up the Timeline from
+        // page_service.
+        debug_assert!(!self.shard_identity.is_key_disposable(&key));
+
+        let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
+        let vectored_res = self
+            .debug_get_vectored_impl(query, reconstruct_state, ctx)
+            .await;
+
+        let key_value = vectored_res?.pop_first();
+        match key_value {
+            Some((got_key, value)) => {
+                if got_key != key {
+                    error!(
+                        "Expected {}, but singular vectored get returned {}",
+                        key, got_key
+                    );
+                    Err(PageReconstructError::Other(anyhow!(
+                        "Singular vectored get returned wrong key"
+                    )))
+                } else {
+                    value
+                }
+            }
+            None => Err(PageReconstructError::MissingKey(Box::new(
+                MissingKeyError {
+                    keyspace: KeySpace::single(key..key.next()),
+                    shard: self.shard_identity.get_shard_number(&key),
+                    original_hwm_lsn: lsn,
+                    ancestor_lsn: None,
+                    backtrace: None,
+                    read_path: None,
+                    query: None,
+                },
+            ))),
+        }
+    }
+
    pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;

    /// Look up multiple page versions at a given LSN
@@ -1554,6 +1598,98 @@ impl Timeline {
        Ok(results)
    }

+    // A copy of the get_vectored_impl method except that we store the image and wal records into `reconstruct_state`.
+    // This is only used in the http getpage call for debugging purpose.
+    pub(super) async fn debug_get_vectored_impl(
+        &self,
+        query: VersionedKeySpaceQuery,
+        reconstruct_state: &mut ValuesReconstructState,
+        ctx: &RequestContext,
+    ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
+        if query.is_empty() {
+            return Ok(BTreeMap::default());
+        }
+
+        let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
+            Some(ReadPath::new(
+                query.total_keyspace(),
+                query.high_watermark_lsn()?,
+            ))
+        } else {
+            None
+        };
+
+        reconstruct_state.read_path = read_path;
+
+        let traversal_res: Result<(), _> = self
+            .get_vectored_reconstruct_data(query.clone(), reconstruct_state, ctx)
+            .await;
+
+        if let Err(err) = traversal_res {
+            // Wait for all the spawned IOs to complete.
+            // See comments on `spawn_io` inside `storage_layer` for more details.
+            let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
+                .into_values()
+                .map(|state| state.collect_pending_ios())
+                .collect::<FuturesUnordered<_>>();
+            while collect_futs.next().await.is_some() {}
+            return Err(err);
+        };
+
+        let reconstruct_state = Arc::new(Mutex::new(reconstruct_state));
+        let futs = FuturesUnordered::new();
+
+        for (key, state) in std::mem::take(&mut reconstruct_state.lock().unwrap().keys) {
+            let req_lsn_for_key = query.map_key_to_lsn(&key);
+            futs.push({
+                let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
+                let rc_clone = Arc::clone(&reconstruct_state);
+
+                async move {
+                    assert_eq!(state.situation, ValueReconstructSituation::Complete);
+
+                    let converted = match state.collect_pending_ios().await {
+                        Ok(ok) => ok,
+                        Err(err) => {
+                            return (key, Err(err));
+                        }
+                    };
+                    DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
+
+                    // The walredo module expects the records to be descending in terms of Lsn.
+                    // And we submit the IOs in that order, so, there shuold be no need to sort here.
+                    debug_assert!(
+                        converted
+                            .records
+                            .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
+                        "{converted:?}"
+                    );
+                    {
+                        let mut guard = rc_clone.lock().unwrap();
+                        guard.set_debug_state(&converted);
+                    }
+                    (
+                        key,
+                        walredo_self
+                            .reconstruct_value(
+                                key,
+                                req_lsn_for_key,
+                                converted,
+                                RedoAttemptType::ReadPage,
+                            )
+                            .await,
+                    )
+                }
+            });
+        }
+
+        let results = futs
+            .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
+            .await;
+
+        Ok(results)
+    }
+
    /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
    pub(crate) fn get_last_record_lsn(&self) -> Lsn {
        self.last_record_lsn.load().last
@@ -1900,6 +2036,8 @@ impl Timeline {
    // an ephemeral layer open forever when idle.  It also freezes layers if the global limit on
    // ephemeral layer bytes has been breached.
    pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
+        debug_assert_current_span_has_tenant_and_timeline_id();
+
        let Ok(mut write_guard) = self.write_lock.try_lock() else {
            // If the write lock is held, there is an active wal receiver: rolling open layers
            // is their responsibility while they hold this lock.
@@ -2854,7 +2992,7 @@ impl Timeline {
    }

    // HADRON
-    fn get_image_creation_timeout(&self) -> Option<Duration> {
+    fn get_image_layer_force_creation_period(&self) -> Option<Duration> {
        let tenant_conf = self.tenant_conf.load();
        tenant_conf
            .tenant_conf
@@ -3134,9 +3272,6 @@ impl Timeline {
                repartition_threshold: 0,
                last_image_layer_creation_check_at: AtomicLsn::new(0),
                last_image_layer_creation_check_instant: Mutex::new(None),
-                // HADRON
-                force_image_creation_lsn: AtomicLsn::new(0),
-                force_image_creation_lsn_computed_at: std::sync::Mutex::new(None),
                last_received_wal: Mutex::new(None),
                rel_size_latest_cache: RwLock::new(HashMap::new()),
                rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
@@ -5381,13 +5516,16 @@ impl Timeline {
        }

        // HADRON
+        // for child timelines, we consider all pages up to ancestor_LSN are redone successfully by the parent timeline
+        min_image_lsn = min_image_lsn.max(self.get_ancestor_lsn());
        if min_image_lsn < force_image_creation_lsn.unwrap_or(Lsn(0)) && max_deltas > 0 {
            info!(
-                "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}",
+                "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}, num deltas: {}",
                partition.ranges[0].start,
                partition.ranges[0].end,
                min_image_lsn,
-                force_image_creation_lsn.unwrap()
+                force_image_creation_lsn.unwrap(),
+                max_deltas
            );
            return true;
        }
@@ -5611,10 +5749,11 @@ impl Timeline {
    /// Predicate function which indicates whether we should check if new image layers
    /// are required. Since checking if new image layers are required is expensive in
    /// terms of CPU, we only do it in the following cases:
-    /// 1. If the timeline has ingested sufficient WAL to justify the cost
+    /// 1. If the timeline has ingested sufficient WAL to justify the cost or ...
    /// 2. If enough time has passed since the last check:
    ///     1. For large tenants, we wish to perform the check more often since they
-    ///        suffer from the lack of image layers
+    ///        suffer from the lack of image layers. Note that we assume sharded tenants
+    ///        to be large since non-zero shards do not track the logical size.
    ///     2. For small tenants (that can mostly fit in RAM), we use a much longer interval
    fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
        let large_timeline_threshold = self.conf.image_layer_generation_large_timeline_threshold;
@@ -5628,30 +5767,39 @@ impl Timeline {

        let distance_based_decision = distance.0 >= min_distance;

-        let mut time_based_decision = false;
        let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
-        if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
-            let check_required_after =
-                if Some(Into::<u64>::into(&logical_size)) >= large_timeline_threshold {
-                    self.get_checkpoint_timeout()
-                } else {
-                    Duration::from_secs(3600 * 48)
-                };
-
-            time_based_decision = match *last_check_instant {
-                Some(last_check) => {
-                    let elapsed = last_check.elapsed();
-                    elapsed >= check_required_after
+        let check_required_after = (|| {
+            if self.shard_identity.is_unsharded() {
+                if let CurrentLogicalSize::Exact(logical_size) =
+                    self.current_logical_size.current_size()
+                {
+                    if Some(Into::<u64>::into(&logical_size)) < large_timeline_threshold {
+                        return Duration::from_secs(3600 * 48);
+                    }
                }
-                None => true,
-            };
-        }
+            }
+
+            self.get_checkpoint_timeout()
+        })();
+
+        let time_based_decision = match *last_check_instant {
+            Some(last_check) => {
+                let elapsed = last_check.elapsed();
+                elapsed >= check_required_after
+            }
+            None => true,
+        };

        // Do the expensive delta layer counting only if this timeline has ingested sufficient
        // WAL since the last check or a checkpoint timeout interval has elapsed since the last
        // check.
        let decision = distance_based_decision || time_based_decision;
-
+        tracing::info!(
+            "Decided to check image layers: {}. Distance-based decision: {}, time-based decision: {}",
+            decision,
+            distance_based_decision,
+            time_based_decision
+        );
        if decision {
            self.last_image_layer_creation_check_at.store(lsn);
            *last_check_instant = Some(Instant::now());
@@ -7153,6 +7301,19 @@ impl Timeline {
            .unwrap()
            .clone()
    }
+
+    /* BEGIN_HADRON */
+    pub(crate) async fn compute_image_consistent_lsn(&self) -> anyhow::Result<Lsn> {
+        let guard = self
+            .layers
+            .read(LayerManagerLockHolder::ComputeImageConsistentLsn)
+            .await;
+        let layer_map = guard.layer_map()?;
+        let disk_consistent_lsn = self.get_disk_consistent_lsn();
+
+        Ok(layer_map.compute_image_consistent_lsn(disk_consistent_lsn))
+    }
+    /* END_HADRON */
 }

 impl Timeline {
--- a/pageserver/src/tenant/timeline/compaction.rs
+++ b/pageserver/src/tenant/timeline/compaction.rs
@@ -8,7 +8,7 @@ use std::cmp::min;
 use std::collections::{BinaryHeap, HashMap, HashSet, VecDeque};
 use std::ops::{Deref, Range};
 use std::sync::Arc;
-use std::time::{Duration, Instant, SystemTime};
+use std::time::{Duration, Instant};

 use super::layer_manager::LayerManagerLockHolder;
 use super::{
@@ -34,7 +34,6 @@ use pageserver_api::models::{CompactInfoResponse, CompactKeyRange};
 use pageserver_api::shard::{ShardCount, ShardIdentity, TenantShardId};
 use pageserver_compaction::helpers::{fully_contains, overlaps_with};
 use pageserver_compaction::interface::*;
-use postgres_ffi::to_pg_timestamp;
 use serde::Serialize;
 use tokio::sync::{OwnedSemaphorePermit, Semaphore};
 use tokio_util::sync::CancellationToken;
@@ -47,7 +46,6 @@ use wal_decoder::models::value::Value;

 use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
 use crate::page_cache;
-use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::statvfs::Statvfs;
 use crate::tenant::checks::check_valid_layermap;
 use crate::tenant::gc_block::GcBlock;
@@ -1271,10 +1269,7 @@ impl Timeline {
        // Define partitioning schema if needed

        // HADRON
-        let force_image_creation_lsn = self
-            .get_or_compute_force_image_creation_lsn(cancel, ctx)
-            .await
-            .map_err(CompactionError::Other)?;
+        let force_image_creation_lsn = self.get_force_image_creation_lsn();

        // 1. L0 Compact
        let l0_outcome = {
@@ -1484,59 +1479,37 @@ impl Timeline {
    }

    /* BEGIN_HADRON */
-    // Get the force image creation LSN. Compute it if the last computed LSN is too old.
-    async fn get_or_compute_force_image_creation_lsn(
-        self: &Arc<Self>,
-        cancel: &CancellationToken,
-        ctx: &RequestContext,
-    ) -> anyhow::Result<Option<Lsn>> {
-        const FORCE_IMAGE_CREATION_LSN_COMPUTE_INTERVAL: Duration = Duration::from_secs(10 * 60); // 10 minutes
-        let image_layer_force_creation_period = self.get_image_creation_timeout();
-        if image_layer_force_creation_period.is_none() {
-            return Ok(None);
+    // Get the force image creation LSN based on gc_cutoff_lsn.
+    // Note that this is an estimation and the workload rate may suddenly change. When that happens,
+    // the force image creation may be too early or too late, but eventually it should be able to catch up.
+    pub(crate) fn get_force_image_creation_lsn(self: &Arc<Self>) -> Option<Lsn> {
+        let image_creation_period = self.get_image_layer_force_creation_period()?;
+        let current_lsn = self.get_last_record_lsn();
+        let pitr_lsn = self.gc_info.read().unwrap().cutoffs.time?;
+        let pitr_interval = self.get_pitr_interval();
+        if pitr_lsn == Lsn::INVALID || pitr_interval.is_zero() {
+            tracing::warn!(
+                "pitr LSN/interval not found, skipping force image creation LSN calculation"
+            );
+            return None;
        }

-        let image_layer_force_creation_period = image_layer_force_creation_period.unwrap();
-        let force_image_creation_lsn_computed_at =
-            *self.force_image_creation_lsn_computed_at.lock().unwrap();
-        if force_image_creation_lsn_computed_at.is_none()
-            || force_image_creation_lsn_computed_at.unwrap().elapsed()
-                > FORCE_IMAGE_CREATION_LSN_COMPUTE_INTERVAL
-        {
-            let now: SystemTime = SystemTime::now();
-            let timestamp = now
-                .checked_sub(image_layer_force_creation_period)
-                .ok_or_else(|| {
-                    anyhow::anyhow!(
-                        "image creation timeout is too large: {image_layer_force_creation_period:?}"
-                    )
-                })?;
-            let timestamp = to_pg_timestamp(timestamp);
-            let force_image_creation_lsn = match self
-                .find_lsn_for_timestamp(timestamp, cancel, ctx)
-                .await?
-            {
-                LsnForTimestamp::Present(lsn) | LsnForTimestamp::Future(lsn) => lsn,
-                _ => {
-                    let gc_lsn = *self.get_applied_gc_cutoff_lsn();
-                    tracing::info!(
-                        "no LSN found for timestamp {timestamp:?}, using latest GC cutoff LSN {}",
-                        gc_lsn
-                    );
-                    gc_lsn
-                }
-            };
-            self.force_image_creation_lsn
-                .store(force_image_creation_lsn);
-            *self.force_image_creation_lsn_computed_at.lock().unwrap() = Some(Instant::now());
-            tracing::info!(
-                "computed force image creation LSN: {}",
-                force_image_creation_lsn
-            );
-            Ok(Some(force_image_creation_lsn))
-        } else {
-            Ok(Some(self.force_image_creation_lsn.load()))
-        }
+        let delta_lsn = current_lsn.checked_sub(pitr_lsn).unwrap().0
+            * image_creation_period.as_secs()
+            / pitr_interval.as_secs();
+        let force_image_creation_lsn = current_lsn.checked_sub(delta_lsn).unwrap_or(Lsn(0));
+
+        tracing::info!(
+            "Tenant shard {} computed force_image_creation_lsn: {}. Current lsn: {}, image_layer_force_creation_period: {:?}, GC cutoff: {}, PITR interval: {:?}",
+            self.tenant_shard_id,
+            force_image_creation_lsn,
+            current_lsn,
+            image_creation_period,
+            pitr_lsn,
+            pitr_interval
+        );
+
+        Some(force_image_creation_lsn)
    }
    /* END_HADRON */

--- a/pageserver/src/tenant/timeline/handle.rs
+++ b/pageserver/src/tenant/timeline/handle.rs
@@ -359,14 +359,14 @@ impl<T: Types> Cache<T> {
                Err(e) => {
                    // Retry on tenant manager error to handle tenant split more gracefully
                    if attempt < GET_MAX_RETRIES {
-                        tracing::warn!(
-                            "Fail to resolve tenant shard in attempt {}: {:?}. Retrying...",
-                            attempt,
-                            e
-                        );
                        tokio::time::sleep(RETRY_BACKOFF).await;
                        continue;
                    } else {
+                        tracing::warn!(
+                            "Failed to resolve tenant shard after {} attempts: {:?}",
+                            GET_MAX_RETRIES,
+                            e
+                        );
                        return Err(e);
                    }
                }
--- a/pageserver/src/tenant/timeline/layer_manager.rs
+++ b/pageserver/src/tenant/timeline/layer_manager.rs
@@ -47,6 +47,7 @@ pub(crate) enum LayerManagerLockHolder {
    ImportPgData,
    DetachAncestor,
    Eviction,
+    ComputeImageConsistentLsn,
    #[cfg(test)]
    Testing,
 }
--- a/pageserver/src/utilization.rs
+++ b/pageserver/src/utilization.rs
@@ -1,6 +1,6 @@
 //! An utilization metric which is used to decide on which pageserver to put next tenant.
 //!
-//! The metric is exposed via `GET /v1/utilization`. Refer and maintain it's openapi spec as the
+//! The metric is exposed via `GET /v1/utilization`. Refer and maintain its openapi spec as the
 //! truth.

 use std::path::Path;
--- a/pageserver/src/walingest.rs
+++ b/pageserver/src/walingest.rs
@@ -32,9 +32,10 @@ use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
 use pageserver_api::shard::ShardIdentity;
 use postgres_ffi::walrecord::*;
 use postgres_ffi::{
-    PgMajorVersion, TimestampTz, TransactionId, dispatch_pgversion, enum_pgversion,
-    enum_pgversion_dispatch, fsm_logical_to_physical, pg_constants,
+    PgMajorVersion, TransactionId, dispatch_pgversion, enum_pgversion, enum_pgversion_dispatch,
+    fsm_logical_to_physical, pg_constants,
 };
+use postgres_ffi_types::TimestampTz;
 use postgres_ffi_types::forknum::{FSM_FORKNUM, INIT_FORKNUM, MAIN_FORKNUM, VISIBILITYMAP_FORKNUM};
 use tracing::*;
 use utils::bin_ser::{DeserializeError, SerializeError};
@@ -1069,7 +1070,7 @@ impl WalIngest {
        // NB: In PostgreSQL, the next-multi-xid stored in the control file is allowed to
        // go to 0, and it's fixed up by skipping to FirstMultiXactId in functions that
        // read it, like GetNewMultiXactId(). This is different from how nextXid is
-        // incremented! nextXid skips over < FirstNormalTransactionId when the the value
+        // incremented! nextXid skips over < FirstNormalTransactionId when the value
        // is stored, so it's never 0 in a checkpoint.
        //
        // I don't know why it's done that way, it seems less error-prone to skip over 0
--- a/pageserver/src/walredo.rs
+++ b/pageserver/src/walredo.rs
@@ -147,6 +147,16 @@ pub enum RedoAttemptType {
    GcCompaction,
 }

+impl std::fmt::Display for RedoAttemptType {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            RedoAttemptType::ReadPage => write!(f, "read page"),
+            RedoAttemptType::LegacyCompaction => write!(f, "legacy compaction"),
+            RedoAttemptType::GcCompaction => write!(f, "gc compaction"),
+        }
+    }
+}
+
 ///
 /// Public interface of WAL redo manager
 ///
@@ -199,6 +209,7 @@ impl PostgresRedoManager {
                        self.conf.wal_redo_timeout,
                        pg_version,
                        max_retry_attempts,
+                        redo_attempt_type,
                    )
                    .await
                };
@@ -221,6 +232,7 @@ impl PostgresRedoManager {
                self.conf.wal_redo_timeout,
                pg_version,
                max_retry_attempts,
+                redo_attempt_type,
            )
            .await
        }
@@ -445,6 +457,7 @@ impl PostgresRedoManager {
        wal_redo_timeout: Duration,
        pg_version: PgMajorVersion,
        max_retry_attempts: u32,
+        redo_attempt_type: RedoAttemptType,
    ) -> Result<Bytes, Error> {
        *(self.last_redo_at.lock().unwrap()) = Some(Instant::now());

@@ -485,17 +498,28 @@ impl PostgresRedoManager {
                );

                if let Err(e) = result.as_ref() {
-                    error!(
-                        "error applying {} WAL records {}..{} ({} bytes) to key {key}, from base image with LSN {} to reconstruct page image at LSN {} n_attempts={}: {:?}",
-                        records.len(),
-                        records.first().map(|p| p.0).unwrap_or(Lsn(0)),
-                        records.last().map(|p| p.0).unwrap_or(Lsn(0)),
-                        nbytes,
-                        base_img_lsn,
-                        lsn,
-                        n_attempts,
-                        e,
-                    );
+                    macro_rules! message {
+                        ($level:tt) => {
+                            $level!(
+                                "error applying {} WAL records {}..{} ({} bytes) to key {} during {}, from base image with LSN {} to reconstruct page image at LSN {} n_attempts={}: {:?}",
+                                records.len(),
+                                records.first().map(|p| p.0).unwrap_or(Lsn(0)),
+                                records.last().map(|p| p.0).unwrap_or(Lsn(0)),
+                                nbytes,
+                                key,
+                                redo_attempt_type,
+                                base_img_lsn,
+                                lsn,
+                                n_attempts,
+                                e,
+                            )
+                        }
+                    }
+                    match redo_attempt_type {
+                        RedoAttemptType::ReadPage => message!(error),
+                        RedoAttemptType::LegacyCompaction => message!(error),
+                        RedoAttemptType::GcCompaction => message!(warn),
+                    }
                }

                result.map_err(Error::Other)
--- a/pgxn/neon/Makefile
+++ b/pgxn/neon/Makefile
@@ -5,7 +5,6 @@ MODULE_big = neon
 OBJS = \
 	$(WIN32RES) \
 	communicator.o \
-	communicator_new.o \
 	extension_server.o \
 	file_cache.o \
 	hll.o \
@@ -30,11 +29,6 @@ PG_CPPFLAGS = -I$(libpq_srcdir)
 SHLIB_LINK_INTERNAL = $(libpq)
 SHLIB_LINK = -lcurl

-UNAME_S := $(shell uname -s)
-ifeq ($(UNAME_S), Darwin)
-    SHLIB_LINK += -framework Security -framework CoreFoundation -framework SystemConfiguration
-endif
-
 EXTENSION = neon
 DATA = \
 	neon--1.0.sql \
@@ -63,7 +57,7 @@ WALPROP_OBJS = \

 # libcommunicator.a is built by cargo from the Rust sources under communicator/
 # subdirectory. `cargo build` also generates communicator_bindings.h.
-communicator_new.o: communicator/communicator_bindings.h
+neon.o: communicator/communicator_bindings.h

 $(NEON_CARGO_ARTIFACT_TARGET_DIR)/libcommunicator.a communicator/communicator_bindings.h &:
 	(cd $(srcdir)/communicator && cargo build $(CARGO_BUILD_FLAGS) $(CARGO_PROFILE))
--- a/pgxn/neon/communicator.c
+++ b/pgxn/neon/communicator.c
@@ -421,7 +421,7 @@ check_getpage_response(PrefetchRequest* slot, NeonResponse* resp)
 {
 	if (resp->tag != T_NeonGetPageResponse && resp->tag != T_NeonErrorResponse)
 	{
-		neon_shard_log(slot->shard_no, PANIC, "Unexpected prefetch response %d, ring_receive=%ld, ring_flush=%ld, ring_unused=%ld",
+		neon_shard_log(slot->shard_no, PANIC, "Unexpected prefetch response %d, ring_receive=" UINT64_FORMAT ", ring_flush=" UINT64_FORMAT ", ring_unused=" UINT64_FORMAT "",
 					   resp->tag, MyPState->ring_receive, MyPState->ring_flush, MyPState->ring_unused);
 	}
 	if (neon_protocol_version >= 3)
@@ -438,7 +438,7 @@ check_getpage_response(PrefetchRequest* slot, NeonResponse* resp)
 				getpage_resp->req.blkno != slot->buftag.blockNum)
 			{
 				NEON_PANIC_CONNECTION_STATE(slot->shard_no, PANIC,
-											"Receive unexpected getpage response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u, block=%u} to get page request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u, block=%u}",
+											"Receive unexpected getpage response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u, block=%u} to get page request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u, block=%u}",
 											resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), RelFileInfoFmt(getpage_resp->req.rinfo), getpage_resp->req.forknum, getpage_resp->req.blkno,
 											slot->reqid, LSN_FORMAT_ARGS(slot->request_lsns.request_lsn), LSN_FORMAT_ARGS(slot->request_lsns.not_modified_since), RelFileInfoFmt(rinfo), slot->buftag.forkNum, slot->buftag.blockNum);
 			}
@@ -447,7 +447,7 @@ check_getpage_response(PrefetchRequest* slot, NeonResponse* resp)
 				 resp->lsn != slot->request_lsns.request_lsn ||
 				 resp->not_modified_since != slot->request_lsns.not_modified_since)
 		{
-			elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match exists request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+			elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match exists request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 				 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 				 slot->reqid, LSN_FORMAT_ARGS(slot->request_lsns.request_lsn), LSN_FORMAT_ARGS(slot->request_lsns.not_modified_since));
 		}
@@ -496,9 +496,9 @@ communicator_prefetch_pump_state(void)
 			slot->my_ring_index != MyPState->ring_receive)
 		{
 			neon_shard_log(slot->shard_no, PANIC,
-						   "Incorrect prefetch slot state after receive: status=%d response=%p my=%lu receive=%lu",
+						   "Incorrect prefetch slot state after receive: status=%d response=%p my=" UINT64_FORMAT " receive=" UINT64_FORMAT "",
 						   slot->status, slot->response,
-						   (long) slot->my_ring_index, (long) MyPState->ring_receive);
+						   slot->my_ring_index, MyPState->ring_receive);
 		}
 		/* update prefetch state */
 		MyPState->n_responses_buffered += 1;
@@ -789,9 +789,9 @@ prefetch_read(PrefetchRequest *slot)
 		slot->my_ring_index != MyPState->ring_receive)
 	{
 		neon_shard_log(slot->shard_no, PANIC,
-					   "Incorrect prefetch read: status=%d response=%p my=%lu receive=%lu",
+					   "Incorrect prefetch read: status=%d response=%p my=" UINT64_FORMAT " receive=" UINT64_FORMAT "",
 					   slot->status, slot->response,
-					   (long)slot->my_ring_index, (long)MyPState->ring_receive);
+					   slot->my_ring_index, MyPState->ring_receive);
 	}

 	/*
@@ -816,9 +816,9 @@ prefetch_read(PrefetchRequest *slot)
 			slot->my_ring_index != MyPState->ring_receive)
 		{
 			neon_shard_log(shard_no, PANIC,
-						   "Incorrect prefetch slot state after receive: status=%d response=%p my=%lu receive=%lu",
+						   "Incorrect prefetch slot state after receive: status=%d response=%p my=" UINT64_FORMAT " receive=" UINT64_FORMAT "",
 						   slot->status, slot->response,
-						   (long) slot->my_ring_index, (long) MyPState->ring_receive);
+						   slot->my_ring_index, MyPState->ring_receive);
 		}

 		/* update prefetch state */
@@ -852,8 +852,8 @@ prefetch_read(PrefetchRequest *slot)
 		 * and the prefetch queue was flushed during the receive call
 		 */
 		neon_shard_log(shard_no, LOG,
-					   "No response from reading prefetch entry %lu: %u/%u/%u.%u block %u. This can be caused by a concurrent disconnect",
-					   (long) my_ring_index,
+					   "No response from reading prefetch entry " UINT64_FORMAT ": %u/%u/%u.%u block %u. This can be caused by a concurrent disconnect",
+					   my_ring_index,
 					   RelFileInfoFmt(BufTagGetNRelFileInfo(buftag)),
 					   buftag.forkNum, buftag.blockNum);
 		return false;
@@ -1844,7 +1844,7 @@ nm_to_string(NeonMessage *msg)
 				NeonDbSizeResponse *msg_resp = (NeonDbSizeResponse *) msg;

 				appendStringInfoString(&s, "{\"type\": \"NeonDbSizeResponse\"");
-				appendStringInfo(&s, ", \"db_size\": %ld}",
+				appendStringInfo(&s, ", \"db_size\": " INT64_FORMAT "}",
 								 msg_resp->db_size);
 				appendStringInfoChar(&s, '}');

@@ -2045,7 +2045,7 @@ communicator_exists(NRelFileInfo rinfo, ForkNumber forkNum, neon_request_lsns *r
 						exists_resp->req.forknum != request.forknum)
 					{
 						NEON_PANIC_CONNECTION_STATE(0, PANIC,
-													"Unexpect response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u} to exits request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u}",
+													"Unexpect response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u} to exits request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u}",
 													resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), RelFileInfoFmt(exists_resp->req.rinfo), exists_resp->req.forknum,
 													request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since), RelFileInfoFmt(request.rinfo), request.forknum);
 					}
@@ -2058,14 +2058,14 @@ communicator_exists(NRelFileInfo rinfo, ForkNumber forkNum, neon_request_lsns *r
 				{
 					if (!equal_requests(resp, &request.hdr))
 					{
-						elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match exists request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+						elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match exists request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 							 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 							 request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since));
 					}
 				}
 				ereport(ERROR,
 						(errcode(ERRCODE_IO_ERROR),
-						 errmsg(NEON_TAG "[reqid %lx] could not read relation existence of rel %u/%u/%u.%u from page server at lsn %X/%08X",
+						 errmsg(NEON_TAG "[reqid " UINT64_HEX_FORMAT "] could not read relation existence of rel %u/%u/%u.%u from page server at lsn %X/%08X",
 								resp->reqid,
 								RelFileInfoFmt(rinfo),
 								forkNum,
@@ -2241,7 +2241,7 @@ Retry:
 			case T_NeonErrorResponse:
 				ereport(ERROR,
 						(errcode(ERRCODE_IO_ERROR),
-						 errmsg(NEON_TAG "[shard %d, reqid %lx] could not read block %u in rel %u/%u/%u.%u from page server at lsn %X/%08X",
+						 errmsg(NEON_TAG "[shard %d, reqid " UINT64_HEX_FORMAT "] could not read block %u in rel %u/%u/%u.%u from page server at lsn %X/%08X",
 								slot->shard_no, resp->reqid, blockno, RelFileInfoFmt(rinfo),
 								forkNum, LSN_FORMAT_ARGS(reqlsns->effective_request_lsn)),
 						 errdetail("page server returned error: %s",
@@ -2294,7 +2294,7 @@ communicator_nblocks(NRelFileInfo rinfo, ForkNumber forknum, neon_request_lsns *
 						relsize_resp->req.forknum != forknum)
 					{
 						NEON_PANIC_CONNECTION_STATE(0, PANIC,
-													"Unexpect response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u} to get relsize request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u}",
+													"Unexpect response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u} to get relsize request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u}",
 													resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), RelFileInfoFmt(relsize_resp->req.rinfo), relsize_resp->req.forknum,
 													request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since), RelFileInfoFmt(request.rinfo), forknum);
 					}
@@ -2307,14 +2307,14 @@ communicator_nblocks(NRelFileInfo rinfo, ForkNumber forknum, neon_request_lsns *
 				{
 					if (!equal_requests(resp, &request.hdr))
 					{
-						elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match get relsize request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+						elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match get relsize request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 							 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 							 request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since));
 					}
 				}
 				ereport(ERROR,
 						(errcode(ERRCODE_IO_ERROR),
-						 errmsg(NEON_TAG "[reqid %lx] could not read relation size of rel %u/%u/%u.%u from page server at lsn %X/%08X",
+						 errmsg(NEON_TAG "[reqid " UINT64_HEX_FORMAT "] could not read relation size of rel %u/%u/%u.%u from page server at lsn %X/%08X",
 								resp->reqid,
 								RelFileInfoFmt(rinfo),
 								forknum,
@@ -2364,7 +2364,7 @@ communicator_dbsize(Oid dbNode, neon_request_lsns *request_lsns)
 						dbsize_resp->req.dbNode != dbNode)
 					{
 						NEON_PANIC_CONNECTION_STATE(0, PANIC,
-													"Unexpect response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, dbNode=%u} to get DB size request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, dbNode=%u}",
+													"Unexpect response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, dbNode=%u} to get DB size request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, dbNode=%u}",
 													resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), dbsize_resp->req.dbNode,
 													request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since), dbNode);
 					}
@@ -2377,14 +2377,14 @@ communicator_dbsize(Oid dbNode, neon_request_lsns *request_lsns)
 				{
 					if (!equal_requests(resp, &request.hdr))
 					{
-						elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match get DB size request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+						elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match get DB size request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 							 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 							 request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since));
 					}
 				}
 				ereport(ERROR,
 						(errcode(ERRCODE_IO_ERROR),
-						 errmsg(NEON_TAG "[reqid %lx] could not read db size of db %u from page server at lsn %X/%08X",
+						 errmsg(NEON_TAG "[reqid " UINT64_HEX_FORMAT "] could not read db size of db %u from page server at lsn %X/%08X",
 								resp->reqid,
 								dbNode, LSN_FORMAT_ARGS(request_lsns->effective_request_lsn)),
 						 errdetail("page server returned error: %s",
@@ -2455,7 +2455,7 @@ communicator_read_slru_segment(SlruKind kind, int64 segno, neon_request_lsns *re
 					slru_resp->req.segno != segno)
 				{
 					NEON_PANIC_CONNECTION_STATE(0, PANIC,
-												"Unexpect response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, kind=%u, segno=%u} to get SLRU segment request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, kind=%u, segno=%lluu}",
+												"Unexpect response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, kind=%u, segno=%u} to get SLRU segment request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, kind=%u, segno=%lluu}",
 												resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), slru_resp->req.kind, slru_resp->req.segno,
 												request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since), kind, (unsigned long long) segno);
 				}
@@ -2469,14 +2469,14 @@ communicator_read_slru_segment(SlruKind kind, int64 segno, neon_request_lsns *re
 			{
 				if (!equal_requests(resp, &request.hdr))
 				{
-					elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match get SLRU segment request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+					elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match get SLRU segment request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 						 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 						 request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since));
 				}
 			}
 			ereport(ERROR,
 					(errcode(ERRCODE_IO_ERROR),
-					 errmsg(NEON_TAG "[reqid %lx] could not read SLRU %d segment %llu at lsn %X/%08X",
+					 errmsg(NEON_TAG "[reqid " UINT64_HEX_FORMAT "] could not read SLRU %d segment %llu at lsn %X/%08X",
 							resp->reqid,
 							kind,
 							(unsigned long long) segno,
--- a/pgxn/neon/communicator/Cargo.lock
+++ b/pgxn/neon/communicator/Cargo.lock
@@ -1,372 +0,0 @@
-# This file is automatically @generated by Cargo.
-# It is not intended for manual editing.
-version = 4
-
-[[package]]
-name = "addr2line"
-version = "0.24.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dfbe277e56a376000877090da837660b4427aad530e3028d44e0bffe4f89a1c1"
-dependencies = [
- "gimli",
-]
-
-[[package]]
-name = "adler2"
-version = "2.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "512761e0bb2578dd7380c6baaa0f4ce03e84f95e960231d1dec8bf4d7d6e2627"
-
-[[package]]
-name = "backtrace"
-version = "0.3.74"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8d82cb332cdfaed17ae235a638438ac4d4839913cc2af585c3c6746e8f8bee1a"
-dependencies = [
- "addr2line",
- "cfg-if",
- "libc",
- "miniz_oxide",
- "object",
- "rustc-demangle",
- "windows-targets",
-]
-
-[[package]]
-name = "base64"
-version = "0.22.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"
-
-[[package]]
-name = "bytes"
-version = "1.10.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d71b6127be86fdcfddb610f7182ac57211d4b18a3e9c82eb2d17662f2227ad6a"
-
-[[package]]
-name = "cfg-if"
-version = "1.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "baf1de4339761588bc0619e3cbc0120ee582ebb74b53b4efbf79117bd2da40fd"
-
-[[package]]
-name = "communicator"
-version = "0.1.0"
-dependencies = [
- "tonic",
-]
-
-[[package]]
-name = "fnv"
-version = "1.0.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1"
-
-[[package]]
-name = "futures-core"
-version = "0.3.31"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "05f29059c0c2090612e8d742178b0580d2dc940c837851ad723096f87af6663e"
-
-[[package]]
-name = "gimli"
-version = "0.31.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "07e28edb80900c19c28f1072f2e8aeca7fa06b23cd4169cefe1af5aa3260783f"
-
-[[package]]
-name = "http"
-version = "1.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f4a85d31aea989eead29a3aaf9e1115a180df8282431156e533de47660892565"
-dependencies = [
- "bytes",
- "fnv",
- "itoa",
-]
-
-[[package]]
-name = "http-body"
-version = "1.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1efedce1fb8e6913f23e0c92de8e62cd5b772a67e7b3946df930a62566c93184"
-dependencies = [
- "bytes",
- "http",
-]
-
-[[package]]
-name = "http-body-util"
-version = "0.1.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b021d93e26becf5dc7e1b75b1bed1fd93124b374ceb73f43d4d4eafec896a64a"
-dependencies = [
- "bytes",
- "futures-core",
- "http",
- "http-body",
- "pin-project-lite",
-]
-
-[[package]]
-name = "itoa"
-version = "1.0.15"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4a5f13b858c8d314ee3e8f639011f7ccefe71f97f96e50151fb991f267928e2c"
-
-[[package]]
-name = "libc"
-version = "0.2.171"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c19937216e9d3aa9956d9bb8dfc0b0c8beb6058fc4f7a4dc4d850edf86a237d6"
-
-[[package]]
-name = "memchr"
-version = "2.7.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "78ca9ab1a0babb1e7d5695e3530886289c18cf2f87ec19a575a0abdce112e3a3"
-
-[[package]]
-name = "miniz_oxide"
-version = "0.8.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ff70ce3e48ae43fa075863cef62e8b43b71a4f2382229920e0df362592919430"
-dependencies = [
- "adler2",
-]
-
-[[package]]
-name = "object"
-version = "0.36.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "62948e14d923ea95ea2c7c86c71013138b66525b86bdc08d2dcc262bdb497b87"
-dependencies = [
- "memchr",
-]
-
-[[package]]
-name = "once_cell"
-version = "1.21.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "42f5e15c9953c5e4ccceeb2e7382a716482c34515315f7b03532b8b4e8393d2d"
-
-[[package]]
-name = "percent-encoding"
-version = "2.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e3148f5046208a5d56bcfc03053e3ca6334e51da8dfb19b6cdc8b306fae3283e"
-
-[[package]]
-name = "pin-project"
-version = "1.1.10"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "677f1add503faace112b9f1373e43e9e054bfdd22ff1a63c1bc485eaec6a6a8a"
-dependencies = [
- "pin-project-internal",
-]
-
-[[package]]
-name = "pin-project-internal"
-version = "1.1.10"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6e918e4ff8c4549eb882f14b3a4bc8c8bc93de829416eacf579f1207a8fbf861"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
-]
-
-[[package]]
-name = "pin-project-lite"
-version = "0.2.16"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3b3cff922bd51709b605d9ead9aa71031d81447142d828eb4a6eba76fe619f9b"
-
-[[package]]
-name = "proc-macro2"
-version = "1.0.94"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a31971752e70b8b2686d7e46ec17fb38dad4051d94024c88df49b667caea9c84"
-dependencies = [
- "unicode-ident",
-]
-
-[[package]]
-name = "quote"
-version = "1.0.40"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1885c039570dc00dcb4ff087a89e185fd56bae234ddc7f056a945bf36467248d"
-dependencies = [
- "proc-macro2",
-]
-
-[[package]]
-name = "rustc-demangle"
-version = "0.1.24"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "719b953e2095829ee67db738b3bfa9fa368c94900df327b3f07fe6e794d2fe1f"
-
-[[package]]
-name = "syn"
-version = "2.0.100"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b09a44accad81e1ba1cd74a32461ba89dee89095ba17b32f5d03683b1b1fc2a0"
-dependencies = [
- "proc-macro2",
- "quote",
- "unicode-ident",
-]
-
-[[package]]
-name = "tokio"
-version = "1.44.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e6b88822cbe49de4185e3a4cbf8321dd487cf5fe0c5c65695fef6346371e9c48"
-dependencies = [
- "backtrace",
- "pin-project-lite",
-]
-
-[[package]]
-name = "tokio-stream"
-version = "0.1.17"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eca58d7bba4a75707817a2c44174253f9236b2d5fbd055602e9d5c07c139a047"
-dependencies = [
- "futures-core",
- "pin-project-lite",
- "tokio",
-]
-
-[[package]]
-name = "tonic"
-version = "0.13.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "85839f0b32fd242bb3209262371d07feda6d780d16ee9d2bc88581b89da1549b"
-dependencies = [
- "base64",
- "bytes",
- "http",
- "http-body",
- "http-body-util",
- "percent-encoding",
- "pin-project",
- "tokio-stream",
- "tower-layer",
- "tower-service",
- "tracing",
-]
-
-[[package]]
-name = "tower-layer"
-version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "121c2a6cda46980bb0fcd1647ffaf6cd3fc79a013de288782836f6df9c48780e"
-
-[[package]]
-name = "tower-service"
-version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
-
-[[package]]
-name = "tracing"
-version = "0.1.41"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "784e0ac535deb450455cbfa28a6f0df145ea1bb7ae51b821cf5e7927fdcfbdd0"
-dependencies = [
- "pin-project-lite",
- "tracing-attributes",
- "tracing-core",
-]
-
-[[package]]
-name = "tracing-attributes"
-version = "0.1.28"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "395ae124c09f9e6918a2310af6038fba074bcf474ac352496d5910dd59a2226d"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
-]
-
-[[package]]
-name = "tracing-core"
-version = "0.1.33"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e672c95779cf947c5311f83787af4fa8fffd12fb27e4993211a84bdfd9610f9c"
-dependencies = [
- "once_cell",
-]
-
-[[package]]
-name = "unicode-ident"
-version = "1.0.18"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5a5f39404a5da50712a4c1eecf25e90dd62b613502b7e925fd4e4d19b5c96512"
-
-[[package]]
-name = "windows-targets"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973"
-dependencies = [
- "windows_aarch64_gnullvm",
- "windows_aarch64_msvc",
- "windows_i686_gnu",
- "windows_i686_gnullvm",
- "windows_i686_msvc",
- "windows_x86_64_gnu",
- "windows_x86_64_gnullvm",
- "windows_x86_64_msvc",
-]
-
-[[package]]
-name = "windows_aarch64_gnullvm"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
-
-[[package]]
-name = "windows_aarch64_msvc"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
-
-[[package]]
-name = "windows_i686_gnu"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b"
-
-[[package]]
-name = "windows_i686_gnullvm"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
-
-[[package]]
-name = "windows_i686_msvc"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
-
-[[package]]
-name = "windows_x86_64_gnu"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78"
-
-[[package]]
-name = "windows_x86_64_gnullvm"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d"
-
-[[package]]
-name = "windows_x86_64_msvc"
-version = "0.52.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec"
--- a/pgxn/neon/communicator/Cargo.toml
+++ b/pgxn/neon/communicator/Cargo.toml
@@ -1,42 +1,19 @@
 [package]
 name = "communicator"
+version = "0.1.0"
 license.workspace = true
 edition.workspace = true

+[lib]
+crate-type = ["staticlib"]
+
 [features]
 # 'testing' feature is currently unused in the communicator, but we accept it for convenience of
 # calling build scripts, so that you can pass the same feature to all packages.
 testing = []

-[lib]
-crate-type = ["staticlib"]
-
 [dependencies]
-axum.workspace = true
-bytes.workspace = true
-clashmap.workspace = true
-http.workspace = true
-libc.workspace = true
-nix.workspace = true
-atomic_enum = "0.3.0"
-prometheus.workspace = true
-prost.workspace = true
-tonic = { version = "0.12.0", default-features = false, features=["codegen", "prost", "transport"] }
-tokio = { version = "1.43.1", features = ["macros", "net", "io-util", "rt", "rt-multi-thread"] }
-tokio-pipe = { version = "0.2.12" }
-thiserror.workspace = true
-tracing.workspace = true
-tracing-subscriber.workspace = true
-
-metrics.workspace = true
-uring-common = { workspace = true, features = ["bytes"] }
-
-pageserver_client_grpc.workspace = true
-pageserver_api.workspace = true
-pageserver_page_api.workspace = true
-
 neon-shmem.workspace = true
-utils.workspace = true
 workspace_hack = { version = "0.1", path = "../../../workspace_hack" }

 [build-dependencies]
--- a/pgxn/neon/communicator/README.md
+++ b/pgxn/neon/communicator/README.md
@@ -1,138 +1,8 @@
-# Communicator
-
-This package provides the so-called "compute-pageserver communicator",
-or just "communicator" in short. It runs in a PostgreSQL server, as
-part of the neon extension, and handles the communication with the
-pageservers. On the PostgreSQL side, the glue code in pgxn/neon/ uses
-the communicator to implement the PostgreSQL Storage Manager (SMGR)
-interface.
-
-## Design criteria
-
- Low latency
- Saturate a 10 Gbit / s network interface without becoming a bottleneck
-
-## Source code view
-
-pgxn/neon/communicator_new.c
-	Contains the glue that interact with PostgreSQL code and the Rust
-	communicator code.
-
-pgxn/neon/communicator/src/backend_interface.rs
-	The entry point for calls from each backend.
-
-pgxn/neon/communicator/src/init.rs
-	Initialization at server startup
-
-pgxn/neon/communicator/src/worker_process/
-    Worker process main loop and glue code
+This package will evolve into a "compute-pageserver communicator"
+process and machinery. For now, it's just a dummy that doesn't do
+anything interesting, but it allows us to test the compilation and
+linking of Rust code into the Postgres extensions.

 At compilation time, pgxn/neon/communicator/ produces a static
 library, libcommunicator.a. It is linked to the neon.so extension
 library.
-
-The real networking code, which is independent of PostgreSQL, is in
-the pageserver/client_grpc crate.
-
-## Process view
-
-The communicator runs in a dedicated background worker process, the
-"communicator process". The communicator uses a multi-threaded Tokio
-runtime to execute the IO requests. So the communicator process has
-multiple threads running. That's unusual for Postgres processes and
-care must be taken to make that work.
-
-### Backend <-> worker communication
-
-Each backend has a number of I/O request slots in shared memory. The
-slots are statically allocated for each backend, and must not be
-accessed by other backends. The worker process reads requests from the
-shared memory slots, and writes responses back to the slots.
-
-Here's an example snapshot of the system, when two requests from two
-different backends are in progress:
-
-```
-Backends           Request slots          Communicator process
---------          -------------          --------------------
-
-Backend 1          1: Idle
-                   2: Idle
-                   3: Processing          tokio task handling request 3
-
-Backend 2          4: Completed
-                   5: Processing          tokio task handling request 5
-                   6: Idle
-
-...                ...
-```
-
-To submit an IO request, the backend first picks one of its Idle
-slots, writes the IO request in the slot, and updates it to
-'Submitted' state. That transfers the ownership of the slot to the
-worker process, until the worker process marks the request as
-Completed. The worker process spawns a separate Tokio task for each
-request.
-
-To inform the worker process that a request slot has a pending IO
-request, there's a pipe shared by the worker process and all backend
-processes. The backend writes the index of the request slot to the
-pipe after changing the slot's state to Submitted. This wakes up the
-worker process.
-
-(Note that the pipe is just used for wakeups, but the worker process
-is free to pick up Submitted IO requests even without receiving the
-wakeup. As of this writing, it doesn't do that, but it might be useful
-in the future to reduce latency even further, for example.)
-
-When the worker process has completed processing the request, it
-writes the result back in the request slot. A GetPage request can also
-contain a pointer to buffer in the shared buffer cache. In that case,
-the worker process writes the resulting page contents directly to the
-buffer, and just a result code in the request slot. It then updates
-the 'state' field to Completed, which passes the owner ship back to
-the originating backend. Finally, it signals the process Latch of the
-originating backend, waking it up.
-
-### Differences between PostgreSQL v16, v17 and v18
-
-PostgreSQL v18 introduced the new AIO mechanism. The PostgreSQL AIO
-mechanism uses a very similar mechanism as described in the previous
-section, for the communication between AIO worker processes and
-backends. With our communicator, the AIO worker processes are not
-used, but we use the same PgAioHandle request slots as in upstream.
-For Neon-specific IO requests like GetDbSize, a neon request slot is
-used. But for the actual IO requests, the request slot merely contains
-a pointer to the PgAioHandle slot. The worker process updates the
-status of that, calls the IO callbacks upon completionetc, just like
-the upstream AIO worker processes do.
-
-## Sequence diagram
-
-                      neon
-    PostgreSQL     extension       backend_interface.rs  worker_process.rs    processor    tonic
-       |               .                    .                   .                 .
-	   | smgr_read()   .                    .                   .                 .
-	   +-------------> +                    .                   .                 .
-	   .               |                    .                   .                 .
-	   .               |  rcommunicator_    .                   .                 .
-	   .               | get_page_at_lsn    .                   .                 .
-	   .               +------------------> +                   .                 .
-                                            |                   .                 .
-                                            | write request to  .                 .                 .
-                                            | slot              .                 .
-                                            |                   .                 .
-                                            |                   .                 .
-											| submit_request()  .                 .
-											+-----------------> +                 .
-											|                   |                 .
-											|					| db_size_request .               .
-																+---------------->.
-																                  . TODO
-
-
-
-### Compute <-> pageserver protocol
-
-The protocol between Compute and the pageserver is based on gRPC. See `protos/`.
-
--- a/Show More
+++ b/Show More