WIP

Squashed commit of the following:

commit 4a1b52c12e
Author: Christian Schwarz <christian@neon.tech>
Date:   Mon May 5 12:45:45 2025 +0200

    WIP

commit 257693e4f2
Author: Christian Schwarz <christian@neon.tech>
Date:   Mon May 5 10:59:45 2025 +0200

    WIP

commit 7aa9beaefd
Author: Christian Schwarz <christian@neon.tech>
Date:   Sun May 4 17:06:46 2025 +0200

    make sk compile

commit 35dbbbaf60
Author: Christian Schwarz <christian@neon.tech>
Date:   Sun May 4 16:50:23 2025 +0200

    move discovery request mechanism into that type as well

    Can't move the policy when we send disovery mechanism because that's
    tied to connection_manager loop state.

commit 6380c9674c
Author: Christian Schwarz <christian@neon.tech>
Date:   Sun May 4 16:22:46 2025 +0200

    move subscription code into new client struct

commit 1f53688189
Author: Christian Schwarz <christian@neon.tech>
Date:   Sun May 4 14:40:44 2025 +0200

    Revert "rip out broker binary target & launch of it in cplane & mention of it in docs"

    This reverts commit 8f201b1580.

commit 8f201b1580
Author: Christian Schwarz <christian@neon.tech>
Date:   Sun May 4 14:38:52 2025 +0200

    rip out broker binary target & launch of it in cplane & mention of it in docs

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

@@ -49,10 +49,6 @@ inputs:
     description: 'A JSON object with project settings'
     required: false
     default: '{}'
-  default_endpoint_settings:
-    description: 'A JSON object with the default endpoint settings'
-    required: false
-    default: '{}'
 
 outputs:
   dsn:
@@ -139,21 +135,6 @@ runs:
             -H "Accept: application/json" -H "Content-Type: application/json" -H "Authorization: Bearer ${ADMIN_API_KEY}" \
             -d "{\"scheduling\": \"Essential\"}"
         fi
-        # XXX
-        # This is a workaround for the default endpoint settings, which currently do not allow some settings in the public API.
-        # https://github.com/neondatabase/cloud/issues/27108
-        if [[ -n ${DEFAULT_ENDPOINT_SETTINGS} && ${DEFAULT_ENDPOINT_SETTINGS} != "{}" ]] ; then
-          PROJECT_DATA=$(curl -X GET \
-              "https://${API_HOST}/regions/${REGION_ID}/api/v1/admin/projects/${project_id}" \
-              -H "Accept: application/json" -H "Content-Type: application/json" -H "Authorization: Bearer ${ADMIN_API_KEY}" \
-              -d "{\"scheduling\": \"Essential\"}"
-          )
-          NEW_DEFAULT_ENDPOINT_SETTINGS=$(echo ${PROJECT_DATA} | jq -rc ".project.default_endpoint_settings + ${DEFAULT_ENDPOINT_SETTINGS}")
-          curl -X POST --fail \
-                "https://${API_HOST}/regions/${REGION_ID}/api/v1/admin/projects/${project_id}/default_endpoint_settings" \
-                -H "Accept: application/json" -H "Content-Type: application/json" -H "Authorization: Bearer ${ADMIN_API_KEY}" \
-                --data "${NEW_DEFAULT_ENDPOINT_SETTINGS}"
-        fi
         
 
       env:
@@ -171,4 +152,3 @@ runs:
         PSQL: ${{ inputs.psql_path }}
         LD_LIBRARY_PATH: ${{ inputs.libpq_lib_path }}
         PROJECT_SETTINGS: ${{ inputs.project_settings }}
-        DEFAULT_ENDPOINT_SETTINGS: ${{ inputs.default_endpoint_settings }}

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

@@ -279,18 +279,14 @@ jobs:
           # run all non-pageserver tests
           ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES -E '!package(pageserver)'
 
-          # run pageserver tests with different settings
-          for get_vectored_concurrent_io in sequential sidecar-task; do
-            for io_engine in std-fs tokio-epoll-uring ; do
-                for io_mode in buffered direct direct-rw ; do
-                  NEON_PAGESERVER_UNIT_TEST_GET_VECTORED_CONCURRENT_IO=$get_vectored_concurrent_io \
-                  NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE=$io_engine \
-                  NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IO_MODE=$io_mode \
-                  ${cov_prefix} \
-                  cargo nextest run $CARGO_FLAGS $CARGO_FEATURES  -E 'package(pageserver)'
-              done
-            done
-          done
+          # run pageserver tests
+          # (When developing new pageserver features gated by config fields, we commonly make the rust
+          # unit tests sensitive to an environment variable NEON_PAGESERVER_UNIT_TEST_FEATURENAME.
+          # Then run the nextest invocation below for all relevant combinations. Singling out the
+          # pageserver tests from non-pageserver tests cuts down the time it takes for this CI step.)
+          NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE=tokio-epoll-uring  \
+          ${cov_prefix} \
+          cargo nextest run $CARGO_FLAGS $CARGO_FEATURES  -E 'package(pageserver)'
 
           # Run separate tests for real S3
           export ENABLE_REAL_S3_REMOTE_STORAGE=nonempty
@@ -405,8 +401,6 @@ jobs:
           CHECK_ONDISK_DATA_COMPATIBILITY: nonempty
           BUILD_TAG: ${{ inputs.build-tag }}
           PAGESERVER_VIRTUAL_FILE_IO_ENGINE: tokio-epoll-uring
-          PAGESERVER_GET_VECTORED_CONCURRENT_IO: sidecar-task
-          PAGESERVER_VIRTUAL_FILE_IO_MODE: direct-rw
           USE_LFC: ${{ matrix.lfc_state == 'with-lfc' && 'true' || 'false' }}
 
       # Temporary disable this step until we figure out why it's so flaky

.github/workflows/build_and_test.yml

@@ -323,8 +323,6 @@ jobs:
           PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
           TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}"
           PAGESERVER_VIRTUAL_FILE_IO_ENGINE: tokio-epoll-uring
-          PAGESERVER_GET_VECTORED_CONCURRENT_IO: sidecar-task
-          PAGESERVER_VIRTUAL_FILE_IO_MODE: direct-rw
           SYNC_BETWEEN_TESTS: true
       # XXX: no coverage data handling here, since benchmarks are run on release builds,
       # while coverage is currently collected for the debug ones
@@ -965,7 +963,7 @@ jobs:
           fi
 
       - name: Verify docker-compose example and test extensions
-        timeout-minutes: 20
+        timeout-minutes: 60
         env:
           TAG: >-
             ${{

.github/workflows/cloud-extensions.yml

@@ -35,7 +35,7 @@ jobs:
       matrix:
         pg-version: [16, 17]
 
-    runs-on: [ self-hosted, small ]
+    runs-on: us-east-2
     container:
       # We use the neon-test-extensions image here as it contains the source code for the extensions.
       image: ghcr.io/neondatabase/neon-test-extensions-v${{ matrix.pg-version }}:latest
@@ -71,20 +71,7 @@ jobs:
           region_id: ${{ inputs.region_id || 'aws-us-east-2' }}
           postgres_version: ${{ matrix.pg-version }}
           project_settings: ${{ steps.project-settings.outputs.settings }}
-          # We need these settings to get the expected output results.
-          # We cannot use the environment variables e.g. PGTZ due to
-          # https://github.com/neondatabase/neon/issues/1287
-          default_endpoint_settings: >
-            {
-              "pg_settings": {
-                "DateStyle": "Postgres,MDY",
-                "TimeZone": "America/Los_Angeles",
-                "compute_query_id": "off",
-                "neon.allow_unstable_extensions": "on"
-              }
-            }
           api_key: ${{ secrets.NEON_STAGING_API_KEY }}
-          admin_api_key: ${{ secrets.NEON_STAGING_ADMIN_API_KEY }}
 
       - name: Run the regression tests
         run: /run-tests.sh -r /ext-src

.github/workflows/neon_extra_builds.yml

@@ -63,8 +63,10 @@ jobs:
 
       - name: Filter out only v-string for build matrix
         id: postgres_changes
+        env:
+          CHANGES: ${{ steps.files_changed.outputs.changes }}
         run: |
-          v_strings_only_as_json_array=$(echo ${{ steps.files_changed.outputs.chnages }} | jq '.[]|select(test("v\\d+"))' | jq --slurp -c)
+          v_strings_only_as_json_array=$(echo ${CHANGES} | jq '.[]|select(test("v\\d+"))' | jq --slurp -c)
           echo "changes=${v_strings_only_as_json_array}" | tee -a "${GITHUB_OUTPUT}"
 
   check-macos-build:

Cargo.lock

@@ -1112,6 +1112,12 @@ version = "1.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "baf1de4339761588bc0619e3cbc0120ee582ebb74b53b4efbf79117bd2da40fd"
 
+[[package]]
+name = "cfg_aliases"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
+
 [[package]]
 name = "cgroups-rs"
 version = "0.3.3"
@@ -1306,7 +1312,7 @@ dependencies = [
  "itertools 0.10.5",
  "jsonwebtoken",
  "metrics",
- "nix 0.27.1",
+ "nix 0.30.1",
  "notify",
  "num_cpus",
  "once_cell",
@@ -1429,7 +1435,7 @@ dependencies = [
  "humantime-serde",
  "hyper 0.14.30",
  "jsonwebtoken",
- "nix 0.27.1",
+ "nix 0.30.1",
  "once_cell",
  "pageserver_api",
  "pageserver_client",
@@ -3512,9 +3518,9 @@ dependencies = [
 
 [[package]]
 name = "libc"
-version = "0.2.169"
+version = "0.2.172"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b5aba8db14291edd000dfcc4d620c7ebfb122c613afb886ca8803fa4e128a20a"
+checksum = "d750af042f7ef4f724306de029d18836c26c1765a54a6a3f094cbd23a7267ffa"
 
 [[package]]
 name = "libloading"
@@ -3788,6 +3794,16 @@ version = "0.8.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e5ce46fe64a9d73be07dcbe690a38ce1b293be448fd8ce1e6c1b8062c9f72c6a"
 
+[[package]]
+name = "neon-shmem"
+version = "0.1.0"
+dependencies = [
+ "nix 0.30.1",
+ "tempfile",
+ "thiserror 1.0.69",
+ "workspace_hack",
+]
+
 [[package]]
 name = "never-say-never"
 version = "6.6.666"
@@ -3821,12 +3837,13 @@ dependencies = [
 
 [[package]]
 name = "nix"
-version = "0.27.1"
+version = "0.30.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2eb04e9c688eff1c89d72b407f168cf79bb9e867a9d3323ed6c01519eb9cc053"
+checksum = "74523f3a35e05aba87a1d978330aef40f67b0304ac79c1c00b294c9830543db6"
 dependencies = [
  "bitflags 2.8.0",
  "cfg-if",
+ "cfg_aliases",
  "libc",
  "memoffset 0.9.0",
 ]
@@ -4280,7 +4297,7 @@ dependencies = [
  "jsonwebtoken",
  "md5",
  "metrics",
- "nix 0.27.1",
+ "nix 0.30.1",
  "num-traits",
  "num_cpus",
  "once_cell",
@@ -4331,6 +4348,7 @@ dependencies = [
  "toml_edit",
  "tracing",
  "tracing-utils",
+ "twox-hash",
  "url",
  "utils",
  "uuid",
@@ -4355,7 +4373,7 @@ dependencies = [
  "humantime",
  "humantime-serde",
  "itertools 0.10.5",
- "nix 0.27.1",
+ "nix 0.30.1",
  "once_cell",
  "postgres_backend",
  "postgres_ffi",
@@ -4416,6 +4434,16 @@ dependencies = [
  "workspace_hack",
 ]
 
+[[package]]
+name = "pageserver_page_api"
+version = "0.1.0"
+dependencies = [
+ "prost 0.13.3",
+ "tonic",
+ "tonic-build",
+ "workspace_hack",
+]
+
 [[package]]
 name = "papaya"
 version = "0.2.1"
@@ -4848,6 +4876,19 @@ dependencies = [
  "workspace_hack",
 ]
 
+[[package]]
+name = "posthog_client_lite"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "reqwest",
+ "serde",
+ "serde_json",
+ "sha2",
+ "thiserror 1.0.69",
+ "workspace_hack",
+]
+
 [[package]]
 name = "powerfmt"
 version = "0.2.0"
@@ -6023,8 +6064,10 @@ dependencies = [
  "serde",
  "serde_json",
  "sha2",
+ "sk_ps_discovery",
  "smallvec",
  "storage_broker",
+ "storage_controller_client",
  "strum",
  "strum_macros",
  "thiserror 1.0.69",
@@ -6036,6 +6079,7 @@ dependencies = [
  "tokio-stream",
  "tokio-tar",
  "tokio-util",
+ "tonic",
  "tracing",
  "tracing-subscriber",
  "url",
@@ -6531,6 +6575,76 @@ version = "0.3.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7bd3e3206899af3f8b12af284fafc038cc1dc2b41d1b89dd17297221c5d225de"
 
+[[package]]
+name = "sk_ps_discovery"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "async-stream",
+ "byteorder",
+ "bytes",
+ "camino",
+ "camino-tempfile",
+ "chrono",
+ "clap",
+ "crc32c",
+ "criterion",
+ "desim",
+ "env_logger",
+ "fail",
+ "futures",
+ "hex",
+ "http 1.1.0",
+ "http-utils",
+ "humantime",
+ "hyper 0.14.30",
+ "itertools 0.10.5",
+ "jsonwebtoken",
+ "metrics",
+ "once_cell",
+ "pageserver_api",
+ "parking_lot 0.12.1",
+ "pem",
+ "postgres-protocol",
+ "postgres_backend",
+ "postgres_ffi",
+ "pprof",
+ "pq_proto",
+ "rand 0.8.5",
+ "regex",
+ "remote_storage",
+ "reqwest",
+ "rustls 0.23.18",
+ "safekeeper_api",
+ "safekeeper_client",
+ "scopeguard",
+ "sd-notify",
+ "serde",
+ "serde_json",
+ "sha2",
+ "smallvec",
+ "storage_broker",
+ "strum",
+ "strum_macros",
+ "thiserror 1.0.69",
+ "tikv-jemallocator",
+ "tokio",
+ "tokio-io-timeout",
+ "tokio-postgres",
+ "tokio-rustls 0.26.0",
+ "tokio-stream",
+ "tokio-tar",
+ "tokio-util",
+ "tonic",
+ "tracing",
+ "tracing-subscriber",
+ "url",
+ "utils",
+ "wal_decoder",
+ "walproposer",
+ "workspace_hack",
+]
+
 [[package]]
 name = "slab"
 version = "0.4.8"
@@ -6637,6 +6751,7 @@ dependencies = [
  "rustls 0.23.18",
  "tokio",
  "tokio-rustls 0.26.0",
+ "tokio-util",
  "tonic",
  "tonic-build",
  "tracing",
@@ -6649,6 +6764,7 @@ name = "storage_controller"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "async-stream",
  "bytes",
  "camino",
  "chrono",
@@ -7885,7 +8001,7 @@ dependencies = [
  "humantime",
  "jsonwebtoken",
  "metrics",
- "nix 0.27.1",
+ "nix 0.30.1",
  "once_cell",
  "pem",
  "pin-project-lite",
@@ -8439,8 +8555,10 @@ dependencies = [
  "fail",
  "form_urlencoded",
  "futures-channel",
+ "futures-core",
  "futures-executor",
  "futures-io",
+ "futures-task",
  "futures-util",
  "generic-array",
  "getrandom 0.2.11",
@@ -8459,6 +8577,7 @@ dependencies = [
  "log",
  "memchr",
  "nix 0.26.4",
+ "nix 0.30.1",
  "nom",
  "num",
  "num-bigint",
@@ -8470,6 +8589,7 @@ dependencies = [
  "once_cell",
  "p256 0.13.2",
  "parquet",
+ "percent-encoding",
  "prettyplease",
  "proc-macro2",
  "prost 0.13.3",

Cargo.toml

@@ -9,6 +9,7 @@ members = [
     "pageserver/ctl",
     "pageserver/client",
     "pageserver/pagebench",
+    "pageserver/page_api",
     "proxy",
     "safekeeper",
     "safekeeper/client",
@@ -23,9 +24,11 @@ members = [
     "libs/postgres_ffi",
     "libs/safekeeper_api",
     "libs/desim",
+    "libs/neon-shmem",
     "libs/utils",
     "libs/consumption_metrics",
     "libs/postgres_backend",
+    "libs/posthog_client_lite",
     "libs/pq_proto",
     "libs/tenant_size_model",
     "libs/metrics",
@@ -40,7 +43,7 @@ members = [
     "libs/proxy/postgres-protocol2",
     "libs/proxy/postgres-types2",
     "libs/proxy/tokio-postgres2",
-    "endpoint_storage",
+    "endpoint_storage", "libs/sk_ps_discovery",
 ]
 
 [workspace.package]
@@ -126,7 +129,7 @@ md5 = "0.7.0"
 measured = { version = "0.0.22", features=["lasso"] }
 measured-process = { version = "0.0.22" }
 memoffset = "0.9"
-nix = { version = "0.27", features = ["dir", "fs", "process", "socket", "signal", "poll"] }
+nix = { version = "0.30.1", features = ["dir", "fs", "mman", "process", "socket", "signal", "poll"] }
 # Do not update to >= 7.0.0, at least. The update will have a significant impact
 # on compute startup metrics (start_postgres_ms), >= 25% degradation.
 notify = "6.0.0"
@@ -250,6 +253,7 @@ pageserver = { path = "./pageserver" }
 pageserver_api = { version = "0.1", path = "./libs/pageserver_api/" }
 pageserver_client = { path = "./pageserver/client" }
 pageserver_compaction = { version = "0.1", path = "./pageserver/compaction/" }
+pageserver_page_api = { path = "./pageserver/page_api" }
 postgres_backend = { version = "0.1", path = "./libs/postgres_backend/" }
 postgres_connection = { version = "0.1", path = "./libs/postgres_connection/" }
 postgres_ffi = { version = "0.1", path = "./libs/postgres_ffi/" }
@@ -258,6 +262,7 @@ pq_proto = { version = "0.1", path = "./libs/pq_proto/" }
 remote_storage = { version = "0.1", path = "./libs/remote_storage/" }
 safekeeper_api = { version = "0.1", path = "./libs/safekeeper_api" }
 safekeeper_client = { path = "./safekeeper/client" }
+sk_ps_discovery = { path = "./libs/sk_ps_discovery" }
 desim = { version = "0.1", path = "./libs/desim" }
 storage_broker = { version = "0.1", path = "./storage_broker/" } # Note: main broker code is inside the binary crate, so linking with the library shouldn't be heavy.
 storage_controller_client = { path = "./storage_controller/client" }

build-tools.Dockerfile

@@ -292,7 +292,7 @@ WORKDIR /home/nonroot
 
 # Rust
 # Please keep the version of llvm (installed above) in sync with rust llvm (`rustc --version --verbose | grep LLVM`)
-ENV RUSTC_VERSION=1.86.0
+ENV RUSTC_VERSION=1.87.0
 ENV RUSTUP_HOME="/home/nonroot/.rustup"
 ENV PATH="/home/nonroot/.cargo/bin:${PATH}"
 ARG RUSTFILT_VERSION=0.2.1

compute/compute-node.Dockerfile

@@ -1117,8 +1117,8 @@ RUN wget https://github.com/microsoft/onnxruntime/archive/refs/tags/v1.18.1.tar.
     mkdir onnxruntime-src && cd onnxruntime-src && tar xzf ../onnxruntime.tar.gz --strip-components=1 -C . && \
     echo "#nothing to test here" > neon-test.sh
 
-RUN wget https://github.com/neondatabase-labs/pgrag/archive/refs/tags/v0.1.1.tar.gz -O pgrag.tar.gz &&  \
-    echo "087b2ecd11ba307dc968042ef2e9e43dc04d9ba60e8306e882c407bbe1350a50 pgrag.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/neondatabase-labs/pgrag/archive/refs/tags/v0.1.2.tar.gz -O pgrag.tar.gz &&  \
+    echo "7361654ea24f08cbb9db13c2ee1c0fe008f6114076401bb871619690dafc5225 pgrag.tar.gz" | sha256sum --check && \
     mkdir pgrag-src && cd pgrag-src && tar xzf ../pgrag.tar.gz --strip-components=1 -C .
 
 FROM rust-extensions-build-pgrx14 AS pgrag-build

compute/patches/rum.patch

@@ -7,7 +7,7 @@ index 255e616..1c6edb7 100644
  			 RelationGetRelationName(index));
  
 +#ifdef NEON_SMGR
-+	smgr_start_unlogged_build(index->rd_smgr);
++	smgr_start_unlogged_build(RelationGetSmgr(index));
 +#endif
 +
  	initRumState(&buildstate.rumstate, index);
@@ -18,7 +18,7 @@ index 255e616..1c6edb7 100644
  	rumUpdateStats(index, &buildstate.buildStats, buildstate.rumstate.isBuild);
  
 +#ifdef NEON_SMGR
-+	smgr_finish_unlogged_build_phase_1(index->rd_smgr);
++	smgr_finish_unlogged_build_phase_1(RelationGetSmgr(index));
 +#endif
 +
  	/*
@@ -29,7 +29,7 @@ index 255e616..1c6edb7 100644
  	}
  
 +#ifdef NEON_SMGR
-+	smgr_end_unlogged_build(index->rd_smgr);
++	smgr_end_unlogged_build(RelationGetSmgr(index));
 +#endif
 +
  	/*

control_plane/src/background_process.rs

@@ -14,7 +14,7 @@
 
 use std::ffi::OsStr;
 use std::io::Write;
-use std::os::unix::prelude::AsRawFd;
+use std::os::fd::AsFd;
 use std::os::unix::process::CommandExt;
 use std::path::Path;
 use std::process::Command;
@@ -356,7 +356,7 @@ where
             let file = pid_file::claim_for_current_process(&path).expect("claim pid file");
             // Remove the FD_CLOEXEC flag on the pidfile descriptor so that the pidfile
             // remains locked after exec.
-            nix::fcntl::fcntl(file.as_raw_fd(), FcntlArg::F_SETFD(FdFlag::empty()))
+            nix::fcntl::fcntl(file.as_fd(), FcntlArg::F_SETFD(FdFlag::empty()))
                 .expect("remove FD_CLOEXEC");
             // Don't run drop(file), it would close the file before we actually exec.
             std::mem::forget(file);

control_plane/src/bin/neon_local.rs

@@ -8,7 +8,6 @@
 use std::borrow::Cow;
 use std::collections::{BTreeSet, HashMap};
 use std::fs::File;
-use std::os::fd::AsRawFd;
 use std::path::PathBuf;
 use std::process::exit;
 use std::str::FromStr;
@@ -31,7 +30,7 @@ use control_plane::safekeeper::SafekeeperNode;
 use control_plane::storage_controller::{
     NeonStorageControllerStartArgs, NeonStorageControllerStopArgs, StorageController,
 };
-use nix::fcntl::{FlockArg, flock};
+use nix::fcntl::{Flock, FlockArg};
 use pageserver_api::config::{
     DEFAULT_HTTP_LISTEN_PORT as DEFAULT_PAGESERVER_HTTP_PORT,
     DEFAULT_PG_LISTEN_PORT as DEFAULT_PAGESERVER_PG_PORT,
@@ -749,16 +748,16 @@ struct TimelineTreeEl {
 
 /// A flock-based guard over the neon_local repository directory
 struct RepoLock {
-    _file: File,
+    _file: Flock<File>,
 }
 
 impl RepoLock {
     fn new() -> Result<Self> {
         let repo_dir = File::open(local_env::base_path())?;
-        let repo_dir_fd = repo_dir.as_raw_fd();
-        flock(repo_dir_fd, FlockArg::LockExclusive)?;
-
-        Ok(Self { _file: repo_dir })
+        match Flock::lock(repo_dir, FlockArg::LockExclusive) {
+            Ok(f) => Ok(Self { _file: f }),
+            Err((_, e)) => Err(e).context("flock error"),
+        }
     }
 }
 

docker-compose/ext-src/alter_db.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+# We need these settings to get the expected output results.
+# We cannot use the environment variables e.g. PGTZ due to
+# https://github.com/neondatabase/neon/issues/1287
+export DATABASE=${1:-contrib_regression}
+psql -c "ALTER DATABASE ${DATABASE} SET neon.allow_unstable_extensions='on'" \
+     -c "ALTER DATABASE ${DATABASE} SET DateStyle='Postgres,MDY'" \
+     -c "ALTER DATABASE ${DATABASE} SET TimeZone='America/Los_Angeles'" \

docker-compose/ext-src/pg_graphql-src/regular-test.sh

@@ -18,6 +18,7 @@ TESTS=${TESTS/row_level_security/}
 TESTS=${TESTS/sqli_connection/}
 dropdb --if-exist contrib_regression
 createdb contrib_regression
+. ../alter_db.sh
 psql -v ON_ERROR_STOP=1 -f test/fixtures.sql -d contrib_regression
 ${REGRESS} --use-existing --dbname=contrib_regression --inputdir=${TESTDIR} ${TESTS}
 

docker-compose/ext-src/pgrag-src/regular-test.sh

@@ -3,6 +3,7 @@ set -ex
 cd "$(dirname "${0}")"
 dropdb --if-exist contrib_regression
 createdb contrib_regression
+. ../alter_db.sh
 psql -d contrib_regression -c "CREATE EXTENSION vector" -c "CREATE EXTENSION rag"
 PG_REGRESS=$(dirname "$(pg_config --pgxs)")/../test/regress/pg_regress
 ${PG_REGRESS} --inputdir=./ --bindir='/usr/local/pgsql/bin'    --use-existing --load-extension=vector --load-extension=rag --dbname=contrib_regression basic_functions text_processing api_keys chunking_functions document_processing embedding_api_functions voyageai_functions

docker-compose/ext-src/pgx_ulid-src/Makefile

@@ -20,5 +20,6 @@ installcheck: regression-test
 regression-test:
 	dropdb --if-exists contrib_regression
 	createdb contrib_regression
+	../alter_db.sh
 	psql -d contrib_regression -c "CREATE EXTENSION $(EXTNAME)"
 	$(PG_REGRESS) --inputdir=. --outputdir=. --use-existing --dbname=contrib_regression $(REGRESS)

docker-compose/ext-src/plv8-src/regular-test.sh

@@ -3,6 +3,7 @@ set -ex
 cd "$(dirname ${0})"
 dropdb --if-exist contrib_regression
 createdb contrib_regression
+. ../alter_db.sh
 PG_REGRESS=$(dirname "$(pg_config --pgxs)")/../test/regress/pg_regress
 REGRESS="$(make -n installcheck | awk '{print substr($0,index($0,"init-extension"));}')"
 REGRESS="${REGRESS/startup_perms/}"

docker-compose/ext-src/rag_bge_small_en_v15-src/Makefile

@@ -11,5 +11,6 @@ PG_REGRESS := $(dir $(PGXS))../../src/test/regress/pg_regress
 installcheck:
 	dropdb --if-exists contrib_regression
 	createdb contrib_regression
+	../alter_db.sh
 	psql -d contrib_regression -c "CREATE EXTENSION vector" -c "CREATE EXTENSION rag_bge_small_en_v15"
 	$(PG_REGRESS) --use-existing --dbname=contrib_regression $(REGRESS)

docker-compose/ext-src/rag_jina_reranker_v1_tiny_en-src/Makefile

@@ -11,5 +11,6 @@ PG_REGRESS := $(dir $(PGXS))../../src/test/regress/pg_regress
 installcheck:
 	dropdb --if-exists contrib_regression
 	createdb contrib_regression
+	../alter_db.sh
 	psql -d contrib_regression -c "CREATE EXTENSION vector" -c "CREATE EXTENSION rag_jina_reranker_v1_tiny_en"
 	$(PG_REGRESS) --use-existing --dbname=contrib_regression $(REGRESS)

docker-compose/ext-src/rum-src/regular-test.sh

@@ -3,5 +3,6 @@ set -ex
 cd "$(dirname ${0})"
 dropdb --if-exist contrib_regression
 createdb contrib_regression
+. ../alter_db.sh
 PG_REGRESS=$(dirname "$(pg_config --pgxs)")/../test/regress/pg_regress
 ${PG_REGRESS} --inputdir=./ --bindir='/usr/local/pgsql/bin' --use-existing --dbname=contrib_regression rum rum_hash ruminv timestamp orderby orderby_hash altorder altorder_hash limits int2 int4 int8 float4 float8 money oid time timetz date interval macaddr inet cidr text varchar char bytea bit varbit numeric rum_weight expr array

docs/rfcs/043-bottom-most-gc-compaction.md

@@ -0,0 +1,194 @@
+# Bottommost Garbage-Collection Compaction
+
+## Summary
+
+The goal of this doc is to propose a way to reliably collect garbages below the GC horizon. This process is called bottom-most garbage-collect-compaction, and is part of the broader legacy-enhanced compaction that we plan to implement in the future.
+
+## Motivation
+
+The current GC algorithm will wait until the covering via image layers before collecting the garbages of a key region. Relying on image layer generation to generate covering images is not reliable. There are prior arts to generate feedbacks from the GC algorithm to the image generation process to accelerate garbage collection, but it slows down the system and creates write amplification.
+
+# Basic Idea
+
+![](images/036-bottom-most-gc-compaction/01-basic-idea.svg)
+
+The idea of bottom-most compaction is simple: we rewrite all layers that are below or intersect with the GC horizon to produce a flat level of image layers at the GC horizon and deltas above the GC horizon. In this process,
+
+- All images and deltas ≤ GC horizon LSN will be dropped. This process collects garbages.
+- We produce images for all keys involved in the compaction process at the GC horizon.
+
+Therefore, it can precisely collect all garbages below the horizon, and reduce the space amplification, i.e., in the staircase pattern (test_gc_feedback).
+
+![The staircase pattern in test_gc_feedback in the original compaction algorithm. The goal is to collect garbage below the red horizontal line.](images/036-bottom-most-gc-compaction/12-staircase-test-gc-feedback.png)
+
+The staircase pattern in test_gc_feedback in the original compaction algorithm. The goal is to collect garbage below the red horizontal line.
+
+# Branches
+
+With branches, the bottom-most compaction should retain a snapshot of the keyspace at the `retain_lsn` so that the child branch can access data at the branch point. This requires some modifications to the basic bottom-most compaction algorithm that we sketched above. 
+
+![](images/036-bottom-most-gc-compaction/03-retain-lsn.svg)
+
+## Single Timeline w/ Snapshots: handle `retain_lsn`
+
+First let’s look into the case where we create branches over the main branch but don’t write any data to them (aka “snapshots”).
+
+The bottom-most compaction algorithm collects all deltas and images of a key and can make decisions on what data to retain. Given that we have a single key’s history as below:
+
+```
+LSN 0x10 -> A
+LSN 0x20 -> append B
+retain_lsn: 0x20
+LSN 0x30 -> append C
+LSN 0x40 -> append D
+retain_lsn: 0x40
+LSN 0x50 -> append E
+GC horizon: 0x50
+LSN 0x60 -> append F
+```
+
+The algorithm will produce:
+
+```
+LSN 0x20 -> AB
+(drop all history below the earliest retain_lsn)
+LSN 0x40 -> ABCD
+(assume the cost of replaying 2 deltas is higher than storing the full image, we generate an image here)
+LSN 0x50 -> append E
+(replay one delta is cheap)
+LSN 0x60 -> append F
+(keep everything as-is above the GC horizon)
+```
+
+![](images/036-bottom-most-gc-compaction/05-btmgc-parent.svg)
+
+What happens is that we balance the space taken by each retain_lsn and the cost of replaying deltas during the bottom-most compaction process. This is controlled by a threshold. If `count(deltas) < $threshold`, the deltas will be retained. Otherwise, an image will be generated and the deltas will be dropped.
+
+In the example above, the `$threshold` is 2.
+
+## Child Branches with data: pull + partial images
+
+In the previous section we have shown how bottom-most compaction respects `retain_lsn` so that all data that was readable at branch creation remains readable. But branches can have data on their own, and that data can fall out of the branch’s PITR window. So, this section explains how we deal with that.
+
+We will run the same bottom-most compaction for these branches, to ensure the space amplification on the child branch is reasonable. 
+
+```
+branch_lsn: 0x20
+LSN 0x30 -> append P
+LSN 0x40 -> append Q
+LSN 0x50 -> append R
+GC horizon: 0x50
+LSN 0x60 -> append S
+```
+
+Note that bottom-most compaction happens on a per-timeline basis. When it processes this key, it only reads the history from LSN 0x30 without a base image. Therefore, on child branches, the bottom-most compaction process will make image creation decisions based on the same `count(deltas) < $threshold` criteria, and if it decides to create an image, the base image will be retrieved from the ancestor branch.
+
+```
+branch_lsn: 0x20
+LSN 0x50 -> ABPQR
+(we pull the image at LSN 0x20 from the ancestor branch to get AB, and then apply append PQ to the page; we replace the record at 0x40 with an image and drop the delta)
+GC horizon: 0x50
+LSN 0x60 -> append S
+```
+
+![](images/036-bottom-most-gc-compaction/06-btmgc-child.svg)
+
+Note that for child branches, we do not create image layers for the images when bottom-most compaction runs. Instead, we drop the 0x30/0x40/0x50 delta records and directly place the image ABPQR@0x50 into the delta layer, which serves as a sparse image layer. For child branches, if we create image layers, we will need to put all keys in the range into the image layer. This causes space bloat and slow compactions. In this proposal, the compaction process will only compact and process keys modified inside the child branch.
+
+# Result
+
+Bottom-most compaction ensures all garbage under the GC horizon gets collected right away (compared with “eventually” in the current algorithm). Meanwhile, it generates images at each of the retain_lsn to ensure branch reads are fast. As we make per-key decisions on whether to generate an image or not, the theoretical lower bound of the storage space we need to retain for a branch is lower than before.
+
+Before: min(sum(logs for each key), sum(image for each key)), for each partition — we always generate image layers on a key range
+
+After: sum(min(logs for each key, image for each key))
+
+# Compaction Trigger
+
+The bottom-most compaction can be automatically triggered. The goal of the trigger is that it should ensure a constant factor for write amplification. Say that the user write 1GB of WAL into the system, we should write 1GB x C data to S3. The legacy compaction algorithm does not have such a constant factor C. The data we write to S3 is quadratic to the logical size of the database (see [A Theoretical View of Neon Storage](https://www.notion.so/A-Theoretical-View-of-Neon-Storage-8d7ad7555b0c41b2a3597fa780911194?pvs=21)).
+
+We propose the following compaction trigger that generates a constant write amplification factor. Write amplification >= total writes to S3 / total user writes. We only analyze the write amplification caused by the bottom-most GC-compaction process, ignoring the legacy create image layers amplification.
+
+Given that we have ***X*** bytes of the delta layers above the GC horizon, ***A*** bytes of the delta layers intersecting with the GC horizon, ***B*** bytes of the delta layers below the GC horizon, and ***C*** bytes of the image layers below the GC horizon.
+
+The legacy GC + compaction loop will always keep ***A*** unchanged, reduce ***B and C*** when there are image layers covering the key range. This yields 0 write amplification (only file deletions) and extra ***B*** bytes of space.
+
+![](images/036-bottom-most-gc-compaction/09-btmgc-analysis-2.svg)
+
+The bottom-most compaction proposed here will split ***A*** into deltas above the GC horizon and below the GC horizon. Everything below the GC horizon will be image layers after the compaction (not considering branches). Therefore, this yields ***A+C*** extra write traffic each iteration, plus 0 extra space.
+
+![](images/036-bottom-most-gc-compaction/07-btmgc-analysis-1.svg)
+
+Also considering read amplification (below the GC horizon). When a read request reaches the GC horizon, the read amplification will be (A+B+C)/C=1+(A+B)/C. Reducing ***A*** and ***B*** can help reduce the read amplification below the GC horizon.
+
+The metrics-based trigger will wait until a point that space amplification is not that large and write amplification is not that large before the compaction gets triggered. The trigger is defined as **(A+B)/C ≥ 1 (or some other ratio)**.
+
+To reason about this trigger, consider the two cases:
+
+**Data Ingestion**
+
+User keeps ingesting data into the database, which indicates that WAL size roughly equals to the database logical size. The compaction gets triggered only when the newly-written WAL roughly equals to the current bottom-most image size (=X). Therefore, it’s triggered when the database size gets doubled. This is a reasonable amount of work. Write amplification is 2X/X=1 for the X amount of data written.
+
+![](images/036-bottom-most-gc-compaction/10-btmgc-analysis-3.svg)
+
+**Updates/Deletion**
+
+In this case, WAL size will be larger than the database logical size ***D***. The compaction gets triggered for every ***D*** bytes of WAL written. Therefore, for every ***D*** bytes of WAL, we rewrite the bottom-most layer, which produces an extra ***D*** bytes of write amplification. This incurs exactly 2x write amplification (by the write of D), 1.5x write amplification (if we count from the start of the process) and no space amplification. 
+
+![](images/036-bottom-most-gc-compaction/11-btmgc-analysis-4.svg)
+
+Note that here I try to reason that write amplification is a constant (i.e., the data we write to S3 is proportional to the data the user write). The main problem with the current legacy compaction algorithm is that write amplification is proportional to the database size.
+
+The next step is to optimize the write amplification above the GC horizon (i.e., change the image creation criteria, top-most compaction, or introduce tiered compaction), to ensure the write amplification of the whole system is a constant factor.
+
+20GB layers → +20GB layers → delete 20GB, need 40GB temporary space
+
+# Sub-Compactions
+
+The gc-compaction algorithm may take a long time and we need to split the job into multiple sub-compaction jobs.
+
+![](images/036-bottom-most-gc-compaction/13-job-split.svg)
+
+As in the figure, the auto-trigger schedules a compaction job covering the full keyspace below a specific LSN. In such case that we cannot finish compacting it in one run in a reasonable amount of time, the algorithm will vertically split it into multiple jobs (in this case, 5).
+
+Each gc-compaction job will create one level of delta layers and one flat level of image layers for each LSN. Those layers will be automatically split based on size, which means that if the sub-compaction job produces 1GB of deltas, it will produce 4 * 256MB delta layers. For those layers that is not fully contained within the sub-compaction job rectangles, it will be rewritten to only contain the keys outside of the key range.
+
+# Implementation
+
+The main implementation of gc-compaction is in `compaction.rs`.
+
+* `compact_with_gc`: The main loop of gc-compaction. It takes a rectangle range of the layer map and compact that specific range. It selects layers intersecting with the rectangle, downloads the layers, creates the k-merge iterator to read those layers in the key-lsn order, and decide which keys to keep or insert a reconstructed page. The process is the basic unit of a gc-compaction and is not interruptable. If the process gets preempted by L0 compaction, it has to be restarted from scratch. For layers overlaps with the rectangle but not fully inside, the main loop will also rewrite them so that the new layer (or two layers if both left and right ends are outside of the rectangle) has the same LSN range as the original one but only contain the keys outside of the compaction range.
+* `gc_compaction_split_jobs`: Splits a big gc-compaction job into sub-compactions based on heuristics in the layer map. The function looks at the layer map and splits the compaction job based on the size of the layers so that each compaction job only pulls ~4GB of layer files.
+* `generate_key_retention` and `KeyHistoryRetention`: Implements the algorithm described in the "basic idea" and "branch" chapter of this RFC. It takes a vector of history of a key (key-lsn-value) and decides which LSNs of the key to retain. If there are too many deltas between two retain_lsns, it will reconstruct the page and insert an image into the compaction result. Also, we implement `KeyHistoryRetention::verify` to ensure the generated result is not corrupted -- all retain_lsns and all LSNs above the gc-horizon should be accessible.
+* `GcCompactionQueue`: the automatic trigger implementation for gc-compaction. `GcCompactionQueue::iteration` is called at the end of the tenant compaction loop. It will then call `trigger_auto_compaction` to decide whether to trigger a gc-compaction job for this tenant. If yes, the compaction-job will be added to the compaction queue, and the queue will be slowly drained once there are no other compaction jobs running. gc-compaction has the lowest priority. If a sub-compaction job is not successful or gets preempted by L0 compaction (see limitations for reasons why a compaction job would fail), it will _not_ be retried.
+* Changes to `index_part.json`: we added a `last_completed_lsn` field to the index part for the auto-trigger to decide when to trigger a compaction.
+* Changes to the read path: when gc-compaction updates the layer map, all reads need to wait. See `gc_compaction_layer_update_lock` and comments in the code path for more information.
+
+Gc-compaction can also be scheduled over the HTTP API. Example:
+
+```
+curl 'localhost:9898/v1/tenant/:tenant_id/timeline/:timeline_id/compact?enhanced_gc_bottom_most_compaction=true&dry_run=true' -X PUT -H "Content-Type: application/json" -d '{"scheduled": true, "compact_key_range": { "start": "000000067F0000A0000002A1CF0100000000", "end": "000000067F0000A0000002A1D70100000000" } }'
+```
+
+The `dry_run` mode can be specified in the query string so that the compaction will go through all layers to estimate how much space can be saved without writing the compaction result into the layer map.
+
+The auto-trigger is controlled by tenant-level flag `gc_compaction_enabled`. If this is set to false, no gc-compaction will be automatically scheduled on this tenant (but manual trigger still works).
+
+# Next Steps
+
+There are still some limitations of gc-compaction itself that needs to be resolved and tested,
+
+- gc-compaction is currently only automatically triggered on root branches. We have not tested gc-compaction on child branches in staging.
+- gc-compaction will skip aux key regions because of the possible conflict with the assumption of aux file tombstones.
+- gc-compaction does not consider keyspaces at retain_lsns and only look at keys in the layers. This also causes us giving up some sub-compaction jobs because a key might have part of its history available due to traditional GC removing part of the history.
+- We limit gc-compaction to run over shards <= 150GB to avoid gc-compaction taking too much time blocking other compaction jobs. The sub-compaction split algorithm needs to be improved to be able to split vertically and horizontally. Also, we need to move the download layer process out of the compaction loop so that we don't block other compaction jobs for too long.
+- The compaction trigger always schedules gc-compaction from the lowest LSN to the gc-horizon. Currently we do not schedule compaction jobs that only selects layers in the middle. Allowing this could potentially reduce the number of layers read/write throughout the process.
+- gc-compaction will give up if there are too many layers to rewrite or if there are not enough disk space for the compaction.
+- gc-compaction sometimes fails with "no key produced during compaction", which means that all existing keys within the compaction range can be collected; but we don't have a way to write this information back to the layer map -- we cannot generate an empty image layer.
+- We limit the maximum size of deltas for a single key to 512MB. If above this size, gc-compaction will give up. This can be resolved by changing `generate_key_retention` to be a stream instead of requiring to collect all the key history.
+
+In the future,
+
+- Top-most compaction: ensure we always have an image coverage for the latest data (or near the latest data), so that reads will be fast at the latest LSN.
+- Tiered compaction on deltas: ensure read from any LSN is fast.
+- Per-timeline compaction → tenant-wide compaction?

docs/rfcs/images/036-bottom-most-gc-compaction/01-basic-idea.svg

@@ -0,0 +1,135 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" version="1.1" viewBox="82 284 863 375" width="863" height="375">
+  <defs/>
+  <g id="01-basic-idea" stroke-opacity="1" stroke-dasharray="none" stroke="none" fill="none" fill-opacity="1">
+    <title>01-basic-idea</title>
+    <rect fill="white" x="82" y="284" width="863" height="375"/>
+    <g id="01-basic-idea_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_2">
+        <rect x="234" y="379.5" width="203.5" height="17.5" fill="white"/>
+        <rect x="234" y="379.5" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_3">
+        <rect x="453.5" y="379.5" width="203.5" height="17.5" fill="white"/>
+        <rect x="453.5" y="379.5" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_4">
+        <rect x="672.5" y="379.5" width="203.5" height="17.5" fill="white"/>
+        <rect x="672.5" y="379.5" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_5">
+        <rect x="234" y="288.5" width="127" height="77.5" fill="white"/>
+        <rect x="234" y="288.5" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_6">
+        <rect x="375" y="288.5" width="127" height="77.5" fill="white"/>
+        <rect x="375" y="288.5" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_7">
+        <rect x="516" y="288.5" width="127" height="77.5" fill="white"/>
+        <rect x="516" y="288.5" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_8">
+        <rect x="657" y="288.5" width="127" height="77.5" fill="white"/>
+        <rect x="657" y="288.5" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_9">
+        <rect x="798" y="288.5" width="78" height="77.5" fill="white"/>
+        <rect x="798" y="288.5" width="78" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_11">
+        <line x1="185.5" y1="326.75" x2="943.7734" y2="326.75" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_12">
+        <text transform="translate(87 318.026)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_13">
+        <text transform="translate(106.41 372.886)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8.39" y="15" xml:space="preserve">Images </tspan>
+          <tspan font-family="Helvetica Neue" font-size="10" fill="black" x="29132252e-19" y="28.447998" xml:space="preserve">at earlier LSN</tspan>
+        </text>
+      </g>
+      <g id="Graphic_14">
+        <text transform="translate(121.92 289.578)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8739676e-19" y="15" xml:space="preserve">Deltas</tspan>
+        </text>
+      </g>
+      <g id="Graphic_15">
+        <path d="M 517.125 423.5 L 553.375 423.5 L 553.375 482 L 571.5 482 L 535.25 512 L 499 482 L 517.125 482 Z" fill="white"/>
+        <path d="M 517.125 423.5 L 553.375 423.5 L 553.375 482 L 571.5 482 L 535.25 512 L 499 482 L 517.125 482 Z" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_26">
+        <rect x="234" y="599.474" width="203.5" height="17.5" fill="white"/>
+        <rect x="234" y="599.474" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_25">
+        <rect x="453.5" y="599.474" width="203.5" height="17.5" fill="white"/>
+        <rect x="453.5" y="599.474" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_24">
+        <rect x="672.5" y="599.474" width="203.5" height="17.5" fill="white"/>
+        <rect x="672.5" y="599.474" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_23">
+        <rect x="234" y="533" width="127" height="52.974" fill="white"/>
+        <rect x="234" y="533" width="127" height="52.974" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_22">
+        <rect x="375" y="533" width="310.5" height="52.974" fill="white"/>
+        <rect x="375" y="533" width="310.5" height="52.974" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_21">
+        <rect x="702.5" y="533" width="173.5" height="52.974" fill="white"/>
+        <rect x="702.5" y="533" width="173.5" height="52.974" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_18">
+        <line x1="185.5" y1="607.724" x2="943.7734" y2="607.724" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_16">
+        <text transform="translate(121.92 538)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8739676e-19" y="15" xml:space="preserve">Deltas</tspan>
+        </text>
+      </g>
+      <g id="Graphic_27">
+        <text transform="translate(114.8 592.86)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="3488765e-18" y="15" xml:space="preserve">Images </tspan>
+          <tspan font-family="Helvetica Neue" font-size="10" fill="black" x="4.01" y="28.447998" xml:space="preserve">at GC LSN</tspan>
+        </text>
+      </g>
+      <g id="Graphic_28">
+        <rect x="243.06836" y="300" width="624.3633" height="17.5" fill="#c0ffc0"/>
+        <text transform="translate(248.06836 301.068)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="13" fill="black" x="233.52364" y="12" xml:space="preserve">Deltas above GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_30">
+        <rect x="243.06836" y="335.5" width="624.3633" height="17.5" fill="#c0ffff"/>
+        <text transform="translate(248.06836 336.568)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="13" fill="black" x="233.89414" y="12" xml:space="preserve">Deltas below GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_32">
+        <rect x="243.06836" y="550.737" width="624.3633" height="17.5" fill="#c0ffc0"/>
+        <text transform="translate(248.06836 551.805)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="13" fill="black" x="233.52364" y="12" xml:space="preserve">Deltas above GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_33">
+        <rect x="304" y="630.474" width="485.5" height="28.447998" fill="#c0ffff"/>
+        <text transform="translate(309 637.016)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="13" fill="black" x="63.095" y="12" xml:space="preserve">Deltas and image below GC Horizon gets garbage-collected</tspan>
+        </text>
+      </g>
+      <g id="Graphic_34">
+        <text transform="translate(576.5 444.0325)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="0" y="11" xml:space="preserve">WAL replay of deltas+image below GC Horizon</tspan>
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="0" y="25.336" xml:space="preserve">Reshuffle deltas</tspan>
+        </text>
+      </g>
+    </g>
+  </g>
+</svg>

docs/rfcs/images/036-bottom-most-gc-compaction/03-retain-lsn.svg

@@ -0,0 +1,141 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" version="1.1" viewBox="-104 215 863 335" width="863" height="335">
+  <defs>
+    <marker orient="auto" overflow="visible" markerUnits="strokeWidth" id="FilledArrow_Marker" stroke-linejoin="miter" stroke-miterlimit="10" viewBox="-1 -4 10 8" markerWidth="10" markerHeight="8" color="#7f8080">
+      <g>
+        <path d="M 8 0 L 0 -3 L 0 3 Z" fill="currentColor" stroke="currentColor" stroke-width="1"/>
+      </g>
+    </marker>
+  </defs>
+  <g id="03-retain-lsn" stroke-opacity="1" stroke-dasharray="none" stroke="none" fill="none" fill-opacity="1">
+    <title>03-retain-lsn</title>
+    <rect fill="white" x="-104" y="215" width="863" height="335"/>
+    <g id="03-retain-lsn_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_28">
+        <rect x="48" y="477" width="203.5" height="9.990005" fill="white"/>
+        <rect x="48" y="477" width="203.5" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_27">
+        <rect x="267.5" y="477" width="203.5" height="9.990005" fill="white"/>
+        <rect x="267.5" y="477" width="203.5" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_26">
+        <rect x="486.5" y="477" width="203.5" height="9.990005" fill="white"/>
+        <rect x="486.5" y="477" width="203.5" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_20">
+        <line x1="-.5" y1="387.172" x2="757.7734" y2="387.172" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_19">
+        <text transform="translate(-99 378.448)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_31">
+        <rect x="48.25" y="410" width="203.5" height="9.990005" fill="white"/>
+        <rect x="48.25" y="410" width="203.5" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_30">
+        <rect x="267.75" y="410" width="203.5" height="9.990005" fill="white"/>
+        <rect x="267.75" y="410" width="203.5" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_29">
+        <rect x="486.75" y="410" width="203.5" height="9.990005" fill="white"/>
+        <rect x="486.75" y="410" width="203.5" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_34">
+        <rect x="48.25" y="431.495" width="113.75" height="34" fill="white"/>
+        <rect x="48.25" y="431.495" width="113.75" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_33">
+        <rect x="172.5" y="431.495" width="203.5" height="34" fill="white"/>
+        <rect x="172.5" y="431.495" width="203.5" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_32">
+        <rect x="386.5" y="431.495" width="303.5" height="34" fill="white"/>
+        <rect x="386.5" y="431.495" width="303.5" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_37">
+        <rect x="48" y="498.495" width="203.5" height="9.990005" fill="white"/>
+        <rect x="48" y="498.495" width="203.5" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_36">
+        <rect x="267.5" y="498.495" width="203.5" height="9.990005" fill="white"/>
+        <rect x="267.5" y="498.495" width="203.5" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_35">
+        <rect x="486.5" y="498.495" width="203.5" height="9.990005" fill="white"/>
+        <rect x="486.5" y="498.495" width="203.5" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_38">
+        <line x1="-10.48" y1="535.5395" x2="39.318294" y2="508.24794" marker-end="url(#FilledArrow_Marker)" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_39">
+        <text transform="translate(-96.984 526.3155)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="40500936e-20" y="15" xml:space="preserve">retain_lsn 1</tspan>
+        </text>
+      </g>
+      <g id="Line_41">
+        <line x1="-10.48" y1="507.0915" x2="38.90236" y2="485.8992" marker-end="url(#FilledArrow_Marker)" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_40">
+        <text transform="translate(-96.984 497.8675)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="40500936e-20" y="15" xml:space="preserve">retain_lsn 2</tspan>
+        </text>
+      </g>
+      <g id="Line_43">
+        <line x1="-10.48" y1="478.6435" x2="39.44267" y2="453.01616" marker-end="url(#FilledArrow_Marker)" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_42">
+        <text transform="translate(-96.984 469.4195)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="40500936e-20" y="15" xml:space="preserve">retain_lsn 3</tspan>
+        </text>
+      </g>
+      <g id="Line_45">
+        <line x1="-10.48" y1="448.495" x2="39.65061" y2="419.90015" marker-end="url(#FilledArrow_Marker)" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_44">
+        <text transform="translate(-96.984 439.271)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="40500936e-20" y="15" xml:space="preserve">retain_lsn 4</tspan>
+        </text>
+      </g>
+      <g id="Graphic_46">
+        <rect x="335.46477" y="215.5" width="353.4299" height="125.495" fill="white"/>
+        <rect x="335.46477" y="215.5" width="353.4299" height="125.495" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_48">
+        <text transform="translate(549.3766 317.547)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="6536993e-19" y="15" xml:space="preserve">Dependent Branch</tspan>
+        </text>
+      </g>
+      <g id="Graphic_50">
+        <text transform="translate(340.43824 317.547)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="40500936e-20" y="15" xml:space="preserve">retain_lsn 3</tspan>
+        </text>
+      </g>
+      <g id="Line_57">
+        <line x1="323.90685" y1="248.8045" x2="714.9232" y2="248.8045" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_56">
+        <text transform="translate(165.91346 240.0805)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="35811354e-19" y="15" xml:space="preserve">Branch GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_58">
+        <rect x="493.9232" y="301.6405" width="107.45294" height="9.990005" fill="white"/>
+        <rect x="493.9232" y="301.6405" width="107.45294" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_59">
+        <text transform="translate(358.9232 277.276)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">Partial Image Coverage</tspan>
+        </text>
+      </g>
+      <g id="Graphic_60">
+        <rect x="354.1732" y="301.6405" width="107.45294" height="9.990005" fill="white"/>
+        <rect x="354.1732" y="301.6405" width="107.45294" height="9.990005" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+    </g>
+  </g>
+</svg>

docs/rfcs/images/036-bottom-most-gc-compaction/05-btmgc-parent.svg

@@ -0,0 +1,187 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" version="1.1" viewBox="-235 426 864 366" width="864" height="366">
+  <defs/>
+  <g id="05-btmgc-parent" stroke-opacity="1" stroke-dasharray="none" stroke="none" fill="none" fill-opacity="1">
+    <title>05-btmgc-parent</title>
+    <rect fill="white" x="-235" y="426" width="864" height="366"/>
+    <g id="05-btmgc-parent_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_23">
+        <rect x="-83" y="510.15" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-83" y="510.15" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-78 516.178)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="51.714" y="11" xml:space="preserve">Append C@0x30</tspan>
+        </text>
+      </g>
+      <g id="Graphic_22">
+        <rect x="136.5" y="510.15" width="203.5" height="26.391998" fill="white"/>
+        <rect x="136.5" y="510.15" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_21">
+        <rect x="355.5" y="510.15" width="203.5" height="26.391998" fill="white"/>
+        <rect x="355.5" y="510.15" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_20">
+        <line x1="-100.448" y1="459.224" x2="626.77344" y2="459.224" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_19">
+        <text transform="translate(-230 450.5)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_18">
+        <rect x="-82.75" y="426.748" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-82.75" y="426.748" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-77.75 432.776)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="52.602" y="11" xml:space="preserve">Append F@0x60</tspan>
+        </text>
+      </g>
+      <g id="Graphic_17">
+        <rect x="136.75" y="426.748" width="203.5" height="26.391998" fill="white"/>
+        <rect x="136.75" y="426.748" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_16">
+        <rect x="355.75" y="426.748" width="203.5" height="26.391998" fill="white"/>
+        <rect x="355.75" y="426.748" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_15">
+        <rect x="-82.75" y="464.645" width="113.75" height="34" fill="white"/>
+        <rect x="-82.75" y="464.645" width="113.75" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-77.75 467.309)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="7.505" y="11" xml:space="preserve">Append E@0x50</tspan>
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="6.947" y="25.336" xml:space="preserve">Append D@0x40</tspan>
+        </text>
+      </g>
+      <g id="Graphic_14">
+        <rect x="41.5" y="464.645" width="203.5" height="34" fill="white"/>
+        <rect x="41.5" y="464.645" width="203.5" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_13">
+        <rect x="255.5" y="464.645" width="303.5" height="34" fill="white"/>
+        <rect x="255.5" y="464.645" width="303.5" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_12">
+        <rect x="-83" y="548.047" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-83" y="548.047" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-78 554.075)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="26.796" y="11" xml:space="preserve">A@0x10, Append B@0x20</tspan>
+        </text>
+      </g>
+      <g id="Graphic_11">
+        <rect x="136.5" y="548.047" width="203.5" height="26.391998" fill="white"/>
+        <rect x="136.5" y="548.047" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_10">
+        <rect x="355.5" y="548.047" width="203.5" height="26.391998" fill="white"/>
+        <rect x="355.5" y="548.047" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_24">
+        <line x1="-104" y1="542" x2="610.5" y2="542" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_25">
+        <text transform="translate(-139.604 534.5)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x20</tspan>
+        </text>
+      </g>
+      <g id="Graphic_28">
+        <text transform="translate(-139.604 452.556)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x50</tspan>
+        </text>
+      </g>
+      <g id="Line_30">
+        <line x1="-100.448" y1="481.145" x2="614.052" y2="481.145" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_29">
+        <text transform="translate(-139.604 473.449)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x40</tspan>
+        </text>
+      </g>
+      <g id="Line_48">
+        <line x1="-99.448" y1="701.513" x2="627.77344" y2="701.513" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_47">
+        <text transform="translate(-229 692.789)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_46">
+        <rect x="-81.75" y="670.496" width="113.75" height="26.391998" fill="white"/>
+        <rect x="-81.75" y="670.496" width="113.75" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-76.75 676.524)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="7.727" y="11" xml:space="preserve">Append F@0x60</tspan>
+        </text>
+      </g>
+      <g id="Graphic_43">
+        <rect x="-81.75" y="708.393" width="113.75" height="34" fill="white"/>
+        <rect x="-81.75" y="708.393" width="113.75" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-76.75 718.225)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="7.505" y="11" xml:space="preserve">Append E@0x50</tspan>
+        </text>
+      </g>
+      <g id="Line_37">
+        <line x1="-101" y1="777.2665" x2="613.5" y2="777.2665" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_36">
+        <text transform="translate(-138.604 769.7665)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x20</tspan>
+        </text>
+      </g>
+      <g id="Graphic_33">
+        <text transform="translate(-138.604 694.845)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x50</tspan>
+        </text>
+      </g>
+      <g id="Line_32">
+        <line x1="-99.448" y1="755.089" x2="615.052" y2="755.089" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_31">
+        <text transform="translate(-138.604 747.393)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x40</tspan>
+        </text>
+      </g>
+      <g id="Graphic_40">
+        <rect x="-82" y="770.909" width="203.5" height="14.107002" fill="white"/>
+        <rect x="-82" y="770.909" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-77 770.7945)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="70.836" y="11" xml:space="preserve">AB@0x20</tspan>
+        </text>
+      </g>
+      <g id="Graphic_39">
+        <rect x="137.5" y="770.909" width="203.5" height="14.107002" fill="white"/>
+        <rect x="137.5" y="770.909" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_38">
+        <rect x="356.5" y="770.909" width="203.5" height="14.107002" fill="white"/>
+        <rect x="356.5" y="770.909" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_54">
+        <rect x="-81.75" y="748.5355" width="203.5" height="14.107002" fill="white"/>
+        <rect x="-81.75" y="748.5355" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-76.75 748.421)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="62.28" y="11" xml:space="preserve">ABCD@0x40</tspan>
+        </text>
+      </g>
+      <g id="Graphic_53">
+        <rect x="137.75" y="748.5355" width="203.5" height="14.107002" fill="white"/>
+        <rect x="137.75" y="748.5355" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_52">
+        <rect x="356.75" y="748.5355" width="203.5" height="14.107002" fill="white"/>
+        <rect x="356.75" y="748.5355" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_57">
+        <path d="M 211.32422 585 L 265.17578 585 L 265.17578 611.332 L 287.84375 611.332 L 238.25 633.117 L 188.65625 611.332 L 211.32422 611.332 Z" fill="white"/>
+        <path d="M 211.32422 585 L 265.17578 585 L 265.17578 611.332 L 287.84375 611.332 L 238.25 633.117 L 188.65625 611.332 L 211.32422 611.332 Z" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_60">
+        <rect x="359" y="692.858" width="203.5" height="14.107002" fill="white"/>
+        <rect x="359" y="692.858" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_59">
+        <rect x="41.5" y="693.858" width="303" height="14.107002" fill="white"/>
+        <rect x="41.5" y="693.858" width="303" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+    </g>
+  </g>
+</svg>

docs/rfcs/images/036-bottom-most-gc-compaction/06-btmgc-child.svg

@@ -0,0 +1,184 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" version="1.1" viewBox="-413 471 931 354" width="931" height="354">
+  <defs/>
+  <g id="06-btmgc-child" stroke-opacity="1" stroke-dasharray="none" stroke="none" fill="none" fill-opacity="1">
+    <title>06-btmgc-child</title>
+    <rect fill="white" x="-413" y="471" width="931" height="354"/>
+    <g id="06-btmgc-child_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_47">
+        <rect x="-412" y="594.402" width="928" height="28.447998" fill="white"/>
+        <rect x="-412" y="594.402" width="928" height="28.447998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_46">
+        <rect x="-205" y="555.552" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-205" y="555.552" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-200 561.58)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="52.158" y="11" xml:space="preserve">Append P@0x30</tspan>
+        </text>
+      </g>
+      <g id="Graphic_45">
+        <rect x="14.5" y="555.552" width="203.5" height="26.391998" fill="white"/>
+        <rect x="14.5" y="555.552" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_44">
+        <rect x="233.5" y="555.552" width="203.5" height="26.391998" fill="white"/>
+        <rect x="233.5" y="555.552" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_43">
+        <line x1="-222.448" y1="504.724" x2="504.77344" y2="504.724" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_42">
+        <text transform="translate(-352 496)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_41">
+        <rect x="-204.75" y="472.15" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-204.75" y="472.15" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-199.75 478.178)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="52.158" y="11" xml:space="preserve">Append S@0x60</tspan>
+        </text>
+      </g>
+      <g id="Graphic_40">
+        <rect x="14.75" y="472.15" width="203.5" height="26.391998" fill="white"/>
+        <rect x="14.75" y="472.15" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_39">
+        <rect x="233.75" y="472.15" width="203.5" height="26.391998" fill="white"/>
+        <rect x="233.75" y="472.15" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_38">
+        <rect x="-204.75" y="510.047" width="113.75" height="34" fill="white"/>
+        <rect x="-204.75" y="510.047" width="113.75" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-199.75 512.711)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="7.061" y="11" xml:space="preserve">Append R@0x50</tspan>
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="6.611" y="25.336" xml:space="preserve">Append Q@0x40</tspan>
+        </text>
+      </g>
+      <g id="Graphic_37">
+        <rect x="-80.5" y="510.047" width="203.5" height="34" fill="white"/>
+        <rect x="-80.5" y="510.047" width="203.5" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_36">
+        <rect x="133.5" y="510.047" width="303.5" height="34" fill="white"/>
+        <rect x="133.5" y="510.047" width="303.5" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_33">
+        <text transform="translate(-261.604 498.056)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x50</tspan>
+        </text>
+      </g>
+      <g id="Line_30">
+        <line x1="-224" y1="607.9115" x2="490.5" y2="607.9115" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_29">
+        <text transform="translate(-261.604 600.4115)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x20</tspan>
+        </text>
+      </g>
+      <g id="Graphic_28">
+        <rect x="-205" y="601.554" width="203.5" height="14.107002" fill="white"/>
+        <rect x="-205" y="601.554" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-200 601.4395)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="70.836" y="11" xml:space="preserve">AB@0x20</tspan>
+        </text>
+      </g>
+      <g id="Graphic_27">
+        <rect x="14.5" y="601.554" width="203.5" height="14.107002" fill="white"/>
+        <rect x="14.5" y="601.554" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_26">
+        <rect x="233.5" y="601.554" width="203.5" height="14.107002" fill="white"/>
+        <rect x="233.5" y="601.554" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_25">
+        <text transform="translate(-407 599.1875)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">Ancestor Branch</tspan>
+        </text>
+      </g>
+      <g id="Graphic_24">
+        <rect x="-411" y="795.46" width="928" height="28.447998" fill="white"/>
+        <rect x="-411" y="795.46" width="928" height="28.447998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_20">
+        <line x1="-221.448" y1="755.528" x2="505.77344" y2="755.528" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_19">
+        <text transform="translate(-351 746.804)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_18">
+        <rect x="-203.75" y="723.579" width="203.25" height="26.391998" fill="white"/>
+        <rect x="-203.75" y="723.579" width="203.25" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-198.75 729.607)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" x="52.033" y="11" xml:space="preserve">Append S@0x60</tspan>
+        </text>
+      </g>
+      <g id="Graphic_10">
+        <text transform="translate(-260.604 748.86)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x50</tspan>
+        </text>
+      </g>
+      <g id="Line_7">
+        <line x1="-223" y1="808.9695" x2="491.5" y2="808.9695" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_6">
+        <text transform="translate(-260.604 801.4695)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x20</tspan>
+        </text>
+      </g>
+      <g id="Graphic_5">
+        <rect x="-204" y="802.612" width="203.5" height="14.107002" fill="white"/>
+        <rect x="-204" y="802.612" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-199 802.4975)" fill="#b1001c">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="#b1001c" x="70.836" y="11" xml:space="preserve">AB</tspan>
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" y="11" xml:space="preserve">@0x20</tspan>
+        </text>
+      </g>
+      <g id="Graphic_4">
+        <rect x="15.5" y="802.612" width="203.5" height="14.107002" fill="white"/>
+        <rect x="15.5" y="802.612" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_3">
+        <rect x="234.5" y="802.612" width="203.5" height="14.107002" fill="white"/>
+        <rect x="234.5" y="802.612" width="203.5" height="14.107002" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_2">
+        <text transform="translate(-406 800.2455)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">Ancestor Branch</tspan>
+        </text>
+      </g>
+      <g id="Graphic_48">
+        <path d="M 89.32422 639.081 L 143.17578 639.081 L 143.17578 665.413 L 165.84375 665.413 L 116.25 687.198 L 66.65625 665.413 L 89.32422 665.413 Z" fill="white"/>
+        <path d="M 89.32422 639.081 L 143.17578 639.081 L 143.17578 665.413 L 165.84375 665.413 L 116.25 687.198 L 66.65625 665.413 L 89.32422 665.413 Z" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_49">
+        <rect x="-204" y="762.428" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-204" y="762.428" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-199 768.456)" fill="#b1001c">
+          <tspan font-family="Helvetica Neue" font-size="12" fill="#b1001c" x="58.278" y="11" xml:space="preserve">AB</tspan>
+          <tspan font-family="Helvetica Neue" font-size="12" fill="black" y="11" xml:space="preserve">PQR@0x50</tspan>
+        </text>
+      </g>
+      <g id="Graphic_59">
+        <rect x="14.5" y="723.579" width="203.5" height="26.391998" fill="white"/>
+        <rect x="14.5" y="723.579" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_58">
+        <rect x="233.5" y="723.579" width="203.5" height="26.391998" fill="white"/>
+        <rect x="233.5" y="723.579" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_63">
+        <rect x="9" y="762.085" width="203.5" height="26.391998" fill="white"/>
+        <rect x="9" y="762.085" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_62">
+        <rect x="225" y="762.085" width="213" height="26.391998" fill="white"/>
+        <rect x="225" y="762.085" width="213" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+    </g>
+  </g>
+</svg>

docs/rfcs/images/036-bottom-most-gc-compaction/07-btmgc-analysis-1.svg

@@ -0,0 +1,180 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" version="1.1" viewBox="-556 476 923 411" width="923" height="411">
+  <defs/>
+  <g id="07-btmgc-analysis-1" stroke-opacity="1" stroke-dasharray="none" stroke="none" fill="none" fill-opacity="1">
+    <title>07-btmgc-analysis-1</title>
+    <rect fill="white" x="-556" y="476" width="923" height="411"/>
+    <g id="07-btmgc-analysis-1_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_85">
+        <rect x="-404" y="609.062" width="203.5" height="17.5" fill="white"/>
+        <rect x="-404" y="609.062" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_84">
+        <rect x="-184.5" y="609.062" width="203.5" height="17.5" fill="white"/>
+        <rect x="-184.5" y="609.062" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_83">
+        <rect x="34.5" y="609.062" width="203.5" height="17.5" fill="white"/>
+        <rect x="34.5" y="609.062" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_82">
+        <rect x="-404" y="479.922" width="127" height="77.5" fill="white"/>
+        <rect x="-404" y="479.922" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_81">
+        <rect x="-263" y="479.922" width="127" height="77.5" fill="white"/>
+        <rect x="-263" y="479.922" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_80">
+        <rect x="-122" y="479.922" width="127" height="77.5" fill="white"/>
+        <rect x="-122" y="479.922" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_79">
+        <rect x="19" y="479.922" width="127" height="77.5" fill="white"/>
+        <rect x="19" y="479.922" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_78">
+        <rect x="160" y="479.922" width="78" height="77.5" fill="white"/>
+        <rect x="160" y="479.922" width="78" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_77">
+        <line x1="-452.5" y1="518.172" x2="251" y2="518.172" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_76">
+        <text transform="translate(-551 509.448)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_75">
+        <text transform="translate(-531.59 602.448)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8.39" y="15" xml:space="preserve">Images </tspan>
+          <tspan font-family="Helvetica Neue" font-size="10" fill="black" x="29132252e-19" y="28.447998" xml:space="preserve">at earlier LSN</tspan>
+        </text>
+      </g>
+      <g id="Graphic_74">
+        <text transform="translate(-516.08 481)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8739676e-19" y="15" xml:space="preserve">Deltas</tspan>
+        </text>
+      </g>
+      <g id="Graphic_73">
+        <path d="M -120.675 651.5 L -84.425 651.5 L -84.425 710 L -66.3 710 L -102.55 740 L -138.8 710 L -120.675 710 Z" fill="white"/>
+        <path d="M -120.675 651.5 L -84.425 651.5 L -84.425 710 L -66.3 710 L -102.55 740 L -138.8 710 L -120.675 710 Z" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_72">
+        <rect x="-403.8" y="827.474" width="203.5" height="17.5" fill="white"/>
+        <rect x="-403.8" y="827.474" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_71">
+        <rect x="-184.3" y="827.474" width="203.5" height="17.5" fill="white"/>
+        <rect x="-184.3" y="827.474" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_70">
+        <rect x="34.7" y="827.474" width="203.5" height="17.5" fill="white"/>
+        <rect x="34.7" y="827.474" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_69">
+        <rect x="-403.8" y="761" width="127" height="52.974" fill="white"/>
+        <rect x="-403.8" y="761" width="127" height="52.974" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_68">
+        <rect x="-262.8" y="761" width="310.5" height="52.974" fill="white"/>
+        <rect x="-262.8" y="761" width="310.5" height="52.974" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_67">
+        <rect x="64.7" y="761" width="173.5" height="52.974" fill="white"/>
+        <rect x="64.7" y="761" width="173.5" height="52.974" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_66">
+        <line x1="-452.3" y1="835.724" x2="251.2" y2="835.724" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_65">
+        <text transform="translate(-515.88 766)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8739676e-19" y="15" xml:space="preserve">Deltas</tspan>
+        </text>
+      </g>
+      <g id="Graphic_64">
+        <text transform="translate(-523 820.86)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="3488765e-18" y="15" xml:space="preserve">Images </tspan>
+          <tspan font-family="Helvetica Neue" font-size="10" fill="black" x="4.01" y="28.447998" xml:space="preserve">at GC LSN</tspan>
+        </text>
+      </g>
+      <g id="Graphic_63">
+        <rect x="-394.93164" y="491.422" width="624.3633" height="17.5" fill="#c0ffc0"/>
+        <text transform="translate(-389.93164 492.49)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="13" fill="black" x="233.52364" y="12" xml:space="preserve">Deltas above GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_62">
+        <rect x="-394.93164" y="526.922" width="624.3633" height="17.5" fill="#c0ffff"/>
+        <text transform="translate(-389.93164 527.99)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="13" fill="black" x="233.89414" y="12" xml:space="preserve">Deltas below GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_61">
+        <rect x="-394.73164" y="778.737" width="624.3633" height="17.5" fill="#c0ffc0"/>
+        <text transform="translate(-389.73164 779.805)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="13" fill="black" x="233.52364" y="12" xml:space="preserve">Deltas above GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_60">
+        <rect x="-333.8" y="858.474" width="485.5" height="28.447998" fill="#c0ffff"/>
+        <text transform="translate(-328.8 865.016)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="13" fill="black" x="63.095" y="12" xml:space="preserve">Deltas and image below GC Horizon gets garbage-collected</tspan>
+        </text>
+      </g>
+      <g id="Graphic_86">
+        <text transform="translate(263 499.724)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="32" fill="black" x="0" y="30" xml:space="preserve">size=A</tspan>
+        </text>
+      </g>
+      <g id="Line_87">
+        <line x1="260.87012" y1="479.068" x2="360.71387" y2="479.068" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_88">
+        <line x1="260.87012" y1="561" x2="360.71387" y2="561" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_89">
+        <rect x="-403.8" y="569" width="161.8" height="28.447998" fill="white"/>
+        <rect x="-403.8" y="569" width="161.8" height="28.447998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_90">
+        <rect x="-229.5" y="569.018" width="277.2" height="28.447998" fill="white"/>
+        <rect x="-229.5" y="569.018" width="277.2" height="28.447998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_91">
+        <rect x="64.7" y="569.018" width="173.5" height="28.447998" fill="white"/>
+        <rect x="64.7" y="569.018" width="173.5" height="28.447998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_92">
+        <line x1="262" y1="602" x2="361.84375" y2="602" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_93">
+        <line x1="263" y1="625.562" x2="362.84375" y2="625.562" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_94">
+        <text transform="translate(264.53787 562.276)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="32" fill="black" x="14210855e-21" y="30" xml:space="preserve">size=B</tspan>
+        </text>
+      </g>
+      <g id="Graphic_95">
+        <text transform="translate(285.12 599.5)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="24" fill="black" x="0" y="23" xml:space="preserve">size=C</tspan>
+        </text>
+      </g>
+      <g id="Graphic_98">
+        <text transform="translate(264.53787 773.772)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="26" fill="black" x="8881784e-19" y="25" xml:space="preserve">A</tspan>
+          <tspan font-family="Lucida Grande" font-size="26" fill="black" y="25" xml:space="preserve">↓</tspan>
+        </text>
+      </g>
+      <g id="Graphic_97">
+        <text transform="translate(265.87013 815.5)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="26" fill="black" x="6536993e-19" y="25" xml:space="preserve">B</tspan>
+          <tspan font-family="Lucida Grande" font-size="26" fill="black" y="25" xml:space="preserve">↓</tspan>
+        </text>
+      </g>
+    </g>
+  </g>
+</svg>

docs/rfcs/images/036-bottom-most-gc-compaction/08-optimization.svg

@@ -0,0 +1,158 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" version="1.1" viewBox="-235 406 586 424" width="586" height="424">
+  <defs/>
+  <g id="08-optimization" stroke-opacity="1" stroke-dasharray="none" stroke="none" fill="none" fill-opacity="1">
+    <title>08-optimization</title>
+    <rect fill="white" x="-235" y="406" width="586" height="424"/>
+    <g id="08-optimization_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_22">
+        <rect x="-100.448" y="509.902" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-100.448" y="509.902" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_21">
+        <rect x="118.552" y="509.902" width="203.5" height="26.391998" fill="white"/>
+        <rect x="118.552" y="509.902" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_20">
+        <line x1="-101.79572" y1="420.322" x2="349.5" y2="420.322" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_19">
+        <text transform="translate(-230 411.598)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_17">
+        <rect x="-100.198" y="426.5" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-100.198" y="426.5" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_16">
+        <rect x="118.802" y="426.5" width="203.5" height="26.391998" fill="white"/>
+        <rect x="118.802" y="426.5" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_14">
+        <rect x="-100.198" y="464.397" width="108.25" height="34" fill="white"/>
+        <rect x="-100.198" y="464.397" width="108.25" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_13">
+        <rect x="18.552" y="464.397" width="303.5" height="34" fill="white"/>
+        <rect x="18.552" y="464.397" width="303.5" height="34" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_11">
+        <rect x="-100.448" y="547.799" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-100.448" y="547.799" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_10">
+        <rect x="118.552" y="547.799" width="203.5" height="26.391998" fill="white"/>
+        <rect x="118.552" y="547.799" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_24">
+        <line x1="-104" y1="542" x2="339.4011" y2="542" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_25">
+        <text transform="translate(-139.604 534.5)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x20</tspan>
+        </text>
+      </g>
+      <g id="Line_27">
+        <line x1="-101.79572" y1="459.098" x2="341.6054" y2="459.098" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_26">
+        <text transform="translate(-139.604 451.402)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x50</tspan>
+        </text>
+      </g>
+      <g id="Graphic_28">
+        <text transform="translate(-139.604 413.654)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x60</tspan>
+        </text>
+      </g>
+      <g id="Line_30">
+        <line x1="-101.79572" y1="481.145" x2="341.6054" y2="481.145" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_29">
+        <text transform="translate(-139.604 473.449)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x40</tspan>
+        </text>
+      </g>
+      <g id="Graphic_77">
+        <rect x="-100.448" y="765.19595" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-100.448" y="765.19595" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_76">
+        <rect x="118.552" y="765.19595" width="203.5" height="26.391998" fill="white"/>
+        <rect x="118.552" y="765.19595" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_75">
+        <line x1="-101.79572" y1="637.317" x2="349.5" y2="637.317" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_74">
+        <text transform="translate(-230 628.593)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_73">
+        <rect x="-100.198" y="681.794" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-100.198" y="681.794" width="203.5" height="26.391998" stroke="#b1001c" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_72">
+        <rect x="118.802" y="681.794" width="203.5" height="26.391998" fill="white"/>
+        <rect x="118.802" y="681.794" width="203.5" height="26.391998" stroke="#b1001c" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_71">
+        <rect x="-100.198" y="719.69096" width="108.25" height="34" fill="white"/>
+        <rect x="-100.198" y="719.69096" width="108.25" height="34" stroke="#b1001c" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_70">
+        <rect x="18.552" y="719.69096" width="303.5" height="34" fill="white"/>
+        <rect x="18.552" y="719.69096" width="303.5" height="34" stroke="#b1001c" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_69">
+        <rect x="-100.448" y="803.09295" width="203.5" height="26.391998" fill="white"/>
+        <rect x="-100.448" y="803.09295" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_68">
+        <rect x="118.552" y="803.09295" width="203.5" height="26.391998" fill="white"/>
+        <rect x="118.552" y="803.09295" width="203.5" height="26.391998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_67">
+        <line x1="-104" y1="797.294" x2="339.4011" y2="797.294" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_66">
+        <text transform="translate(-139.604 789.794)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x20</tspan>
+        </text>
+      </g>
+      <g id="Graphic_63">
+        <text transform="translate(-139.604 630.649)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x70</tspan>
+        </text>
+      </g>
+      <g id="Line_62">
+        <line x1="-101.79572" y1="736.439" x2="341.6054" y2="736.439" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="4.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_61">
+        <text transform="translate(-139.604 728.743)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="black" x="0" y="13" xml:space="preserve">0x40</tspan>
+        </text>
+      </g>
+      <g id="Graphic_79">
+        <rect x="-100.198" y="644.393" width="168.198" height="26.391998" fill="white"/>
+        <rect x="-100.198" y="644.393" width="168.198" height="26.391998" stroke="#b1001c" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_78">
+        <rect x="80" y="644.393" width="242.302" height="26.391998" fill="white"/>
+        <rect x="80" y="644.393" width="242.302" height="26.391998" stroke="#b1001c" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_81">
+        <line x1="-101.79572" y1="714.139" x2="341.6054" y2="714.139" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="1.0,4.0" stroke-width="1"/>
+      </g>
+      <g id="Graphic_80">
+        <text transform="translate(-139.604 706.443)" fill="#a5a5a5">
+          <tspan font-family="Helvetica Neue" font-size="14" fill="#a5a5a5" x="0" y="13" xml:space="preserve">0x50</tspan>
+        </text>
+      </g>
+    </g>
+  </g>
+</svg>

docs/rfcs/images/036-bottom-most-gc-compaction/09-btmgc-analysis-2.svg

@@ -0,0 +1,184 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" version="1.1" viewBox="-562 479 876 429" width="876" height="429">
+  <defs/>
+  <g id="09-btmgc-analysis-2" stroke-opacity="1" stroke-dasharray="none" stroke="none" fill="none" fill-opacity="1">
+    <title>09-btmgc-analysis-2</title>
+    <rect fill="white" x="-562" y="479" width="876" height="429"/>
+    <g id="09-btmgc-analysis-2_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_85">
+        <rect x="-404" y="622.386" width="203.5" height="17.5" fill="white"/>
+        <rect x="-404" y="622.386" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-399 621.912)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="90.974" y="15" xml:space="preserve">C</tspan>
+        </text>
+      </g>
+      <g id="Graphic_84">
+        <rect x="-184.5" y="622.386" width="203.5" height="17.5" fill="white"/>
+        <rect x="-184.5" y="622.386" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-179.5 621.912)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="90.974" y="15" xml:space="preserve">C</tspan>
+        </text>
+      </g>
+      <g id="Graphic_83">
+        <rect x="34.5" y="622.386" width="203.5" height="17.5" fill="white"/>
+        <rect x="34.5" y="622.386" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(39.5 621.912)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="90.974" y="15" xml:space="preserve">C</tspan>
+        </text>
+      </g>
+      <g id="Graphic_82">
+        <rect x="-404" y="479.922" width="127" height="77.5" fill="white"/>
+        <rect x="-404" y="479.922" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-399 509.448)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="53.316" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Graphic_81">
+        <rect x="-263" y="479.922" width="127" height="77.5" fill="white"/>
+        <rect x="-263" y="479.922" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-258 509.448)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="53.316" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Graphic_80">
+        <rect x="-122" y="479.922" width="127" height="77.5" fill="white"/>
+        <rect x="-122" y="479.922" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-117 509.448)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="53.316" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Graphic_79">
+        <rect x="19" y="479.922" width="127" height="77.5" fill="white"/>
+        <rect x="19" y="479.922" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(24 509.448)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="53.316" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Graphic_78">
+        <rect x="160" y="479.922" width="78" height="77.5" fill="white"/>
+        <rect x="160" y="479.922" width="78" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(165 509.448)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="28.816" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Line_77">
+        <line x1="-452.5" y1="518.172" x2="251" y2="518.172" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_76">
+        <text transform="translate(-551 509.448)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_73">
+        <path d="M -120.675 651.5 L -84.425 651.5 L -84.425 710 L -66.3 710 L -102.55 740 L -138.8 710 L -120.675 710 Z" fill="white"/>
+        <path d="M -120.675 651.5 L -84.425 651.5 L -84.425 710 L -66.3 710 L -102.55 740 L -138.8 710 L -120.675 710 Z" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_89">
+        <rect x="-403.8" y="582.324" width="161.8" height="28.447998" fill="white"/>
+        <rect x="-403.8" y="582.324" width="161.8" height="28.447998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-398.8 587.324)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="70.42" y="15" xml:space="preserve">B</tspan>
+        </text>
+      </g>
+      <g id="Graphic_90">
+        <rect x="-229.5" y="582.342" width="277.2" height="28.447998" fill="white"/>
+        <rect x="-229.5" y="582.342" width="277.2" height="28.447998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-224.5 587.342)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="128.12" y="15" xml:space="preserve">B</tspan>
+        </text>
+      </g>
+      <g id="Graphic_91">
+        <rect x="64.7" y="582.342" width="173.5" height="28.447998" fill="white"/>
+        <rect x="64.7" y="582.342" width="173.5" height="28.447998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(69.7 587.342)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="76.27" y="15" xml:space="preserve">B</tspan>
+        </text>
+      </g>
+      <g id="Graphic_97">
+        <rect x="-403.8" y="564.842" width="490.8" height="12.157997" fill="white"/>
+        <rect x="-403.8" y="564.842" width="490.8" height="12.157997" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-398.8 561.697)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="234.624" y="15" xml:space="preserve">C</tspan>
+        </text>
+      </g>
+      <g id="Graphic_109">
+        <rect x="28.6" y="889.964" width="203.5" height="17.5" fill="white"/>
+        <rect x="28.6" y="889.964" width="203.5" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(33.6 889.49)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="90.974" y="15" xml:space="preserve">C</tspan>
+        </text>
+      </g>
+      <g id="Graphic_108">
+        <rect x="-409.9" y="747.5" width="127" height="77.5" fill="white"/>
+        <rect x="-409.9" y="747.5" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-404.9 777.026)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="53.316" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Graphic_107">
+        <rect x="-268.9" y="747.5" width="127" height="77.5" fill="white"/>
+        <rect x="-268.9" y="747.5" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-263.9 777.026)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="53.316" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Graphic_106">
+        <rect x="-127.9" y="747.5" width="127" height="77.5" fill="white"/>
+        <rect x="-127.9" y="747.5" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-122.9 777.026)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="53.316" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Graphic_105">
+        <rect x="13.1" y="747.5" width="127" height="77.5" fill="white"/>
+        <rect x="13.1" y="747.5" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(18.1 777.026)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="53.316" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Graphic_104">
+        <rect x="154.1" y="747.5" width="78" height="77.5" fill="white"/>
+        <rect x="154.1" y="747.5" width="78" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(159.1 777.026)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="28.816" y="15" xml:space="preserve">A</tspan>
+        </text>
+      </g>
+      <g id="Line_103">
+        <line x1="-458.4" y1="785.75" x2="245.1" y2="785.75" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_102">
+        <text transform="translate(-556.9 777.026)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_99">
+        <rect x="58.8" y="849.92" width="173.5" height="28.447998" fill="white"/>
+        <rect x="58.8" y="849.92" width="173.5" height="28.447998" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(63.8 854.92)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="76.27" y="15" xml:space="preserve">B</tspan>
+        </text>
+      </g>
+      <g id="Graphic_98">
+        <rect x="-409.7" y="832.42" width="490.8" height="12.157997" fill="white"/>
+        <rect x="-409.7" y="832.42" width="490.8" height="12.157997" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(-404.7 829.275)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="234.624" y="15" xml:space="preserve">C</tspan>
+        </text>
+      </g>
+      <g id="Graphic_112">
+        <text transform="translate(273 797.5)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="26" fill="black" x="6536993e-19" y="25" xml:space="preserve">B</tspan>
+          <tspan font-family="Lucida Grande" font-size="26" fill="black" y="25" xml:space="preserve">↓</tspan>
+        </text>
+      </g>
+      <g id="Graphic_113">
+        <text transform="translate(273 833.974)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="26" fill="black" x="42277293e-20" y="25" xml:space="preserve">C</tspan>
+          <tspan font-family="Lucida Grande" font-size="26" fill="black" y="25" xml:space="preserve">↓</tspan>
+        </text>
+      </g>
+    </g>
+  </g>
+</svg>

docs/rfcs/images/036-bottom-most-gc-compaction/10-btmgc-analysis-3.svg

@@ -0,0 +1,81 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" version="1.1" viewBox="-12 920 809 269" width="809" height="269">
+  <defs/>
+  <g id="10-btmgc-analysis-3" stroke-opacity="1" stroke-dasharray="none" stroke="none" fill="none" fill-opacity="1">
+    <title>10-btmgc-analysis-3</title>
+    <rect fill="white" x="-12" y="920" width="809" height="269"/>
+    <g id="10-btmgc-analysis-3_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_13">
+        <rect x="433.7" y="949" width="63.559346" height="77.5" fill="white"/>
+        <rect x="433.7" y="949" width="63.559346" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(438.7 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8.107673" y="15" xml:space="preserve">1/5 X</tspan>
+        </text>
+      </g>
+      <g id="Graphic_12">
+        <rect x="503.7654" y="949" width="63.559346" height="77.5" fill="white"/>
+        <rect x="503.7654" y="949" width="63.559346" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(508.7654 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8.107673" y="15" xml:space="preserve">1/5 X</tspan>
+        </text>
+      </g>
+      <g id="Graphic_11">
+        <rect x="574.8318" y="949" width="63.559346" height="77.5" fill="white"/>
+        <rect x="574.8318" y="949" width="63.559346" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(579.8318 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8.107673" y="15" xml:space="preserve">1/5 X</tspan>
+        </text>
+      </g>
+      <g id="Graphic_10">
+        <rect x="645.3977" y="949" width="63.559346" height="77.5" fill="white"/>
+        <rect x="645.3977" y="949" width="63.559346" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(650.3977 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8.107673" y="15" xml:space="preserve">1/5 X</tspan>
+        </text>
+      </g>
+      <g id="Line_8">
+        <line x1="92" y1="934.276" x2="795.5" y2="934.276" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_7">
+        <text transform="translate(-6.500003 925.552)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_2">
+        <rect x="113.2" y="1033.92" width="321.3" height="12.157997" fill="white"/>
+        <rect x="113.2" y="1033.92" width="321.3" height="12.157997" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(118.2 1030.775)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="150.762" y="15" xml:space="preserve">X</tspan>
+        </text>
+      </g>
+      <g id="Graphic_17">
+        <path d="M 420.125 1062 L 456.375 1062 L 456.375 1120.5 L 474.5 1120.5 L 438.25 1150.5 L 402 1120.5 L 420.125 1120.5 Z" fill="white"/>
+        <path d="M 420.125 1062 L 456.375 1062 L 456.375 1120.5 L 474.5 1120.5 L 438.25 1150.5 L 402 1120.5 L 420.125 1120.5 Z" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_27">
+        <line x1="93" y1="1164.224" x2="796.5" y2="1164.224" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_26">
+        <text transform="translate(-5.5000034 1155.5)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_25">
+        <rect x="114" y="1173.5" width="641.8" height="12.157997" fill="white"/>
+        <rect x="114" y="1173.5" width="641.8" height="12.157997" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(119 1170.355)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="306.564" y="15" xml:space="preserve">2X</tspan>
+        </text>
+      </g>
+      <g id="Graphic_33">
+        <rect x="715.96355" y="949" width="63.559346" height="77.5" fill="white"/>
+        <rect x="715.96355" y="949" width="63.559346" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(720.96355 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="8.107673" y="15" xml:space="preserve">1/5 X</tspan>
+        </text>
+      </g>
+    </g>
+  </g>
+</svg>

docs/rfcs/images/036-bottom-most-gc-compaction/11-btmgc-analysis-4.svg

@@ -0,0 +1,81 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" version="1.1" viewBox="-12 920 809 269" width="809" height="269">
+  <defs/>
+  <g id="11-btmgc-analysis-4" stroke-opacity="1" stroke-dasharray="none" stroke="none" fill="none" fill-opacity="1">
+    <title>11-btmgc-analysis-4</title>
+    <rect fill="white" x="-12" y="920" width="809" height="269"/>
+    <g id="11-btmgc-analysis-4_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_13">
+        <rect x="113" y="949" width="127" height="77.5" fill="white"/>
+        <rect x="113" y="949" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(118 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="39.084" y="15" xml:space="preserve">1/5 D</tspan>
+        </text>
+      </g>
+      <g id="Graphic_12">
+        <rect x="253" y="949" width="127" height="77.5" fill="white"/>
+        <rect x="253" y="949" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(258 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="39.084" y="15" xml:space="preserve">1/5 D</tspan>
+        </text>
+      </g>
+      <g id="Graphic_11">
+        <rect x="395" y="949" width="127" height="77.5" fill="white"/>
+        <rect x="395" y="949" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(400 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="39.084" y="15" xml:space="preserve">1/5 D</tspan>
+        </text>
+      </g>
+      <g id="Graphic_10">
+        <rect x="536" y="949" width="127" height="77.5" fill="white"/>
+        <rect x="536" y="949" width="127" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(541 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="39.084" y="15" xml:space="preserve">1/5 D</tspan>
+        </text>
+      </g>
+      <g id="Graphic_9">
+        <rect x="677" y="949" width="78" height="77.5" fill="white"/>
+        <rect x="677" y="949" width="78" height="77.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(682 978.526)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="14.584" y="15" xml:space="preserve">1/5 D</tspan>
+        </text>
+      </g>
+      <g id="Line_8">
+        <line x1="92" y1="934.276" x2="795.5" y2="934.276" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_7">
+        <text transform="translate(-6.500003 925.552)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_2">
+        <rect x="113.2" y="1033.92" width="641.8" height="12.157997" fill="white"/>
+        <rect x="113.2" y="1033.92" width="641.8" height="12.157997" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(118.2 1030.775)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="310.268" y="15" xml:space="preserve">D</tspan>
+        </text>
+      </g>
+      <g id="Graphic_17">
+        <path d="M 420.125 1062 L 456.375 1062 L 456.375 1120.5 L 474.5 1120.5 L 438.25 1150.5 L 402 1120.5 L 420.125 1120.5 Z" fill="white"/>
+        <path d="M 420.125 1062 L 456.375 1062 L 456.375 1120.5 L 474.5 1120.5 L 438.25 1150.5 L 402 1120.5 L 420.125 1120.5 Z" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_27">
+        <line x1="93" y1="1164.224" x2="796.5" y2="1164.224" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_26">
+        <text transform="translate(-5.5000034 1155.5)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">GC Horizon</tspan>
+        </text>
+      </g>
+      <g id="Graphic_25">
+        <rect x="114" y="1173.5" width="641.8" height="12.157997" fill="white"/>
+        <rect x="114" y="1173.5" width="641.8" height="12.157997" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(119 1170.355)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="310.268" y="15" xml:space="preserve">D</tspan>
+        </text>
+      </g>
+    </g>
+  </g>
+</svg>

docs/rfcs/images/036-bottom-most-gc-compaction/13-job-split.svg

@@ -0,0 +1,176 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg version="1.1" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns="http://www.w3.org/2000/svg" xmlns:xl="http://www.w3.org/1999/xlink" viewBox="210 271 870 514" width="870" height="514">
+  <defs/>
+  <g id="gc-compaction-split" stroke-dasharray="none" fill-opacity="1" stroke="none" fill="none" stroke-opacity="1">
+    <title>gc-compaction-split</title>
+    <rect fill="white" x="210" y="271" width="870" height="514"/>
+    <g id="gc-compaction-split_Layer_1">
+      <title>Layer 1</title>
+      <g id="Graphic_12">
+        <rect x="241" y="272" width="213" height="50.5" fill="white"/>
+        <rect x="241" y="272" width="213" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_11">
+        <rect x="468.72266" y="272" width="213" height="50.5" fill="white"/>
+        <rect x="468.72266" y="272" width="213" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_10">
+        <rect x="695.72266" y="272" width="213" height="50.5" fill="white"/>
+        <rect x="695.72266" y="272" width="213" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_9">
+        <rect x="241" y="337.3711" width="303.5" height="50.5" fill="white"/>
+        <rect x="241" y="337.3711" width="303.5" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_8">
+        <rect x="556.2617" y="337.3711" width="352.46094" height="50.5" fill="white"/>
+        <rect x="556.2617" y="337.3711" width="352.46094" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_7">
+        <rect x="241" y="402.7422" width="667.72266" height="50.5" fill="white"/>
+        <rect x="241" y="402.7422" width="667.72266" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_6">
+        <line x1="211" y1="355.5" x2="947.4961" y2="355.5" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_5">
+        <text transform="translate(952.4961 346.776)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">branch point</tspan>
+        </text>
+      </g>
+      <g id="Line_4">
+        <line x1="212" y1="438.5182" x2="948.4961" y2="438.5182" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_3">
+        <text transform="translate(953.4961 429.7942)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">last branch point</tspan>
+        </text>
+      </g>
+      <g id="Graphic_13">
+        <rect x="241" y="272" width="127.99101" height="181.24219" fill="#3a8eed" fill-opacity=".5"/>
+        <text transform="translate(246 353.3971)" fill="white">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="white" x="38.835502" y="15" xml:space="preserve">Job 1</tspan>
+        </text>
+      </g>
+      <g id="Graphic_57">
+        <rect x="359" y="647.96484" width="551.72266" height="50.5" fill="white"/>
+        <rect x="359" y="647.96484" width="551.72266" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_54">
+        <rect x="359" y="517.22266" width="96" height="50.5" fill="white"/>
+        <rect x="359" y="517.22266" width="96" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_53">
+        <rect x="469.72266" y="517.22266" width="213" height="50.5" fill="white"/>
+        <rect x="469.72266" y="517.22266" width="213" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_52">
+        <rect x="696.72266" y="517.22266" width="213" height="50.5" fill="white"/>
+        <rect x="696.72266" y="517.22266" width="213" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_51">
+        <rect x="359" y="582.59375" width="186.5" height="50.5" fill="white"/>
+        <rect x="359" y="582.59375" width="186.5" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_50">
+        <rect x="557.2617" y="582.59375" width="352.46094" height="50.5" fill="white"/>
+        <rect x="557.2617" y="582.59375" width="352.46094" height="50.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Line_49">
+        <line x1="212" y1="600.72266" x2="948.4961" y2="600.72266" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_48">
+        <text transform="translate(953.4961 591.99866)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">branch point</tspan>
+        </text>
+      </g>
+      <g id="Line_47">
+        <line x1="213" y1="683.74084" x2="949.4961" y2="683.74084" stroke="#7f8080" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_46">
+        <text transform="translate(954.4961 675.01685)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="0" y="15" xml:space="preserve">last branch point</tspan>
+        </text>
+      </g>
+      <g id="Graphic_63">
+        <rect x="376.72525" y="272" width="127.99101" height="181.24219" fill="#3a8eed" fill-opacity=".5"/>
+        <text transform="translate(381.72525 353.3971)" fill="white">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="white" x="38.835502" y="15" xml:space="preserve">Job 2</tspan>
+        </text>
+      </g>
+      <g id="Graphic_64">
+        <rect x="511.39405" y="272" width="127.99101" height="181.24219" fill="#3a8eed" fill-opacity=".5"/>
+        <text transform="translate(516.39405 353.3971)" fill="white">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="white" x="38.835502" y="15" xml:space="preserve">Job 3</tspan>
+        </text>
+      </g>
+      <g id="Graphic_65">
+        <rect x="646.06285" y="272" width="127.99101" height="181.24219" fill="#3a8eed" fill-opacity=".5"/>
+        <text transform="translate(651.06285 353.3971)" fill="white">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="white" x="38.835502" y="15" xml:space="preserve">Job 4</tspan>
+        </text>
+      </g>
+      <g id="Graphic_66">
+        <rect x="780.73165" y="272" width="127.99101" height="181.24219" fill="#3a8eed" fill-opacity=".5"/>
+        <text transform="translate(785.73165 353.3971)" fill="white">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="white" x="38.835502" y="15" xml:space="preserve">Job 5</tspan>
+        </text>
+      </g>
+      <g id="Graphic_56">
+        <rect x="243.5" y="517.22266" width="125.49101" height="181.24219" fill="#ccc"/>
+        <rect x="243.5" y="517.22266" width="125.49101" height="181.24219" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_55">
+        <rect x="243.5" y="673.46484" width="125.49101" height="17.5" fill="#6b7ca5"/>
+        <rect x="243.5" y="673.46484" width="125.49101" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_68">
+        <rect x="379.22525" y="517.22266" width="125.49101" height="181.24219" fill="#ccc"/>
+        <rect x="379.22525" y="517.22266" width="125.49101" height="181.24219" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_67">
+        <rect x="379.22525" y="673.46484" width="125.49101" height="17.5" fill="#6b7ca5"/>
+        <rect x="379.22525" y="673.46484" width="125.49101" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_70">
+        <rect x="514.22525" y="517.22266" width="125.49101" height="181.24219" fill="#ccc"/>
+        <rect x="514.22525" y="517.22266" width="125.49101" height="181.24219" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_69">
+        <rect x="514.22525" y="673.46484" width="125.49101" height="17.5" fill="#6b7ca5"/>
+        <rect x="514.22525" y="673.46484" width="125.49101" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_72">
+        <rect x="649.22525" y="517.22266" width="125.49101" height="181.24219" fill="#ccc"/>
+        <rect x="649.22525" y="517.22266" width="125.49101" height="181.24219" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_71">
+        <rect x="649.22525" y="673.46484" width="125.49101" height="17.5" fill="#6b7ca5"/>
+        <rect x="649.22525" y="673.46484" width="125.49101" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_74">
+        <rect x="785.23165" y="517.22266" width="125.49101" height="181.24219" fill="#ccc"/>
+        <rect x="785.23165" y="517.22266" width="125.49101" height="181.24219" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_73">
+        <rect x="785.23165" y="673.46484" width="125.49101" height="17.5" fill="#6b7ca5"/>
+        <rect x="785.23165" y="673.46484" width="125.49101" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+      </g>
+      <g id="Graphic_78">
+        <rect x="241" y="731.3359" width="125.49101" height="27.26953" fill="#ccc"/>
+        <rect x="241" y="731.3359" width="125.49101" height="27.26953" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(246 735.7467)" fill="black">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="black" x="17.297502" y="15" xml:space="preserve">Delta Layer</tspan>
+        </text>
+      </g>
+      <g id="Graphic_79">
+        <rect x="241" y="766.759" width="125.49101" height="17.5" fill="#6b7ca5"/>
+        <rect x="241" y="766.759" width="125.49101" height="17.5" stroke="black" stroke-linecap="round" stroke-linejoin="round" stroke-width="1"/>
+        <text transform="translate(246 766.285)" fill="white">
+          <tspan font-family="Helvetica Neue" font-size="16" fill="white" x="13.737502" y="15" xml:space="preserve">Image Layer</tspan>
+        </text>
+      </g>
+    </g>
+  </g>
+</svg>

libs/neon-shmem/Cargo.toml

@@ -0,0 +1,13 @@
+[package]
+name = "neon-shmem"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+thiserror.workspace = true
+nix.workspace=true
+workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+
+[target.'cfg(target_os = "macos")'.dependencies]
+tempfile = "3.14.0"

libs/neon-shmem/src/lib.rs

@@ -0,0 +1,418 @@
+//! Shared memory utilities for neon communicator
+
+use std::num::NonZeroUsize;
+use std::os::fd::{AsFd, BorrowedFd, OwnedFd};
+use std::ptr::NonNull;
+use std::sync::atomic::{AtomicUsize, Ordering};
+
+use nix::errno::Errno;
+use nix::sys::mman::MapFlags;
+use nix::sys::mman::ProtFlags;
+use nix::sys::mman::mmap as nix_mmap;
+use nix::sys::mman::munmap as nix_munmap;
+use nix::unistd::ftruncate as nix_ftruncate;
+
+/// ShmemHandle represents a shared memory area that can be shared by processes over fork().
+/// Unlike shared memory allocated by Postgres, this area is resizable, up to 'max_size' that's
+/// specified at creation.
+///
+/// The area is backed by an anonymous file created with memfd_create(). The full address space for
+/// 'max_size' is reserved up-front with mmap(), but whenever you call [`ShmemHandle::set_size`],
+/// the underlying file is resized. Do not access the area beyond the current size. Currently, that
+/// will cause the file to be expanded, but we might use mprotect() etc. to enforce that in the
+/// future.
+pub struct ShmemHandle {
+    /// memfd file descriptor
+    fd: OwnedFd,
+
+    max_size: usize,
+
+    // Pointer to the beginning of the shared memory area. The header is stored there.
+    shared_ptr: NonNull<SharedStruct>,
+
+    // Pointer to the beginning of the user data
+    pub data_ptr: NonNull<u8>,
+}
+
+/// This is stored at the beginning in the shared memory area.
+struct SharedStruct {
+    max_size: usize,
+
+    /// Current size of the backing file. The high-order bit is used for the RESIZE_IN_PROGRESS flag
+    current_size: AtomicUsize,
+}
+
+const RESIZE_IN_PROGRESS: usize = 1 << 63;
+
+const HEADER_SIZE: usize = std::mem::size_of::<SharedStruct>();
+
+/// Error type returned by the ShmemHandle functions.
+#[derive(thiserror::Error, Debug)]
+#[error("{msg}: {errno}")]
+pub struct Error {
+    pub msg: String,
+    pub errno: Errno,
+}
+
+impl Error {
+    fn new(msg: &str, errno: Errno) -> Error {
+        Error {
+            msg: msg.to_string(),
+            errno,
+        }
+    }
+}
+
+impl ShmemHandle {
+    /// Create a new shared memory area. To communicate between processes, the processes need to be
+    /// fork()'d after calling this, so that the ShmemHandle is inherited by all processes.
+    ///
+    /// If the ShmemHandle is dropped, the memory is unmapped from the current process. Other
+    /// processes can continue using it, however.
+    pub fn new(name: &str, initial_size: usize, max_size: usize) -> Result<ShmemHandle, Error> {
+        // create the backing anonymous file.
+        let fd = create_backing_file(name)?;
+
+        Self::new_with_fd(fd, initial_size, max_size)
+    }
+
+    fn new_with_fd(
+        fd: OwnedFd,
+        initial_size: usize,
+        max_size: usize,
+    ) -> Result<ShmemHandle, Error> {
+        // We reserve the high-order bit for the RESIZE_IN_PROGRESS flag, and the actual size
+        // is a little larger than this because of the SharedStruct header. Make the upper limit
+        // somewhat smaller than that, because with anything close to that, you'll run out of
+        // memory anyway.
+        if max_size >= 1 << 48 {
+            panic!("max size {} too large", max_size);
+        }
+        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();
+        }
+    }
+}

libs/pageserver_api/src/config.rs

@@ -305,6 +305,7 @@ impl From<OtelExporterProtocol> for tracing_utils::Protocol {
 pub struct TimelineImportConfig {
     pub import_job_concurrency: NonZeroUsize,
     pub import_job_soft_size_limit: NonZeroUsize,
+    pub import_job_checkpoint_threshold: NonZeroUsize,
 }
 
 pub mod statvfs {
@@ -639,23 +640,15 @@ impl Default for ConfigToml {
             tenant_config: TenantConfigToml::default(),
             no_sync: None,
             wal_receiver_protocol: DEFAULT_WAL_RECEIVER_PROTOCOL,
-            page_service_pipelining: if !cfg!(test) {
-                PageServicePipeliningConfig::Serial
-            } else {
-                // Do not turn this into the default until scattered reads have been
-                // validated and rolled-out fully.
-                PageServicePipeliningConfig::Pipelined(PageServicePipeliningConfigPipelined {
+            page_service_pipelining: PageServicePipeliningConfig::Pipelined(
+                PageServicePipeliningConfigPipelined {
                     max_batch_size: NonZeroUsize::new(32).unwrap(),
                     execution: PageServiceProtocolPipelinedExecutionStrategy::ConcurrentFutures,
                     batching: PageServiceProtocolPipelinedBatchingStrategy::ScatteredLsn,
-                })
-            },
-            get_vectored_concurrent_io: if !cfg!(test) {
-                GetVectoredConcurrentIo::Sequential
-            } else {
-                GetVectoredConcurrentIo::SidecarTask
-            },
-            enable_read_path_debugging: if cfg!(test) || cfg!(feature = "testing") {
+                },
+            ),
+            get_vectored_concurrent_io: GetVectoredConcurrentIo::SidecarTask,
+            enable_read_path_debugging: if cfg!(feature = "testing") {
                 Some(true)
             } else {
                 None
@@ -669,6 +662,7 @@ impl Default for ConfigToml {
             timeline_import_config: TimelineImportConfig {
                 import_job_concurrency: NonZeroUsize::new(128).unwrap(),
                 import_job_soft_size_limit: NonZeroUsize::new(1024 * 1024 * 1024).unwrap(),
+                import_job_checkpoint_threshold: NonZeroUsize::new(128).unwrap(),
             },
         }
     }

libs/pageserver_api/src/key.rs

@@ -910,6 +910,11 @@ impl Key {
         self.field1 == 0x00 && self.field4 != 0 && self.field6 != 0xffffffff
     }
 
+    #[inline(always)]
+    pub fn is_rel_block_of_rel(&self, rel: Oid) -> bool {
+        self.is_rel_block_key() && self.field4 == rel
+    }
+
     #[inline(always)]
     pub fn is_rel_dir_key(&self) -> bool {
         self.field1 == 0x00

libs/pageserver_api/src/models.rs

@@ -336,14 +336,30 @@ impl TimelineCreateRequest {
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Eq)]
 pub enum ShardImportStatus {
-    InProgress,
+    InProgress(Option<ShardImportProgress>),
     Done,
     Error(String),
 }
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Eq)]
+pub enum ShardImportProgress {
+    V1(ShardImportProgressV1),
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Eq)]
+pub struct ShardImportProgressV1 {
+    /// Total number of jobs in the import plan
+    pub jobs: usize,
+    /// Number of jobs completed
+    pub completed: usize,
+    /// Hash of the plan
+    pub import_plan_hash: u64,
+}
+
 impl ShardImportStatus {
     pub fn is_terminal(&self) -> bool {
         match self {
-            ShardImportStatus::InProgress => false,
+            ShardImportStatus::InProgress(_) => false,
             ShardImportStatus::Done | ShardImportStatus::Error(_) => true,
         }
     }
@@ -1803,7 +1819,6 @@ pub struct TopTenantShardsResponse {
 }
 
 pub mod virtual_file {
-    use std::sync::LazyLock;
 
     #[derive(
         Copy,
@@ -1851,15 +1866,7 @@ pub mod virtual_file {
 
     impl IoMode {
         pub fn preferred() -> Self {
-            // The default behavior when running Rust unit tests without any further
-            // flags is to use the newest behavior (DirectRw).
-            // The CI uses the environment variable to unit tests for all different modes.
-            // NB: the Python regression & perf tests have their own defaults management
-            // that writes pageserver.toml; they do not use this variable.
-            static ENV_OVERRIDE: LazyLock<Option<IoMode>> = LazyLock::new(|| {
-                utils::env::var_serde_json_string("NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IO_MODE")
-            });
-            ENV_OVERRIDE.unwrap_or(IoMode::DirectRw)
+            IoMode::DirectRw
         }
     }
 

libs/pageserver_api/src/upcall_api.rs

@@ -4,6 +4,7 @@
 //! See docs/rfcs/025-generation-numbers.md
 
 use serde::{Deserialize, Serialize};
+use utils::generation::Generation;
 use utils::id::{NodeId, TimelineId};
 
 use crate::controller_api::NodeRegisterRequest;
@@ -63,9 +64,17 @@ pub struct ValidateResponseTenant {
     pub valid: bool,
 }
 
+#[derive(Serialize, Deserialize)]
+pub struct TimelineImportStatusRequest {
+    pub tenant_shard_id: TenantShardId,
+    pub timeline_id: TimelineId,
+    pub generation: Generation,
+}
+
 #[derive(Serialize, Deserialize)]
 pub struct PutTimelineImportStatusRequest {
     pub tenant_shard_id: TenantShardId,
     pub timeline_id: TimelineId,
     pub status: ShardImportStatus,
+    pub generation: Generation,
 }

libs/pageserver_api/src/value.rs

@@ -36,6 +36,24 @@ impl Value {
             Value::WalRecord(rec) => rec.will_init(),
         }
     }
+
+    #[inline(always)]
+    pub fn estimated_size(&self) -> usize {
+        match self {
+            Value::Image(image) => image.len(),
+            Value::WalRecord(NeonWalRecord::AuxFile {
+                content: Some(content),
+                ..
+            }) => content.len(),
+            Value::WalRecord(NeonWalRecord::Postgres { rec, .. }) => rec.len(),
+            Value::WalRecord(NeonWalRecord::ClogSetAborted { xids }) => xids.len() * 4,
+            Value::WalRecord(NeonWalRecord::ClogSetCommitted { xids, .. }) => xids.len() * 4,
+            Value::WalRecord(NeonWalRecord::MultixactMembersCreate { members, .. }) => {
+                members.len() * 8
+            }
+            _ => 8192, /* use image size as the estimation */
+        }
+    }
 }
 
 #[derive(Debug, PartialEq)]

libs/posthog_client_lite/Cargo.toml

@@ -0,0 +1,14 @@
+[package]
+name = "posthog_client_lite"
+version = "0.1.0"
+edition = "2024"
+license.workspace = true
+
+[dependencies]
+anyhow.workspace = true
+reqwest.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+sha2.workspace = true
+workspace_hack.workspace = true
+thiserror.workspace = true

libs/posthog_client_lite/src/lib.rs

@@ -0,0 +1,634 @@
+//! A lite version of the PostHog client that only supports local evaluation of feature flags.
+
+use std::collections::HashMap;
+
+use serde::{Deserialize, Serialize};
+use serde_json::json;
+use sha2::Digest;
+
+#[derive(Debug, thiserror::Error)]
+pub enum PostHogEvaluationError {
+    /// The feature flag is not available, for example, because the local evaluation data is not populated yet.
+    #[error("Feature flag not available: {0}")]
+    NotAvailable(String),
+    #[error("No condition group is matched")]
+    NoConditionGroupMatched,
+    /// Real errors, e.g., the rollout percentage does not add up to 100.
+    #[error("Failed to evaluate feature flag: {0}")]
+    Internal(String),
+}
+
+#[derive(Deserialize)]
+pub struct LocalEvaluationResponse {
+    #[allow(dead_code)]
+    flags: Vec<LocalEvaluationFlag>,
+}
+
+#[derive(Deserialize)]
+pub struct LocalEvaluationFlag {
+    key: String,
+    filters: LocalEvaluationFlagFilters,
+    active: bool,
+}
+
+#[derive(Deserialize)]
+pub struct LocalEvaluationFlagFilters {
+    groups: Vec<LocalEvaluationFlagFilterGroup>,
+    multivariate: LocalEvaluationFlagMultivariate,
+}
+
+#[derive(Deserialize)]
+pub struct LocalEvaluationFlagFilterGroup {
+    variant: Option<String>,
+    properties: Option<Vec<LocalEvaluationFlagFilterProperty>>,
+    rollout_percentage: i64,
+}
+
+#[derive(Deserialize)]
+pub struct LocalEvaluationFlagFilterProperty {
+    key: String,
+    value: PostHogFlagFilterPropertyValue,
+    operator: String,
+}
+
+#[derive(Debug, Serialize, Deserialize)]
+#[serde(untagged)]
+pub enum PostHogFlagFilterPropertyValue {
+    String(String),
+    Number(f64),
+    Boolean(bool),
+    List(Vec<String>),
+}
+
+#[derive(Deserialize)]
+pub struct LocalEvaluationFlagMultivariate {
+    variants: Vec<LocalEvaluationFlagMultivariateVariant>,
+}
+
+#[derive(Deserialize)]
+pub struct LocalEvaluationFlagMultivariateVariant {
+    key: String,
+    rollout_percentage: i64,
+}
+
+pub struct FeatureStore {
+    flags: HashMap<String, LocalEvaluationFlag>,
+}
+
+impl Default for FeatureStore {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+enum GroupEvaluationResult {
+    MatchedAndOverride(String),
+    MatchedAndEvaluate,
+    Unmatched,
+}
+
+impl FeatureStore {
+    pub fn new() -> Self {
+        Self {
+            flags: HashMap::new(),
+        }
+    }
+
+    pub fn set_flags(&mut self, flags: Vec<LocalEvaluationFlag>) {
+        self.flags.clear();
+        for flag in flags {
+            self.flags.insert(flag.key.clone(), flag);
+        }
+    }
+
+    /// Generate a consistent hash for a user ID (e.g., tenant ID).
+    ///
+    /// The implementation is different from PostHog SDK. In PostHog SDK, it is sha1 of `user_id.distinct_id.salt`.
+    /// However, as we do not upload all of our tenant IDs to PostHog, we do not have the PostHog distinct_id for a
+    /// tenant. Therefore, the way we compute it is sha256 of `user_id.feature_id.salt`.
+    fn consistent_hash(user_id: &str, flag_key: &str, salt: &str) -> f64 {
+        let mut hasher = sha2::Sha256::new();
+        hasher.update(user_id);
+        hasher.update(".");
+        hasher.update(flag_key);
+        hasher.update(".");
+        hasher.update(salt);
+        let hash = hasher.finalize();
+        let hash_int = u64::from_le_bytes(hash[..8].try_into().unwrap());
+        hash_int as f64 / u64::MAX as f64
+    }
+
+    /// Evaluate a condition. Returns an error if the condition cannot be evaluated due to parsing error or missing
+    /// property.
+    fn evaluate_condition(
+        &self,
+        operator: &str,
+        provided: &PostHogFlagFilterPropertyValue,
+        requested: &PostHogFlagFilterPropertyValue,
+    ) -> Result<bool, PostHogEvaluationError> {
+        match operator {
+            "exact" => {
+                let PostHogFlagFilterPropertyValue::String(provided) = provided else {
+                    // Left should be a string
+                    return Err(PostHogEvaluationError::Internal(format!(
+                        "The left side of the condition is not a string: {:?}",
+                        provided
+                    )));
+                };
+                let PostHogFlagFilterPropertyValue::List(requested) = requested else {
+                    // Right should be a list of string
+                    return Err(PostHogEvaluationError::Internal(format!(
+                        "The right side of the condition is not a list: {:?}",
+                        requested
+                    )));
+                };
+                Ok(requested.contains(provided))
+            }
+            "lt" | "gt" => {
+                let PostHogFlagFilterPropertyValue::String(requested) = requested else {
+                    // Right should be a string
+                    return Err(PostHogEvaluationError::Internal(format!(
+                        "The right side of the condition is not a string: {:?}",
+                        requested
+                    )));
+                };
+                let Ok(requested) = requested.parse::<f64>() else {
+                    return Err(PostHogEvaluationError::Internal(format!(
+                        "Can not parse the right side of the condition as a number: {:?}",
+                        requested
+                    )));
+                };
+                // Left can either be a number or a string
+                let provided = match provided {
+                    PostHogFlagFilterPropertyValue::Number(provided) => *provided,
+                    PostHogFlagFilterPropertyValue::String(provided) => {
+                        let Ok(provided) = provided.parse::<f64>() else {
+                            return Err(PostHogEvaluationError::Internal(format!(
+                                "Can not parse the left side of the condition as a number: {:?}",
+                                provided
+                            )));
+                        };
+                        provided
+                    }
+                    _ => {
+                        return Err(PostHogEvaluationError::Internal(format!(
+                            "The left side of the condition is not a number or a string: {:?}",
+                            provided
+                        )));
+                    }
+                };
+                match operator {
+                    "lt" => Ok(provided < requested),
+                    "gt" => Ok(provided > requested),
+                    op => Err(PostHogEvaluationError::Internal(format!(
+                        "Unsupported operator: {}",
+                        op
+                    ))),
+                }
+            }
+            _ => Err(PostHogEvaluationError::Internal(format!(
+                "Unsupported operator: {}",
+                operator
+            ))),
+        }
+    }
+
+    /// Evaluate a percentage.
+    fn evaluate_percentage(&self, mapped_user_id: f64, percentage: i64) -> bool {
+        mapped_user_id <= percentage as f64 / 100.0
+    }
+
+    /// Evaluate a filter group for a feature flag. Returns an error if there are errors during the evaluation.
+    ///
+    /// Return values:
+    /// Ok(GroupEvaluationResult::MatchedAndOverride(variant)): matched and evaluated to this value
+    /// Ok(GroupEvaluationResult::MatchedAndEvaluate): condition matched but no variant override, use the global rollout percentage
+    /// Ok(GroupEvaluationResult::Unmatched): condition unmatched
+    fn evaluate_group(
+        &self,
+        group: &LocalEvaluationFlagFilterGroup,
+        hash_on_group_rollout_percentage: f64,
+        provided_properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
+    ) -> Result<GroupEvaluationResult, PostHogEvaluationError> {
+        if let Some(ref properties) = group.properties {
+            for property in properties {
+                if let Some(value) = provided_properties.get(&property.key) {
+                    // The user provided the property value
+                    if !self.evaluate_condition(
+                        property.operator.as_ref(),
+                        value,
+                        &property.value,
+                    )? {
+                        return Ok(GroupEvaluationResult::Unmatched);
+                    }
+                } else {
+                    // We cannot evaluate, the property is not available
+                    return Err(PostHogEvaluationError::NotAvailable(format!(
+                        "The required property in the condition is not available: {}",
+                        property.key
+                    )));
+                }
+            }
+        }
+
+        // The group has no condition matchers or we matched the properties
+        if self.evaluate_percentage(hash_on_group_rollout_percentage, group.rollout_percentage) {
+            if let Some(ref variant_override) = group.variant {
+                Ok(GroupEvaluationResult::MatchedAndOverride(
+                    variant_override.clone(),
+                ))
+            } else {
+                Ok(GroupEvaluationResult::MatchedAndEvaluate)
+            }
+        } else {
+            Ok(GroupEvaluationResult::Unmatched)
+        }
+    }
+
+    /// Evaluate a multivariate feature flag. Returns `None` if the flag is not available or if there are errors
+    /// during the evaluation.
+    ///
+    /// The parsing logic is as follows:
+    ///
+    /// * Match each filter group.
+    ///   - If a group is matched, it will first determine whether the user is in the range of the group's rollout
+    ///     percentage. We will generate a consistent hash for the user ID on the group rollout percentage. This hash
+    ///     is shared across all groups.
+    ///   - If the hash falls within the group's rollout percentage, return the variant if it's overridden, or
+    ///   - Evaluate the variant using the global config and the global rollout percentage.
+    /// * Otherwise, continue with the next group until all groups are evaluated and no group is within the
+    ///   rollout percentage.
+    /// * If there are no matching groups, return an error.
+    ///
+    /// Example: we have a multivariate flag with 3 groups of the configured global rollout percentage: A (10%), B (20%), C (70%).
+    /// There is a single group with a condition that has a rollout percentage of 10% and it does not have a variant override.
+    /// Then, we will have 1% of the users evaluated to A, 2% to B, and 7% to C.
+    pub fn evaluate_multivariate(
+        &self,
+        flag_key: &str,
+        user_id: &str,
+    ) -> Result<String, PostHogEvaluationError> {
+        let hash_on_global_rollout_percentage =
+            Self::consistent_hash(user_id, flag_key, "multivariate");
+        let hash_on_group_rollout_percentage =
+            Self::consistent_hash(user_id, flag_key, "within_group");
+        self.evaluate_multivariate_inner(
+            flag_key,
+            hash_on_global_rollout_percentage,
+            hash_on_group_rollout_percentage,
+            &HashMap::new(),
+        )
+    }
+
+    /// Evaluate a multivariate feature flag. Note that we directly take the mapped user ID
+    /// (a consistent hash ranging from 0 to 1) so that it is easier to use it in the tests
+    /// and avoid duplicate computations.
+    ///
+    /// Use a different consistent hash for evaluating the group rollout percentage.
+    /// The behavior: if the condition is set to rolling out to 10% of the users, and
+    /// we set the variant A to 20% in the global config, then 2% of the total users will
+    /// be evaluated to variant A.
+    ///
+    /// Note that the hash to determine group rollout percentage is shared across all groups. So if we have two
+    /// exactly-the-same conditions with 10% and 20% rollout percentage respectively, a total of 20% of the users
+    /// will be evaluated (versus 30% if group evaluation is done independently).
+    pub(crate) fn evaluate_multivariate_inner(
+        &self,
+        flag_key: &str,
+        hash_on_global_rollout_percentage: f64,
+        hash_on_group_rollout_percentage: f64,
+        properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
+    ) -> Result<String, PostHogEvaluationError> {
+        if let Some(flag_config) = self.flags.get(flag_key) {
+            if !flag_config.active {
+                return Err(PostHogEvaluationError::NotAvailable(format!(
+                    "The feature flag is not active: {}",
+                    flag_key
+                )));
+            }
+            // TODO: sort the groups so that variant overrides always get evaluated first and it follows the PostHog
+            // Python SDK behavior; for now we do not configure conditions without variant overrides in Neon so it
+            // does not matter.
+            for group in &flag_config.filters.groups {
+                match self.evaluate_group(group, hash_on_group_rollout_percentage, properties)? {
+                    GroupEvaluationResult::MatchedAndOverride(variant) => return Ok(variant),
+                    GroupEvaluationResult::MatchedAndEvaluate => {
+                        let mut percentage = 0;
+                        for variant in &flag_config.filters.multivariate.variants {
+                            percentage += variant.rollout_percentage;
+                            if self
+                                .evaluate_percentage(hash_on_global_rollout_percentage, percentage)
+                            {
+                                return Ok(variant.key.clone());
+                            }
+                        }
+                        // This should not happen because the rollout percentage always adds up to 100, but just in case that PostHog
+                        // returned invalid spec, we return an error.
+                        return Err(PostHogEvaluationError::Internal(format!(
+                            "Rollout percentage does not add up to 100: {}",
+                            flag_key
+                        )));
+                    }
+                    GroupEvaluationResult::Unmatched => continue,
+                }
+            }
+            // If no group is matched, the feature is not available, and up to the caller to decide what to do.
+            Err(PostHogEvaluationError::NoConditionGroupMatched)
+        } else {
+            // The feature flag is not available yet
+            Err(PostHogEvaluationError::NotAvailable(format!(
+                "Not found in the local evaluation spec: {}",
+                flag_key
+            )))
+        }
+    }
+}
+
+/// A lite PostHog client.
+///
+/// At the point of writing this code, PostHog does not have a functional Rust client with feature flag support.
+/// This is a lite version that only supports local evaluation of feature flags and only supports those JSON specs
+/// that will be used within Neon.
+///
+/// PostHog is designed as a browser-server system: the browser (client) side uses the client key and is exposed
+/// to the end users; the server side uses a server key and is not exposed to the end users. The client and the
+/// server has different API keys and provide a different set of APIs. In Neon, we only have the server (that is
+/// pageserver), and it will use both the client API and the server API. So we need to store two API keys within
+/// our PostHog client.
+///
+/// The server API is used to fetch the feature flag specs. The client API is used to capture events in case we
+/// want to report the feature flag usage back to PostHog. The current plan is to use PostHog only as an UI to
+/// configure feature flags so it is very likely that the client API will not be used.
+pub struct PostHogClient {
+    /// The server API key.
+    server_api_key: String,
+    /// The client API key.
+    client_api_key: String,
+    /// The project ID.
+    project_id: String,
+    /// The private API URL.
+    private_api_url: String,
+    /// The public API URL.
+    public_api_url: String,
+    /// The HTTP client.
+    client: reqwest::Client,
+}
+
+impl PostHogClient {
+    pub fn new(
+        server_api_key: String,
+        client_api_key: String,
+        project_id: String,
+        private_api_url: String,
+        public_api_url: String,
+    ) -> Self {
+        let client = reqwest::Client::new();
+        Self {
+            server_api_key,
+            client_api_key,
+            project_id,
+            private_api_url,
+            public_api_url,
+            client,
+        }
+    }
+
+    pub fn new_with_us_region(
+        server_api_key: String,
+        client_api_key: String,
+        project_id: String,
+    ) -> Self {
+        Self::new(
+            server_api_key,
+            client_api_key,
+            project_id,
+            "https://us.posthog.com".to_string(),
+            "https://us.i.posthog.com".to_string(),
+        )
+    }
+
+    /// Fetch the feature flag specs from the server.
+    ///
+    /// This is unfortunately an undocumented API at:
+    /// - <https://posthog.com/docs/api/feature-flags#get-api-projects-project_id-feature_flags-local_evaluation>
+    /// - <https://posthog.com/docs/feature-flags/local-evaluation>
+    ///
+    /// The handling logic in [`FeatureStore`] mostly follows the Python API implementation.
+    /// See `_compute_flag_locally` in <https://github.com/PostHog/posthog-python/blob/master/posthog/client.py>
+    pub async fn get_feature_flags_local_evaluation(
+        &self,
+    ) -> anyhow::Result<LocalEvaluationResponse> {
+        // BASE_URL/api/projects/:project_id/feature_flags/local_evaluation
+        // with bearer token of self.server_api_key
+        let url = format!(
+            "{}/api/projects/{}/feature_flags/local_evaluation",
+            self.private_api_url, self.project_id
+        );
+        let response = self
+            .client
+            .get(url)
+            .bearer_auth(&self.server_api_key)
+            .send()
+            .await?;
+        let body = response.text().await?;
+        Ok(serde_json::from_str(&body)?)
+    }
+
+    /// Capture an event. This will only be used to report the feature flag usage back to PostHog, though
+    /// it also support a lot of other functionalities.
+    ///
+    /// <https://posthog.com/docs/api/capture>
+    pub async fn capture_event(
+        &self,
+        event: &str,
+        distinct_id: &str,
+        properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
+    ) -> anyhow::Result<()> {
+        // PUBLIC_URL/capture/
+        // with bearer token of self.client_api_key
+        let url = format!("{}/capture/", self.public_api_url);
+        self.client
+            .post(url)
+            .body(serde_json::to_string(&json!({
+                "api_key": self.client_api_key,
+                "distinct_id": distinct_id,
+                "event": event,
+                "properties": properties,
+            }))?)
+            .send()
+            .await?;
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn data() -> &'static str {
+        r#"{
+            "flags": [
+                {
+                    "id": 132794,
+                    "team_id": 152860,
+                    "name": "",
+                    "key": "gc-compaction",
+                    "filters": {
+                        "groups": [
+                            {
+                                "variant": "enabled-stage-2",
+                                "properties": [
+                                    {
+                                        "key": "plan_type",
+                                        "type": "person",
+                                        "value": [
+                                            "free"
+                                        ],
+                                        "operator": "exact"
+                                    },
+                                    {
+                                        "key": "pageserver_remote_size",
+                                        "type": "person",
+                                        "value": "10000000",
+                                        "operator": "lt"
+                                    }
+                                ],
+                                "rollout_percentage": 50
+                            },
+                            {
+                                "properties": [
+                                    {
+                                        "key": "plan_type",
+                                        "type": "person",
+                                        "value": [
+                                            "free"
+                                        ],
+                                        "operator": "exact"
+                                    },
+                                    {
+                                        "key": "pageserver_remote_size",
+                                        "type": "person",
+                                        "value": "10000000",
+                                        "operator": "lt"
+                                    }
+                                ],
+                                "rollout_percentage": 80
+                            }
+                        ],
+                        "payloads": {},
+                        "multivariate": {
+                            "variants": [
+                                {
+                                    "key": "disabled",
+                                    "name": "",
+                                    "rollout_percentage": 90
+                                },
+                                {
+                                    "key": "enabled-stage-1",
+                                    "name": "",
+                                    "rollout_percentage": 10
+                                },
+                                {
+                                    "key": "enabled-stage-2",
+                                    "name": "",
+                                    "rollout_percentage": 0
+                                },
+                                {
+                                    "key": "enabled-stage-3",
+                                    "name": "",
+                                    "rollout_percentage": 0
+                                },
+                                {
+                                    "key": "enabled",
+                                    "name": "",
+                                    "rollout_percentage": 0
+                                }
+                            ]
+                        }
+                    },
+                    "deleted": false,
+                    "active": true,
+                    "ensure_experience_continuity": false,
+                    "has_encrypted_payloads": false,
+                    "version": 6
+                }
+            ],
+            "group_type_mapping": {},
+            "cohorts": {}
+        }"#
+    }
+
+    #[test]
+    fn parse_local_evaluation() {
+        let data = data();
+        let _: LocalEvaluationResponse = serde_json::from_str(data).unwrap();
+    }
+
+    #[test]
+    fn evaluate_multivariate() {
+        let mut store = FeatureStore::new();
+        let response: LocalEvaluationResponse = serde_json::from_str(data()).unwrap();
+        store.set_flags(response.flags);
+
+        // This lacks the required properties and cannot be evaluated.
+        let variant =
+            store.evaluate_multivariate_inner("gc-compaction", 1.00, 0.40, &HashMap::new());
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NotAvailable(_))
+        ),);
+
+        let properties_unmatched = HashMap::from([
+            (
+                "plan_type".to_string(),
+                PostHogFlagFilterPropertyValue::String("paid".to_string()),
+            ),
+            (
+                "pageserver_remote_size".to_string(),
+                PostHogFlagFilterPropertyValue::Number(1000.0),
+            ),
+        ]);
+
+        // This does not match any group so there will be an error.
+        let variant =
+            store.evaluate_multivariate_inner("gc-compaction", 1.00, 0.40, &properties_unmatched);
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NoConditionGroupMatched)
+        ),);
+        let variant =
+            store.evaluate_multivariate_inner("gc-compaction", 0.80, 0.80, &properties_unmatched);
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NoConditionGroupMatched)
+        ),);
+
+        let properties = HashMap::from([
+            (
+                "plan_type".to_string(),
+                PostHogFlagFilterPropertyValue::String("free".to_string()),
+            ),
+            (
+                "pageserver_remote_size".to_string(),
+                PostHogFlagFilterPropertyValue::Number(1000.0),
+            ),
+        ]);
+
+        // It matches the first group as 0.10 <= 0.50 and the properties are matched. Then it gets evaluated to the variant override.
+        let variant = store.evaluate_multivariate_inner("gc-compaction", 0.10, 0.10, &properties);
+        assert_eq!(variant.unwrap(), "enabled-stage-2".to_string());
+
+        // It matches the second group as 0.50 <= 0.60 <= 0.80 and the properties are matched. Then it gets evaluated using the global percentage.
+        let variant = store.evaluate_multivariate_inner("gc-compaction", 0.99, 0.60, &properties);
+        assert_eq!(variant.unwrap(), "enabled-stage-1".to_string());
+        let variant = store.evaluate_multivariate_inner("gc-compaction", 0.80, 0.60, &properties);
+        assert_eq!(variant.unwrap(), "disabled".to_string());
+
+        // It matches the group conditions but not the group rollout percentage.
+        let variant = store.evaluate_multivariate_inner("gc-compaction", 1.00, 0.90, &properties);
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NoConditionGroupMatched)
+        ),);
+    }
+}

libs/remote_storage/src/azure_blob.rs

@@ -330,11 +330,18 @@ impl AzureBlobStorage {
                 if let Err(DownloadError::Timeout) = &next_item {
                     timeout_try_cnt += 1;
                     if timeout_try_cnt <= 5 {
-                        continue;
+                        continue 'outer;
                     }
                 }
 
-                let next_item = next_item?;
+                let next_item = match next_item {
+                    Ok(next_item) => next_item,
+                    Err(e) => {
+                        // The error is potentially retryable, so we must rewind the loop after yielding.
+                        yield Err(e);
+                        continue 'outer;
+                    },
+                };
 
                 // Log a warning if we saw two timeouts in a row before a successful request
                 if timeout_try_cnt > 2 {

libs/remote_storage/src/s3_bucket.rs

@@ -657,7 +657,14 @@ impl RemoteStorage for S3Bucket {
                     res = request => Ok(res),
                     _ = tokio::time::sleep(self.timeout) => Err(DownloadError::Timeout),
                     _ = cancel.cancelled() => Err(DownloadError::Cancelled),
-                }?;
+                };
+
+                if let Err(DownloadError::Timeout) = &response {
+                    yield Err(DownloadError::Timeout);
+                    continue 'outer;
+                }
+
+                let response = response?; // always yield cancellation errors and stop the stream
 
                 let response = response
                     .context("Failed to list S3 prefixes")

libs/safekeeper_api/src/models.rs

@@ -6,9 +6,11 @@ use pageserver_api::shard::ShardIdentity;
 use postgres_ffi::TimestampTz;
 use serde::{Deserialize, Serialize};
 use tokio::time::Instant;
+use utils::generation::Generation;
 use utils::id::{NodeId, TenantId, TenantTimelineId, TimelineId};
 use utils::lsn::Lsn;
 use utils::pageserver_feedback::PageserverFeedback;
+use utils::shard::ShardIndex;
 
 use crate::membership::Configuration;
 use crate::{ServerInfo, Term};
@@ -309,3 +311,29 @@ pub struct PullTimelineResponse {
     pub safekeeper_host: Option<String>,
     // TODO: add more fields?
 }
+
+#[derive(Clone, Debug, Serialize, Deserialize)]
+#[serde(tag = "action")]
+pub enum TenantShardPageserverAttachmentChange {
+    Attach(TenantShardPageserverAttachment),
+    Detach(TenantShardPageserverAttachment),
+}
+
+impl TenantShardPageserverAttachmentChange {
+    pub fn attachment(&self) -> &TenantShardPageserverAttachment {
+        match self {
+            TenantShardPageserverAttachmentChange::Attach(a) => a,
+            TenantShardPageserverAttachmentChange::Detach(a) => a,
+        }
+    }
+}
+
+#[derive(Clone, Debug, Serialize, Deserialize)]
+pub struct TenantShardPageserverAttachment {
+    pub shard_id: ShardIndex,
+    pub generation: Generation,
+    pub ps_id: NodeId,
+    // TODO: avoid transmitting this with every request.
+    // How nice things could be if there were simple DNS records for ps-$node_id.$cell.$region.$cloud.neon.tech
+    pub ps_hostname: String, // TODO: some type safety
+}

libs/sk_ps_discovery/Cargo.toml

@@ -0,0 +1,81 @@
+[package]
+name = "sk_ps_discovery"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+workspace_hack.workspace = true
+
+async-stream.workspace = true
+anyhow.workspace = true
+byteorder.workspace = true
+bytes.workspace = true
+camino.workspace = true
+camino-tempfile.workspace = true
+chrono.workspace = true
+clap = { workspace = true, features = ["derive"] }
+crc32c.workspace = true
+fail.workspace = true
+hex.workspace = true
+humantime.workspace = true
+http.workspace = true
+hyper0.workspace = true
+itertools.workspace = true
+jsonwebtoken.workspace = true
+futures.workspace = true
+once_cell.workspace = true
+parking_lot.workspace = true
+pageserver_api.workspace = true
+postgres-protocol.workspace = true
+pprof.workspace = true
+rand.workspace = true
+regex.workspace = true
+reqwest = { workspace = true, features = ["json"] }
+rustls.workspace = true
+scopeguard.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+smallvec.workspace = true
+strum.workspace = true
+strum_macros.workspace = true
+thiserror.workspace = true
+tikv-jemallocator.workspace = true
+tokio = { workspace = true, features = ["fs"] }
+tokio-io-timeout.workspace = true
+tokio-postgres.workspace = true
+tokio-rustls.workspace = true
+tokio-tar.workspace = true
+tokio-util = { workspace = true }
+tonic = { workspace = true }
+tracing.workspace = true
+url.workspace = true
+metrics.workspace = true
+pem.workspace = true
+postgres_backend.workspace = true
+postgres_ffi.workspace = true
+pq_proto.workspace = true
+remote_storage.workspace = true
+safekeeper_api.workspace = true
+safekeeper_client.workspace = true
+sha2.workspace = true
+sd-notify.workspace = true
+storage_broker.workspace = true
+tokio-stream.workspace = true
+http-utils.workspace = true
+utils.workspace = true
+wal_decoder.workspace = true
+env_logger.workspace = true
+
+[dev-dependencies]
+criterion.workspace = true
+itertools.workspace = true
+walproposer.workspace = true
+rand.workspace = true
+desim.workspace = true
+tracing.workspace = true
+tracing-subscriber = { workspace = true, features = ["json"] }
+
+[[bench]]
+name = "bench"
+harness = false

libs/sk_ps_discovery/benches/bench.rs

@@ -0,0 +1,97 @@
+//! WAL ingestion benchmarks.
+
+use std::time::Instant;
+
+use criterion::{Criterion, criterion_group, criterion_main};
+use pprof::criterion::{Output, PProfProfiler};
+use sk_ps_discovery::{
+    AttachmentUpdate, RemoteConsistentLsnAdv, TenantShardAttachmentId, TimelineAttachmentId,
+};
+use utils::{
+    generation::Generation,
+    id::{NodeId, TenantId, TenantTimelineId, TimelineId},
+    lsn::Lsn,
+    shard::ShardIndex,
+};
+
+/// Use jemalloc and enable profiling, to mirror bin/safekeeper.rs.
+#[global_allocator]
+static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
+
+#[allow(non_upper_case_globals)]
+#[unsafe(export_name = "malloc_conf")]
+pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:21\0";
+
+// Register benchmarks with Criterion.
+criterion_group!(
+    name = benches;
+    config = Criterion::default().with_profiler(PProfProfiler::new(100, Output::Flamegraph(None)));
+    targets = bench_simple,
+);
+criterion_main!(benches);
+
+fn bench_simple(c: &mut Criterion) {
+    let mut g = c.benchmark_group("simple");
+
+    // setup
+    let mut world = sk_ps_discovery::World::default();
+
+    // Simplified view: lots of unsharded tenants with one timeline each
+    let n_pageservers = 20;
+    let n_tenant_shards_per_pageserver = 2000;
+    for ps_id in 1..=n_pageservers {
+        for _ in ..n_tenant_shards_per_pageserver {
+            let tenant_id = TenantId::generate();
+            let timeline_id = TimelineId::generate();
+            for generation in 10..=11 {
+                let tenant_shard_attachment_id = TenantShardAttachmentId {
+                    tenant_id,
+                    shard_id: ShardIndex::unsharded(),
+                    generation: Generation::Valid(generation),
+                };
+                let timeline_attachment = TimelineAttachmentId {
+                    tenant_timeline_id: TenantTimelineId {
+                        tenant_id,
+                        timeline_id,
+                    },
+                    shard_id: ShardIndex::unsharded(),
+                    generation: Generation::Valid(generation),
+                };
+                world.update_attachment(AttachmentUpdate {
+                    tenant_shard_attachment_id,
+                    action: sk_ps_discovery::AttachmentUpdateAction::Attach {
+                        ps_id: NodeId(ps_id),
+                    },
+                });
+                world.handle_remote_consistent_lsn_advertisement(RemoteConsistentLsnAdv {
+                    remote_consistent_lsn: Lsn(23),
+                    attachment: timeline_attachment,
+                });
+            }
+            world.handle_commit_lsn_advancement(
+                TenantTimelineId {
+                    tenant_id,
+                    timeline_id,
+                },
+                Lsn(42),
+            );
+        }
+    }
+
+    // setup done
+    let world = world;
+    g.bench_function("get_commit_lsn_advertisements", |bencher| {
+        bencher.iter_custom(|iters| {
+            let started = Instant::now();
+
+            for _ in 0..iters {
+                criterion::black_box(world.get_commit_lsn_advertisements());
+            }
+
+            let elapsed = started.elapsed();
+            elapsed
+        });
+    });
+
+    g.finish();
+}

libs/sk_ps_discovery/src/lib.rs

@@ -0,0 +1,515 @@
+#[cfg(test)]
+mod tests;
+
+use std::{
+    collections::{BTreeMap, BTreeSet, HashMap, HashSet, btree_map, hash_map},
+    ops::RangeInclusive,
+};
+
+use tracing::{info, warn};
+use utils::{
+    generation::Generation,
+    id::{NodeId, TenantId, TenantTimelineId, TimelineId},
+    lsn::Lsn,
+    merge_join,
+    shard::ShardIndex,
+};
+
+#[derive(Debug, Default)]
+pub struct World {
+    attachments: BTreeMap<TenantShardAttachmentId, NodeId>,
+    attachment_count: HashMap<TenantId, u16>,
+    nodes_timelines: HashMap<NodeId, HashMap<TenantTimelineId, u16>>, // u16 is a refcount from each timeline attachment id
+    // continously maintained aggregate for efficient decisionmaking on quiescing;
+    // quiesced timelines are always caught up
+    // can quiesce one == attachment_count (TODO: this requires enforcing foreign key relationship between attachments and remote_consistent_lsn)
+    caught_up_count: HashMap<TenantTimelineId, u16>,
+
+    // BEGIN quiescing/active split
+    quiesced_timelines: BTreeMap<TenantTimelineId, Lsn>,
+    // ^
+    // either a timeline is in quiesced_timelines
+    // or it is below
+    // v
+    commit_lsns: BTreeMap<TenantTimelineId, Lsn>,
+    remote_consistent_lsns: BTreeMap<TimelineAttachmentId, Lsn>,
+    // END quiescing/active split
+
+    // other fields
+}
+
+#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, PartialOrd, Ord)]
+pub struct TenantShardAttachmentId {
+    pub tenant_id: TenantId,
+    pub shard_id: ShardIndex,
+    pub generation: Generation,
+}
+
+#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, PartialOrd, Ord)]
+pub struct TimelineAttachmentId {
+    pub tenant_timeline_id: TenantTimelineId,
+    pub shard_id: ShardIndex,
+    pub generation: Generation,
+}
+
+pub struct AttachmentUpdate {
+    pub tenant_shard_attachment_id: TenantShardAttachmentId,
+    pub action: AttachmentUpdateAction,
+}
+
+pub enum AttachmentUpdateAction {
+    Attach { ps_id: NodeId },
+    Detach,
+}
+
+pub struct RemoteConsistentLsnAdv {
+    pub attachment: TimelineAttachmentId,
+    pub remote_consistent_lsn: Lsn,
+}
+
+impl World {
+    fn check_invariants(&self) {
+        if !cfg!(debug_assertions) {
+            return;
+        }
+
+        // caught_up_count maintenance
+        {
+            for (tenant_timeline_id, caught_up_count) in
+                self.caught_up_count.iter().map(|(k, v)| (*k, *v))
+            {
+                let attachment_count = *self
+                    .attachment_count
+                    .get(&tenant_timeline_id.tenant_id)
+                    .unwrap();
+                assert!(caught_up_count <= attachment_count);
+                if caught_up_count == attachment_count {
+                    self.quiesced_timelines.contains_key(&tenant_timeline_id);
+                    // remote_consistent_lsn and commit_lsns is empty, checked by "quiescing XOR ..." below
+                } else {
+                    let commit_lsn = self.commit_lsns[&&tenant_timeline_id];
+                    let mut validate_caught_up = 0;
+                    let mut validate_not_caught_up = 0;
+                    for (_, r_c_lsn) in self
+                        .remote_consistent_lsns
+                        .range(TimelineAttachmentId::timeline_range(tenant_timeline_id))
+                        .map(|(k, v)| (*k, *v))
+                    {
+                        if r_c_lsn == commit_lsn {
+                            validate_caught_up += 1;
+                        } else {
+                            assert!(r_c_lsn < commit_lsn);
+                            validate_not_caught_up += 1;
+                        }
+                    }
+                    assert_eq!(validate_caught_up, caught_up_count);
+                    assert_eq!(
+                        validate_caught_up + validate_not_caught_up,
+                        attachment_count
+                    );
+                }
+            }
+        }
+
+        // quiescing XOR ...
+        {
+            let quiesced_timelines: HashSet<TenantTimelineId> =
+                self.quiesced_timelines.keys().cloned().collect();
+            let commit_lsn_timelines: HashSet<TenantTimelineId> =
+                self.commit_lsns.keys().cloned().collect();
+            let remote_consistent_lsn_timelines: HashSet<TenantTimelineId> = self
+                .remote_consistent_lsns
+                .keys()
+                .map(|tlaid: &TimelineAttachmentId| tlaid.tenant_timeline_id)
+                .collect();
+            #[rustfmt::skip]
+            assert_eq!(0, quiesced_timelines.intersection(&commit_lsn_timelines).count());
+            #[rustfmt::skip]
+            assert_eq!(0, quiesced_timelines.intersection(&remote_consistent_lsn_timelines).count());
+        }
+
+        // nodes_timelines maintenance
+        {
+            let mut expect: HashMap<NodeId, HashMap<TenantTimelineId, u16>> = HashMap::new();
+            let all_ttids: BTreeSet<TenantTimelineId> = self
+                .quiesced_timelines
+                .keys()
+                .cloned()
+                .chain(
+                    self.remote_consistent_lsns
+                        .keys()
+                        .cloned()
+                        .map(|tlaid| tlaid.tenant_timeline_id),
+                )
+                .collect();
+            for ttid in all_ttids {
+                for (_, node_id) in self
+                    .attachments
+                    .range(TenantShardAttachmentId::tenant_range(ttid.tenant_id))
+                    .map(|(k, v)| (*k, *v))
+                {
+                    let expect = expect.entry(node_id).or_default();
+                    let refcount = expect.entry(ttid).or_default();
+                    *refcount += 1;
+                }
+            }
+            assert_eq!(expect, self.nodes_timelines);
+        }
+    }
+    pub fn update_attachment(&mut self, upd: AttachmentUpdate) {
+        self.check_invariants();
+        use AttachmentUpdateAction::*;
+        use btree_map::Entry::*;
+        let AttachmentUpdate {
+            tenant_shard_attachment_id,
+            action,
+        } = upd;
+        match (action, self.attachments.entry(tenant_shard_attachment_id)) {
+            (Attach { ps_id }, Occupied(e)) if *e.get() == ps_id => {
+                info!("attachment is already known")
+            }
+            (Attach { ps_id }, Occupied(e)) => {
+                warn!(current_node=%e.get(), proposed_node=%ps_id, "ignoring update that moves attachment to a different pageserver");
+            }
+            (Attach { ps_id }, Vacant(e)) => {
+                e.insert(ps_id);
+                // Keep attachmount_count up to date
+                let attachment_count = self
+                    .attachment_count
+                    .entry(tenant_shard_attachment_id.tenant_id)
+                    .or_default();
+                *attachment_count += attachment_count.checked_add(1).unwrap();
+                // Keep nodes_timelines up to date
+                let nodes_timelines = self.nodes_timelines.entry(ps_id).or_default();
+                for (ttid, _) in self.commit_lsns.range(TenantTimelineId::tenant_range(
+                    tenant_shard_attachment_id.tenant_id,
+                )) {
+                    let refcount = nodes_timelines.entry(*ttid).or_default();
+                    *refcount = refcount.checked_add(1).unwrap();
+                }
+                if nodes_timelines.is_empty() {
+                    self.nodes_timelines.remove(&ps_id);
+                }
+                // New shards may start at an older LSN than where we quiesced => activate all quiesced timelines.
+                let activate_range =
+                    TenantTimelineId::tenant_range(tenant_shard_attachment_id.tenant_id);
+                let activate: HashSet<TenantTimelineId> = self
+                    .quiesced_timelines
+                    .range(activate_range)
+                    .map(|(ttid, _quiesced_lsn)| *ttid)
+                    .collect();
+                for tenant_timeline_id in activate {
+                    self.activate_timeline(tenant_timeline_id);
+                }
+            }
+            (Detach, Occupied(e)) => {
+                let ps_id = e.remove();
+                // Keep attachment count up to date
+                let attachment_count = self
+                    .attachment_count
+                    .get_mut(&tenant_shard_attachment_id.tenant_id)
+                    .expect("attachment action initializes the hasmap entry");
+                *attachment_count = attachment_count.checked_sub(1).unwrap();
+                // Keep nodes_timelines up to date
+                let nodes_timelines = self
+                    .nodes_timelines
+                    .get_mut(&ps_id)
+                    .expect("attachment action initializes hashmap entry");
+                for (ttid, _) in self.commit_lsns.range(TenantTimelineId::tenant_range(
+                    tenant_shard_attachment_id.tenant_id,
+                )) {
+                    let refcount = nodes_timelines.entry(*ttid).or_default();
+                    *refcount = refcount.checked_sub(1).unwrap();
+                }
+            }
+            (Detach, Vacant(_)) => {
+                info!("detachment is already known");
+            }
+        }
+        self.check_invariants();
+    }
+    pub fn handle_remote_consistent_lsn_advertisement(&mut self, adv: RemoteConsistentLsnAdv) {
+        self.check_invariants();
+        let RemoteConsistentLsnAdv {
+            attachment,
+            remote_consistent_lsn,
+        } = adv;
+
+        match self.remote_consistent_lsns.entry(attachment) {
+            btree_map::Entry::Occupied(mut occupied_entry) => {
+                let current = occupied_entry.get_mut();
+                use std::cmp::Ordering::*;
+                match (*current).cmp(&remote_consistent_lsn) {
+                    Less => {
+                        *current = remote_consistent_lsn;
+                        let caught_up_count = self
+                            .caught_up_count
+                            .get_mut(&attachment.tenant_timeline_id)
+                            .unwrap();
+                        *caught_up_count = caught_up_count.checked_add(1).unwrap();
+                        if *caught_up_count
+                            == self.attachment_count[&attachment.tenant_timeline_id.tenant_id]
+                        {
+                            self.quiesce_timeline(attachment.tenant_timeline_id);
+                        }
+                    }
+                    Equal => {
+                        info!("ignoring no-op update, likely duplicate delivery");
+                    }
+                    Greater => {
+                        warn!(
+                            "ignoring advertisement because remote_consistent_lsn is moving backwards"
+                        );
+                    }
+                }
+            }
+            btree_map::Entry::Vacant(_) => {
+                let ttid = attachment.tenant_timeline_id;
+                match self.quiesced_timelines.get(&ttid).cloned() {
+                    Some(quiesced_lsn) if quiesced_lsn == remote_consistent_lsn => {
+                        info!("ignoring no-op update for quiesced timeline");
+                    }
+                    Some(_) => {
+                        self.activate_timeline(ttid);
+                        // recurse one level, guarnateed to hit `Occupied` case above
+                        self.handle_remote_consistent_lsn_advertisement(adv);
+                    }
+                    None => {
+                        info!("ignoring advertisement because timeline is not known");
+                    }
+                }
+            }
+        }
+        self.check_invariants();
+    }
+    pub fn handle_commit_lsn_advancement(&mut self, ttid: TenantTimelineId, update: Lsn) {
+        self.check_invariants();
+        match self.commit_lsns.entry(ttid) {
+            btree_map::Entry::Occupied(mut entry) => {
+                let current = entry.get_mut();
+                use std::cmp::Ordering::*;
+                match (*current).cmp(&update) {
+                    Less => {
+                        *current = update;
+                        // We never allow remote_consistent_lsn to be ahead of commit_lsn.
+                        // Therefore, it is safe to say nothing is caught up anymore.
+                        let caught_up_count = self.caught_up_count.get_mut(&ttid).unwrap();
+                        *caught_up_count = 0;
+                    }
+                    Equal => {
+                        // This code runs in safekeeper impl, no reason why there would be duplicate delivery.
+                        warn!("ignoring no-op update; why is this happening?");
+                    }
+                    Greater => {
+                        panic!(
+                            "proposed commit_lsn would move it backwards: current={} update={}",
+                            current, update
+                        );
+                    }
+                }
+            }
+
+            btree_map::Entry::Vacant(entry) => {
+                match self.quiesced_timelines.get(&ttid).cloned() {
+                    Some(quiesced_lsn) if quiesced_lsn == update => {
+                        info!("ignoring no-op update for quiesced timeline");
+                    }
+                    Some(_) => {
+                        self.activate_timeline(ttid);
+                        // recurse one level, guarnateed to hit `Occupied` case above
+                        self.handle_commit_lsn_advancement(ttid, update);
+                    }
+                    None => {
+                        info!("first time hearing about this timeline, initializing");
+                        entry.insert(update);
+                        let replaced = self.caught_up_count.insert(ttid, 0);
+                        // only commit_lsn advancement makes timelines known to world
+                        assert_eq!(None, replaced);
+                        for (attachment, node_id) in self
+                            .attachments
+                            .range(TenantShardAttachmentId::tenant_range(ttid.tenant_id))
+                        {
+                            let replaced = self.remote_consistent_lsns.insert(
+                                attachment.timeline_attachment_id(ttid.timeline_id),
+                                Lsn(0),
+                            );
+                            // only commit_lsn advancement makes timelines known to World
+                            assert_eq!(None, replaced);
+
+                            let nodes_timelines = self.nodes_timelines.entry(*node_id).or_default();
+                            let refcount = nodes_timelines.entry(ttid).or_default();
+                            *refcount = refcount.checked_add(1).unwrap();
+                        }
+                    }
+                }
+            }
+        }
+        self.check_invariants();
+    }
+
+    pub fn get_commit_lsn_advertisements(&self) -> HashMap<NodeId, HashMap<TenantTimelineId, Lsn>> {
+        let mut commit_lsn_advertisements_by_node: HashMap<NodeId, HashMap<TenantTimelineId, Lsn>> =
+            HashMap::with_capacity(self.nodes_timelines.len());
+        let commit_lsns_iter = self.commit_lsns.iter().map(|(k, v)| (*k, *v));
+        let attachments_iter = self.attachments.iter().map(|(k, v)| (*k, *v));
+        let remote_consistent_lsns_iter = self.remote_consistent_lsns.iter().map(|(k, v)| (*k, *v));
+
+        let join = merge_join::inner_equi_join_with_merge_strategy(
+            commit_lsns_iter,
+            attachments_iter,
+            |(tenant_timeline_id, _)| tenant_timeline_id.tenant_id,
+            |(shard_attachment_id, _)| shard_attachment_id.tenant_id,
+        );
+        let join = merge_join::left_equi_join_with_merge_strategy(
+            join,
+            remote_consistent_lsns_iter,
+            |((ttid, _), _)| ttid.tenant_id,
+            |(tlaid, _)| tlaid.tenant_timeline_id.tenant_id,
+        );
+        for ((c, a), r) in join {
+            let (tenant_timeline_id, commit_lsn): (TenantTimelineId, Lsn) = c;
+            let (_, node_id): (TenantShardAttachmentId, NodeId) = a;
+            match r {
+                // TODO: can > ever happen?
+                Some((_, remote_consistent_lsn)) if remote_consistent_lsn >= commit_lsn => {
+                    // this timeline shard attachment is already caught up
+                    continue;
+                }
+                Some(_) | None => {
+                    // need to advertise
+                    // -> fallthrough
+                }
+            };
+            // DISTINCT node_id, array_agg(DISTINCT tenant_shard_id )
+            let for_node = commit_lsn_advertisements_by_node
+                .entry(node_id)
+                .or_insert_with(|| HashMap::with_capacity(self.nodes_timelines[&node_id].len()));
+            match for_node.entry(tenant_timeline_id) {
+                hash_map::Entry::Vacant(vacant_entry) => {
+                    vacant_entry.insert(commit_lsn);
+                }
+                hash_map::Entry::Occupied(occupied_entry) => {
+                    assert_eq!(*occupied_entry.get(), commit_lsn);
+                }
+            }
+        }
+        commit_lsn_advertisements_by_node
+    }
+
+    fn activate_timeline(&mut self, tenant_timeline_id: TenantTimelineId) {
+        let quiesced_lsn = self
+            .quiesced_timelines
+            .remove(&tenant_timeline_id)
+            .expect("must call this function only on quiesced tenant_timeline_id");
+        let replaced = self.commit_lsns.insert(tenant_timeline_id, quiesced_lsn);
+        assert_eq!(None, replaced);
+        let reconstruct_remote_consistent_lsn_entries = self
+            .attachments
+            .range(TenantShardAttachmentId::tenant_range(
+                tenant_timeline_id.tenant_id,
+            ))
+            .map(|(k, _)| *k)
+            .map(|tenant_shard_attachment_id| {
+                (
+                    tenant_shard_attachment_id
+                        .timeline_attachment_id(tenant_timeline_id.timeline_id),
+                    quiesced_lsn,
+                )
+            });
+        for (key, value) in reconstruct_remote_consistent_lsn_entries {
+            let replaced = self.remote_consistent_lsns.insert(key, value);
+            assert_eq!(None, replaced);
+        }
+    }
+
+    fn quiesce_timeline(&mut self, tenant_timeline_id: TenantTimelineId) {
+        self.check_invariants();
+        if self.quiesced_timelines.contains_key(&tenant_timeline_id) {
+            panic!("only call this function on active timelines");
+        }
+        let quiesced_lsn = self
+            .commit_lsns
+            .remove(&tenant_timeline_id)
+            .expect("inconsistent: we checked it's not in quiesced_timelines, so, must be active");
+        let caught_up_count = self
+            .caught_up_count
+            .remove(&tenant_timeline_id)
+            .expect("inconsistent: we checked it's not in quiesced_timleines, so, must be active");
+        let mut remove_remote_consistent_lsns = Vec::new();
+        for (k, remote_consistent_lsn) in self
+            .remote_consistent_lsns
+            .range(TimelineAttachmentId::timeline_range(tenant_timeline_id))
+        {
+            assert_eq!(*remote_consistent_lsn, quiesced_lsn);
+            remove_remote_consistent_lsns.push(*k);
+        }
+        assert_eq!(
+            caught_up_count,
+            u16::try_from(remove_remote_consistent_lsns.len()).unwrap()
+        );
+        for k in remove_remote_consistent_lsns {
+            let removed = self.remote_consistent_lsns.remove(&k);
+            assert!(removed.is_some(), "we just added");
+        }
+        let replaced = self
+            .quiesced_timelines
+            .insert(tenant_timeline_id, quiesced_lsn);
+        assert_eq!(None, replaced); // we checked at function entry
+        self.check_invariants();
+    }
+}
+
+impl TimelineAttachmentId {
+    pub fn timeline_range(ttid: TenantTimelineId) -> RangeInclusive<Self> {
+        let shard_index_range: RangeInclusive<_> = ShardIndex::RANGE;
+        let generation_range: RangeInclusive<_> = Generation::RANGE;
+        RangeInclusive::new(
+            TimelineAttachmentId {
+                tenant_timeline_id: ttid,
+                shard_id: *shard_index_range.start(),
+                generation: *generation_range.start(),
+            },
+            TimelineAttachmentId {
+                tenant_timeline_id: ttid,
+                shard_id: *shard_index_range.end(),
+                generation: *generation_range.end(),
+            },
+        )
+    }
+    pub fn tenant_shard_attachment_id(self) -> TenantShardAttachmentId {
+        TenantShardAttachmentId {
+            tenant_id: self.tenant_timeline_id.tenant_id,
+            shard_id: self.shard_id,
+            generation: self.generation,
+        }
+    }
+}
+
+impl TenantShardAttachmentId {
+    pub fn timeline_attachment_id(self, timeline_id: TimelineId) -> TimelineAttachmentId {
+        TimelineAttachmentId {
+            tenant_timeline_id: TenantTimelineId {
+                tenant_id: self.tenant_id,
+                timeline_id,
+            },
+            shard_id: self.shard_id,
+            generation: self.generation,
+        }
+    }
+    pub fn tenant_range(tenant_id: TenantId) -> RangeInclusive<Self> {
+        let shard_index_range: RangeInclusive<_> = ShardIndex::RANGE;
+        let generation_range: RangeInclusive<_> = Generation::RANGE;
+        RangeInclusive::new(
+            Self {
+                tenant_id,
+                shard_id: *shard_index_range.start(),
+                generation: *generation_range.start(),
+            },
+            Self {
+                tenant_id,
+                shard_id: *shard_index_range.end(),
+                generation: *generation_range.end(),
+            },
+        )
+    }
+}

libs/sk_ps_discovery/src/tests.rs

@@ -0,0 +1,224 @@
+use utils::{id::TenantId, logging};
+
+use super::*;
+use crate::World;
+
+#[track_caller]
+fn validate_advertisements(
+    actual: HashMap<NodeId, HashMap<TenantTimelineId, Lsn>>,
+    expect: Vec<(NodeId, Vec<(TenantTimelineId, Lsn)>)>,
+) {
+    let expect: HashMap<_, _> = expect
+        .into_iter()
+        .map(|(node_id, innermap)| (node_id, innermap.into_iter().collect()))
+        .collect();
+    assert_eq!(actual, expect);
+}
+
+#[test]
+fn basic() {
+    let mut world = World::default();
+
+    let tenant_id = TenantId::from_array([0xff; 16]);
+    let timeline_id = TimelineId::from_array([1; 16]);
+    let timeline2 = TimelineId::from_array([2; 16]);
+
+    let attachment1 = TenantShardAttachmentId {
+        tenant_id,
+        shard_id: ShardIndex::unsharded(),
+        generation: Generation::Valid(2),
+    };
+    let attachment2 = TenantShardAttachmentId {
+        tenant_id,
+        shard_id: ShardIndex::unsharded(),
+        generation: Generation::Valid(3),
+    };
+
+    let ps1 = NodeId(0x100);
+
+    // Out of order; in happy path, commit_lsn advances first, but let's test the
+    // case where safekeeper doesn't know about the attachments yet first, before
+    // we extend the case to the happy path.
+
+    world.handle_remote_consistent_lsn_advertisement(RemoteConsistentLsnAdv {
+        attachment: attachment1.timeline_attachment_id(timeline_id),
+        remote_consistent_lsn: Lsn(0x23),
+    });
+    world.handle_remote_consistent_lsn_advertisement(RemoteConsistentLsnAdv {
+        attachment: attachment2.timeline_attachment_id(timeline_id),
+        remote_consistent_lsn: Lsn(0x42),
+    });
+    // SK authoritative info on which advertisements ought exist is still empty
+    assert_eq!(world.get_commit_lsn_advertisements(), HashMap::default());
+    world.update_attachment(AttachmentUpdate {
+        tenant_shard_attachment_id: attachment1,
+        action: AttachmentUpdateAction::Attach { ps_id: ps1 },
+    });
+    // We have not inserted any commit_lsn info yet, so, still no advs expected
+    assert_eq!(world.get_commit_lsn_advertisements(), HashMap::default());
+    // insert commit_lsn info for different timeline
+    world.handle_commit_lsn_advancement(
+        TenantTimelineId {
+            tenant_id,
+            timeline_id: timeline2,
+        },
+        Lsn(0x66),
+    );
+    // Advs should still be empty
+    validate_advertisements(
+        world.get_commit_lsn_advertisements(),
+        vec![(
+            ps1,
+            vec![(
+                TenantTimelineId {
+                    tenant_id,
+                    timeline_id: timeline2,
+                },
+                Lsn(0x66),
+            )],
+        )],
+    );
+
+    // Ok, out of order part tested. Now Safekeeper learns about the attachments.
+
+    // insert commit_lsn info for the timeline we have remote_consistent_lsn info for
+    world.handle_commit_lsn_advancement(
+        TenantTimelineId {
+            tenant_id,
+            timeline_id,
+        },
+        Lsn(0x55),
+    );
+    dbg!(&world);
+    // Now advertisements to attachment1 will be sent out, but attachment2  is still not known, so, no advertisements to it.
+    validate_advertisements(
+        world.get_commit_lsn_advertisements(),
+        vec![(
+            ps1,
+            vec![(
+                TenantTimelineId {
+                    tenant_id,
+                    timeline_id,
+                },
+                Lsn(0x55),
+            )],
+        )],
+    );
+}
+
+#[test]
+fn advertisement_for_new_timeline() {
+    let mut world = World::default();
+
+    let tenant_id = TenantId::generate();
+    let timeline_id = TimelineId::generate();
+    let ttid = TenantTimelineId {
+        tenant_id,
+        timeline_id,
+    };
+
+    let tenant_shard_attachment_id = TenantShardAttachmentId {
+        tenant_id,
+        shard_id: ShardIndex::unsharded(),
+        generation: Generation::Valid(2),
+    };
+
+    let ps_id = NodeId(0x100);
+
+    world.update_attachment(AttachmentUpdate {
+        tenant_shard_attachment_id,
+        action: AttachmentUpdateAction::Attach { ps_id },
+    });
+    world.handle_commit_lsn_advancement(ttid, Lsn(23));
+
+    let advs = world.get_commit_lsn_advertisements();
+    validate_advertisements(advs, vec![(ps_id, vec![(ttid, Lsn(23))])]);
+}
+
+#[test]
+fn quiescing_timeline_catchup() {
+    let _guard = logging::init(
+        logging::LogFormat::Test,
+        logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
+        logging::Output::Stdout,
+    )
+    .unwrap();
+
+    let mut world = World::default();
+
+    let tenant_id = TenantId::generate();
+    let timeline_id = TimelineId::generate();
+    let ttid = TenantTimelineId {
+        tenant_id,
+        timeline_id,
+    };
+
+    let tenant_shard_attachment_id = TenantShardAttachmentId {
+        tenant_id,
+        shard_id: ShardIndex::unsharded(),
+        generation: Generation::Valid(2),
+    };
+
+    let ps_id = NodeId(0x100);
+
+    world.update_attachment(AttachmentUpdate {
+        tenant_shard_attachment_id,
+        action: AttachmentUpdateAction::Attach { ps_id },
+    });
+    world.handle_commit_lsn_advancement(ttid, Lsn(23));
+
+    assert!(world.quiesced_timelines.is_empty());
+
+    world.handle_remote_consistent_lsn_advertisement(RemoteConsistentLsnAdv {
+        attachment: tenant_shard_attachment_id.timeline_attachment_id(timeline_id),
+        remote_consistent_lsn: Lsn(23),
+    });
+
+    assert!(world.quiesced_timelines.contains_key(&ttid));
+}
+
+#[test]
+fn nodes_timelines() {
+    let mut world = World::default();
+
+    let tenant_id = TenantId::generate();
+    let timeline_id = TimelineId::from_array([0x1; 16]);
+    let ttid = TenantTimelineId {
+        tenant_id,
+        timeline_id,
+    };
+
+    let tenant_shard_attachment_id = TenantShardAttachmentId {
+        tenant_id,
+        shard_id: ShardIndex::unsharded(),
+        generation: Generation::Valid(2),
+    };
+
+    let ps_id = NodeId(0x100);
+
+    world.update_attachment(AttachmentUpdate {
+        tenant_shard_attachment_id,
+        action: AttachmentUpdateAction::Attach { ps_id },
+    });
+
+    assert!(world.nodes_timelines.get(&ps_id).is_none());
+
+    world.handle_commit_lsn_advancement(ttid, Lsn(0x23));
+
+    assert_eq!(world.nodes_timelines[&ps_id].len(), 1);
+
+    let timeline2 = TimelineId::from_array([0x2; 16]);
+    world.handle_remote_consistent_lsn_advertisement(RemoteConsistentLsnAdv {
+        attachment: TimelineAttachmentId {
+            tenant_timeline_id: TenantTimelineId {
+                tenant_id,
+                timeline_id: timeline2,
+            },
+            shard_id: ShardIndex::unsharded(),
+            generation: Generation::Valid(2),
+        },
+        remote_consistent_lsn: Lsn(0x42),
+    });
+}
+
+// TODO: need more tests, esp for the removal path

libs/utils/src/crashsafe.rs

@@ -1,7 +1,7 @@
 use std::borrow::Cow;
 use std::fs::{self, File};
 use std::io::{self, Write};
-use std::os::fd::AsRawFd;
+use std::os::fd::AsFd;
 
 use camino::{Utf8Path, Utf8PathBuf};
 
@@ -210,13 +210,13 @@ pub fn overwrite(
 
 /// Syncs the filesystem for the given file descriptor.
 #[cfg_attr(target_os = "macos", allow(unused_variables))]
-pub fn syncfs(fd: impl AsRawFd) -> anyhow::Result<()> {
+pub fn syncfs(fd: impl AsFd) -> anyhow::Result<()> {
     // Linux guarantees durability for syncfs.
     // POSIX doesn't have syncfs, and further does not actually guarantee durability of sync().
     #[cfg(target_os = "linux")]
     {
         use anyhow::Context;
-        nix::unistd::syncfs(fd.as_raw_fd()).context("syncfs")?;
+        nix::unistd::syncfs(fd).context("syncfs")?;
     }
     #[cfg(target_os = "macos")]
     {

libs/utils/src/fs_ext/rename_noreplace.rs

@@ -11,9 +11,9 @@ pub fn rename_noreplace<P1: ?Sized + NixPath, P2: ?Sized + NixPath>(
         #[cfg(all(target_os = "linux", target_env = "gnu"))]
         {
             nix::fcntl::renameat2(
-                None,
+                nix::fcntl::AT_FDCWD,
                 src,
-                None,
+                nix::fcntl::AT_FDCWD,
                 dst,
                 nix::fcntl::RenameFlags::RENAME_NOREPLACE,
             )

libs/utils/src/generation.rs

@@ -1,4 +1,4 @@
-use std::fmt::Debug;
+use std::{fmt::Debug, ops::RangeInclusive};
 
 use serde::{Deserialize, Serialize};
 
@@ -25,7 +25,9 @@ pub enum Generation {
 /// scenarios where pageservers might otherwise issue conflicting writes to
 /// remote storage
 impl Generation {
+    pub const MIN: Self = Self::None;
     pub const MAX: Self = Self::Valid(u32::MAX);
+    pub const RANGE: RangeInclusive<Self> = RangeInclusive::new(Self::MIN, Self::MAX);
 
     /// Create a new Generation that represents a legacy key format with
     /// no generation suffix

libs/utils/src/id.rs

@@ -1,5 +1,6 @@
 use std::fmt;
 use std::num::ParseIntError;
+use std::ops::RangeInclusive;
 use std::str::FromStr;
 
 use anyhow::Context;
@@ -320,6 +321,19 @@ impl TenantTimelineId {
     pub fn empty() -> Self {
         Self::new(TenantId::from([0u8; 16]), TimelineId::from([0u8; 16]))
     }
+
+    pub fn tenant_range(tenant_id: TenantId) -> RangeInclusive<Self> {
+        RangeInclusive::new(
+            Self {
+                tenant_id,
+                timeline_id: TimelineId::from_array([u8::MIN; 16]),
+            },
+            Self {
+                tenant_id,
+                timeline_id: TimelineId::from_array([u8::MAX; 16]),
+            },
+        )
+    }
 }
 
 impl fmt::Display for TenantTimelineId {

libs/utils/src/lib.rs

@@ -95,6 +95,9 @@ pub mod guard_arc_swap;
 
 pub mod elapsed_accum;
 
+pub mod merge_join;
+
+
 #[cfg(target_os = "linux")]
 pub mod linux_socket_ioctl;
 

libs/utils/src/lock_file.rs

@@ -1,6 +1,6 @@
 //! A module to create and read lock files.
 //!
-//! File locking is done using [`fcntl::flock`] exclusive locks.
+//! File locking is done using [`nix::fcntl::Flock`] exclusive locks.
 //! The only consumer of this module is currently
 //! [`pid_file`](crate::pid_file). See the module-level comment
 //! there for potential pitfalls with lock files that are used
@@ -9,26 +9,25 @@
 use std::fs;
 use std::io::{Read, Write};
 use std::ops::Deref;
-use std::os::unix::prelude::AsRawFd;
 
 use anyhow::Context;
 use camino::{Utf8Path, Utf8PathBuf};
 use nix::errno::Errno::EAGAIN;
-use nix::fcntl;
+use nix::fcntl::{Flock, FlockArg};
 
 use crate::crashsafe;
 
-/// A handle to an open and unlocked, but not-yet-written lock file.
+/// A handle to an open and flocked, but not-yet-written lock file.
 /// Returned by [`create_exclusive`].
 #[must_use]
 pub struct UnwrittenLockFile {
     path: Utf8PathBuf,
-    file: fs::File,
+    file: Flock<fs::File>,
 }
 
 /// Returned by [`UnwrittenLockFile::write_content`].
 #[must_use]
-pub struct LockFileGuard(fs::File);
+pub struct LockFileGuard(Flock<fs::File>);
 
 impl Deref for LockFileGuard {
     type Target = fs::File;
@@ -67,17 +66,14 @@ pub fn create_exclusive(lock_file_path: &Utf8Path) -> anyhow::Result<UnwrittenLo
         .open(lock_file_path)
         .context("open lock file")?;
 
-    let res = fcntl::flock(
-        lock_file.as_raw_fd(),
-        fcntl::FlockArg::LockExclusiveNonblock,
-    );
+    let res = Flock::lock(lock_file, FlockArg::LockExclusiveNonblock);
     match res {
-        Ok(()) => Ok(UnwrittenLockFile {
+        Ok(lock_file) => Ok(UnwrittenLockFile {
             path: lock_file_path.to_owned(),
             file: lock_file,
         }),
-        Err(EAGAIN) => anyhow::bail!("file is already locked"),
-        Err(e) => Err(e).context("flock error"),
+        Err((_, EAGAIN)) => anyhow::bail!("file is already locked"),
+        Err((_, e)) => Err(e).context("flock error"),
     }
 }
 
@@ -105,32 +101,37 @@ pub enum LockFileRead {
 /// Check the [`LockFileRead`] variants for details.
 pub fn read_and_hold_lock_file(path: &Utf8Path) -> anyhow::Result<LockFileRead> {
     let res = fs::OpenOptions::new().read(true).open(path);
-    let mut lock_file = match res {
+    let lock_file = match res {
         Ok(f) => f,
         Err(e) => match e.kind() {
             std::io::ErrorKind::NotFound => return Ok(LockFileRead::NotExist),
             _ => return Err(e).context("open lock file"),
         },
     };
-    let res = fcntl::flock(
-        lock_file.as_raw_fd(),
-        fcntl::FlockArg::LockExclusiveNonblock,
-    );
+    let res = Flock::lock(lock_file, FlockArg::LockExclusiveNonblock);
     // We need the content regardless of lock success / failure.
     // But, read it after flock so that, if it succeeded, the content is consistent.
-    let mut content = String::new();
-    lock_file
-        .read_to_string(&mut content)
-        .context("read lock file")?;
     match res {
-        Ok(()) => Ok(LockFileRead::NotHeldByAnyProcess(
-            LockFileGuard(lock_file),
-            content,
-        )),
-        Err(EAGAIN) => Ok(LockFileRead::LockedByOtherProcess {
-            not_locked_file: lock_file,
-            content,
-        }),
-        Err(e) => Err(e).context("flock error"),
+        Ok(mut locked_file) => {
+            let mut content = String::new();
+            locked_file
+                .read_to_string(&mut content)
+                .context("read lock file")?;
+            Ok(LockFileRead::NotHeldByAnyProcess(
+                LockFileGuard(locked_file),
+                content,
+            ))
+        }
+        Err((mut not_locked_file, EAGAIN)) => {
+            let mut content = String::new();
+            not_locked_file
+                .read_to_string(&mut content)
+                .context("read lock file")?;
+            Ok(LockFileRead::LockedByOtherProcess {
+                not_locked_file,
+                content,
+            })
+        }
+        Err((_, e)) => Err(e).context("flock error"),
     }
 }

libs/utils/src/merge_join.rs

@@ -0,0 +1,164 @@
+pub fn inner_equi_join_with_merge_strategy<L, LI, R, RI, K, FL, FR>(
+    l: L,
+    r: R,
+    key_l: FL,
+    key_r: FR,
+) -> impl Iterator<Item = (LI, RI)>
+where
+    L: Iterator<Item = LI>, // + Sorted
+    R: Iterator<Item = RI>, // + Sorted
+    FL: 'static + Fn(&LI) -> K,
+    FR: 'static + Fn(&RI) -> K,
+    LI: Copy,
+    RI: Copy,
+    K: PartialEq + Eq + Ord,
+{
+    let mut l = l.map(move |i| (i, key_l(&i))).peekable();
+    let mut r = r.map(move |i| (i, key_r(&i))).peekable();
+    std::iter::from_fn(move || {
+        loop {
+            match (l.peek(), r.peek()) {
+                (Some((_, lk)), Some((_, rk))) if lk < rk => {
+                    drop(l.next());
+                    continue;
+                }
+                (Some((_, lk)), Some((_, rk))) if lk > rk => {
+                    drop(r.next());
+                    continue;
+                }
+                (Some((lv, lk)), Some((_, rk))) => {
+                    assert!(lk == rk);
+                    let (rv, _) = r.next().unwrap();
+                    return Some((lv.clone(), rv));
+                }
+                (None, None) | (None, Some(_)) | (Some(_), None) => return None,
+            }
+        }
+    })
+}
+
+pub fn left_equi_join_with_merge_strategy<L, LI, R, RI, K, FL, FR>(
+    l: L,
+    r: R,
+    key_l: FL,
+    key_r: FR,
+) -> impl Iterator<Item = (LI, Option<RI>)>
+where
+    L: Iterator<Item = LI>, // + Sorted
+    R: Iterator<Item = RI>, // + Sorted
+    FL: 'static + Fn(&LI) -> K,
+    FR: 'static + Fn(&RI) -> K,
+    LI: Copy,
+    RI: Copy,
+    K: PartialEq + Eq + Ord,
+{
+    let mut l = l.map(move |i| (i, key_l(&i))).peekable();
+    let mut r = r.map(move |i| (i, key_r(&i))).peekable();
+    let mut l_had_match = false;
+    std::iter::from_fn(move || {
+        loop {
+            match (l.peek(), r.peek()) {
+                (Some((_, lk)), Some((_, rk))) if lk < rk => {
+                    let (lv, _) = l.next().unwrap();
+                    if l_had_match {
+                        l_had_match = false;
+                        continue;
+                    } else {
+                        return Some((lv, None));
+                    }
+                }
+                (Some((_, _)), None) => {
+                    let (lv, _) = l.next().unwrap();
+                    if l_had_match {
+                        l_had_match = false;
+                        continue;
+                    } else {
+                        return Some((lv, None));
+                    }
+                }
+                (Some((_, lk)), Some((_, rk))) if lk > rk => {
+                    drop(r.next());
+                    continue;
+                }
+                (Some((lv, lk)), Some((_, rk))) => {
+                    l_had_match = true;
+                    assert!(lk == rk);
+                    let (rv, _) = r.next().unwrap();
+                    return Some((lv.clone(), Some(rv)));
+                }
+                (None, None) | (None, Some(_)) => return None,
+            }
+        }
+    })
+}
+#[cfg(test)]
+mod tests {
+
+    #[test]
+    fn inner_equi_basic() {
+        let l = vec![b"a", b"c"];
+        let r = vec![b"aa", b"ad", b"ba", b"bb", b"ca", b"cb", b"cd", b"dd"];
+
+        let res: Vec<_> = super::inner_equi_join_with_merge_strategy(
+            l.into_iter(),
+            r.into_iter(),
+            |l| &l[0..1],
+            |r| &r[0..1],
+        )
+        .collect();
+
+        assert_eq!(
+            res,
+            vec![
+                (b"a", b"aa"),
+                (b"a", b"ad"),
+                (b"c", b"ca"),
+                (b"c", b"cb"),
+                (b"c", b"cd"),
+            ]
+        );
+    }
+
+    #[test]
+    fn left_equi_basic() {
+        /*
+        create table aleft (id text, aleft text);
+        create table aright (id text, aright text);
+        insert into aleft values ('a', 'a'), ('b', 'b');
+        insert into aright values ('a', 'aa'), ('a', 'ab'), ('c', 'cd');
+        select * from aleft left join aright using ("id");
+        */
+
+        let l = vec![b"a", b"b"];
+        let r = vec![b"aa", b"ab", b"cd"];
+
+        let res: Vec<_> = super::left_equi_join_with_merge_strategy(
+            l.into_iter(),
+            r.into_iter(),
+            |l| &l[0..1],
+            |r| &r[0..1],
+        )
+        .collect();
+
+        assert_eq!(
+            res,
+            vec![(b"a", Some(b"aa")), (b"a", Some(b"ab")), (b"b", None)]
+        );
+    }
+
+    #[test]
+    fn left_equi_basic_2() {
+        let l = vec![b"b"];
+        let r = vec![b"aa", b"ab", b"bb"];
+
+        let res: Vec<_> = super::left_equi_join_with_merge_strategy(
+            l.into_iter(),
+            r.into_iter(),
+            |l| &l[0..1],
+            |r| &r[0..1],
+        )
+        .collect();
+
+        assert_eq!(res, vec![(b"b", Some(b"bb"))])
+    }
+}

libs/utils/src/shard.rs

@@ -52,6 +52,7 @@ pub struct TenantShardId {
 impl ShardCount {
     pub const MAX: Self = Self(u8::MAX);
     pub const MIN: Self = Self(0);
+    pub const RANGE: RangeInclusive<Self> = RangeInclusive::new(Self::MIN, Self::MAX);
 
     /// The internal value of a ShardCount may be zero, which means "1 shard, but use
     /// legacy format for TenantShardId that excludes the shard suffix", also known
@@ -85,7 +86,9 @@ impl ShardCount {
 }
 
 impl ShardNumber {
+    pub const MIN: Self = Self(0);
     pub const MAX: Self = Self(u8::MAX);
+    pub const RANGE: RangeInclusive<Self> = RangeInclusive::new(Self::MIN, Self::MAX);
 }
 
 impl TenantShardId {
@@ -100,16 +103,17 @@ impl TenantShardId {
     /// The range of all TenantShardId that belong to a particular TenantId.  This is useful when
     /// you have a BTreeMap of TenantShardId, and are querying by TenantId.
     pub fn tenant_range(tenant_id: TenantId) -> RangeInclusive<Self> {
+        let shard_index_range: RangeInclusive<_> = ShardIndex::RANGE;
         RangeInclusive::new(
             Self {
                 tenant_id,
-                shard_number: ShardNumber(0),
-                shard_count: ShardCount(0),
+                shard_number: shard_index_range.start().shard_number,
+                shard_count: shard_index_range.start().shard_count,
             },
             Self {
                 tenant_id,
-                shard_number: ShardNumber::MAX,
-                shard_count: ShardCount::MAX,
+                shard_number: shard_index_range.end().shard_number,
+                shard_count: shard_index_range.end().shard_count,
             },
         )
     }
@@ -241,6 +245,16 @@ impl From<[u8; 18]> for TenantShardId {
 }
 
 impl ShardIndex {
+    pub const MIN: Self = ShardIndex {
+        shard_number: ShardNumber::MIN,
+        shard_count: ShardCount::MIN,
+    };
+    pub const MAX: Self = ShardIndex {
+        shard_number: ShardNumber::MAX,
+        shard_count: ShardCount::MAX,
+    };
+    pub const RANGE: RangeInclusive<Self> = RangeInclusive::new(Self::MIN, Self::MAX);
+
     pub fn new(number: ShardNumber, count: ShardCount) -> Self {
         Self {
             shard_number: number,

libs/utils/src/sync.rs

@@ -4,3 +4,5 @@ pub mod duplex;
 pub mod gate;
 
 pub mod spsc_fold;
+
+pub mod spsc_watch;

libs/utils/src/sync/spsc_fold.rs

@@ -56,7 +56,7 @@ impl<T: Send> Sender<T> {
     /// # Panics
     ///
     /// If `try_fold` panics,  any subsequent call to `send` panic.
-    pub async fn send<F>(&mut self, value: T, try_fold: F) -> Result<(), SendError>
+    pub async fn send<F>(&mut self, value: T, try_fold: F) -> Result<(), (T, SendError)>
     where
         F: Fn(&mut T, T) -> Result<(), T>,
     {
@@ -104,7 +104,9 @@ impl<T: Send> Sender<T> {
                     }
                     Poll::Pending
                 }
-                State::ReceiverGone => Poll::Ready(Err(SendError::ReceiverGone)),
+                State::ReceiverGone => {
+                    Poll::Ready(Err((value.take().unwrap(), SendError::ReceiverGone)))
+                }
                 State::SenderGone(_)
                 | State::AllGone
                 | State::SenderDropping

libs/utils/src/sync/spsc_watch.rs

@@ -0,0 +1,55 @@
+//! watch is probably not the right word, because we do take out
+
+use tokio_util::sync::CancellationToken;
+
+use crate::sync::spsc_fold;
+
+pub fn channel<T: Send>() -> (Sender<T>, Receiver<T>) {
+    let (tx, rx) = spsc_fold::channel();
+    let cancel = CancellationToken::new();
+    (
+        Sender {
+            tx,
+            _cancel: cancel.clone().drop_guard(),
+        },
+        Receiver { rx, cancel },
+    )
+}
+
+pub struct Sender<T> {
+    tx: spsc_fold::Sender<T>,
+    _cancel: tokio_util::sync::DropGuard,
+}
+
+pub struct Receiver<T> {
+    rx: spsc_fold::Receiver<T>,
+    cancel: CancellationToken,
+}
+
+impl<T: Send> Sender<T> {
+    pub fn send_replace(&mut self, value: T) -> Result<(), (T, spsc_fold::SendError)> {
+        poll_ready(self.tx.send(value, |old, new| {
+            *old = new;
+            Ok(())
+        }))
+    }
+}
+
+impl<T: Send> Receiver<T> {
+    pub async fn recv(&mut self) -> Result<T, spsc_fold::RecvError> {
+        self.rx.recv().await
+    }
+    pub async fn cancelled(&mut self) {
+        self.cancel.cancelled().await
+    }
+}
+
+fn poll_ready<F: Future<Output = O>, O>(f: F) -> O {
+    futures::executor::block_on(async move {
+        let f = std::pin::pin!(f);
+        match futures::poll!(f) {
+            std::task::Poll::Ready(r) => r,
+            std::task::Poll::Pending => unreachable!("expecting future to always return Ready"),
+        }
+    })
+}

libs/utils/src/tracing_span_assert.rs

@@ -127,12 +127,12 @@ macro_rules! __check_fields_present {
 
             match check_fields_present0($extractors) {
                 Ok(FoundEverything) => Ok(()),
-                Ok(Unconfigured) if cfg!(test) => {
+                Ok(Unconfigured) if cfg!(feature = "testing") => {
                     // allow unconfigured in tests
                     Ok(())
                 },
                 Ok(Unconfigured) => {
-                    panic!("utils::tracing_span_assert: outside of #[cfg(test)] expected tracing to be configured with tracing_error::ErrorLayer")
+                    panic!(r#"utils::tracing_span_assert: outside of #[cfg(feature = "testing")] expected tracing to be configured with tracing_error::ErrorLayer"#)
                 },
                 Err(missing) => Err(missing)
             }

pageserver/Cargo.toml

@@ -96,6 +96,7 @@ strum.workspace = true
 strum_macros.workspace = true
 wal_decoder.workspace = true
 smallvec.workspace = true
+twox-hash.workspace = true
 
 [target.'cfg(target_os = "linux")'.dependencies]
 procfs.workspace = true

pageserver/client/src/mgmt_api.rs

@@ -1,5 +1,6 @@
 use std::collections::HashMap;
 use std::error::Error as _;
+use std::time::Duration;
 
 use bytes::Bytes;
 use detach_ancestor::AncestorDetached;
@@ -819,4 +820,25 @@ impl Client {
             .await
             .map(|resp| resp.status())
     }
+
+    pub async fn activate_post_import(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        activate_timeline_timeout: Duration,
+    ) -> Result<TimelineInfo> {
+        let uri = format!(
+            "{}/v1/tenant/{}/timeline/{}/activate_post_import?timeline_activate_timeout_ms={}",
+            self.mgmt_api_endpoint,
+            tenant_shard_id,
+            timeline_id,
+            activate_timeline_timeout.as_millis()
+        );
+
+        self.request(Method::PUT, uri, ())
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
 }

pageserver/page_api/Cargo.toml

@@ -0,0 +1,13 @@
+[package]
+name = "pageserver_page_api"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+prost.workspace = true
+tonic.workspace = true
+workspace_hack.workspace = true
+
+[build-dependencies]
+tonic-build.workspace = true

pageserver/page_api/build.rs

@@ -0,0 +1,7 @@
+fn main() -> Result<(), Box<dyn std::error::Error>> {
+    // Generates Rust code from .proto Protobuf schemas.
+    tonic_build::configure()
+        .bytes(["."])
+        .compile_protos(&["proto/page_service.proto"], &["proto"])
+        .map_err(|err| err.into())
+}

pageserver/page_api/proto/page_service.proto

@@ -0,0 +1,220 @@
+// Page service, presented by pageservers for computes.
+//
+// This is the compute read path. It primarily serves page versions at given
+// LSNs, but also base backups, SLRU segments, and relation metadata.
+//
+// EXPERIMENTAL: this is still under development and subject to change.
+//
+// Request metadata headers:
+// - authorization: JWT token ("Bearer <token>"), if auth is enabled
+// - neon-tenant-id: tenant ID ("7c4a1f9e3bd6470c8f3e21a65bd2e980")
+// - neon-shard-id: shard ID, as <number><count> in hex ("0b10" = shard 11 of 16, 0-based)
+// - neon-timeline-id: timeline ID ("f08c4e9a2d5f76b1e3a7c2d8910f4b3e")
+//
+// TODO: consider adding neon-compute-mode ("primary", "static", "replica").
+// However, this will require reconnecting when changing modes.
+//
+// TODO: write implementation guidance on
+// - Health checks
+// - Tracing, OpenTelemetry
+// - Compression
+
+syntax = "proto3";
+package page_service;
+
+service PageService {
+  // Returns whether a relation exists.
+  rpc CheckRelExists(CheckRelExistsRequest) returns (CheckRelExistsResponse);
+
+  // Fetches a base backup.
+  rpc GetBaseBackup (GetBaseBackupRequest) returns (stream GetBaseBackupResponseChunk);
+
+  // Returns the total size of a database, as # of bytes.
+  rpc GetDbSize (GetDbSizeRequest) returns (GetDbSizeResponse);
+
+  // Fetches pages.
+  //
+  // This is implemented as a bidirectional streaming RPC for performance. Unary
+  // requests incur costs for e.g. HTTP/2 stream setup, header parsing,
+  // authentication, and so on -- with streaming, we only pay these costs during
+  // the initial stream setup. This ~doubles throughput in benchmarks. Other
+  // RPCs use regular unary requests, since they are not as frequent and
+  // performance-critical, and this simplifies implementation.
+  //
+  // NB: a status response (e.g. errors) will terminate the stream. The stream
+  // may be shared by e.g. multiple Postgres backends, so we should avoid this.
+  // Most errors are therefore sent as GetPageResponse.status instead.
+  rpc GetPages (stream GetPageRequest) returns (stream GetPageResponse);
+
+  // Returns the size of a relation, as # of blocks.
+  rpc GetRelSize (GetRelSizeRequest) returns (GetRelSizeResponse);
+
+  // Fetches an SLRU segment.
+  rpc GetSlruSegment (GetSlruSegmentRequest) returns (GetSlruSegmentResponse);
+}
+
+// The LSN a request should read at.
+message ReadLsn {
+  // The request's read LSN. Required.
+  uint64 request_lsn = 1;
+  // If given, the caller guarantees that the page has not been modified since
+  // this LSN. Must be smaller than or equal to request_lsn. This allows the
+  // Pageserver to serve an old page without waiting for the request LSN to
+  // arrive. Valid for all request types.
+  //
+  // It is undefined behaviour to make a request such that the page was, in
+  // fact, modified between request_lsn and not_modified_since_lsn. The
+  // Pageserver might detect it and return an error, or it might return the old
+  // page version or the new page version. Setting not_modified_since_lsn equal
+  // to request_lsn is always safe, but can lead to unnecessary waiting.
+  uint64 not_modified_since_lsn = 2;
+}
+
+// A relation identifier.
+message RelTag {
+    uint32 spc_oid = 1;
+    uint32 db_oid = 2;
+    uint32 rel_number = 3;
+    uint32 fork_number = 4;
+}
+
+// Checks whether a relation exists, at the given LSN. Only valid on shard 0,
+// other shards will error.
+message CheckRelExistsRequest {
+  ReadLsn read_lsn = 1;
+  RelTag rel = 2;
+}
+
+message CheckRelExistsResponse {
+  bool exists = 1;
+}
+
+// Requests a base backup at a given LSN.
+message GetBaseBackupRequest {
+  // The LSN to fetch a base backup at.
+  ReadLsn read_lsn = 1;
+  // If true, logical replication slots will not be created.
+  bool replica = 2;
+}
+
+// Base backup response chunk, returned as an ordered stream.
+message GetBaseBackupResponseChunk {
+  // A basebackup data chunk. The size is undefined, but bounded by the 4 MB
+  // gRPC message size limit.
+  bytes chunk = 1;
+}
+
+// Requests the size of a database, as # of bytes. Only valid on shard 0, other
+// shards will error.
+message GetDbSizeRequest {
+  ReadLsn read_lsn = 1;
+  uint32 db_oid = 2;
+}
+
+message GetDbSizeResponse {
+  uint64 num_bytes = 1;
+}
+
+// Requests one or more pages.
+message GetPageRequest {
+  // A request ID. Will be included in the response. Should be unique for
+  // in-flight requests on the stream.
+  uint64 request_id = 1;
+  // The request class.
+  GetPageClass request_class = 2;
+  // The LSN to read at.
+  ReadLsn read_lsn = 3;
+  // The relation to read from.
+  RelTag rel = 4;
+  // Page numbers to read. Must belong to the remote shard.
+  //
+  // Multiple pages will be executed as a single batch by the Pageserver,
+  // amortizing layer access costs and parallelizing them. This may increase the
+  // latency of any individual request, but improves the overall latency and
+  // throughput of the batch as a whole.
+  //
+  // TODO: this causes an allocation in the common single-block case. The sender
+  // can use a SmallVec to stack-allocate it, but Prost will always deserialize
+  // into a heap-allocated Vec. Consider optimizing this.
+  //
+  // TODO: we might be able to avoid a sort or something if we mandate that these
+  // are always in order. But we can't currenly rely on this on the server, because
+  // of compatibility with the libpq protocol handler.
+  repeated uint32 block_number = 5;
+}
+
+// A GetPageRequest class. Primarily intended for observability, but may also be
+// used for prioritization in the future.
+enum GetPageClass {
+  // Unknown class. For forwards compatibility: used when the client sends a
+  // class that the server doesn't know about.
+  GET_PAGE_CLASS_UNKNOWN = 0;
+  // A normal request. This is the default.
+  GET_PAGE_CLASS_NORMAL = 1;
+  // A prefetch request. NB: can only be classified on pg < 18.
+  GET_PAGE_CLASS_PREFETCH = 2;
+  // A background request (e.g. vacuum).
+  GET_PAGE_CLASS_BACKGROUND = 3;
+}
+
+// A GetPage response.
+//
+// A batch response will contain all of the requested pages. We could eagerly
+// emit individual pages as soon as they are ready, but on a readv() Postgres
+// holds buffer pool locks on all pages in the batch and we'll only return once
+// the entire batch is ready, so no one can make use of the individual pages.
+message GetPageResponse {
+  // The original request's ID.
+  uint64 request_id = 1;
+  // The response status code.
+  GetPageStatus status = 2;
+  // A string describing the status, if any.
+  string reason = 3;
+  // The 8KB page images, in the same order as the request. Empty if status != OK.
+  repeated bytes page_image = 4;
+}
+
+// A GetPageResponse status code. Since we use a bidirectional stream, we don't
+// want to send errors as gRPC statuses, since this would terminate the stream.
+enum GetPageStatus {
+  // Unknown status. For forwards compatibility: used when the server sends a
+  // status code that the client doesn't know about.
+  GET_PAGE_STATUS_UNKNOWN = 0;
+  // The request was successful.
+  GET_PAGE_STATUS_OK = 1;
+  // The page did not exist. The tenant/timeline/shard has already been
+  // validated during stream setup.
+  GET_PAGE_STATUS_NOT_FOUND = 2;
+  // The request was invalid.
+  GET_PAGE_STATUS_INVALID = 3;
+  // The tenant is rate limited. Slow down and retry later.
+  GET_PAGE_STATUS_SLOW_DOWN = 4;
+  // TODO: consider adding a GET_PAGE_STATUS_LAYER_DOWNLOAD in the case of a
+  // layer download. This could free up the server task to process other
+  // requests while the layer download is in progress.
+}
+
+// Fetches the size of a relation at a given LSN, as # of blocks. Only valid on
+// shard 0, other shards will error.
+message GetRelSizeRequest {
+  ReadLsn read_lsn = 1;
+  RelTag rel = 2;
+}
+
+message GetRelSizeResponse {
+  uint32 num_blocks = 1;
+}
+
+// Requests an SLRU segment. Only valid on shard 0, other shards will error.
+message GetSlruSegmentRequest {
+  ReadLsn read_lsn = 1;
+  uint32 kind = 2;
+  uint32 segno = 3;
+}
+
+// Returns an SLRU segment.
+//
+// These are up 32 pages (256 KB), so we can send them as a single response.
+message GetSlruSegmentResponse {
+  bytes segment = 1;
+}

pageserver/page_api/src/lib.rs

@@ -0,0 +1,14 @@
+//! This crate provides the Pageserver's page API. It contains:
+//!
+//! * proto/page_service.proto: the Protobuf schema for the page API.
+//! * proto: auto-generated Protobuf types for gRPC.
+//!
+//! This crate is used by both the client and the server. Try to keep it slim.
+
+// Code generated by protobuf.
+pub mod proto {
+    tonic::include_proto!("page_service");
+
+    pub use page_service_client::PageServiceClient;
+    pub use page_service_server::{PageService, PageServiceServer};
+}

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -65,6 +65,9 @@ pub(crate) struct Args {
     #[clap(long, default_value = "1")]
     queue_depth: NonZeroUsize,
 
+    #[clap(long)]
+    only_relnode: Option<u32>,
+
     targets: Option<Vec<TenantTimelineId>>,
 }
 
@@ -206,7 +209,12 @@ async fn main_impl(
                     for r in partitioning.keys.ranges.iter() {
                         let mut i = r.start;
                         while i != r.end {
-                            if i.is_rel_block_key() {
+                            let mut include = true;
+                            include &= i.is_rel_block_key();
+                            if let Some(only_relnode) = args.only_relnode {
+                                include &= i.is_rel_block_of_rel(only_relnode);
+                            }
+                            if include {
                                 filtered.add_key(i);
                             }
                             i = i.next();

pageserver/src/bin/pageserver.rs

@@ -423,11 +423,14 @@ fn start_pageserver(
                     .map(storage_broker::Certificate::from_pem),
             );
             // Note: we do not attempt connecting here (but validate endpoints sanity).
-            storage_broker::connect(
+            let service_client = storage_broker::connect(
                 conf.broker_endpoint.clone(),
                 conf.broker_keepalive_interval,
                 tls_config,
-            )
+            )?;
+            anyhow::Ok(storage_broker::TimelineUpdatesSubscriber::new(
+                service_client,
+            ))
         })
         .with_context(|| {
             format!(

pageserver/src/controller_upcall_client.rs

@@ -7,7 +7,7 @@ use pageserver_api::models::ShardImportStatus;
 use pageserver_api::shard::TenantShardId;
 use pageserver_api::upcall_api::{
     PutTimelineImportStatusRequest, ReAttachRequest, ReAttachResponse, ReAttachResponseTenant,
-    ValidateRequest, ValidateRequestTenant, ValidateResponse,
+    TimelineImportStatusRequest, ValidateRequest, ValidateRequestTenant, ValidateResponse,
 };
 use reqwest::Certificate;
 use serde::Serialize;
@@ -51,8 +51,15 @@ pub trait StorageControllerUpcallApi {
         &self,
         tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
+        generation: Generation,
         status: ShardImportStatus,
     ) -> impl Future<Output = Result<(), RetryForeverError>> + Send;
+    fn get_timeline_import_status(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        generation: Generation,
+    ) -> impl Future<Output = Result<ShardImportStatus, RetryForeverError>> + Send;
 }
 
 impl StorageControllerUpcallClient {
@@ -97,6 +104,7 @@ impl StorageControllerUpcallClient {
         &self,
         url: &url::Url,
         request: R,
+        method: reqwest::Method,
     ) -> Result<T, RetryForeverError>
     where
         R: Serialize,
@@ -106,7 +114,7 @@ impl StorageControllerUpcallClient {
             || async {
                 let response = self
                     .http_client
-                    .post(url.clone())
+                    .request(method.clone(), url.clone())
                     .json(&request)
                     .send()
                     .await?;
@@ -215,7 +223,9 @@ impl StorageControllerUpcallApi for StorageControllerUpcallClient {
             register: register.clone(),
         };
 
-        let response: ReAttachResponse = self.retry_http_forever(&url, request).await?;
+        let response: ReAttachResponse = self
+            .retry_http_forever(&url, request, reqwest::Method::POST)
+            .await?;
         tracing::info!(
             "Received re-attach response with {} tenants (node {}, register: {:?})",
             response.tenants.len(),
@@ -268,7 +278,9 @@ impl StorageControllerUpcallApi for StorageControllerUpcallClient {
                 return Err(RetryForeverError::ShuttingDown);
             }
 
-            let response: ValidateResponse = self.retry_http_forever(&url, request).await?;
+            let response: ValidateResponse = self
+                .retry_http_forever(&url, request, reqwest::Method::POST)
+                .await?;
             for rt in response.tenants {
                 result.insert(rt.id, rt.valid);
             }
@@ -287,6 +299,7 @@ impl StorageControllerUpcallApi for StorageControllerUpcallClient {
         &self,
         tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
+        generation: Generation,
         status: ShardImportStatus,
     ) -> Result<(), RetryForeverError> {
         let url = self
@@ -297,9 +310,35 @@ impl StorageControllerUpcallApi for StorageControllerUpcallClient {
         let request = PutTimelineImportStatusRequest {
             tenant_shard_id,
             timeline_id,
+            generation,
             status,
         };
 
-        self.retry_http_forever(&url, request).await
+        self.retry_http_forever(&url, request, reqwest::Method::POST)
+            .await
+    }
+
+    #[tracing::instrument(skip_all)] // so that warning logs from retry_http_forever have context
+    async fn get_timeline_import_status(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        generation: Generation,
+    ) -> Result<ShardImportStatus, RetryForeverError> {
+        let url = self
+            .base_url
+            .join("timeline_import_status")
+            .expect("Failed to build path");
+
+        let request = TimelineImportStatusRequest {
+            tenant_shard_id,
+            timeline_id,
+            generation,
+        };
+
+        let response: ShardImportStatus = self
+            .retry_http_forever(&url, request, reqwest::Method::GET)
+            .await?;
+        Ok(response)
     }
 }

pageserver/src/deletion_queue.rs

@@ -663,6 +663,7 @@ mod test {
     use camino::Utf8Path;
     use hex_literal::hex;
     use pageserver_api::key::Key;
+    use pageserver_api::models::ShardImportStatus;
     use pageserver_api::shard::ShardIndex;
     use pageserver_api::upcall_api::ReAttachResponseTenant;
     use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
@@ -792,10 +793,20 @@ mod test {
             &self,
             _tenant_shard_id: TenantShardId,
             _timeline_id: TimelineId,
+            _generation: Generation,
             _status: pageserver_api::models::ShardImportStatus,
         ) -> Result<(), RetryForeverError> {
             unimplemented!()
         }
+
+        async fn get_timeline_import_status(
+            &self,
+            _tenant_shard_id: TenantShardId,
+            _timeline_id: TimelineId,
+            _generation: Generation,
+        ) -> Result<ShardImportStatus, RetryForeverError> {
+            unimplemented!()
+        }
     }
 
     async fn setup(test_name: &str) -> anyhow::Result<TestSetup> {

pageserver/src/http/routes.rs

@@ -100,7 +100,7 @@ pub struct State {
     auth: Option<Arc<SwappableJwtAuth>>,
     allowlist_routes: &'static [&'static str],
     remote_storage: GenericRemoteStorage,
-    broker_client: storage_broker::BrokerClientChannel,
+    broker_client: storage_broker::TimelineUpdatesSubscriber,
     disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
     deletion_queue_client: DeletionQueueClient,
     secondary_controller: SecondaryController,
@@ -114,7 +114,7 @@ impl State {
         tenant_manager: Arc<TenantManager>,
         auth: Option<Arc<SwappableJwtAuth>>,
         remote_storage: GenericRemoteStorage,
-        broker_client: storage_broker::BrokerClientChannel,
+        broker_client: storage_broker::TimelineUpdatesSubscriber,
         disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
         deletion_queue_client: DeletionQueueClient,
         secondary_controller: SecondaryController,
@@ -3500,6 +3500,107 @@ async fn put_tenant_timeline_import_wal(
     }.instrument(span).await
 }
 
+/// Activate a timeline after its import has completed
+///
+/// The endpoint is idempotent and callers are expected to retry all
+/// errors until a successful response.
+async fn activate_post_import_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    const DEFAULT_ACTIVATE_TIMEOUT: Duration = Duration::from_secs(1);
+    let activate_timeout = parse_query_param(&request, "timeline_activate_timeout_ms")?
+        .map(Duration::from_millis)
+        .unwrap_or(DEFAULT_ACTIVATE_TIMEOUT);
+
+    let span = info_span!(
+        "activate_post_import_handler",
+        tenant_id=%tenant_shard_id.tenant_id,
+        timeline_id=%timeline_id,
+        shard_id=%tenant_shard_id.shard_slug()
+    );
+
+    async move {
+        let state = get_state(&request);
+        let tenant = state
+            .tenant_manager
+            .get_attached_tenant_shard(tenant_shard_id)?;
+
+        tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+
+        tenant
+            .finalize_importing_timeline(timeline_id)
+            .await
+            .map_err(ApiError::InternalServerError)?;
+
+        match tenant.get_timeline(timeline_id, false) {
+            Ok(_timeline) => {
+                // Timeline is already visible. Reset not required: fall through.
+            }
+            Err(GetTimelineError::NotFound { .. }) => {
+                // This is crude: we reset the whole tenant such that the new timeline is detected
+                // and activated. We can come up with something more granular in the future.
+                //
+                // Note that we only reset the tenant if required: when the timeline is
+                // not present in [`Tenant::timelines`].
+                let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
+                state
+                    .tenant_manager
+                    .reset_tenant(tenant_shard_id, false, &ctx)
+                    .await
+                    .map_err(ApiError::InternalServerError)?;
+            }
+            Err(GetTimelineError::ShuttingDown) => {
+                return Err(ApiError::ShuttingDown);
+            }
+            Err(GetTimelineError::NotActive { .. }) => {
+                unreachable!("Called get_timeline with active_only=false");
+            }
+        }
+
+        let timeline = tenant.get_timeline(timeline_id, false)?;
+
+        let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn)
+            .with_scope_timeline(&timeline);
+
+        let result =
+            tokio::time::timeout(activate_timeout, timeline.wait_to_become_active(&ctx)).await;
+        match result {
+            Ok(Ok(())) => {
+                // fallthrough
+            }
+            // Timeline reached some other state that's not active
+            // TODO(vlad): if the tenant is broken, return a permananet error
+            Ok(Err(_timeline_state)) => {
+                return Err(ApiError::InternalServerError(anyhow::anyhow!(
+                    "Timeline activation failed"
+                )));
+            }
+            // Activation timed out
+            Err(_) => {
+                return Err(ApiError::Timeout("Timeline activation timed out".into()));
+            }
+        }
+
+        let timeline_info = build_timeline_info(
+            &timeline, false, // include_non_incremental_logical_size,
+            false, // force_await_initial_logical_size
+            &ctx,
+        )
+        .await
+        .context("get local timeline info")
+        .map_err(ApiError::InternalServerError)?;
+
+        json_response(StatusCode::OK, timeline_info)
+    }
+    .instrument(span)
+    .await
+}
+
 /// Read the end of a tar archive.
 ///
 /// A tar archive normally ends with two consecutive blocks of zeros, 512 bytes each.
@@ -3924,5 +4025,9 @@ pub fn make_router(
             "/v1/tenant/:tenant_id/timeline/:timeline_id/import_wal",
             |r| api_handler(r, put_tenant_timeline_import_wal),
         )
+        .put(
+            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/activate_post_import",
+            |r| api_handler(r, activate_post_import_handler),
+        )
         .any(handler_404))
 }

pageserver/src/page_service.rs

@@ -1278,7 +1278,7 @@ impl PageServerHandler {
     }
 
     #[instrument(level = tracing::Level::DEBUG, skip_all)]
-    async fn pagesteam_handle_batched_message<IO>(
+    async fn pagestream_handle_batched_message<IO>(
         &mut self,
         pgb_writer: &mut PostgresBackend<IO>,
         batch: BatchedFeMessage,
@@ -1733,7 +1733,7 @@ impl PageServerHandler {
             };
 
             let result = self
-                .pagesteam_handle_batched_message(
+                .pagestream_handle_batched_message(
                     pgb_writer,
                     msg,
                     io_concurrency.clone(),
@@ -1909,7 +1909,7 @@ impl PageServerHandler {
                             return Err(e);
                         }
                     };
-                    self.pagesteam_handle_batched_message(
+                    self.pagestream_handle_batched_message(
                         pgb_writer,
                         batch,
                         io_concurrency.clone(),

pageserver/src/tenant.rs

@@ -48,8 +48,8 @@ use remote_timeline_client::{
     download_tenant_manifest,
 };
 use secondary::heatmap::{HeatMapTenant, HeatMapTimeline};
-use storage_broker::BrokerClientChannel;
 use timeline::compaction::{CompactionOutcome, GcCompactionQueue};
+use timeline::import_pgdata::ImportingTimeline;
 use timeline::offload::{OffloadError, offload_timeline};
 use timeline::{
     CompactFlags, CompactOptions, CompactionError, PreviousHeatmap, ShutdownMode, import_pgdata,
@@ -152,7 +152,7 @@ pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
 /// as the shared remote storage client and process initialization state.
 #[derive(Clone)]
 pub struct TenantSharedResources {
-    pub broker_client: storage_broker::BrokerClientChannel,
+    pub broker_client: storage_broker::TimelineUpdatesSubscriber,
     pub remote_storage: GenericRemoteStorage,
     pub deletion_queue_client: DeletionQueueClient,
     pub l0_flush_global_state: L0FlushGlobalState,
@@ -284,6 +284,19 @@ pub struct TenantShard {
     /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
     timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
 
+    /// Tracks the timelines that are currently importing into this tenant shard.
+    ///
+    /// Note that importing timelines are also present in [`Self::timelines_creating`].
+    /// Keep this in mind when ordering lock acquisition.
+    ///
+    /// Lifetime:
+    /// * An imported timeline is created while scanning the bucket on tenant attach
+    ///   if the index part contains an `import_pgdata` entry and said field marks the import
+    ///   as in progress.
+    /// * Imported timelines are removed when the storage controller calls the post timeline
+    ///   import activation endpoint.
+    timelines_importing: std::sync::Mutex<HashMap<TimelineId, ImportingTimeline>>,
+
     /// The last tenant manifest known to be in remote storage. None if the manifest has not yet
     /// been either downloaded or uploaded. Always Some after tenant attach.
     ///
@@ -923,19 +936,10 @@ enum StartCreatingTimelineResult {
 
 #[allow(clippy::large_enum_variant, reason = "TODO")]
 enum TimelineInitAndSyncResult {
-    ReadyToActivate(Arc<Timeline>),
+    ReadyToActivate,
     NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
 }
 
-impl TimelineInitAndSyncResult {
-    fn ready_to_activate(self) -> Option<Arc<Timeline>> {
-        match self {
-            Self::ReadyToActivate(timeline) => Some(timeline),
-            _ => None,
-        }
-    }
-}
-
 #[must_use]
 struct TimelineInitAndSyncNeedsSpawnImportPgdata {
     timeline: Arc<Timeline>,
@@ -1012,10 +1016,6 @@ enum CreateTimelineCause {
 enum LoadTimelineCause {
     Attach,
     Unoffload,
-    ImportPgdata {
-        create_guard: TimelineCreateGuard,
-        activate: ActivateTimelineArgs,
-    },
 }
 
 #[derive(thiserror::Error, Debug)]
@@ -1097,7 +1097,7 @@ impl TenantShard {
         self: &Arc<Self>,
         timeline_id: TimelineId,
         resources: TimelineResources,
-        mut index_part: IndexPart,
+        index_part: IndexPart,
         metadata: TimelineMetadata,
         previous_heatmap: Option<PreviousHeatmap>,
         ancestor: Option<Arc<Timeline>>,
@@ -1106,7 +1106,7 @@ impl TenantShard {
     ) -> anyhow::Result<TimelineInitAndSyncResult> {
         let tenant_id = self.tenant_shard_id;
 
-        let import_pgdata = index_part.import_pgdata.take();
+        let import_pgdata = index_part.import_pgdata.clone();
         let idempotency = match &import_pgdata {
             Some(import_pgdata) => {
                 CreateTimelineIdempotency::ImportPgdata(CreatingTimelineIdempotencyImportPgdata {
@@ -1127,7 +1127,7 @@ impl TenantShard {
             }
         };
 
-        let (timeline, timeline_ctx) = self.create_timeline_struct(
+        let (timeline, _timeline_ctx) = self.create_timeline_struct(
             timeline_id,
             &metadata,
             previous_heatmap,
@@ -1197,14 +1197,6 @@ impl TenantShard {
 
         match import_pgdata {
             Some(import_pgdata) if !import_pgdata.is_done() => {
-                match cause {
-                    LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
-                    LoadTimelineCause::ImportPgdata { .. } => {
-                        unreachable!(
-                            "ImportPgdata should not be reloading timeline import is done and persisted as such in s3"
-                        )
-                    }
-                }
                 let mut guard = self.timelines_creating.lock().unwrap();
                 if !guard.insert(timeline_id) {
                     // We should never try and load the same timeline twice during startup
@@ -1260,26 +1252,7 @@ impl TenantShard {
                     "Timeline has no ancestor and no layer files"
                 );
 
-                match cause {
-                    LoadTimelineCause::Attach | LoadTimelineCause::Unoffload => (),
-                    LoadTimelineCause::ImportPgdata {
-                        create_guard,
-                        activate,
-                    } => {
-                        // TODO: see the comment in the task code above how I'm not so certain
-                        // it is safe to activate here because of concurrent shutdowns.
-                        match activate {
-                            ActivateTimelineArgs::Yes { broker_client } => {
-                                info!("activating timeline after reload from pgdata import task");
-                                timeline.activate(self.clone(), broker_client, None, &timeline_ctx);
-                            }
-                            ActivateTimelineArgs::No => (),
-                        }
-                        drop(create_guard);
-                    }
-                }
-
-                Ok(TimelineInitAndSyncResult::ReadyToActivate(timeline))
+                Ok(TimelineInitAndSyncResult::ReadyToActivate)
             }
         }
     }
@@ -1768,7 +1741,7 @@ impl TenantShard {
                 })?;
 
             match effect {
-                TimelineInitAndSyncResult::ReadyToActivate(_) => {
+                TimelineInitAndSyncResult::ReadyToActivate => {
                     // activation happens later, on Tenant::activate
                 }
                 TimelineInitAndSyncResult::NeedsSpawnImportPgdata(
@@ -1778,13 +1751,24 @@ impl TenantShard {
                         guard,
                     },
                 ) => {
-                    tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
-                        timeline,
-                        import_pgdata,
-                        ActivateTimelineArgs::No,
-                        guard,
-                        ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
-                    ));
+                    let timeline_id = timeline.timeline_id;
+                    let import_task_handle =
+                        tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
+                            timeline.clone(),
+                            import_pgdata,
+                            guard,
+                            ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
+                        ));
+
+                    let prev = self.timelines_importing.lock().unwrap().insert(
+                        timeline_id,
+                        ImportingTimeline {
+                            timeline: timeline.clone(),
+                            import_task_handle,
+                        },
+                    );
+
+                    assert!(prev.is_none());
                 }
             }
         }
@@ -2122,7 +2106,7 @@ impl TenantShard {
     async fn unoffload_timeline(
         self: &Arc<Self>,
         timeline_id: TimelineId,
-        broker_client: storage_broker::BrokerClientChannel,
+        broker_client: storage_broker::TimelineUpdatesSubscriber,
         ctx: RequestContext,
     ) -> Result<Arc<Timeline>, TimelineArchivalError> {
         info!("unoffloading timeline");
@@ -2257,7 +2241,7 @@ impl TenantShard {
         self: &Arc<Self>,
         timeline_id: TimelineId,
         new_state: TimelineArchivalState,
-        broker_client: storage_broker::BrokerClientChannel,
+        broker_client: storage_broker::TimelineUpdatesSubscriber,
         ctx: RequestContext,
     ) -> Result<(), TimelineArchivalError> {
         info!("setting timeline archival config");
@@ -2586,7 +2570,7 @@ impl TenantShard {
     pub(crate) async fn create_timeline(
         self: &Arc<TenantShard>,
         params: CreateTimelineParams,
-        broker_client: storage_broker::BrokerClientChannel,
+        broker_client: storage_broker::TimelineUpdatesSubscriber,
         ctx: &RequestContext,
     ) -> Result<Arc<Timeline>, CreateTimelineError> {
         if !self.is_active() {
@@ -2678,14 +2662,7 @@ impl TenantShard {
                     .await?
             }
             CreateTimelineParams::ImportPgdata(params) => {
-                self.create_timeline_import_pgdata(
-                    params,
-                    ActivateTimelineArgs::Yes {
-                        broker_client: broker_client.clone(),
-                    },
-                    ctx,
-                )
-                .await?
+                self.create_timeline_import_pgdata(params, ctx).await?
             }
         };
 
@@ -2759,7 +2736,6 @@ impl TenantShard {
     async fn create_timeline_import_pgdata(
         self: &Arc<Self>,
         params: CreateTimelineParamsImportPgdata,
-        activate: ActivateTimelineArgs,
         ctx: &RequestContext,
     ) -> Result<CreateTimelineResult, CreateTimelineError> {
         let CreateTimelineParamsImportPgdata {
@@ -2840,24 +2816,71 @@ impl TenantShard {
 
         let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
 
-        tokio::spawn(self.clone().create_timeline_import_pgdata_task(
+        let import_task_handle = tokio::spawn(self.clone().create_timeline_import_pgdata_task(
             timeline.clone(),
             index_part,
-            activate,
             timeline_create_guard,
             timeline_ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
         ));
 
+        let prev = self.timelines_importing.lock().unwrap().insert(
+            timeline.timeline_id,
+            ImportingTimeline {
+                timeline: timeline.clone(),
+                import_task_handle,
+            },
+        );
+
+        // Idempotency is enforced higher up the stack
+        assert!(prev.is_none());
+
         // NB: the timeline doesn't exist in self.timelines at this point
         Ok(CreateTimelineResult::ImportSpawned(timeline))
     }
 
+    /// Finalize the import of a timeline on this shard by marking it complete in
+    /// the index part. If the import task hasn't finished yet, returns an error.
+    ///
+    /// This method is idempotent. If the import was finalized once, the next call
+    /// will be a no-op.
+    pub(crate) async fn finalize_importing_timeline(
+        &self,
+        timeline_id: TimelineId,
+    ) -> anyhow::Result<()> {
+        let timeline = {
+            let locked = self.timelines_importing.lock().unwrap();
+            match locked.get(&timeline_id) {
+                Some(importing_timeline) => {
+                    if !importing_timeline.import_task_handle.is_finished() {
+                        return Err(anyhow::anyhow!("Import task not done yet"));
+                    }
+
+                    importing_timeline.timeline.clone()
+                }
+                None => {
+                    return Ok(());
+                }
+            }
+        };
+
+        timeline
+            .remote_client
+            .schedule_index_upload_for_import_pgdata_finalize()?;
+        timeline.remote_client.wait_completion().await?;
+
+        self.timelines_importing
+            .lock()
+            .unwrap()
+            .remove(&timeline_id);
+
+        Ok(())
+    }
+
     #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline.timeline_id))]
     async fn create_timeline_import_pgdata_task(
         self: Arc<TenantShard>,
         timeline: Arc<Timeline>,
         index_part: import_pgdata::index_part_format::Root,
-        activate: ActivateTimelineArgs,
         timeline_create_guard: TimelineCreateGuard,
         ctx: RequestContext,
     ) {
@@ -2869,7 +2892,6 @@ impl TenantShard {
             .create_timeline_import_pgdata_task_impl(
                 timeline,
                 index_part,
-                activate,
                 timeline_create_guard,
                 ctx,
             )
@@ -2885,60 +2907,15 @@ impl TenantShard {
         self: Arc<TenantShard>,
         timeline: Arc<Timeline>,
         index_part: import_pgdata::index_part_format::Root,
-        activate: ActivateTimelineArgs,
-        timeline_create_guard: TimelineCreateGuard,
+        _timeline_create_guard: TimelineCreateGuard,
         ctx: RequestContext,
     ) -> Result<(), anyhow::Error> {
         info!("importing pgdata");
+        let ctx = ctx.with_scope_timeline(&timeline);
         import_pgdata::doit(&timeline, index_part, &ctx, self.cancel.clone())
             .await
             .context("import")?;
-        info!("import done");
-
-        //
-        // Reload timeline from remote.
-        // This proves that the remote state is attachable, and it reuses the code.
-        //
-        // TODO: think about whether this is safe to do with concurrent TenantShard::shutdown.
-        // timeline_create_guard hols the tenant gate open, so, shutdown cannot _complete_ until we exit.
-        // But our activate() call might launch new background tasks after TenantShard::shutdown
-        // already went past shutting down the TenantShard::timelines, which this timeline here is no part of.
-        // I think the same problem exists with the bootstrap & branch mgmt API tasks (tenant shutting
-        // down while bootstrapping/branching + activating), but, the race condition is much more likely
-        // to manifest because of the long runtime of this import task.
-
-        //        in theory this shouldn't even .await anything except for coop yield
-        info!("shutting down timeline");
-        timeline.shutdown(ShutdownMode::Hard).await;
-        info!("timeline shut down, reloading from remote");
-        // TODO: we can't do the following check because create_timeline_import_pgdata must return an Arc<Timeline>
-        // let Some(timeline) = Arc::into_inner(timeline) else {
-        //     anyhow::bail!("implementation error: timeline that we shut down was still referenced from somewhere");
-        // };
-        let timeline_id = timeline.timeline_id;
-
-        // load from object storage like TenantShard::attach does
-        let resources = self.build_timeline_resources(timeline_id);
-        let index_part = resources
-            .remote_client
-            .download_index_file(&self.cancel)
-            .await?;
-        let index_part = match index_part {
-            MaybeDeletedIndexPart::Deleted(_) => {
-                // likely concurrent delete call, cplane should prevent this
-                anyhow::bail!(
-                    "index part says deleted but we are not done creating yet, this should not happen but"
-                )
-            }
-            MaybeDeletedIndexPart::IndexPart(p) => p,
-        };
-        let metadata = index_part.metadata.clone();
-        self
-            .load_remote_timeline(timeline_id, index_part, metadata, None, resources, LoadTimelineCause::ImportPgdata{
-                create_guard: timeline_create_guard, activate, }, &ctx)
-            .await?
-            .ready_to_activate()
-            .context("implementation error: reloaded timeline still needs import after import reported success")?;
+        info!("import done - waiting for activation");
 
         anyhow::Ok(())
     }
@@ -3321,7 +3298,7 @@ impl TenantShard {
     /// to delay background jobs. Background jobs can be started right away when None is given.
     fn activate(
         self: &Arc<Self>,
-        broker_client: BrokerClientChannel,
+        broker_client: storage_broker::TimelineUpdatesSubscriber,
         background_jobs_can_start: Option<&completion::Barrier>,
         ctx: &RequestContext,
     ) {
@@ -3475,6 +3452,14 @@ impl TenantShard {
                 timeline.defuse_for_tenant_drop();
             });
         }
+        {
+            let mut timelines_importing = self.timelines_importing.lock().unwrap();
+            timelines_importing
+                .drain()
+                .for_each(|(_timeline_id, importing_timeline)| {
+                    importing_timeline.shutdown();
+                });
+        }
         // test_long_timeline_create_then_tenant_delete is leaning on this message
         tracing::info!("Waiting for timelines...");
         while let Some(res) = js.join_next().await {
@@ -3949,13 +3934,6 @@ where
     Ok(result)
 }
 
-enum ActivateTimelineArgs {
-    Yes {
-        broker_client: storage_broker::BrokerClientChannel,
-    },
-    No,
-}
-
 impl TenantShard {
     pub fn tenant_specific_overrides(&self) -> pageserver_api::models::TenantConfig {
         self.tenant_conf.load().tenant_conf.clone()
@@ -4322,6 +4300,7 @@ impl TenantShard {
             timelines: Mutex::new(HashMap::new()),
             timelines_creating: Mutex::new(HashSet::new()),
             timelines_offloaded: Mutex::new(HashMap::new()),
+            timelines_importing: Mutex::new(HashMap::new()),
             remote_tenant_manifest: Default::default(),
             gc_cs: tokio::sync::Mutex::new(()),
             walredo_mgr,

pageserver/src/tenant/remote_timeline_client.rs

@@ -949,6 +949,35 @@ impl RemoteTimelineClient {
         Ok(())
     }
 
+    /// If the `import_pgdata` field marks the timeline as having an import in progress,
+    /// launch an index-file upload operation that transitions it to done in the background
+    pub(crate) fn schedule_index_upload_for_import_pgdata_finalize(
+        self: &Arc<Self>,
+    ) -> anyhow::Result<()> {
+        use import_pgdata::index_part_format;
+
+        let mut guard = self.upload_queue.lock().unwrap();
+        let upload_queue = guard.initialized_mut()?;
+        let to_update = match &upload_queue.dirty.import_pgdata {
+            Some(import) if !import.is_done() => Some(import),
+            Some(_) | None => None,
+        };
+
+        if let Some(old) = to_update {
+            let new =
+                index_part_format::Root::V1(index_part_format::V1::Done(index_part_format::Done {
+                    idempotency_key: old.idempotency_key().clone(),
+                    started_at: *old.started_at(),
+                    finished_at: chrono::Utc::now().naive_utc(),
+                }));
+
+            upload_queue.dirty.import_pgdata = Some(new);
+            self.schedule_index_upload(upload_queue);
+        }
+
+        Ok(())
+    }
+
     /// Launch an index-file upload operation in the background, setting `gc_compaction_state` field.
     pub(crate) fn schedule_index_upload_for_gc_compaction_state_update(
         self: &Arc<Self>,

pageserver/src/tenant/secondary/downloader.rs

@@ -668,7 +668,9 @@ impl From<DownloadError> for UpdateError {
 
 impl From<std::io::Error> for UpdateError {
     fn from(value: std::io::Error) -> Self {
-        if let Some(nix::errno::Errno::ENOSPC) = value.raw_os_error().map(nix::errno::from_i32) {
+        if let Some(nix::errno::Errno::ENOSPC) =
+            value.raw_os_error().map(nix::errno::Errno::from_raw)
+        {
             UpdateError::NoSpace
         } else if value
             .get_ref()

pageserver/src/tenant/timeline.rs

@@ -61,7 +61,6 @@ use postgres_ffi::{WAL_SEGMENT_SIZE, to_pg_timestamp};
 use rand::Rng;
 use remote_storage::DownloadError;
 use serde_with::serde_as;
-use storage_broker::BrokerClientChannel;
 use tokio::runtime::Handle;
 use tokio::sync::mpsc::Sender;
 use tokio::sync::{Notify, oneshot, watch};
@@ -2080,7 +2079,7 @@ impl Timeline {
     pub(crate) fn activate(
         self: &Arc<Self>,
         parent: Arc<crate::tenant::TenantShard>,
-        broker_client: BrokerClientChannel,
+        broker_client: storage_broker::TimelineUpdatesSubscriber,
         background_jobs_can_start: Option<&completion::Barrier>,
         ctx: &RequestContext,
     ) {
@@ -3114,7 +3113,7 @@ impl Timeline {
     fn launch_wal_receiver(
         self: &Arc<Self>,
         ctx: &RequestContext,
-        broker_client: BrokerClientChannel,
+        broker_client: storage_broker::TimelineUpdatesSubscriber,
     ) {
         info!(
             "launching WAL receiver for timeline {} of tenant {}",

pageserver/src/tenant/timeline/compaction.rs

@@ -3435,6 +3435,7 @@ impl Timeline {
 
         // Step 2: Produce images+deltas.
         let mut accumulated_values = Vec::new();
+        let mut accumulated_values_estimated_size = 0;
         let mut last_key: Option<Key> = None;
 
         // Only create image layers when there is no ancestor branches. TODO: create covering image layer
@@ -3611,12 +3612,16 @@ impl Timeline {
                 if last_key.is_none() {
                     last_key = Some(key);
                 }
+                accumulated_values_estimated_size += val.estimated_size();
                 accumulated_values.push((key, lsn, val));
 
-                if accumulated_values.len() >= 65536 {
-                    // Assume all of them are images, that would be 512MB of data in memory for a single key.
+                // Accumulated values should never exceed 512MB.
+                if accumulated_values_estimated_size >= 1024 * 1024 * 512 {
                     return Err(CompactionError::Other(anyhow!(
-                        "too many values for a single key, giving up gc-compaction"
+                        "too many values for a single key: {} for key {}, {} items",
+                        accumulated_values_estimated_size,
+                        key,
+                        accumulated_values.len()
                     )));
                 }
             } else {
@@ -3651,6 +3656,7 @@ impl Timeline {
                     .map_err(CompactionError::Other)?;
                 accumulated_values.clear();
                 *last_key = key;
+                accumulated_values_estimated_size = val.estimated_size();
                 accumulated_values.push((key, lsn, val));
             }
         }

pageserver/src/tenant/timeline/import_pgdata.rs

@@ -1,8 +1,10 @@
 use std::sync::Arc;
 
 use anyhow::{Context, bail};
+use importbucket_client::{ControlFile, RemoteStorageWrapper};
 use pageserver_api::models::ShardImportStatus;
 use remote_storage::RemotePath;
+use tokio::task::JoinHandle;
 use tokio_util::sync::CancellationToken;
 use tracing::info;
 use utils::lsn::Lsn;
@@ -17,6 +19,17 @@ mod importbucket_client;
 mod importbucket_format;
 pub(crate) mod index_part_format;
 
+pub(crate) struct ImportingTimeline {
+    pub import_task_handle: JoinHandle<()>,
+    pub timeline: Arc<Timeline>,
+}
+
+impl ImportingTimeline {
+    pub(crate) fn shutdown(self) {
+        self.import_task_handle.abort();
+    }
+}
+
 pub async fn doit(
     timeline: &Arc<Timeline>,
     index_part: index_part_format::Root,
@@ -26,173 +39,225 @@ pub async fn doit(
     let index_part_format::Root::V1(v1) = index_part;
     let index_part_format::InProgress {
         location,
-        idempotency_key,
-        started_at,
+        idempotency_key: _,
+        started_at: _,
     } = match v1 {
         index_part_format::V1::Done(_) => return Ok(()),
         index_part_format::V1::InProgress(in_progress) => in_progress,
     };
 
-    let storage = importbucket_client::new(timeline.conf, &location, cancel.clone()).await?;
+    let storcon_client = StorageControllerUpcallClient::new(timeline.conf, &cancel);
 
-    let status_prefix = RemotePath::from_string("status").unwrap();
+    let shard_status = storcon_client
+        .get_timeline_import_status(
+            timeline.tenant_shard_id,
+            timeline.timeline_id,
+            timeline.generation,
+        )
+        .await
+        .map_err(|_err| anyhow::anyhow!("Shut down while getting timeline import status"))?;
 
-    //
-    // See if shard is done.
-    // TODO: incorporate generations into status key for split brain safety. Figure out together with checkpointing.
-    //
-    let shard_status_key =
-        status_prefix.join(format!("shard-{}", timeline.tenant_shard_id.shard_slug()));
-    let shard_status: Option<importbucket_format::ShardStatus> =
-        storage.get_json(&shard_status_key).await?;
     info!(?shard_status, "peeking shard status");
-    if shard_status.map(|st| st.done).unwrap_or(false) {
-        info!("shard status indicates that the shard is done, skipping import");
-    } else {
-        // TODO: checkpoint the progress into the IndexPart instead of restarting
-        // from the beginning.
+    match shard_status {
+        ShardImportStatus::InProgress(maybe_progress) => {
+            let storage =
+                importbucket_client::new(timeline.conf, &location, cancel.clone()).await?;
 
-        //
-        // Wipe the slate clean - the flow does not allow resuming.
-        // We can implement resuming in the future by checkpointing the progress into the IndexPart.
-        //
-        info!("wipe the slate clean");
-        {
-            // TODO: do we need to hold GC lock for this?
-            let mut guard = timeline.layers.write().await;
-            assert!(
-                guard.layer_map()?.open_layer.is_none(),
-                "while importing, there should be no in-memory layer" // this just seems like a good place to assert it
-            );
-            let all_layers_keys = guard.all_persistent_layers();
-            let all_layers: Vec<_> = all_layers_keys
-                .iter()
-                .map(|key| guard.get_from_key(key))
-                .collect();
-            let open = guard.open_mut().context("open_mut")?;
+            let control_file_res = if maybe_progress.is_none() {
+                // Only prepare the import once when there's no progress.
+                prepare_import(timeline, storage.clone(), &cancel).await
+            } else {
+                storage.get_control_file().await
+            };
 
-            timeline.remote_client.schedule_gc_update(&all_layers)?;
-            open.finish_gc_timeline(&all_layers);
-        }
-
-        //
-        // Wait for pgdata to finish uploading
-        //
-        info!("wait for pgdata to reach status 'done'");
-        let pgdata_status_key = status_prefix.join("pgdata");
-        loop {
-            let res = async {
-                let pgdata_status: Option<importbucket_format::PgdataStatus> = storage
-                    .get_json(&pgdata_status_key)
-                    .await
-                    .context("get pgdata status")?;
-                info!(?pgdata_status, "peeking pgdata status");
-                if pgdata_status.map(|st| st.done).unwrap_or(false) {
-                    Ok(())
-                } else {
-                    Err(anyhow::anyhow!("pgdata not done yet"))
-                }
-            }
-            .await;
-            match res {
-                Ok(_) => break,
+            let control_file = match control_file_res {
+                Ok(cf) => cf,
                 Err(err) => {
-                    info!(?err, "indefinitely waiting for pgdata to finish");
-                    if tokio::time::timeout(std::time::Duration::from_secs(10), cancel.cancelled())
-                        .await
-                        .is_ok()
-                    {
-                        bail!("cancelled while waiting for pgdata");
-                    }
+                    return Err(
+                        terminate_flow_with_error(timeline, err, &storcon_client, &cancel).await,
+                    );
                 }
-            }
-        }
+            };
 
-        //
-        // Do the import
-        //
-        info!("do the import");
-        let control_file = storage.get_control_file().await?;
-        let base_lsn = control_file.base_lsn();
-
-        info!("update TimelineMetadata based on LSNs from control file");
-        {
-            let pg_version = control_file.pg_version();
-            let _ctx: &RequestContext = ctx;
-            async move {
-                // FIXME: The 'disk_consistent_lsn' should be the LSN at the *end* of the
-                // checkpoint record, and prev_record_lsn should point to its beginning.
-                // We should read the real end of the record from the WAL, but here we
-                // just fake it.
-                let disk_consistent_lsn = Lsn(base_lsn.0 + 8);
-                let prev_record_lsn = base_lsn;
-                let metadata = TimelineMetadata::new(
-                    disk_consistent_lsn,
-                    Some(prev_record_lsn),
-                    None,     // no ancestor
-                    Lsn(0),   // no ancestor lsn
-                    base_lsn, // latest_gc_cutoff_lsn
-                    base_lsn, // initdb_lsn
-                    pg_version,
+            let res = flow::run(
+                timeline.clone(),
+                control_file,
+                storage.clone(),
+                maybe_progress,
+                ctx,
+            )
+            .await;
+            if let Err(err) = res {
+                return Err(
+                    terminate_flow_with_error(timeline, err, &storcon_client, &cancel).await,
                 );
-
-                let _start_lsn = disk_consistent_lsn + 1;
-
-                timeline
-                    .remote_client
-                    .schedule_index_upload_for_full_metadata_update(&metadata)?;
-
-                timeline.remote_client.wait_completion().await?;
-
-                anyhow::Ok(())
             }
+
+            // Communicate that shard is done.
+            // Ensure at-least-once delivery of the upcall to storage controller
+            // before we mark the task as done and never come here again.
+            //
+            // Note that we do not mark the import complete in the index part now.
+            // This happens in [`Tenant::finalize_importing_timeline`] in response
+            // to the storage controller calling
+            // `/v1/tenant/:tenant_id/timeline/:timeline_id/activate_post_import`.
+            storcon_client
+                .put_timeline_import_status(
+                    timeline.tenant_shard_id,
+                    timeline.timeline_id,
+                    timeline.generation,
+                    ShardImportStatus::Done,
+                )
+                .await
+                .map_err(|_err| {
+                    anyhow::anyhow!("Shut down while putting timeline import status")
+                })?;
+        }
+        ShardImportStatus::Error(err) => {
+            info!(
+                "shard status indicates that the shard is done (error), skipping import {}",
+                err
+            );
+        }
+        ShardImportStatus::Done => {
+            info!("shard status indicates that the shard is done (success), skipping import");
         }
-        .await?;
-
-        flow::run(timeline.clone(), control_file, storage.clone(), ctx).await?;
-
-        //
-        // Communicate that shard is done.
-        // Ensure at-least-once delivery of the upcall to storage controller
-        // before we mark the task as done and never come here again.
-        //
-        let storcon_client = StorageControllerUpcallClient::new(timeline.conf, &cancel);
-        storcon_client
-            .put_timeline_import_status(
-                timeline.tenant_shard_id,
-                timeline.timeline_id,
-                // TODO(vlad): What about import errors?
-                ShardImportStatus::Done,
-            )
-            .await
-            .map_err(|_err| anyhow::anyhow!("Shut down while putting timeline import status"))?;
-
-        storage
-            .put_json(
-                &shard_status_key,
-                &importbucket_format::ShardStatus { done: true },
-            )
-            .await
-            .context("put shard status")?;
     }
 
-    //
-    // Mark as done in index_part.
-    // This makes subsequent timeline loads enter the normal load code path
-    // instead of spawning the import task and calling this here function.
-    //
-    info!("mark import as complete in index part");
-    timeline
-        .remote_client
-        .schedule_index_upload_for_import_pgdata_state_update(Some(index_part_format::Root::V1(
-            index_part_format::V1::Done(index_part_format::Done {
-                idempotency_key,
-                started_at,
-                finished_at: chrono::Utc::now().naive_utc(),
-            }),
-        )))?;
-
-    timeline.remote_client.wait_completion().await?;
-
     Ok(())
 }
+
+async fn prepare_import(
+    timeline: &Arc<Timeline>,
+    storage: RemoteStorageWrapper,
+    cancel: &CancellationToken,
+) -> anyhow::Result<ControlFile> {
+    // Wipe the slate clean before starting the import as a precaution.
+    // This method is only called when there's no recorded checkpoint for the import
+    // in the storage controller.
+    //
+    // Note that this is split-brain safe (two imports for same timeline shards running in
+    // different generations) because we go through the usual deletion path, including deletion queue.
+    info!("wipe the slate clean");
+    {
+        // TODO: do we need to hold GC lock for this?
+        let mut guard = timeline.layers.write().await;
+        assert!(
+            guard.layer_map()?.open_layer.is_none(),
+            "while importing, there should be no in-memory layer" // this just seems like a good place to assert it
+        );
+        let all_layers_keys = guard.all_persistent_layers();
+        let all_layers: Vec<_> = all_layers_keys
+            .iter()
+            .map(|key| guard.get_from_key(key))
+            .collect();
+        let open = guard.open_mut().context("open_mut")?;
+
+        timeline.remote_client.schedule_gc_update(&all_layers)?;
+        open.finish_gc_timeline(&all_layers);
+    }
+
+    //
+    // Wait for pgdata to finish uploading
+    //
+    info!("wait for pgdata to reach status 'done'");
+    let status_prefix = RemotePath::from_string("status").unwrap();
+    let pgdata_status_key = status_prefix.join("pgdata");
+    loop {
+        let res = async {
+            let pgdata_status: Option<importbucket_format::PgdataStatus> = storage
+                .get_json(&pgdata_status_key)
+                .await
+                .context("get pgdata status")?;
+            info!(?pgdata_status, "peeking pgdata status");
+            if pgdata_status.map(|st| st.done).unwrap_or(false) {
+                Ok(())
+            } else {
+                Err(anyhow::anyhow!("pgdata not done yet"))
+            }
+        }
+        .await;
+        match res {
+            Ok(_) => break,
+            Err(err) => {
+                info!(?err, "indefinitely waiting for pgdata to finish");
+                if tokio::time::timeout(std::time::Duration::from_secs(10), cancel.cancelled())
+                    .await
+                    .is_ok()
+                {
+                    bail!("cancelled while waiting for pgdata");
+                }
+            }
+        }
+    }
+
+    let control_file = storage.get_control_file().await?;
+    let base_lsn = control_file.base_lsn();
+
+    info!("update TimelineMetadata based on LSNs from control file");
+    {
+        let pg_version = control_file.pg_version();
+        async move {
+            // FIXME: The 'disk_consistent_lsn' should be the LSN at the *end* of the
+            // checkpoint record, and prev_record_lsn should point to its beginning.
+            // We should read the real end of the record from the WAL, but here we
+            // just fake it.
+            let disk_consistent_lsn = Lsn(base_lsn.0 + 8);
+            let prev_record_lsn = base_lsn;
+            let metadata = TimelineMetadata::new(
+                disk_consistent_lsn,
+                Some(prev_record_lsn),
+                None,     // no ancestor
+                Lsn(0),   // no ancestor lsn
+                base_lsn, // latest_gc_cutoff_lsn
+                base_lsn, // initdb_lsn
+                pg_version,
+            );
+
+            let _start_lsn = disk_consistent_lsn + 1;
+
+            timeline
+                .remote_client
+                .schedule_index_upload_for_full_metadata_update(&metadata)?;
+
+            timeline.remote_client.wait_completion().await?;
+
+            anyhow::Ok(())
+        }
+    }
+    .await?;
+
+    Ok(control_file)
+}
+
+async fn terminate_flow_with_error(
+    timeline: &Arc<Timeline>,
+    error: anyhow::Error,
+    storcon_client: &StorageControllerUpcallClient,
+    cancel: &CancellationToken,
+) -> anyhow::Error {
+    // The import task is a aborted on tenant shutdown, so in principle, it should
+    // never be cancelled. To be on the safe side, check the cancellation tokens
+    // before marking the import as failed.
+    if !(cancel.is_cancelled() || timeline.cancel.is_cancelled()) {
+        let notify_res = storcon_client
+            .put_timeline_import_status(
+                timeline.tenant_shard_id,
+                timeline.timeline_id,
+                timeline.generation,
+                ShardImportStatus::Error(format!("{error:#}")),
+            )
+            .await;
+
+        if let Err(_notify_error) = notify_res {
+            // The [`StorageControllerUpcallClient::put_timeline_import_status`] retries
+            // forever internally, so errors returned by it can only be due to cancellation.
+            info!("failed to notify storcon about permanent import error");
+        }
+
+        // Will be logged by [`Tenant::create_timeline_import_pgdata_task`]
+        error
+    } else {
+        anyhow::anyhow!("Import task cancelled")
+    }
+}

pageserver/src/tenant/timeline/import_pgdata/flow.rs

@@ -29,10 +29,11 @@
 //! - version-specific CheckPointData (=> pgv abstraction, already exists for regular walingest)
 
 use std::collections::HashSet;
+use std::hash::{Hash, Hasher};
 use std::ops::Range;
 use std::sync::Arc;
 
-use anyhow::{bail, ensure};
+use anyhow::ensure;
 use bytes::Bytes;
 use futures::stream::FuturesOrdered;
 use itertools::Itertools;
@@ -43,6 +44,7 @@ use pageserver_api::key::{
     slru_segment_size_to_key,
 };
 use pageserver_api::keyspace::{contiguous_range_len, is_contiguous_range, singleton_range};
+use pageserver_api::models::{ShardImportProgress, ShardImportProgressV1, ShardImportStatus};
 use pageserver_api::reltag::{RelTag, SlruKind};
 use pageserver_api::shard::ShardIdentity;
 use postgres_ffi::relfile_utils::parse_relfilename;
@@ -53,21 +55,42 @@ use tokio_stream::StreamExt;
 use tracing::{debug, instrument};
 use utils::bin_ser::BeSer;
 use utils::lsn::Lsn;
+use utils::pausable_failpoint;
 
 use super::Timeline;
 use super::importbucket_client::{ControlFile, RemoteStorageWrapper};
 use crate::assert_u64_eq_usize::UsizeIsU64;
 use crate::context::{DownloadBehavior, RequestContext};
+use crate::controller_upcall_client::{StorageControllerUpcallApi, StorageControllerUpcallClient};
 use crate::pgdatadir_mapping::{
     DbDirectory, RelDirectory, SlruSegmentDirectory, TwoPhaseDirectory,
 };
 use crate::task_mgr::TaskKind;
-use crate::tenant::storage_layer::{ImageLayerWriter, Layer};
+use crate::tenant::storage_layer::{AsLayerDesc, ImageLayerWriter, Layer};
 
 pub async fn run(
     timeline: Arc<Timeline>,
     control_file: ControlFile,
     storage: RemoteStorageWrapper,
+    import_progress: Option<ShardImportProgress>,
+    ctx: &RequestContext,
+) -> anyhow::Result<()> {
+    // Match how we run the import based on the progress version.
+    // If there's no import progress, it means that this is a new import
+    // and we can use whichever version we want.
+    match import_progress {
+        Some(ShardImportProgress::V1(progress)) => {
+            run_v1(timeline, control_file, storage, Some(progress), ctx).await
+        }
+        None => run_v1(timeline, control_file, storage, None, ctx).await,
+    }
+}
+
+async fn run_v1(
+    timeline: Arc<Timeline>,
+    control_file: ControlFile,
+    storage: RemoteStorageWrapper,
+    import_progress: Option<ShardImportProgressV1>,
     ctx: &RequestContext,
 ) -> anyhow::Result<()> {
     let planner = Planner {
@@ -79,7 +102,32 @@ pub async fn run(
 
     let import_config = &timeline.conf.timeline_import_config;
     let plan = planner.plan(import_config).await?;
-    plan.execute(timeline, import_config, ctx).await
+
+    // Hash the plan and compare with the hash of the plan we got back from the storage controller.
+    // If the two match, it means that the planning stage had the same output.
+    //
+    // This is not intended to be a cryptographically secure hash.
+    const SEED: u64 = 42;
+    let mut hasher = twox_hash::XxHash64::with_seed(SEED);
+    plan.hash(&mut hasher);
+    let plan_hash = hasher.finish();
+
+    if let Some(progress) = &import_progress {
+        if plan_hash != progress.import_plan_hash {
+            anyhow::bail!("Import plan does not match storcon metadata");
+        }
+
+        // Handle collisions on jobs of unequal length
+        if progress.jobs != plan.jobs.len() {
+            anyhow::bail!("Import plan job length does not match storcon metadata")
+        }
+    }
+
+    pausable_failpoint!("import-timeline-pre-execute-pausable");
+
+    let start_from_job_idx = import_progress.map(|progress| progress.completed);
+    plan.execute(timeline, start_from_job_idx, plan_hash, import_config, ctx)
+        .await
 }
 
 struct Planner {
@@ -89,8 +137,11 @@ struct Planner {
     tasks: Vec<AnyImportTask>,
 }
 
+#[derive(Hash)]
 struct Plan {
     jobs: Vec<ChunkProcessingJob>,
+    // Included here such that it ends up in the hash for the plan
+    shard: ShardIdentity,
 }
 
 impl Planner {
@@ -194,7 +245,10 @@ impl Planner {
             pgdata_lsn,
         ));
 
-        Ok(Plan { jobs })
+        Ok(Plan {
+            jobs,
+            shard: self.shard,
+        })
     }
 
     #[instrument(level = tracing::Level::DEBUG, skip_all, fields(dboid=%db.dboid, tablespace=%db.spcnode, path=%db.path))]
@@ -323,25 +377,45 @@ impl Plan {
     async fn execute(
         self,
         timeline: Arc<Timeline>,
+        start_after_job_idx: Option<usize>,
+        import_plan_hash: u64,
         import_config: &TimelineImportConfig,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
+        let storcon_client = StorageControllerUpcallClient::new(timeline.conf, &timeline.cancel);
+
         let mut work = FuturesOrdered::new();
         let semaphore = Arc::new(Semaphore::new(import_config.import_job_concurrency.into()));
 
         let jobs_in_plan = self.jobs.len();
 
-        let mut jobs = self.jobs.into_iter().enumerate().peekable();
-        let mut results = Vec::new();
+        let mut jobs = self
+            .jobs
+            .into_iter()
+            .enumerate()
+            .map(|(idx, job)| (idx + 1, job))
+            .filter(|(idx, _job)| {
+                // Filter out any jobs that have been done already
+                if let Some(start_after) = start_after_job_idx {
+                    *idx > start_after
+                } else {
+                    true
+                }
+            })
+            .peekable();
+
+        let mut last_completed_job_idx = start_after_job_idx.unwrap_or(0);
+        let checkpoint_every: usize = import_config.import_job_checkpoint_threshold.into();
 
         // Run import jobs concurrently up to the limit specified by the pageserver configuration.
         // Note that we process completed futures in the oreder of insertion. This will be the
         // building block for resuming imports across pageserver restarts or tenant migrations.
-        while results.len() < jobs_in_plan {
+        while last_completed_job_idx < jobs_in_plan {
             tokio::select! {
                 permit = semaphore.clone().acquire_owned(), if jobs.peek().is_some() => {
                     let permit = permit.expect("never closed");
                     let (job_idx, job) = jobs.next().expect("we peeked");
+
                     let job_timeline = timeline.clone();
                     let ctx = ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Error);
 
@@ -353,13 +427,35 @@ impl Plan {
                 },
                 maybe_complete_job_idx = work.next() => {
                     match maybe_complete_job_idx {
-                        Some(Ok((_job_idx, res))) => {
-                            results.push(res);
+                        Some(Ok((job_idx, res))) => {
+                            assert!(last_completed_job_idx.checked_add(1).unwrap() == job_idx);
+
+                            res?;
+                            last_completed_job_idx = job_idx;
+
+                            if last_completed_job_idx % checkpoint_every == 0 {
+                                let progress = ShardImportProgressV1 {
+                                    jobs: jobs_in_plan,
+                                    completed: last_completed_job_idx,
+                                    import_plan_hash,
+                                };
+
+                                storcon_client.put_timeline_import_status(
+                                    timeline.tenant_shard_id,
+                                    timeline.timeline_id,
+                                    timeline.generation,
+                                    ShardImportStatus::InProgress(Some(ShardImportProgress::V1(progress)))
+                                )
+                                .await
+                                .map_err(|_err| {
+                                    anyhow::anyhow!("Shut down while putting timeline import status")
+                                })?;
+                            }
                         },
                         Some(Err(_)) => {
-                            results.push(Err(anyhow::anyhow!(
-                                "parallel job panicked or cancelled, check pageserver logs"
-                            )));
+                            anyhow::bail!(
+                                "import job panicked or cancelled"
+                            );
                         }
                         None => {}
                     }
@@ -367,17 +463,7 @@ impl Plan {
             }
         }
 
-        if results.iter().all(|r| r.is_ok()) {
-            Ok(())
-        } else {
-            let mut msg = String::new();
-            for result in results {
-                if let Err(err) = result {
-                    msg.push_str(&format!("{err:?}\n\n"));
-                }
-            }
-            bail!("Some parallel jobs failed:\n\n{msg}");
-        }
+        Ok(())
     }
 }
 
@@ -549,6 +635,15 @@ struct ImportSingleKeyTask {
     buf: Bytes,
 }
 
+impl Hash for ImportSingleKeyTask {
+    fn hash<H: Hasher>(&self, state: &mut H) {
+        let ImportSingleKeyTask { key, buf } = self;
+
+        key.hash(state);
+        buf.hash(state);
+    }
+}
+
 impl ImportSingleKeyTask {
     fn new(key: Key, buf: Bytes) -> Self {
         ImportSingleKeyTask { key, buf }
@@ -577,6 +672,20 @@ struct ImportRelBlocksTask {
     storage: RemoteStorageWrapper,
 }
 
+impl Hash for ImportRelBlocksTask {
+    fn hash<H: Hasher>(&self, state: &mut H) {
+        let ImportRelBlocksTask {
+            shard_identity: _,
+            key_range,
+            path,
+            storage: _,
+        } = self;
+
+        key_range.hash(state);
+        path.hash(state);
+    }
+}
+
 impl ImportRelBlocksTask {
     fn new(
         shard_identity: ShardIdentity,
@@ -661,6 +770,19 @@ struct ImportSlruBlocksTask {
     storage: RemoteStorageWrapper,
 }
 
+impl Hash for ImportSlruBlocksTask {
+    fn hash<H: Hasher>(&self, state: &mut H) {
+        let ImportSlruBlocksTask {
+            key_range,
+            path,
+            storage: _,
+        } = self;
+
+        key_range.hash(state);
+        path.hash(state);
+    }
+}
+
 impl ImportSlruBlocksTask {
     fn new(key_range: Range<Key>, path: &RemotePath, storage: RemoteStorageWrapper) -> Self {
         ImportSlruBlocksTask {
@@ -703,6 +825,7 @@ impl ImportTask for ImportSlruBlocksTask {
     }
 }
 
+#[derive(Hash)]
 enum AnyImportTask {
     SingleKey(ImportSingleKeyTask),
     RelBlocks(ImportRelBlocksTask),
@@ -749,6 +872,7 @@ impl From<ImportSlruBlocksTask> for AnyImportTask {
     }
 }
 
+#[derive(Hash)]
 struct ChunkProcessingJob {
     range: Range<Key>,
     tasks: Vec<AnyImportTask>,
@@ -786,17 +910,51 @@ impl ChunkProcessingJob {
 
         let resident_layer = if nimages > 0 {
             let (desc, path) = writer.finish(ctx).await?;
+
+            {
+                let guard = timeline.layers.read().await;
+                let existing_layer = guard.try_get_from_key(&desc.key());
+                if let Some(layer) = existing_layer {
+                    if layer.metadata().generation != timeline.generation {
+                        return Err(anyhow::anyhow!(
+                            "Import attempted to rewrite layer file in the same generation: {}",
+                            layer.local_path()
+                        ));
+                    }
+                }
+            }
+
             Layer::finish_creating(timeline.conf, &timeline, desc, &path)?
         } else {
             // dropping the writer cleans up
             return Ok(());
         };
 
-        // this is sharing the same code as create_image_layers
+        // The same import job might run multiple times since not each job is checkpointed.
+        // Hence, we must support the cases where the layer already exists. We cannot be
+        // certain that the existing layer is identical to the new one, so in that case
+        // we replace the old layer with the one we just generated.
+
         let mut guard = timeline.layers.write().await;
-        guard
-            .open_mut()?
-            .track_new_image_layers(&[resident_layer.clone()], &timeline.metrics);
+
+        let existing_layer = guard
+            .try_get_from_key(&resident_layer.layer_desc().key())
+            .cloned();
+        match existing_layer {
+            Some(existing) => {
+                guard.open_mut()?.rewrite_layers(
+                    &[(existing.clone(), resident_layer.clone())],
+                    &[],
+                    &timeline.metrics,
+                );
+            }
+            None => {
+                guard
+                    .open_mut()?
+                    .track_new_image_layers(&[resident_layer.clone()], &timeline.metrics);
+            }
+        }
+
         crate::tenant::timeline::drop_wlock(guard);
 
         timeline

pageserver/src/tenant/timeline/import_pgdata/importbucket_client.rs

@@ -190,31 +190,6 @@ impl RemoteStorageWrapper {
         Ok(Some(res))
     }
 
-    #[instrument(level = tracing::Level::DEBUG, skip_all, fields(%path))]
-    pub async fn put_json<T>(&self, path: &RemotePath, value: &T) -> anyhow::Result<()>
-    where
-        T: serde::Serialize,
-    {
-        let buf = serde_json::to_vec(value)?;
-        let bytes = Bytes::from(buf);
-        utils::backoff::retry(
-            || async {
-                let size = bytes.len();
-                let bytes = futures::stream::once(futures::future::ready(Ok(bytes.clone())));
-                self.storage
-                    .upload_storage_object(bytes, size, path, &self.cancel)
-                    .await
-            },
-            remote_storage::TimeoutOrCancel::caused_by_cancel,
-            1,
-            u32::MAX,
-            &format!("put json {path}"),
-            &self.cancel,
-        )
-        .await
-        .expect("practically infinite retries")
-    }
-
     #[instrument(level = tracing::Level::DEBUG, skip_all, fields(%path))]
     pub async fn get_range(
         &self,

pageserver/src/tenant/timeline/import_pgdata/importbucket_format.rs

@@ -5,9 +5,3 @@ pub struct PgdataStatus {
     pub done: bool,
     // TODO: remaining fields
 }
-
-#[derive(Deserialize, Serialize, Debug, Clone, PartialEq, Eq)]
-pub struct ShardStatus {
-    pub done: bool,
-    // TODO: remaining fields
-}

pageserver/src/tenant/timeline/import_pgdata/index_part_format.rs

@@ -64,4 +64,12 @@ impl Root {
             },
         }
     }
+    pub fn started_at(&self) -> &chrono::NaiveDateTime {
+        match self {
+            Root::V1(v1) => match v1 {
+                V1::InProgress(in_progress) => &in_progress.started_at,
+                V1::Done(done) => &done.started_at,
+            },
+        }
+    }
 }

pageserver/src/tenant/timeline/uninit.rs

@@ -161,7 +161,7 @@ impl<'t> UninitializedTimeline<'t> {
         tenant: Arc<TenantShard>,
         copyin_read: &mut (impl tokio::io::AsyncRead + Send + Sync + Unpin),
         base_lsn: Lsn,
-        broker_client: storage_broker::BrokerClientChannel,
+        broker_client: storage_broker::TimelineUpdatesSubscriber,
         ctx: &RequestContext,
     ) -> anyhow::Result<Arc<Timeline>> {
         self.write(|raw_timeline| async move {

pageserver/src/tenant/timeline/walreceiver.rs

@@ -28,7 +28,6 @@ use std::num::NonZeroU64;
 use std::sync::Arc;
 use std::time::Duration;
 
-use storage_broker::BrokerClientChannel;
 use tokio::sync::watch;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -70,7 +69,7 @@ impl WalReceiver {
     pub fn start(
         timeline: Arc<Timeline>,
         conf: WalReceiverConf,
-        mut broker_client: BrokerClientChannel,
+        mut broker_client: storage_broker::TimelineUpdatesSubscriber,
         ctx: &RequestContext,
     ) -> Self {
         let tenant_shard_id = timeline.tenant_shard_id;

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

@@ -17,19 +17,12 @@ use std::time::Duration;
 
 use anyhow::Context;
 use chrono::{NaiveDateTime, Utc};
+use futures::StreamExt;
 use pageserver_api::models::TimelineState;
 use postgres_connection::PgConnectionConfig;
-use storage_broker::proto::{
-    FilterTenantTimelineId, MessageType, SafekeeperDiscoveryRequest, SafekeeperDiscoveryResponse,
-    SubscribeByFilterRequest, TenantTimelineId as ProtoTenantTimelineId, TypeSubscription,
-    TypedMessage,
-};
-use storage_broker::{BrokerClientChannel, Code, Streaming};
+use storage_broker::proto::SafekeeperDiscoveryResponse;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
-use utils::backoff::{
-    DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS, exponential_backoff,
-};
 use utils::id::{NodeId, TenantTimelineId};
 use utils::lsn::Lsn;
 use utils::postgres_client::{
@@ -56,7 +49,7 @@ pub(crate) struct Cancelled;
 ///
 /// Not cancellation-safe. Use `cancel` token to request cancellation.
 pub(super) async fn connection_manager_loop_step(
-    broker_client: &mut BrokerClientChannel,
+    broker_client: &mut storage_broker::TimelineUpdatesSubscriber,
     connection_manager_state: &mut ConnectionManagerState,
     ctx: &RequestContext,
     cancel: &CancellationToken,
@@ -81,11 +74,6 @@ pub(super) async fn connection_manager_loop_step(
         WALRECEIVER_ACTIVE_MANAGERS.dec();
     }
 
-    let id = TenantTimelineId {
-        tenant_id: connection_manager_state.timeline.tenant_shard_id.tenant_id,
-        timeline_id: connection_manager_state.timeline.timeline_id,
-    };
-
     let mut timeline_state_updates = connection_manager_state
         .timeline
         .subscribe_for_state_updates();
@@ -101,7 +89,12 @@ pub(super) async fn connection_manager_loop_step(
     // Subscribe to the broker updates. Stream shares underlying TCP connection
     // with other streams on this client (other connection managers). When
     // object goes out of scope, stream finishes in drop() automatically.
-    let mut broker_subscription = subscribe_for_timeline_updates(broker_client, id, cancel).await?;
+    let (timeline_updates, mut discovery_requester) = broker_client.subscribe(
+        connection_manager_state.timeline.tenant_shard_id,
+        connection_manager_state.timeline.timeline_id,
+        cancel,
+    );
+    let mut timeline_updates = Box::pin(timeline_updates);
     debug!("Subscribed for broker timeline updates");
 
     loop {
@@ -155,29 +148,10 @@ pub(super) async fn connection_manager_loop_step(
                 }
             },
 
-            // Got a new update from the broker
-            broker_update = broker_subscription.message() /* TODO: review cancellation-safety */ => {
-                match broker_update {
-                    Ok(Some(broker_update)) => connection_manager_state.register_timeline_update(broker_update),
-                    Err(status) => {
-                        match status.code() {
-                            Code::Unknown if status.message().contains("stream closed because of a broken pipe") || status.message().contains("connection reset") || status.message().contains("error reading a body from connection") => {
-                                // tonic's error handling doesn't provide a clear code for disconnections: we get
-                                // "h2 protocol error: error reading a body from connection: stream closed because of a broken pipe"
-                                // => https://github.com/neondatabase/neon/issues/9562
-                                info!("broker disconnected: {status}");
-                            },
-                            _ => {
-                                warn!("broker subscription failed: {status}");
-                            }
-                        }
-                        return Ok(());
-                    }
-                    Ok(None) => {
-                        error!("broker subscription stream ended"); // can't happen
-                        return Ok(());
-                    }
-                }
+            // Got a new update from the broker.
+            // The stream ends with None if and only if `cancel` is cancelled.
+            Some(timeline_update) = timeline_updates.next() => {
+                connection_manager_state.register_timeline_update(timeline_update)
             },
 
             new_event = async {
@@ -258,32 +232,11 @@ pub(super) async fn connection_manager_loop_step(
                     tokio::time::sleep(next_discovery_ts - now).await;
                 }
 
-                let tenant_timeline_id = Some(ProtoTenantTimelineId {
-                    tenant_id: id.tenant_id.as_ref().to_owned(),
-                    timeline_id: id.timeline_id.as_ref().to_owned(),
-                });
-                let request = SafekeeperDiscoveryRequest { tenant_timeline_id };
-                let msg = TypedMessage {
-                    r#type: MessageType::SafekeeperDiscoveryRequest as i32,
-                    safekeeper_timeline_info: None,
-                    safekeeper_discovery_request: Some(request),
-                    safekeeper_discovery_response: None,
-                    };
+                info!("No active connection and no candidates, sending discovery request to the broker");
+                discovery_requester.request().await;
 
                 last_discovery_ts = Some(std::time::Instant::now());
-                info!("No active connection and no candidates, sending discovery request to the broker");
 
-                // Cancellation safety: we want to send a message to the broker, but publish_one()
-                // function can get cancelled by the other select! arm. This is absolutely fine, because
-                // we just want to receive broker updates and discovery is not important if we already
-                // receive updates.
-                //
-                // It is possible that `last_discovery_ts` will be updated, but the message will not be sent.
-                // This is totally fine because of the reason above.
-
-                // This is a fire-and-forget request, we don't care about the response
-                let _ = broker_client.publish_one(msg).await;
-                debug!("Discovery request sent to the broker");
                 None
             } => {}
         }
@@ -298,63 +251,6 @@ pub(super) async fn connection_manager_loop_step(
     }
 }
 
-/// Endlessly try to subscribe for broker updates for a given timeline.
-async fn subscribe_for_timeline_updates(
-    broker_client: &mut BrokerClientChannel,
-    id: TenantTimelineId,
-    cancel: &CancellationToken,
-) -> Result<Streaming<TypedMessage>, Cancelled> {
-    let mut attempt = 0;
-    loop {
-        exponential_backoff(
-            attempt,
-            DEFAULT_BASE_BACKOFF_SECONDS,
-            DEFAULT_MAX_BACKOFF_SECONDS,
-            cancel,
-        )
-        .await;
-        attempt += 1;
-
-        // subscribe to the specific timeline
-        let request = SubscribeByFilterRequest {
-            types: vec![
-                TypeSubscription {
-                    r#type: MessageType::SafekeeperTimelineInfo as i32,
-                },
-                TypeSubscription {
-                    r#type: MessageType::SafekeeperDiscoveryResponse as i32,
-                },
-            ],
-            tenant_timeline_id: Some(FilterTenantTimelineId {
-                enabled: true,
-                tenant_timeline_id: Some(ProtoTenantTimelineId {
-                    tenant_id: id.tenant_id.as_ref().to_owned(),
-                    timeline_id: id.timeline_id.as_ref().to_owned(),
-                }),
-            }),
-        };
-
-        match {
-            tokio::select! {
-                r = broker_client.subscribe_by_filter(request) => { r }
-                _ = cancel.cancelled() => { return Err(Cancelled); }
-            }
-        } {
-            Ok(resp) => {
-                return Ok(resp.into_inner());
-            }
-            Err(e) => {
-                // Safekeeper nodes can stop pushing timeline updates to the broker, when no new writes happen and
-                // entire WAL is streamed. Keep this noticeable with logging, but do not warn/error.
-                info!(
-                    "Attempt #{attempt}, failed to subscribe for timeline {id} updates in broker: {e:#}"
-                );
-                continue;
-            }
-        }
-    }
-}
-
 const WALCONNECTION_RETRY_MIN_BACKOFF_SECONDS: f64 = 0.1;
 const WALCONNECTION_RETRY_MAX_BACKOFF_SECONDS: f64 = 15.0;
 const WALCONNECTION_RETRY_BACKOFF_MULTIPLIER: f64 = 1.5;
@@ -695,44 +591,14 @@ impl ConnectionManagerState {
     }
 
     /// Adds another broker timeline into the state, if its more recent than the one already added there for the same key.
-    fn register_timeline_update(&mut self, typed_msg: TypedMessage) {
-        let mut is_discovery = false;
-        let timeline_update = match typed_msg.r#type() {
-            MessageType::SafekeeperTimelineInfo => {
-                let info = match typed_msg.safekeeper_timeline_info {
-                    Some(info) => info,
-                    None => {
-                        warn!("bad proto message from broker: no safekeeper_timeline_info");
-                        return;
-                    }
-                };
-                SafekeeperDiscoveryResponse {
-                    safekeeper_id: info.safekeeper_id,
-                    tenant_timeline_id: info.tenant_timeline_id,
-                    commit_lsn: info.commit_lsn,
-                    safekeeper_connstr: info.safekeeper_connstr,
-                    availability_zone: info.availability_zone,
-                    standby_horizon: info.standby_horizon,
-                }
-            }
-            MessageType::SafekeeperDiscoveryResponse => {
-                is_discovery = true;
-                match typed_msg.safekeeper_discovery_response {
-                    Some(response) => response,
-                    None => {
-                        warn!("bad proto message from broker: no safekeeper_discovery_response");
-                        return;
-                    }
-                }
-            }
-            _ => {
-                // unexpected message
-                return;
-            }
-        };
-
+    fn register_timeline_update(&mut self, timeline_update: storage_broker::TimelineShardUpdate) {
         WALRECEIVER_BROKER_UPDATES.inc();
 
+        let storage_broker::TimelineShardUpdate {
+            is_discovery,
+            inner: timeline_update,
+        } = timeline_update;
+
         trace!(
             "safekeeper info update: standby_horizon(cutoff)={}",
             timeline_update.standby_horizon
@@ -1013,7 +879,7 @@ impl ConnectionManagerState {
                     shard_stripe_size,
                     listen_pg_addr_str: info.safekeeper_connstr.as_ref(),
                     auth_token: self.conf.auth_token.as_ref().map(|t| t.as_str()),
-                    availability_zone: self.conf.availability_zone.as_deref()
+                    availability_zone: self.conf.availability_zone.as_deref(),
                 };
 
                 match wal_stream_connection_config(connection_conf_args) {

pageserver/src/virtual_file.rs

@@ -408,7 +408,7 @@ impl OpenFiles {
 /// error types may be elegible for retry.
 pub(crate) fn is_fatal_io_error(e: &std::io::Error) -> bool {
     use nix::errno::Errno::*;
-    match e.raw_os_error().map(nix::errno::from_i32) {
+    match e.raw_os_error().map(nix::errno::Errno::from_raw) {
         Some(EIO) => {
             // Terminate on EIO because we no longer trust the device to store
             // data safely, or to uphold persistence guarantees on fsync.

pageserver/src/virtual_file/io_engine.rs

@@ -124,9 +124,7 @@ pub(super) fn epoll_uring_error_to_std(
 ) -> std::io::Error {
     match e {
         tokio_epoll_uring::Error::Op(e) => e,
-        tokio_epoll_uring::Error::System(system) => {
-            std::io::Error::new(std::io::ErrorKind::Other, system)
-        }
+        tokio_epoll_uring::Error::System(system) => std::io::Error::other(system),
     }
 }
 

pgxn/neon/file_cache.c

@@ -936,6 +936,44 @@ lfc_prewarm_main(Datum main_arg)
 	lfc_ctl->prewarm_workers[worker_id].completed = GetCurrentTimestamp();
 }
 
+void
+lfc_invalidate(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber nblocks)
+{
+	BufferTag	tag;
+	FileCacheEntry *entry;
+	uint32		hash;
+
+	if (lfc_maybe_disabled())	/* fast exit if file cache is disabled */
+		return;
+
+	CopyNRelFileInfoToBufTag(tag, rinfo);
+	tag.forkNum = forkNum;
+
+	CriticalAssert(BufTagGetRelNumber(&tag) != InvalidRelFileNumber);
+
+	LWLockAcquire(lfc_lock, LW_EXCLUSIVE);
+	if (LFC_ENABLED())
+	{
+		for (BlockNumber blkno = 0; blkno < nblocks; blkno += lfc_blocks_per_chunk)
+		{
+			tag.blockNum = blkno;
+			hash = get_hash_value(lfc_hash, &tag);
+			entry = hash_search_with_hash_value(lfc_hash, &tag, hash, HASH_FIND, NULL);
+			if (entry != NULL)
+			{
+				for (int i = 0; i < lfc_blocks_per_chunk; i++)
+				{
+					if (GET_STATE(entry, i) == AVAILABLE)
+					{
+						lfc_ctl->used_pages -= 1;
+						SET_STATE(entry, i, UNAVAILABLE);
+					}
+				}
+			}
+		}
+	}
+	LWLockRelease(lfc_lock);
+}
 
 /*
  * Check if page is present in the cache.

pgxn/neon/file_cache.h

@@ -28,6 +28,7 @@ typedef struct FileCacheState
 extern bool lfc_store_prefetch_result;
 
 /* functions for local file cache */
+extern void lfc_invalidate(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber nblocks);
 extern void lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum,
 					   BlockNumber blkno, const void *const *buffers,
 					   BlockNumber nblocks);

pgxn/neon/neon_pgversioncompat.h

@@ -86,7 +86,7 @@ InitBufferTag(BufferTag *tag, const RelFileNode *rnode,
 
 #define InvalidRelFileNumber InvalidOid
 
-#define SMgrRelGetRelInfo(reln) \
+#define SMgrRelGetRelInfo(reln)				\
 	(reln->smgr_rnode.node)
 
 #define DropRelationAllLocalBuffers DropRelFileNodeAllLocalBuffers
@@ -148,6 +148,12 @@ InitBufferTag(BufferTag *tag, const RelFileNode *rnode,
 #define DropRelationAllLocalBuffers DropRelationAllLocalBuffers
 #endif
 
+#define NRelFileInfoInvalidate(rinfo) do { \
+		NInfoGetSpcOid(rinfo) = InvalidOid; \
+		NInfoGetDbOid(rinfo) = InvalidOid; \
+		NInfoGetRelNumber(rinfo) = InvalidRelFileNumber; \
+	} while (0)
+
 #if PG_MAJORVERSION_NUM < 17
 #define ProcNumber BackendId
 #define INVALID_PROC_NUMBER InvalidBackendId

pgxn/neon/pagestore_smgr.c

@@ -108,7 +108,7 @@ typedef enum
 	UNLOGGED_BUILD_NOT_PERMANENT
 } UnloggedBuildPhase;
 
-static SMgrRelation unlogged_build_rel = NULL;
+static NRelFileInfo unlogged_build_rel_info;
 static UnloggedBuildPhase unlogged_build_phase = UNLOGGED_BUILD_NOT_IN_PROGRESS;
 
 static bool neon_redo_read_buffer_filter(XLogReaderState *record, uint8 block_id);
@@ -912,16 +912,19 @@ neon_extend(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno,
 	{
 		case 0:
 			neon_log(ERROR, "cannot call smgrextend() on rel with unknown persistence");
+			break;
 
 		case RELPERSISTENCE_PERMANENT:
+			if (RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)))
+			{
+				mdextend(reln, forkNum, blkno, buffer, skipFsync);
+				return;
+			}
 			break;
 
 		case RELPERSISTENCE_TEMP:
 		case RELPERSISTENCE_UNLOGGED:
 			mdextend(reln, forkNum, blkno, buffer, skipFsync);
-			/* Update LFC in case of unlogged index build */
-			if (reln == unlogged_build_rel && unlogged_build_phase == UNLOGGED_BUILD_PHASE_2)
-				lfc_write(InfoFromSMgrRel(reln), forkNum, blkno, buffer);
 			return;
 
 		default:
@@ -1003,21 +1006,19 @@ neon_zeroextend(SMgrRelation reln, ForkNumber forkNum, BlockNumber blocknum,
 	{
 		case 0:
 			neon_log(ERROR, "cannot call smgrextend() on rel with unknown persistence");
+			break;
 
 		case RELPERSISTENCE_PERMANENT:
+			if (RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)))
+			{
+				mdzeroextend(reln, forkNum, blocknum, nblocks, skipFsync);
+				return;
+			}
 			break;
 
 		case RELPERSISTENCE_TEMP:
 		case RELPERSISTENCE_UNLOGGED:
 			mdzeroextend(reln, forkNum, blocknum, nblocks, skipFsync);
-			/* Update LFC in case of unlogged index build */
-			if (reln == unlogged_build_rel && unlogged_build_phase == UNLOGGED_BUILD_PHASE_2)
-			{
-				for (int i = 0; i < nblocks; i++)
-				{
-					lfc_write(InfoFromSMgrRel(reln), forkNum, blocknum + i, buffer.data);
-				}
-			}
 			return;
 
 		default:
@@ -1281,75 +1282,24 @@ neon_read_at_lsn(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 	communicator_read_at_lsnv(rinfo, forkNum, blkno, &request_lsns, &buffer, 1, NULL);
 }
 
-#if PG_MAJORVERSION_NUM < 17
-/*
- *	neon_read() -- Read the specified block from a relation.
- */
-#if PG_MAJORVERSION_NUM < 16
-static void
-neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, char *buffer)
-#else
-static void
-neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer)
-#endif
-{
-	neon_request_lsns request_lsns;
-	bits8		present;
-	void	   *bufferp;
-
-	switch (reln->smgr_relpersistence)
-	{
-		case 0:
-			neon_log(ERROR, "cannot call smgrread() on rel with unknown persistence");
-
-		case RELPERSISTENCE_PERMANENT:
-			break;
-
-		case RELPERSISTENCE_TEMP:
-		case RELPERSISTENCE_UNLOGGED:
-			mdread(reln, forkNum, blkno, buffer);
-			return;
-
-		default:
-			neon_log(ERROR, "unknown relpersistence '%c'", reln->smgr_relpersistence);
-	}
-
-	/* Try to read PS results if they are available */
-	communicator_prefetch_pump_state();
-
-	neon_get_request_lsns(InfoFromSMgrRel(reln), forkNum, blkno, &request_lsns, 1);
-
-	present = 0;
-	bufferp = buffer;
-	if (communicator_prefetch_lookupv(InfoFromSMgrRel(reln), forkNum, blkno, &request_lsns, 1, &bufferp, &present))
-	{
-		/* Prefetch hit */
-		return;
-	}
-
-	/* Try to read from local file cache */
-	if (lfc_read(InfoFromSMgrRel(reln), forkNum, blkno, buffer))
-	{
-		MyNeonCounters->file_cache_hits_total++;
-		return;
-	}
-
-	neon_read_at_lsn(InfoFromSMgrRel(reln), forkNum, blkno, request_lsns, buffer);
-
-	/*
-	 * Try to receive prefetch results once again just to make sure we don't leave the smgr code while the OS might still have buffered bytes.
-	 */
-	communicator_prefetch_pump_state();
-
 #ifdef DEBUG_COMPARE_LOCAL
+static void
+compare_with_local(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void* buffer, XLogRecPtr request_lsn)
+{
 	if (forkNum == MAIN_FORKNUM && IS_LOCAL_REL(reln))
 	{
 		char		pageserver_masked[BLCKSZ];
 		PGIOAlignedBlock mdbuf;
 		PGIOAlignedBlock mdbuf_masked;
-		XLogRecPtr  request_lsn = request_lsns.request_lsn;
 
+#if PG_MAJORVERSION_NUM >= 17
+		{
+			void* mdbuffers[1] = { mdbuf.data };
+			mdreadv(reln, forkNum, blkno, mdbuffers, 1);
+		}
+#else
 		mdread(reln, forkNum, blkno, mdbuf.data);
+#endif
 
 		memcpy(pageserver_masked, buffer, BLCKSZ);
 		memcpy(mdbuf_masked.data, mdbuf.data, BLCKSZ);
@@ -1413,11 +1363,111 @@ neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer
 			}
 		}
 	}
+}
+#endif
+
+
+#if PG_MAJORVERSION_NUM < 17
+
+/*
+ *	neon_read() -- Read the specified block from a relation.
+ */
+#if PG_MAJORVERSION_NUM < 16
+static void
+neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, char *buffer)
+#else
+static void
+neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer)
+#endif
+{
+	neon_request_lsns request_lsns;
+	bits8		present;
+	void	   *bufferp;
+
+	switch (reln->smgr_relpersistence)
+	{
+		case 0:
+			neon_log(ERROR, "cannot call smgrread() on rel with unknown persistence");
+			break;
+
+		case RELPERSISTENCE_PERMANENT:
+			if (RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)))
+			{
+				mdread(reln, forkNum, blkno, buffer);
+				return;
+			}
+			break;
+
+		case RELPERSISTENCE_TEMP:
+		case RELPERSISTENCE_UNLOGGED:
+			mdread(reln, forkNum, blkno, buffer);
+			return;
+
+		default:
+			neon_log(ERROR, "unknown relpersistence '%c'", reln->smgr_relpersistence);
+	}
+
+	/* Try to read PS results if they are available */
+	communicator_prefetch_pump_state();
+
+	neon_get_request_lsns(InfoFromSMgrRel(reln), forkNum, blkno, &request_lsns, 1);
+
+	present = 0;
+	bufferp = buffer;
+	if (communicator_prefetch_lookupv(InfoFromSMgrRel(reln), forkNum, blkno, &request_lsns, 1, &bufferp, &present))
+	{
+		/* Prefetch hit */
+#ifdef DEBUG_COMPARE_LOCAL
+		compare_with_local(reln, forkNum, blkno, buffer, request_lsns.request_lsn);
+#else
+		return;
+#endif
+	}
+
+	/* Try to read from local file cache */
+	if (lfc_read(InfoFromSMgrRel(reln), forkNum, blkno, buffer))
+	{
+		MyNeonCounters->file_cache_hits_total++;
+#ifdef DEBUG_COMPARE_LOCAL
+		compare_with_local(reln, forkNum, blkno, buffer, request_lsns.request_lsn);
+#else
+		return;
+#endif
+	}
+
+	neon_read_at_lsn(InfoFromSMgrRel(reln), forkNum, blkno, request_lsns, buffer);
+
+	/*
+	 * Try to receive prefetch results once again just to make sure we don't leave the smgr code while the OS might still have buffered bytes.
+	 */
+	communicator_prefetch_pump_state();
+
+#ifdef DEBUG_COMPARE_LOCAL
+	compare_with_local(reln, forkNum, blkno, buffer, request_lsns.request_lsn);
 #endif
 }
 #endif /* PG_MAJORVERSION_NUM <= 16 */
 
 #if PG_MAJORVERSION_NUM >= 17
+
+#ifdef DEBUG_COMPARE_LOCAL
+static void
+compare_with_localv(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void** buffers, BlockNumber nblocks, neon_request_lsns* request_lsns, bits8* read_pages)
+{
+	if (forkNum == MAIN_FORKNUM && IS_LOCAL_REL(reln))
+	{
+		for (BlockNumber i = 0; i < nblocks; i++)
+		{
+			if (BITMAP_ISSET(read_pages, i))
+			{
+				compare_with_local(reln, forkNum, blkno + i, buffers[i], request_lsns[i].request_lsn);
+			}
+		}
+	}
+}
+#endif
+
+
 static void
 neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 		   void **buffers, BlockNumber nblocks)
@@ -1431,8 +1481,14 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 	{
 		case 0:
 			neon_log(ERROR, "cannot call smgrread() on rel with unknown persistence");
+			break;
 
 		case RELPERSISTENCE_PERMANENT:
+			if (RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)))
+			{
+				mdreadv(reln, forknum, blocknum, buffers, nblocks);
+				return;
+			}
 			break;
 
 		case RELPERSISTENCE_TEMP:
@@ -1460,8 +1516,13 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 													blocknum, request_lsns, nblocks,
 													buffers, read_pages);
 
+#ifdef DEBUG_COMPARE_LOCAL
+	compare_with_localv(reln, forknum, blocknum, buffers, nblocks, request_lsns, read_pages);
+	memset(read_pages, 0, sizeof(read_pages));
+#else
 	if (prefetch_result == nblocks)
 		return;
+#endif
 
 	/* Try to read from local file cache */
 	lfc_result = lfc_readv_select(InfoFromSMgrRel(reln), forknum, blocknum, buffers,
@@ -1470,9 +1531,14 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 	if (lfc_result > 0)
 		MyNeonCounters->file_cache_hits_total += lfc_result;
 
+#ifdef DEBUG_COMPARE_LOCAL
+	compare_with_localv(reln, forknum, blocknum, buffers, nblocks, request_lsns, read_pages);
+	memset(read_pages, 0, sizeof(read_pages));
+#else
 	/* Read all blocks from LFC, so we're done */
 	if (prefetch_result + lfc_result == nblocks)
 		return;
+#endif
 
 	communicator_read_at_lsnv(InfoFromSMgrRel(reln), forknum, blocknum, request_lsns,
 							  buffers, nblocks, read_pages);
@@ -1483,91 +1549,8 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 	communicator_prefetch_pump_state();
 
 #ifdef DEBUG_COMPARE_LOCAL
-	if (forknum == MAIN_FORKNUM && IS_LOCAL_REL(reln))
-	{
-		char		pageserver_masked[BLCKSZ];
-		PGIOAlignedBlock mdbuf;
-		PGIOAlignedBlock mdbuf_masked;
-		XLogRecPtr  request_lsn = request_lsns->request_lsn;
-
-		for (int i = 0; i < nblocks; i++)
-		{
-			BlockNumber blkno = blocknum + i;
-			if (!BITMAP_ISSET(read_pages, i))
-				continue;
-
-#if PG_MAJORVERSION_NUM >= 17
-			{
-				void* mdbuffers[1] = { mdbuf.data };
-				mdreadv(reln, forknum, blkno, mdbuffers, 1);
-			}
-#else
-			mdread(reln, forknum, blkno, mdbuf.data);
-#endif
-
-			memcpy(pageserver_masked, buffers[i], BLCKSZ);
-			memcpy(mdbuf_masked.data, mdbuf.data, BLCKSZ);
-
-			if (PageIsNew((Page) mdbuf.data))
-			{
-				if (!PageIsNew((Page) pageserver_masked))
-				{
-					neon_log(PANIC, "page is new in MD but not in Page Server at blk %u in rel %u/%u/%u fork %u (request LSN %X/%08X):\n%s\n",
-						 blkno,
-						 RelFileInfoFmt(InfoFromSMgrRel(reln)),
-						 forknum,
-						 (uint32) (request_lsn >> 32), (uint32) request_lsn,
-						 hexdump_page(buffers[i]));
-				}
-			}
-			else if (PageIsNew((Page) buffers[i]))
-			{
-				neon_log(PANIC, "page is new in Page Server but not in MD at blk %u in rel %u/%u/%u fork %u (request LSN %X/%08X):\n%s\n",
-					 blkno,
-					 RelFileInfoFmt(InfoFromSMgrRel(reln)),
-					 forknum,
-					 (uint32) (request_lsn >> 32), (uint32) request_lsn,
-					 hexdump_page(mdbuf.data));
-			}
-			else if (PageGetSpecialSize(mdbuf.data) == 0)
-			{
-				/* assume heap */
-				RmgrTable[RM_HEAP_ID].rm_mask(mdbuf_masked.data, blkno);
-				RmgrTable[RM_HEAP_ID].rm_mask(pageserver_masked, blkno);
-
-				if (memcmp(mdbuf_masked.data, pageserver_masked, BLCKSZ) != 0)
-				{
-					neon_log(PANIC, "heap buffers differ at blk %u in rel %u/%u/%u fork %u (request LSN %X/%08X):\n------ MD ------\n%s\n------ Page Server ------\n%s\n",
-						 blkno,
-						 RelFileInfoFmt(InfoFromSMgrRel(reln)),
-						 forknum,
-						 (uint32) (request_lsn >> 32), (uint32) request_lsn,
-						 hexdump_page(mdbuf_masked.data),
-						 hexdump_page(pageserver_masked));
-				}
-			}
-			else if (PageGetSpecialSize(mdbuf.data) == MAXALIGN(sizeof(BTPageOpaqueData)))
-			{
-				if (((BTPageOpaqueData *) PageGetSpecialPointer(mdbuf.data))->btpo_cycleid < MAX_BT_CYCLE_ID)
-				{
-					/* assume btree */
-					RmgrTable[RM_BTREE_ID].rm_mask(mdbuf_masked.data, blkno);
-					RmgrTable[RM_BTREE_ID].rm_mask(pageserver_masked, blkno);
-	
-					if (memcmp(mdbuf_masked.data, pageserver_masked, BLCKSZ) != 0)
-					{
-						neon_log(PANIC, "btree buffers differ at blk %u in rel %u/%u/%u fork %u (request LSN %X/%08X):\n------ MD ------\n%s\n------ Page Server ------\n%s\n",
-							 blkno,
-							 RelFileInfoFmt(InfoFromSMgrRel(reln)),
-							 forknum,
-							 (uint32) (request_lsn >> 32), (uint32) request_lsn,
-							 hexdump_page(mdbuf_masked.data),
-							 hexdump_page(pageserver_masked));
-					}
-				}
-			}
-		}
-	}
+	memset(read_pages, 0xFF, sizeof(read_pages));
+	compare_with_localv(reln, forknum, blocknum, buffers, nblocks, request_lsns, read_pages);
 #endif
 }
 #endif
@@ -1638,6 +1621,15 @@ neon_write(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, const vo
 			break;
 
 		case RELPERSISTENCE_PERMANENT:
+			if (RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)))
+			{
+#if PG_MAJORVERSION_NUM >= 17
+				mdwritev(reln, forknum, blocknum, &buffer, 1, skipFsync);
+#else
+				mdwrite(reln, forknum, blocknum, buffer, skipFsync);
+#endif
+				return;
+			}
 			break;
 
 		case RELPERSISTENCE_TEMP:
@@ -1647,9 +1639,6 @@ neon_write(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, const vo
 			#else
 			mdwrite(reln, forknum, blocknum, buffer, skipFsync);
 			#endif
-			/* Update LFC in case of unlogged index build */
-			if (reln == unlogged_build_rel && unlogged_build_phase == UNLOGGED_BUILD_PHASE_2)
-				lfc_write(InfoFromSMgrRel(reln), forknum, blocknum, buffer);
 			return;
 		default:
 			neon_log(ERROR, "unknown relpersistence '%c'", reln->smgr_relpersistence);
@@ -1710,14 +1699,16 @@ neon_writev(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
 			break;
 
 		case RELPERSISTENCE_PERMANENT:
+			if (RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)))
+			{
+				mdwritev(reln, forknum, blkno, buffers, nblocks, skipFsync);
+				return;
+			}
 			break;
 
 		case RELPERSISTENCE_TEMP:
 		case RELPERSISTENCE_UNLOGGED:
 			mdwritev(reln, forknum, blkno, buffers, nblocks, skipFsync);
-			/* Update LFC in case of unlogged index build */
-			if (reln == unlogged_build_rel && unlogged_build_phase == UNLOGGED_BUILD_PHASE_2)
-				lfc_writev(InfoFromSMgrRel(reln), forknum, blkno, buffers, nblocks);
 			return;
 		default:
 			neon_log(ERROR, "unknown relpersistence '%c'", reln->smgr_relpersistence);
@@ -1753,6 +1744,10 @@ neon_nblocks(SMgrRelation reln, ForkNumber forknum)
 			break;
 
 		case RELPERSISTENCE_PERMANENT:
+			if (RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)))
+			{
+				return mdnblocks(reln, forknum);
+			}
 			break;
 
 		case RELPERSISTENCE_TEMP:
@@ -1822,6 +1817,11 @@ neon_truncate(SMgrRelation reln, ForkNumber forknum, BlockNumber old_blocks, Blo
 			break;
 
 		case RELPERSISTENCE_PERMANENT:
+			if (RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)))
+			{
+				mdtruncate(reln, forknum, old_blocks, nblocks);
+				return;
+			}
 			break;
 
 		case RELPERSISTENCE_TEMP:
@@ -1960,7 +1960,6 @@ neon_start_unlogged_build(SMgrRelation reln)
 	 */
 	if (unlogged_build_phase != UNLOGGED_BUILD_NOT_IN_PROGRESS)
 		neon_log(ERROR, "unlogged relation build is already in progress");
-	Assert(unlogged_build_rel == NULL);
 
 	ereport(SmgrTrace,
 			(errmsg(NEON_TAG "starting unlogged build of relation %u/%u/%u",
@@ -1977,7 +1976,7 @@ neon_start_unlogged_build(SMgrRelation reln)
 
 		case RELPERSISTENCE_TEMP:
 		case RELPERSISTENCE_UNLOGGED:
-			unlogged_build_rel = reln;
+			unlogged_build_rel_info = InfoFromSMgrRel(reln);
 			unlogged_build_phase = UNLOGGED_BUILD_NOT_PERMANENT;
 #ifdef DEBUG_COMPARE_LOCAL
 			if (!IsParallelWorker())
@@ -1998,12 +1997,9 @@ neon_start_unlogged_build(SMgrRelation reln)
 		neon_log(ERROR, "cannot perform unlogged index build, index is not empty ");
 #endif
 
-	unlogged_build_rel = reln;
+	unlogged_build_rel_info = InfoFromSMgrRel(reln);
 	unlogged_build_phase = UNLOGGED_BUILD_PHASE_1;
 
-	/* Make the relation look like it's unlogged */
-	reln->smgr_relpersistence = RELPERSISTENCE_UNLOGGED;
-
 	/*
 	 * Create the local file. In a parallel build, the leader is expected to
 	 * call this first and do it.
@@ -2030,17 +2026,16 @@ neon_start_unlogged_build(SMgrRelation reln)
 static void
 neon_finish_unlogged_build_phase_1(SMgrRelation reln)
 {
-	Assert(unlogged_build_rel == reln);
+	Assert(RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)));
 
 	ereport(SmgrTrace,
 			(errmsg(NEON_TAG "finishing phase 1 of unlogged build of relation %u/%u/%u",
-					RelFileInfoFmt(InfoFromSMgrRel(reln)))));
+					RelFileInfoFmt((unlogged_build_rel_info)))));
 
 	if (unlogged_build_phase == UNLOGGED_BUILD_NOT_PERMANENT)
 		return;
 
 	Assert(unlogged_build_phase == UNLOGGED_BUILD_PHASE_1);
-	Assert(reln->smgr_relpersistence == RELPERSISTENCE_UNLOGGED);
 
 	/*
 	 * In a parallel build, (only) the leader process performs the 2nd
@@ -2048,7 +2043,7 @@ neon_finish_unlogged_build_phase_1(SMgrRelation reln)
 	 */
 	if (IsParallelWorker())
 	{
-		unlogged_build_rel = NULL;
+		NRelFileInfoInvalidate(unlogged_build_rel_info);
 		unlogged_build_phase = UNLOGGED_BUILD_NOT_IN_PROGRESS;
 	}
 	else
@@ -2069,11 +2064,11 @@ neon_end_unlogged_build(SMgrRelation reln)
 {
 	NRelFileInfoBackend rinfob = InfoBFromSMgrRel(reln);
 
-	Assert(unlogged_build_rel == reln);
+	Assert(RelFileInfoEquals(unlogged_build_rel_info, InfoFromSMgrRel(reln)));
 
 	ereport(SmgrTrace,
 			(errmsg(NEON_TAG "ending unlogged build of relation %u/%u/%u",
-					RelFileInfoFmt(InfoFromNInfoB(rinfob)))));
+					RelFileInfoFmt(unlogged_build_rel_info))));
 
 	if (unlogged_build_phase != UNLOGGED_BUILD_NOT_PERMANENT)
 	{
@@ -2081,7 +2076,6 @@ neon_end_unlogged_build(SMgrRelation reln)
 		BlockNumber nblocks;
 
 		Assert(unlogged_build_phase == UNLOGGED_BUILD_PHASE_2);
-		Assert(reln->smgr_relpersistence == RELPERSISTENCE_UNLOGGED);
 
 		/*
 		 * Update the last-written LSN cache.
@@ -2102,9 +2096,6 @@ neon_end_unlogged_build(SMgrRelation reln)
 								InfoFromNInfoB(rinfob),
 								MAIN_FORKNUM);
 
-		/* Make the relation look permanent again */
-		reln->smgr_relpersistence = RELPERSISTENCE_PERMANENT;
-
 		/* Remove local copy */
 		for (int forknum = 0; forknum <= MAX_FORKNUM; forknum++)
 		{
@@ -2113,6 +2104,8 @@ neon_end_unlogged_build(SMgrRelation reln)
 				 forknum);
 
 			forget_cached_relsize(InfoFromNInfoB(rinfob), forknum);
+			lfc_invalidate(InfoFromNInfoB(rinfob), forknum, nblocks);
+
 			mdclose(reln, forknum);
 #ifndef DEBUG_COMPARE_LOCAL
 			/* use isRedo == true, so that we drop it immediately */
@@ -2123,7 +2116,7 @@ neon_end_unlogged_build(SMgrRelation reln)
 		mdunlink(rinfob, INIT_FORKNUM, true);
 #endif
 	}
-	unlogged_build_rel = NULL;
+	NRelFileInfoInvalidate(unlogged_build_rel_info);
 	unlogged_build_phase = UNLOGGED_BUILD_NOT_IN_PROGRESS;
 }
 
@@ -2196,7 +2189,7 @@ AtEOXact_neon(XactEvent event, void *arg)
 			 * Forget about any build we might have had in progress. The local
 			 * file will be unlinked by smgrDoPendingDeletes()
 			 */
-			unlogged_build_rel = NULL;
+			NRelFileInfoInvalidate(unlogged_build_rel_info);
 			unlogged_build_phase = UNLOGGED_BUILD_NOT_IN_PROGRESS;
 			break;
 
@@ -2208,7 +2201,7 @@ AtEOXact_neon(XactEvent event, void *arg)
 		case XACT_EVENT_PRE_PREPARE:
 			if (unlogged_build_phase != UNLOGGED_BUILD_NOT_IN_PROGRESS)
 			{
-				unlogged_build_rel = NULL;
+				NRelFileInfoInvalidate(unlogged_build_rel_info);
 				unlogged_build_phase = UNLOGGED_BUILD_NOT_IN_PROGRESS;
 				ereport(ERROR,
 						(errcode(ERRCODE_INTERNAL_ERROR),

poetry.lock

@@ -1,4 +1,4 @@
-# This file is automatically @generated by Poetry 2.1.2 and should not be changed by hand.
+# This file is automatically @generated by Poetry 2.1.1 and should not be changed by hand.
 
 [[package]]
 name = "aiohappyeyeballs"
@@ -1145,18 +1145,19 @@ dotenv = ["python-dotenv"]
 
 [[package]]
 name = "flask-cors"
-version = "5.0.0"
-description = "A Flask extension adding a decorator for CORS support"
+version = "6.0.0"
+description = "A Flask extension simplifying CORS support"
 optional = false
-python-versions = "*"
+python-versions = "<4.0,>=3.9"
 groups = ["main"]
 files = [
-    {file = "Flask_Cors-5.0.0-py2.py3-none-any.whl", hash = "sha256:b9e307d082a9261c100d8fb0ba909eec6a228ed1b60a8315fd85f783d61910bc"},
-    {file = "flask_cors-5.0.0.tar.gz", hash = "sha256:5aadb4b950c4e93745034594d9f3ea6591f734bb3662e16e255ffbf5e89c88ef"},
+    {file = "flask_cors-6.0.0-py3-none-any.whl", hash = "sha256:6332073356452343a8ccddbfec7befdc3fdd040141fe776ec9b94c262f058657"},
+    {file = "flask_cors-6.0.0.tar.gz", hash = "sha256:4592c1570246bf7beee96b74bc0adbbfcb1b0318f6ba05c412e8909eceec3393"},
 ]
 
 [package.dependencies]
-Flask = ">=0.9"
+flask = ">=0.9"
+Werkzeug = ">=0.7"
 
 [[package]]
 name = "frozenlist"

proxy/src/auth/backend/console_redirect.rs

@@ -207,8 +207,15 @@ async fn authenticate(
     ctx.set_project(db_info.aux.clone());
     info!("woken up a compute node");
 
-    // we need TLS connection with SNI info to properly route it
-    config.ssl_mode(SslMode::Require);
+    // Backwards compatibility. pg_sni_proxy uses "--" in domain names
+    // while direct connections do not. Once we migrate to pg_sni_proxy
+    // everywhere, we can remove this.
+    if db_info.host.contains("--") {
+        // we need TLS connection with SNI info to properly route it
+        config.ssl_mode(SslMode::Require);
+    } else {
+        config.ssl_mode(SslMode::Disable);
+    }
 
     if let Some(password) = db_info.password {
         config.password(password.as_ref());

proxy/src/bin/proxy.rs

@@ -1,6 +1,10 @@
 #[global_allocator]
 static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
 
+#[allow(non_upper_case_globals)]
+#[unsafe(export_name = "malloc_conf")]
+pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:21\0";
+
 #[tokio::main]
 async fn main() -> anyhow::Result<()> {
     proxy::binary::proxy::run().await

proxy/src/binary/pg_sni_router.rs

@@ -394,6 +394,7 @@ async fn handle_client(
     }
 }
 
+#[allow(clippy::large_enum_variant)]
 enum Connection {
     Raw(tokio::net::TcpStream),
     Tls(tokio_rustls::client::TlsStream<tokio::net::TcpStream>),

proxy/src/binary/proxy.rs

@@ -43,11 +43,12 @@ project_build_tag!(BUILD_TAG);
 use clap::{Parser, ValueEnum};
 
 #[derive(Clone, Debug, ValueEnum)]
+#[clap(rename_all = "kebab-case")]
 enum AuthBackendType {
-    #[value(name("cplane-v1"), alias("control-plane"))]
-    ControlPlaneV1,
+    #[clap(alias("cplane-v1"))]
+    ControlPlane,
 
-    #[value(name("link"), alias("control-redirect"))]
+    #[clap(alias("link"))]
     ConsoleRedirect,
 
     #[cfg(any(test, feature = "testing"))]
@@ -707,7 +708,7 @@ fn build_auth_backend(
     args: &ProxyCliArgs,
 ) -> anyhow::Result<Either<&'static auth::Backend<'static, ()>, &'static ConsoleRedirectBackend>> {
     match &args.auth_backend {
-        AuthBackendType::ControlPlaneV1 => {
+        AuthBackendType::ControlPlane => {
             let wake_compute_cache_config: CacheOptions = args.wake_compute_cache.parse()?;
             let project_info_cache_config: ProjectInfoCacheOptions =
                 args.project_info_cache.parse()?;
@@ -862,7 +863,7 @@ async fn configure_redis(
         ("irsa", _) => match (&args.redis_host, args.redis_port) {
             (Some(host), Some(port)) => Some(
                 ConnectionWithCredentialsProvider::new_with_credentials_provider(
-                    host.to_string(),
+                    host.clone(),
                     port,
                     elasticache::CredentialsProvider::new(
                         args.aws_region.clone(),

proxy/src/cancellation.rs

@@ -6,12 +6,12 @@ use ipnet::{IpNet, Ipv4Net, Ipv6Net};
 use postgres_client::CancelToken;
 use postgres_client::tls::MakeTlsConnect;
 use pq_proto::CancelKeyData;
-use redis::{FromRedisValue, Pipeline, Value, pipe};
+use redis::{Cmd, FromRedisValue, Value};
 use serde::{Deserialize, Serialize};
 use thiserror::Error;
 use tokio::net::TcpStream;
 use tokio::sync::{mpsc, oneshot};
-use tracing::{debug, info, warn};
+use tracing::{debug, error, info, warn};
 
 use crate::auth::backend::ComputeUserInfo;
 use crate::auth::{AuthError, check_peer_addr_is_in_list};
@@ -56,8 +56,70 @@ pub enum CancelKeyOp {
     },
 }
 
+pub struct Pipeline {
+    inner: redis::Pipeline,
+    replies: Vec<CancelReplyOp>,
+}
+
+impl Pipeline {
+    fn with_capacity(n: usize) -> Self {
+        Self {
+            inner: redis::Pipeline::with_capacity(n),
+            replies: Vec::with_capacity(n),
+        }
+    }
+
+    async fn execute(&mut self, client: &mut RedisKVClient) {
+        let responses = self.replies.len();
+        let batch_size = self.inner.len();
+
+        match client.query(&self.inner).await {
+            // for each reply, we expect that many values.
+            Ok(Value::Array(values)) if values.len() == responses => {
+                debug!(
+                    batch_size,
+                    responses, "successfully completed cancellation jobs",
+                );
+                for (value, reply) in std::iter::zip(values, self.replies.drain(..)) {
+                    reply.send_value(value);
+                }
+            }
+            Ok(value) => {
+                error!(batch_size, ?value, "unexpected redis return value");
+                for reply in self.replies.drain(..) {
+                    reply.send_err(anyhow!("incorrect response type from redis"));
+                }
+            }
+            Err(err) => {
+                for reply in self.replies.drain(..) {
+                    reply.send_err(anyhow!("could not send cmd to redis: {err}"));
+                }
+            }
+        }
+
+        self.inner.clear();
+        self.replies.clear();
+    }
+
+    fn add_command_with_reply(&mut self, cmd: Cmd, reply: CancelReplyOp) {
+        self.inner.add_command(cmd);
+        self.replies.push(reply);
+    }
+
+    fn add_command_no_reply(&mut self, cmd: Cmd) {
+        self.inner.add_command(cmd).ignore();
+    }
+
+    fn add_command(&mut self, cmd: Cmd, reply: Option<CancelReplyOp>) {
+        match reply {
+            Some(reply) => self.add_command_with_reply(cmd, reply),
+            None => self.add_command_no_reply(cmd),
+        }
+    }
+}
+
 impl CancelKeyOp {
-    fn register(self, pipe: &mut Pipeline) -> Option<CancelReplyOp> {
+    fn register(self, pipe: &mut Pipeline) {
         #[allow(clippy::used_underscore_binding)]
         match self {
             CancelKeyOp::StoreCancelKey {
@@ -68,18 +130,18 @@ impl CancelKeyOp {
                 _guard,
                 expire,
             } => {
-                pipe.hset(&key, field, value);
-                pipe.expire(key, expire);
-                let resp_tx = resp_tx?;
-                Some(CancelReplyOp::StoreCancelKey { resp_tx, _guard })
+                let reply =
+                    resp_tx.map(|resp_tx| CancelReplyOp::StoreCancelKey { resp_tx, _guard });
+                pipe.add_command(Cmd::hset(&key, field, value), reply);
+                pipe.add_command_no_reply(Cmd::expire(key, expire));
             }
             CancelKeyOp::GetCancelData {
                 key,
                 resp_tx,
                 _guard,
             } => {
-                pipe.hgetall(key);
-                Some(CancelReplyOp::GetCancelData { resp_tx, _guard })
+                let reply = CancelReplyOp::GetCancelData { resp_tx, _guard };
+                pipe.add_command_with_reply(Cmd::hgetall(key), reply);
             }
             CancelKeyOp::RemoveCancelKey {
                 key,
@@ -87,9 +149,9 @@ impl CancelKeyOp {
                 resp_tx,
                 _guard,
             } => {
-                pipe.hdel(key, field);
-                let resp_tx = resp_tx?;
-                Some(CancelReplyOp::RemoveCancelKey { resp_tx, _guard })
+                let reply =
+                    resp_tx.map(|resp_tx| CancelReplyOp::RemoveCancelKey { resp_tx, _guard });
+                pipe.add_command(Cmd::hdel(key, field), reply);
             }
         }
     }
@@ -170,8 +232,8 @@ pub async fn handle_cancel_messages(
     client: &mut RedisKVClient,
     mut rx: mpsc::Receiver<CancelKeyOp>,
 ) -> anyhow::Result<()> {
-    let mut batch = Vec::new();
-    let mut replies = vec![];
+    let mut batch = Vec::with_capacity(BATCH_SIZE);
+    let mut pipeline = Pipeline::with_capacity(BATCH_SIZE);
 
     loop {
         if rx.recv_many(&mut batch, BATCH_SIZE).await == 0 {
@@ -182,42 +244,11 @@ pub async fn handle_cancel_messages(
         let batch_size = batch.len();
         debug!(batch_size, "running cancellation jobs");
 
-        let mut pipe = pipe();
         for msg in batch.drain(..) {
-            if let Some(reply) = msg.register(&mut pipe) {
-                replies.push(reply);
-            } else {
-                pipe.ignore();
-            }
+            msg.register(&mut pipeline);
         }
 
-        let responses = replies.len();
-
-        match client.query(pipe).await {
-            // for each reply, we expect that many values.
-            Ok(Value::Array(values)) if values.len() == responses => {
-                debug!(
-                    batch_size,
-                    responses, "successfully completed cancellation jobs",
-                );
-                for (value, reply) in std::iter::zip(values, replies.drain(..)) {
-                    reply.send_value(value);
-                }
-            }
-            Ok(value) => {
-                debug!(?value, "unexpected redis return value");
-                for reply in replies.drain(..) {
-                    reply.send_err(anyhow!("incorrect response type from redis"));
-                }
-            }
-            Err(err) => {
-                for reply in replies.drain(..) {
-                    reply.send_err(anyhow!("could not send cmd to redis: {err}"));
-                }
-            }
-        }
-
-        replies.clear();
+        pipeline.execute(client).await;
     }
 }
 

proxy/src/context/mod.rs

@@ -78,7 +78,7 @@ struct RequestContextInner {
 
 #[derive(Clone, Debug)]
 pub(crate) enum AuthMethod {
-    // aka passwordless, fka link
+    // aka link
     ConsoleRedirect,
     ScramSha256,
     ScramSha256Plus,

proxy/src/http/health_server.rs

@@ -3,7 +3,7 @@ use std::net::TcpListener;
 use std::sync::{Arc, Mutex};
 
 use anyhow::{anyhow, bail};
-use http_utils::endpoint::{self, request_span};
+use http_utils::endpoint::{self, profile_cpu_handler, profile_heap_handler, request_span};
 use http_utils::error::ApiError;
 use http_utils::json::json_response;
 use http_utils::{RouterBuilder, RouterService};
@@ -33,6 +33,12 @@ fn make_router(metrics: AppMetrics) -> RouterBuilder<hyper0::Body, ApiError> {
             request_span(r, move |b| prometheus_metrics_handler(b, state))
         })
         .get("/v1/status", status_handler)
+        .get("/profile/cpu", move |r| {
+            request_span(r, profile_cpu_handler)
+        })
+        .get("/profile/heap", move |r| {
+            request_span(r, profile_heap_handler)
+        })
 }
 
 pub async fn task_main(

proxy/src/redis/kv_ops.rs

@@ -47,7 +47,7 @@ impl RedisKVClient {
 
     pub(crate) async fn query<T: FromRedisValue>(
         &mut self,
-        q: impl Queryable,
+        q: &impl Queryable,
     ) -> anyhow::Result<T> {
         if !self.limiter.check() {
             tracing::info!("Rate limit exceeded. Skipping query");