wip

## Refs

- Epic: https://github.com/neondatabase/neon/issues/9378

Co-authored-by: Vlad Lazar <vlad@neon.tech>
Co-authored-by: Christian Schwarz <christian@neon.tech>

## Problem

The read path does its IOs sequentially.
This means that if N values need to be read to reconstruct a page,
we will do N IOs and getpage latency is `O(N*IoLatency)`.

## Solution

With this PR we gain the ability to issue IO concurrently within one
layer visit **and** to move on to the next layer without waiting for IOs
from the previous visit to complete.

This is an evolved version of the work done at the Lisbon hackathon,
cf https://github.com/neondatabase/neon/pull/9002.

## Design

### `will_init` now sourced from disk btree index keys

On the algorithmic level, the only change is that the
`get_values_reconstruct_data`
now sources `will_init` from the disk btree index key (which is
PS-page_cache'd), instead
of from the `Value`, which is only available after the IO completes.

### Concurrent IOs, Submission & Completion 

To separate IO submission from waiting for its completion, while
simultaneously
feature-gating the change, we introduce the notion of an `IoConcurrency`
struct
through which IO futures are "spawned".

An IO is an opaque future, and waiting for completions is handled
through
`tokio::sync::oneshot` channels.
The oneshot Receiver's take the place of the `img` and `records` fields
inside `VectoredValueReconstructState`.

When we're done visiting all the layers and submitting all the IOs along
the way
we concurrently `collect_pending_ios` for each value, which means
for each value there is a future that awaits all the oneshot receivers
and then calls into walredo to reconstruct the page image.
Walredo is now invoked concurrently for each value instead of
sequentially.
Walredo itself remains unchanged.

The spawned IO futures are driven to completion by a sidecar tokio task
that
is separate from the task that performs all the layer visiting and
spawning of IOs.
That tasks receives the IO futures via an unbounded mpsc channel and
drives them to completion inside a `FuturedUnordered`.

(The behavior from before this PR is available through
`IoConcurrency::Sequential`,
which awaits the IO futures in place, without "spawning" or "submitting"
them
anywhere.)

#### Alternatives Explored

A few words on the rationale behind having a sidecar *task* and what
alternatives were considered.

One option is to queue up all IO futures in a FuturesUnordered that is
polled
the first time when we `collect_pending_ios`.

Firstly, the IO futures are opaque, compiler-generated futures that need
to be polled at least once to submit their IO. "At least once" because
tokio-epoll-uring may not be able to submit the IO to the kernel on
first
poll right away.

Second, there are deadlocks if we don't drive the IO futures to
completion
independently of the spawning task.
The reason is that both the IO futures and the spawning task may hold
some
_and_ try to acquire _more_ shared limited resources.
For example, both spawning task and IO future may try to acquire
* a VirtualFile file descriptor cache slot async mutex (observed during
impl)
* a tokio-epoll-uring submission slot (observed during impl)
* a PageCache slot (currently this is not the case but we may move more
code into the IO futures in the future)

Another option is to spawn a short-lived `tokio::task` for each IO
future.
We implemented and benchmarked it during development, but found little
throughput improvement and moderate mean & tail latency degradation.
Concerns about pressure on the tokio scheduler made us discard this
variant.

The sidecar task could be obsoleted if the IOs were not arbitrary code
but a well-defined struct.
However,
1. the opaque futures approach taken in this PR allows leaving the
existing
   code unchanged, which
2. allows us to implement the `IoConcurrency::Sequential` mode for
feature-gating
   the change.

Once the new mode sidecar task implementation is rolled out everywhere,
and `::Sequential` removed, we can think about a descriptive submission
& completion interface.
The problems around deadlocks pointed out earlier will need to be solved
then.
For example, we could eliminate VirtualFile file descriptor cache and
tokio-epoll-uring slots.
The latter has been drafted in
https://github.com/neondatabase/tokio-epoll-uring/pull/63.

See the lengthy doc comment on `spawn_io()` for more details.

### Error handling

There are two error classes during reconstruct data retrieval:
* traversal errors: index lookup, move to next layer, and the like
* value read IO errors

A traversal error fails the entire get_vectored request, as before this
PR.
A value read error only fails that value.

In any case, we preserve the existing behavior that once
`get_vectored` returns, all IOs are done. Panics and failing
to poll `get_vectored` to completion will leave the IOs dangling,
which is safe but shouldn't happen, and so, a rate-limited
log statement will be emitted at warning level.
There is a doc comment on `collect_pending_ios` giving more code-level
details and rationale.

### Feature Gating

The new behavior is opt-in via pageserver config.
The `Sequential` mode is the default.
The only significant change in `Sequential` mode compared to before
this PR is the buffering of results in the `oneshot`s.

## Code-Level Changes

Prep work:
  * Make `GateGuard` clonable.

Core Feature:
* Traversal code: track  `will_init` in `BlobMeta` and source it from
the Delta/Image/InMemory layer index, instead of determining `will_init`
  after we've read the value. This avoids having to read the value to
  determine whether traversal can stop.
* Introduce `IoConcurrency` & its sidecar task.
  * `IoConcurrency` is the clonable handle.
  * It connects to the sidecar task via an `mpsc`.
* Plumb through `IoConcurrency` from high level code to the
  individual layer implementations' `get_values_reconstruct_data`.
  We piggy-back on the `ValuesReconstructState` for this.
   * The sidecar task should be long-lived, so, `IoConcurrency` needs
     to be rooted up "high" in the call stack.
   * Roots as of this PR:
     * `page_service`: outside of pagestream loop
     * `create_image_layers`: when it is called
     * `basebackup`(only auxfiles + replorigin + SLRU segments)
   * Code with no roots that uses `IoConcurrency::sequential`
     * any `Timeline::get` call
       * `collect_keyspace` is a good example
       * follow-up: https://github.com/neondatabase/neon/issues/10460
* `TimelineAdaptor` code used by the compaction simulator, unused in
practive
     * `ingest_xlog_dbase_create`
* Transform Delta/Image/InMemoryLayer to
  * do their values IO in a distinct `async {}` block
  * extend the residence of the Delta/Image layer until the IO is done
  * buffer their results in a `oneshot` channel instead of straight
    in `ValuesReconstructState` 
* the `oneshot` channel is wrapped in `OnDiskValueIo` /
`OnDiskValueIoWaiter`
    types that aid in expressiveness and are used to keep track of
    in-flight IOs so we can print warnings if we leave them dangling.
* Change `ValuesReconstructState` to hold the receiving end of the
 `oneshot` channel aka `OnDiskValueIoWaiter`.
* Change `get_vectored_impl` to `collect_pending_ios` and issue walredo
concurrently, in a `FuturesUnordered`.

Testing / Benchmarking:
* Support queue-depth in pagebench for manual benchmarkinng.
* Add test suite support for setting concurrency mode ps config
   field via a) an env var and b) via NeonEnvBuilder.
* Hacky helper to have sidecar-based IoConcurrency in tests.
   This will be cleaned up later.

More benchmarking will happen post-merge in nightly benchmarks, plus in
staging/pre-prod.
Some intermediate helpers for manual benchmarking have been preserved in
https://github.com/neondatabase/neon/pull/10466 and will be landed in
later PRs.
(L0 layer stack generator!)

Drive-By:
* test suite actually didn't enable batching by default because
`config.compatibility_neon_binpath` is always Truthy in our CI
environment
  => https://neondb.slack.com/archives/C059ZC138NR/p1737490501941309
* initial logical size calculation wasn't always polled to completion,
which was
  surfaced through the added WARN logs emitted when dropping a 
  `ValuesReconstructState` that still has inflight IOs.
* remove the timing histograms
`pageserver_getpage_get_reconstruct_data_seconds`
and `pageserver_getpage_reconstruct_seconds` because with planning,
value read
IO, and walredo happening concurrently, one can no longer attribute
latency
to any one of them; we'll revisit this when Vlad's work on
tracing/sampling
  through RequestContext lands.
* remove code related to `get_cached_lsn()`.
  The logic around this has been dead at runtime for a long time,
  ever since the removal of the materialized page cache in #8105.

## Testing

Unit tests use the sidecar task by default and run both modes in CI.
Python regression tests and benchmarks also use the sidecar task by
default.
We'll test more in staging and possibly preprod.

# Future Work

Please refer to the parent epic for the full plan.

The next step will be to fold the plumbing of IoConcurrency
into RequestContext so that the function signatures get cleaned up.

Once `Sequential` isn't used anymore, we can take the next
big leap which is replacing the opaque IOs with structs
that have well-defined semantics.

---------

Co-authored-by: Christian Schwarz <christian@neon.tech>

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

@@ -229,8 +229,13 @@ jobs:
           ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES -E '!package(pageserver)'
 
           # run pageserver tests with different settings
-          for io_engine in std-fs tokio-epoll-uring ; do
-            NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE=$io_engine ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES  -E 'package(pageserver)'
+          for get_vectored_concurrent_io in sequential sidecar-task; do
+            for io_engine in std-fs tokio-epoll-uring ; do
+              NEON_PAGESERVER_UNIT_TEST_GET_VECTORED_CONCURRENT_IO=$get_vectored_concurrent_io \
+                NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE=$io_engine \
+                ${cov_prefix} \
+                cargo nextest run $CARGO_FLAGS $CARGO_FEATURES  -E 'package(pageserver)'
+            done
           done
 
           # Run separate tests for real S3
@@ -314,6 +319,7 @@ 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
           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

@@ -824,7 +824,7 @@ jobs:
           docker compose -f ./docker-compose/docker-compose.yml down
 
   promote-images-dev:
-    needs: [ check-permissions, tag, vm-compute-node-image ]
+    needs: [ check-permissions, tag, vm-compute-node-image, neon-image ]
     runs-on: ubuntu-22.04
 
     permissions:
@@ -1078,12 +1078,6 @@ jobs:
               console.log(`Tag ${tag} created successfully.`);
             }
 
-            // TODO: check how GitHub releases looks for proxy/compute releases and enable them if they're ok
-            if (context.ref !== 'refs/heads/release') {
-              console.log(`GitHub release skipped for ${context.ref}.`);
-              return;
-            }
-
             try {
               const existingRelease = await github.rest.repos.getReleaseByTag({
                 owner: context.repo.owner,
@@ -1102,7 +1096,8 @@ jobs:
                 owner: context.repo.owner,
                 repo: context.repo.repo,
                 tag_name: tag,
-                generate_release_notes: true,
+                // TODO: Automate release notes properly
+                generate_release_notes: false,
               });
               console.log(`Release for tag ${tag} created successfully.`);
             }

Cargo.lock

@@ -7004,12 +7004,9 @@ version = "0.12.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "877c5b330756d856ffcc4553ab34a5684481ade925ecc54bcd1bf02b1d0d4d52"
 dependencies = [
- "async-stream",
  "async-trait",
- "axum",
  "base64 0.22.1",
  "bytes",
- "h2 0.4.4",
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
@@ -7021,7 +7018,6 @@ dependencies = [
  "prost",
  "rustls-native-certs 0.8.0",
  "rustls-pemfile 2.1.1",
- "socket2",
  "tokio",
  "tokio-rustls 0.26.0",
  "tokio-stream",
@@ -7582,7 +7578,6 @@ dependencies = [
  "tikv-jemallocator",
  "tokio",
  "tokio-util",
- "tonic",
  "tonic-build",
  "tracing",
  "utils",
@@ -7991,8 +7986,6 @@ version = "0.1.0"
 dependencies = [
  "ahash",
  "anyhow",
- "axum",
- "axum-core",
  "base64 0.13.1",
  "base64 0.21.1",
  "base64ct",
@@ -8073,7 +8066,6 @@ dependencies = [
  "toml_edit",
  "tonic",
  "tower 0.4.13",
- "tower 0.5.2",
  "tracing",
  "tracing-core",
  "url",

Cargo.toml

@@ -187,7 +187,7 @@ tokio-tar = "0.3"
 tokio-util = { version = "0.7.10", features = ["io", "rt"] }
 toml = "0.8"
 toml_edit = "0.22"
-tonic = {version = "0.12.3", features = ["tls", "tls-roots"]}
+tonic = {version = "0.12.3", default-features = false, features = ["channel", "tls", "tls-roots"]}
 tower = { version = "0.5.2", default-features = false }
 tower-http = { version = "0.6.2", features = ["request-id", "trace"] }
 tower-service = "0.3.3"

compute/compute-node.Dockerfile

@@ -360,6 +360,8 @@ COPY compute/patches/pgvector.patch /pgvector.patch
 RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.8.0.tar.gz -O pgvector.tar.gz && \
     echo "867a2c328d4928a5a9d6f052cd3bc78c7d60228a9b914ad32aa3db88e9de27b0 pgvector.tar.gz" | sha256sum --check && \
     mkdir pgvector-src && cd pgvector-src && tar xzf ../pgvector.tar.gz --strip-components=1 -C . && \
+    wget https://github.com/pgvector/pgvector/raw/refs/tags/v0.7.4/sql/vector.sql -O ./sql/vector--0.7.4.sql && \
+    echo "10218d05dc02299562252a9484775178b14a1d8edb92a2d1672ef488530f7778 ./sql/vector--0.7.4.sql" | sha256sum --check && \
     patch -p1 < /pgvector.patch && \
     make -j $(getconf _NPROCESSORS_ONLN) OPTFLAGS="" && \
     make -j $(getconf _NPROCESSORS_ONLN) OPTFLAGS="" install && \
@@ -1347,9 +1349,6 @@ COPY --from=pg-roaringbitmap-pg-build /pg_roaringbitmap.tar.gz /ext-src
 COPY --from=pg-semver-pg-build /pg_semver.tar.gz /ext-src
 #COPY --from=pg-embedding-pg-build /home/nonroot/pg_embedding-src/ /ext-src
 #COPY --from=wal2json-pg-build /wal2json_2_5.tar.gz /ext-src
-#pg_anon is not supported yet for pg v17 so, don't fail if nothing found
-COPY --from=pg-anon-pg-build /pg_anon.tar.g? /ext-src
-COPY compute/patches/pg_anon.patch /ext-src
 COPY --from=pg-ivm-build /pg_ivm.tar.gz /ext-src
 COPY --from=pg-partman-build /pg_partman.tar.gz /ext-src
 RUN cd /ext-src/ && for f in *.tar.gz; \
@@ -1360,9 +1359,6 @@ RUN cd /ext-src/rum-src && patch -p1 <../rum.patch
 RUN cd /ext-src/pgvector-src && patch -p1 <../pgvector.patch
 RUN cd /ext-src/pg_hint_plan-src && patch -p1 < /ext-src/pg_hint_plan_${PG_VERSION}.patch
 COPY --chmod=755 docker-compose/run-tests.sh /run-tests.sh
-RUN case "${PG_VERSION}" in "v17") \
-    echo "postgresql_anonymizer does not yet support PG17" && exit 0;; \
-    esac && patch -p1 </ext-src/pg_anon.patch
 RUN patch -p1 </ext-src/pg_cron.patch
 ENV PATH=/usr/local/pgsql/bin:$PATH
 ENV PGHOST=compute

compute/patches/pgvector.patch

@@ -1,8 +1,24 @@
+diff --git a/Makefile b/Makefile
+index 7a4b88c..56678af 100644
+--- a/Makefile
++++ b/Makefile
+@@ -3,7 +3,10 @@ EXTVERSION = 0.8.0
+ 
+ MODULE_big = vector
+ DATA = $(wildcard sql/*--*--*.sql)
+-DATA_built = sql/$(EXTENSION)--$(EXTVERSION).sql
++# This change is needed to install different per-version SQL files
++# like pgvector--0.8.0.sql and pgvector--0.7.4.sql
++# The corresponding file is downloaded during the Docker image build process
++DATA_built = sql/$(EXTENSION)--$(EXTVERSION).sql sql/vector--0.7.4.sql
+ OBJS = src/bitutils.o src/bitvec.o src/halfutils.o src/halfvec.o src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/sparsevec.o src/vector.o
+ HEADERS = src/halfvec.h src/sparsevec.h src/vector.h
+ 
 diff --git a/src/hnswbuild.c b/src/hnswbuild.c
-index dcfb2bd..d5189ee 100644
+index b667478..fc1897c 100644
 --- a/src/hnswbuild.c
 +++ b/src/hnswbuild.c
-@@ -860,9 +860,17 @@ HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc)
+@@ -843,9 +843,17 @@ HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc)
  
  	hnswarea = shm_toc_lookup(toc, PARALLEL_KEY_HNSW_AREA, false);
  
@@ -20,7 +36,7 @@ index dcfb2bd..d5189ee 100644
  	/* Close relations within worker */
  	index_close(indexRel, indexLockmode);
  	table_close(heapRel, heapLockmode);
-@@ -1117,12 +1125,38 @@ BuildIndex(Relation heap, Relation index, IndexInfo *indexInfo,
+@@ -1100,12 +1108,38 @@ BuildIndex(Relation heap, Relation index, IndexInfo *indexInfo,
  	SeedRandom(42);
  #endif
  

compute_tools/src/bin/fast_import.rs

@@ -31,6 +31,7 @@ use camino::{Utf8Path, Utf8PathBuf};
 use clap::Parser;
 use compute_tools::extension_server::{get_pg_version, PostgresMajorVersion};
 use nix::unistd::Pid;
+use tokio::io::BufReader;
 use tracing::{error, info, info_span, warn, Instrument};
 use utils::fs_ext::is_directory_empty;
 
@@ -52,12 +53,16 @@ struct Args {
     s3_prefix: Option<s3_uri::S3Uri>,
     #[clap(long)]
     source_connection_string: Option<String>,
+    #[clap(long)]
+    restore_connection_string: Option<String>, // will not run postgres if specified, will do pg_restore to this connection string
     #[clap(short, long)]
     interactive: bool,
     #[clap(long)]
     pg_bin_dir: Utf8PathBuf,
     #[clap(long)]
     pg_lib_dir: Utf8PathBuf,
+    #[clap(long)]
+    pg_port: Option<u16>, // port to run postgres on, 5432 is default
 }
 
 #[serde_with::serde_as]
@@ -66,6 +71,8 @@ struct Spec {
     encryption_secret: EncryptionSecret,
     #[serde_as(as = "serde_with::base64::Base64")]
     source_connstring_ciphertext_base64: Vec<u8>,
+    #[serde_as(as = "Option<serde_with::base64::Base64>")]
+    restore_connstring_ciphertext_base64: Option<Vec<u8>>,
 }
 
 #[derive(serde::Deserialize)]
@@ -74,6 +81,196 @@ enum EncryptionSecret {
     KMS { key_id: String },
 }
 
+// copied from pageserver_api::config::defaults::DEFAULT_LOCALE to avoid dependency just for a constant
+const DEFAULT_LOCALE: &str = if cfg!(target_os = "macos") {
+    "C"
+} else {
+    "C.UTF-8"
+};
+
+async fn decode_connstring(
+    kms_client: &aws_sdk_kms::Client,
+    key_id: &String,
+    connstring_ciphertext_base64: Vec<u8>,
+) -> Result<String, anyhow::Error> {
+    let mut output = kms_client
+        .decrypt()
+        .key_id(key_id)
+        .ciphertext_blob(aws_sdk_s3::primitives::Blob::new(
+            connstring_ciphertext_base64,
+        ))
+        .send()
+        .await
+        .context("decrypt connection string")?;
+
+    let plaintext = output
+        .plaintext
+        .take()
+        .context("get plaintext connection string")?;
+
+    String::from_utf8(plaintext.into_inner()).context("parse connection string as utf8")
+}
+
+struct PostgresProcess {
+    pgdata_dir: Utf8PathBuf,
+    pg_bin_dir: Utf8PathBuf,
+    pgbin: Utf8PathBuf,
+    pg_lib_dir: Utf8PathBuf,
+    postgres_proc: Option<tokio::process::Child>,
+}
+
+impl PostgresProcess {
+    fn new(pgdata_dir: Utf8PathBuf, pg_bin_dir: Utf8PathBuf, pg_lib_dir: Utf8PathBuf) -> Self {
+        Self {
+            pgdata_dir,
+            pgbin: pg_bin_dir.join("postgres"),
+            pg_bin_dir,
+            pg_lib_dir,
+            postgres_proc: None,
+        }
+    }
+
+    async fn prepare(&self, initdb_user: &str) -> Result<(), anyhow::Error> {
+        tokio::fs::create_dir(&self.pgdata_dir)
+            .await
+            .context("create pgdata directory")?;
+
+        let pg_version = match get_pg_version(self.pgbin.as_ref()) {
+            PostgresMajorVersion::V14 => 14,
+            PostgresMajorVersion::V15 => 15,
+            PostgresMajorVersion::V16 => 16,
+            PostgresMajorVersion::V17 => 17,
+        };
+        postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
+            superuser: initdb_user,
+            locale: DEFAULT_LOCALE, // XXX: this shouldn't be hard-coded,
+            pg_version,
+            initdb_bin: self.pg_bin_dir.join("initdb").as_ref(),
+            library_search_path: &self.pg_lib_dir, // TODO: is this right? Prob works in compute image, not sure about neon_local.
+            pgdata: &self.pgdata_dir,
+        })
+        .await
+        .context("initdb")
+    }
+
+    async fn start(
+        &mut self,
+        initdb_user: &str,
+        port: u16,
+        nproc: usize,
+    ) -> Result<&tokio::process::Child, anyhow::Error> {
+        self.prepare(initdb_user).await?;
+
+        //
+        // Launch postgres process
+        //
+        let mut proc = tokio::process::Command::new(&self.pgbin)
+            .arg("-D")
+            .arg(&self.pgdata_dir)
+            .args(["-p", &format!("{port}")])
+            .args(["-c", "wal_level=minimal"])
+            .args(["-c", "shared_buffers=10GB"])
+            .args(["-c", "max_wal_senders=0"])
+            .args(["-c", "fsync=off"])
+            .args(["-c", "full_page_writes=off"])
+            .args(["-c", "synchronous_commit=off"])
+            .args(["-c", "maintenance_work_mem=8388608"])
+            .args(["-c", &format!("max_parallel_maintenance_workers={nproc}")])
+            .args(["-c", &format!("max_parallel_workers={nproc}")])
+            .args(["-c", &format!("max_parallel_workers_per_gather={nproc}")])
+            .args(["-c", &format!("max_worker_processes={nproc}")])
+            .args(["-c", "effective_io_concurrency=100"])
+            .env_clear()
+            .env("LD_LIBRARY_PATH", &self.pg_lib_dir)
+            .stdout(std::process::Stdio::piped())
+            .stderr(std::process::Stdio::piped())
+            .spawn()
+            .context("spawn postgres")?;
+
+        info!("spawned postgres, waiting for it to become ready");
+        tokio::spawn(
+            child_stdio_to_log::relay_process_output(proc.stdout.take(), proc.stderr.take())
+                .instrument(info_span!("postgres")),
+        );
+
+        self.postgres_proc = Some(proc);
+        Ok(self.postgres_proc.as_ref().unwrap())
+    }
+
+    async fn shutdown(&mut self) -> Result<(), anyhow::Error> {
+        let proc: &mut tokio::process::Child = self.postgres_proc.as_mut().unwrap();
+        info!("shutdown postgres");
+        {
+            nix::sys::signal::kill(
+                Pid::from_raw(i32::try_from(proc.id().unwrap()).expect("convert child pid to i32")),
+                nix::sys::signal::SIGTERM,
+            )
+            .context("signal postgres to shut down")?;
+            proc.wait()
+                .await
+                .context("wait for postgres to shut down")?;
+        }
+        Ok(())
+    }
+}
+
+async fn wait_until_ready(connstring: String, create_dbname: String) {
+    // Create neondb database in the running postgres
+    let start_time = std::time::Instant::now();
+
+    loop {
+        if start_time.elapsed() > PG_WAIT_TIMEOUT {
+            error!(
+                "timeout exceeded: failed to poll postgres and create database within 10 minutes"
+            );
+            std::process::exit(1);
+        }
+
+        match tokio_postgres::connect(
+            &connstring.replace("dbname=neondb", "dbname=postgres"),
+            tokio_postgres::NoTls,
+        )
+        .await
+        {
+            Ok((client, connection)) => {
+                // Spawn the connection handling task to maintain the connection
+                tokio::spawn(async move {
+                    if let Err(e) = connection.await {
+                        warn!("connection error: {}", e);
+                    }
+                });
+
+                match client
+                    .simple_query(format!("CREATE DATABASE {create_dbname};").as_str())
+                    .await
+                {
+                    Ok(_) => {
+                        info!("created {} database", create_dbname);
+                        break;
+                    }
+                    Err(e) => {
+                        warn!(
+                            "failed to create database: {}, retying in {}s",
+                            e,
+                            PG_WAIT_RETRY_INTERVAL.as_secs_f32()
+                        );
+                        tokio::time::sleep(PG_WAIT_RETRY_INTERVAL).await;
+                        continue;
+                    }
+                }
+            }
+            Err(_) => {
+                info!(
+                    "postgres not ready yet, retrying in {}s",
+                    PG_WAIT_RETRY_INTERVAL.as_secs_f32()
+                );
+                tokio::time::sleep(PG_WAIT_RETRY_INTERVAL).await;
+                continue;
+            }
+        }
+    }
+}
+
 #[tokio::main]
 pub(crate) async fn main() -> anyhow::Result<()> {
     utils::logging::init(
@@ -107,8 +304,18 @@ pub(crate) async fn main() -> anyhow::Result<()> {
         (None, None)
     };
 
-    // Get source connection string either from S3 spec or direct argument
-    let source_connection_string = if let Some(s3_prefix) = &args.s3_prefix {
+    let superuser = "cloud_admin";
+    let pg_port = || {
+        args.pg_port.unwrap_or_else(|| {
+            info!("pg_port not specified, using default 5432");
+            5432
+        })
+    };
+
+    let mut run_postgres = true;
+
+    // Get connection strings either from S3 spec or direct arguments
+    let (source_connstring, restore_connstring) = if let Some(s3_prefix) = &args.s3_prefix {
         let spec: Spec = {
             let spec_key = s3_prefix.append("/spec.json");
             let s3_client = aws_sdk_s3::Client::new(aws_config.as_ref().unwrap());
@@ -128,28 +335,47 @@ pub(crate) async fn main() -> anyhow::Result<()> {
 
         match spec.encryption_secret {
             EncryptionSecret::KMS { key_id } => {
-                let mut output = kms_client
-                    .unwrap()
-                    .decrypt()
-                    .key_id(key_id)
-                    .ciphertext_blob(aws_sdk_s3::primitives::Blob::new(
-                        spec.source_connstring_ciphertext_base64,
-                    ))
-                    .send()
-                    .await
-                    .context("decrypt source connection string")?;
-                let plaintext = output
-                    .plaintext
-                    .take()
-                    .context("get plaintext source connection string")?;
-                String::from_utf8(plaintext.into_inner())
-                    .context("parse source connection string as utf8")?
+                let source = decode_connstring(
+                    kms_client.as_ref().unwrap(),
+                    &key_id,
+                    spec.source_connstring_ciphertext_base64,
+                )
+                .await?;
+
+                let restore =
+                    if let Some(restore_ciphertext) = spec.restore_connstring_ciphertext_base64 {
+                        run_postgres = false;
+                        decode_connstring(kms_client.as_ref().unwrap(), &key_id, restore_ciphertext)
+                            .await?
+                    } else {
+                        // restoring to local postgres otherwise
+                        format!(
+                            "host=localhost port={} user={} dbname=neondb",
+                            pg_port(),
+                            superuser
+                        )
+                    };
+
+                (source, restore)
             }
         }
     } else {
-        args.source_connection_string.unwrap()
+        (
+            args.source_connection_string.unwrap(),
+            if let Some(val) = args.restore_connection_string {
+                run_postgres = false;
+                val
+            } else {
+                format!(
+                    "host=localhost port={} user={} dbname=neondb",
+                    pg_port(),
+                    superuser
+                )
+            },
+        )
     };
 
+    // unused if run_postgres is false, but needed for shutdown
     match tokio::fs::create_dir(&working_directory).await {
         Ok(()) => {}
         Err(e) if e.kind() == std::io::ErrorKind::AlreadyExists => {
@@ -164,125 +390,21 @@ pub(crate) async fn main() -> anyhow::Result<()> {
         }
         Err(e) => return Err(anyhow::Error::new(e).context("create working directory")),
     }
-
     let pgdata_dir = working_directory.join("pgdata");
-    tokio::fs::create_dir(&pgdata_dir)
-        .await
-        .context("create pgdata directory")?;
 
-    let pgbin = pg_bin_dir.join("postgres");
-    let pg_version = match get_pg_version(pgbin.as_ref()) {
-        PostgresMajorVersion::V14 => 14,
-        PostgresMajorVersion::V15 => 15,
-        PostgresMajorVersion::V16 => 16,
-        PostgresMajorVersion::V17 => 17,
+    let postgres_proc = if run_postgres {
+        assert!(restore_connstring.contains("host=localhost"));
+        let mut proc =
+            PostgresProcess::new(pgdata_dir.clone(), pg_bin_dir.clone(), pg_lib_dir.clone());
+        let nproc = num_cpus::get();
+        proc.start(superuser, pg_port(), nproc).await?;
+        wait_until_ready(restore_connstring.clone(), "neondb".to_string()).await;
+
+        Some(proc)
+    } else {
+        info!("restore_connection_string specified, not running postgres process");
+        None
     };
-    let superuser = "cloud_admin"; // XXX: this shouldn't be hard-coded
-    postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
-        superuser,
-        locale: "en_US.UTF-8", // XXX: this shouldn't be hard-coded,
-        pg_version,
-        initdb_bin: pg_bin_dir.join("initdb").as_ref(),
-        library_search_path: &pg_lib_dir, // TODO: is this right? Prob works in compute image, not sure about neon_local.
-        pgdata: &pgdata_dir,
-    })
-    .await
-    .context("initdb")?;
-
-    let nproc = num_cpus::get();
-
-    //
-    // Launch postgres process
-    //
-    let mut postgres_proc = tokio::process::Command::new(pgbin)
-        .arg("-D")
-        .arg(&pgdata_dir)
-        .args(["-c", "wal_level=minimal"])
-        .args(["-c", "shared_buffers=10GB"])
-        .args(["-c", "max_wal_senders=0"])
-        .args(["-c", "fsync=off"])
-        .args(["-c", "full_page_writes=off"])
-        .args(["-c", "synchronous_commit=off"])
-        .args(["-c", "maintenance_work_mem=8388608"])
-        .args(["-c", &format!("max_parallel_maintenance_workers={nproc}")])
-        .args(["-c", &format!("max_parallel_workers={nproc}")])
-        .args(["-c", &format!("max_parallel_workers_per_gather={nproc}")])
-        .args(["-c", &format!("max_worker_processes={nproc}")])
-        .args([
-            "-c",
-            &format!(
-                "effective_io_concurrency={}",
-                if cfg!(target_os = "macos") { 0 } else { 100 }
-            ),
-        ])
-        .env_clear()
-        .stdout(std::process::Stdio::piped())
-        .stderr(std::process::Stdio::piped())
-        .spawn()
-        .context("spawn postgres")?;
-
-    info!("spawned postgres, waiting for it to become ready");
-    tokio::spawn(
-        child_stdio_to_log::relay_process_output(
-            postgres_proc.stdout.take(),
-            postgres_proc.stderr.take(),
-        )
-        .instrument(info_span!("postgres")),
-    );
-
-    // Create neondb database in the running postgres
-    let restore_pg_connstring =
-        format!("host=localhost port=5432 user={superuser} dbname=postgres");
-
-    let start_time = std::time::Instant::now();
-
-    loop {
-        if start_time.elapsed() > PG_WAIT_TIMEOUT {
-            error!(
-                "timeout exceeded: failed to poll postgres and create database within 10 minutes"
-            );
-            std::process::exit(1);
-        }
-
-        match tokio_postgres::connect(&restore_pg_connstring, tokio_postgres::NoTls).await {
-            Ok((client, connection)) => {
-                // Spawn the connection handling task to maintain the connection
-                tokio::spawn(async move {
-                    if let Err(e) = connection.await {
-                        warn!("connection error: {}", e);
-                    }
-                });
-
-                match client.simple_query("CREATE DATABASE neondb;").await {
-                    Ok(_) => {
-                        info!("created neondb database");
-                        break;
-                    }
-                    Err(e) => {
-                        warn!(
-                            "failed to create database: {}, retying in {}s",
-                            e,
-                            PG_WAIT_RETRY_INTERVAL.as_secs_f32()
-                        );
-                        tokio::time::sleep(PG_WAIT_RETRY_INTERVAL).await;
-                        continue;
-                    }
-                }
-            }
-            Err(_) => {
-                info!(
-                    "postgres not ready yet, retrying in {}s",
-                    PG_WAIT_RETRY_INTERVAL.as_secs_f32()
-                );
-                tokio::time::sleep(PG_WAIT_RETRY_INTERVAL).await;
-                continue;
-            }
-        }
-    }
-
-    let restore_pg_connstring = restore_pg_connstring.replace("dbname=postgres", "dbname=neondb");
-
-    let dumpdir = working_directory.join("dumpdir");
 
     let common_args = [
         // schema mapping (prob suffices to specify them on one side)
@@ -294,62 +416,66 @@ pub(crate) async fn main() -> anyhow::Result<()> {
         "--no-tablespaces".to_string(),
         // format
         "--format".to_string(),
-        "directory".to_string(),
+        "custom".to_string(),
         // concurrency
-        "--jobs".to_string(),
-        num_cpus::get().to_string(),
+        // "--jobs".to_string(),
+        // num_cpus::get().to_string(),
         // progress updates
         "--verbose".to_string(),
     ];
 
     info!("dump into the working directory");
-    {
-        let mut pg_dump = tokio::process::Command::new(pg_bin_dir.join("pg_dump"))
-            .args(&common_args)
-            .arg("-f")
-            .arg(&dumpdir)
-            .arg("--no-sync")
-            // POSITIONAL args
-            // source db (db name included in connection string)
-            .arg(&source_connection_string)
-            // how we run it
-            .env_clear()
-            .kill_on_drop(true)
-            .stdout(std::process::Stdio::piped())
-            .stderr(std::process::Stdio::piped())
-            .spawn()
-            .context("spawn pg_dump")?;
+    let mut pg_dump = tokio::process::Command::new(pg_bin_dir.join("pg_dump"))
+        .args(&common_args)
+        .arg("--no-sync")
+        // POSITIONAL args
+        // source db (db name included in connection string)
+        .arg(&source_connstring)
+        // how we run it
+        .env_clear()
+        .env("LD_LIBRARY_PATH", &pg_lib_dir)
+        .kill_on_drop(true)
+        .stdout(std::process::Stdio::piped())
+        .stderr(std::process::Stdio::piped())
+        .spawn()
+        .context("spawn pg_dump")?;
 
+    const BUF_SIZE: usize = 64 * 1024 * 1024;
+    let mut buf = BufReader::with_capacity(BUF_SIZE, pg_dump.stdout.take().unwrap());
+
+    tokio::spawn(async move {
         info!(pid=%pg_dump.id().unwrap(), "spawned pg_dump");
-
-        tokio::spawn(
-            child_stdio_to_log::relay_process_output(pg_dump.stdout.take(), pg_dump.stderr.take())
-                .instrument(info_span!("pg_dump")),
-        );
-
-        let st = pg_dump.wait().await.context("wait for pg_dump")?;
+        let st = pg_dump.wait().await.expect("wait for pg_dump");
         info!(status=?st, "pg_dump exited");
         if !st.success() {
             warn!(status=%st, "pg_dump failed, restore will likely fail as well");
         }
-    }
+
+        tokio::spawn(
+            child_stdio_to_log::relay_process_output(
+                None,
+                pg_dump.stderr.take(),
+            )
+            .instrument(info_span!("pg_dump")),
+        );
+    });
 
     // TODO: do it in a streaming way, plenty of internal research done on this already
     // TODO: do the unlogged table trick
-
-    info!("restore from working directory into vanilla postgres");
     {
         let mut pg_restore = tokio::process::Command::new(pg_bin_dir.join("pg_restore"))
             .args(&common_args)
             .arg("-d")
-            .arg(&restore_pg_connstring)
-            // POSITIONAL args
-            .arg(&dumpdir)
+            .arg(&restore_connstring)
+            .arg("--clean")
+            .arg("--if-exists")
             // how we run it
             .env_clear()
+            .env("LD_LIBRARY_PATH", &pg_lib_dir)
             .kill_on_drop(true)
             .stdout(std::process::Stdio::piped())
             .stderr(std::process::Stdio::piped())
+            .stdin(std::process::Stdio::piped())
             .spawn()
             .context("spawn pg_restore")?;
 
@@ -361,51 +487,48 @@ pub(crate) async fn main() -> anyhow::Result<()> {
             )
             .instrument(info_span!("pg_restore")),
         );
+
+        let mut restore_stdin = pg_restore.stdin.take().unwrap();
+
+        tokio::spawn(async move {
+            tokio::io::copy_buf(&mut buf, &mut restore_stdin).await.expect("pg_restore failed to read from pg_dump");
+        });
+
         let st = pg_restore.wait().await.context("wait for pg_restore")?;
         info!(status=?st, "pg_restore exited");
         if !st.success() {
             warn!(status=%st, "pg_restore failed, restore will likely fail as well");
         }
+
     }
 
-    // If interactive mode, wait for Ctrl+C
-    if args.interactive {
-        info!("Running in interactive mode. Press Ctrl+C to shut down.");
-        tokio::signal::ctrl_c().await.context("wait for ctrl-c")?;
-    }
+    if let Some(mut proc) = postgres_proc {
+        // If interactive mode, wait for Ctrl+C
+        if args.interactive {
+            info!("Running in interactive mode. Press Ctrl+C to shut down.");
+            tokio::signal::ctrl_c().await.context("wait for ctrl-c")?;
+        }
 
-    info!("shutdown postgres");
-    {
-        nix::sys::signal::kill(
-            Pid::from_raw(
-                i32::try_from(postgres_proc.id().unwrap()).expect("convert child pid to i32"),
-            ),
-            nix::sys::signal::SIGTERM,
-        )
-        .context("signal postgres to shut down")?;
-        postgres_proc
-            .wait()
-            .await
-            .context("wait for postgres to shut down")?;
-    }
+        proc.shutdown().await?;
 
-    // Only sync if s3_prefix was specified
-    if let Some(s3_prefix) = args.s3_prefix {
-        info!("upload pgdata");
-        aws_s3_sync::sync(Utf8Path::new(&pgdata_dir), &s3_prefix.append("/pgdata/"))
-            .await
-            .context("sync dump directory to destination")?;
-
-        info!("write status");
-        {
-            let status_dir = working_directory.join("status");
-            std::fs::create_dir(&status_dir).context("create status directory")?;
-            let status_file = status_dir.join("pgdata");
-            std::fs::write(&status_file, serde_json::json!({"done": true}).to_string())
-                .context("write status file")?;
-            aws_s3_sync::sync(&status_dir, &s3_prefix.append("/status/"))
+        // Only sync if s3_prefix was specified
+        if let Some(s3_prefix) = args.s3_prefix {
+            info!("upload pgdata");
+            aws_s3_sync::sync(Utf8Path::new(&pgdata_dir), &s3_prefix.append("/pgdata/"))
                 .await
-                .context("sync status directory to destination")?;
+                .context("sync dump directory to destination")?;
+
+            info!("write status");
+            {
+                let status_dir = working_directory.join("status");
+                std::fs::create_dir(&status_dir).context("create status directory")?;
+                let status_file = status_dir.join("pgdata");
+                std::fs::write(&status_file, serde_json::json!({"done": true}).to_string())
+                    .context("write status file")?;
+                aws_s3_sync::sync(&status_dir, &s3_prefix.append("/status/"))
+                    .await
+                    .context("sync status directory to destination")?;
+            }
         }
     }
 

control_plane/src/pageserver.rs

@@ -418,6 +418,26 @@ impl PageServerNode {
                 .map(serde_json::from_str)
                 .transpose()
                 .context("parse `wal_receiver_protocol_override` from json")?,
+            rel_size_v2_enabled: settings
+                .remove("rel_size_v2_enabled")
+                .map(|x| x.parse::<bool>())
+                .transpose()
+                .context("Failed to parse 'rel_size_v2_enabled' as bool")?,
+            gc_compaction_enabled: settings
+                .remove("gc_compaction_enabled")
+                .map(|x| x.parse::<bool>())
+                .transpose()
+                .context("Failed to parse 'gc_compaction_enabled' as bool")?,
+            gc_compaction_initial_threshold_kb: settings
+                .remove("gc_compaction_initial_threshold_kb")
+                .map(|x| x.parse::<u64>())
+                .transpose()
+                .context("Failed to parse 'gc_compaction_initial_threshold_kb' as integer")?,
+            gc_compaction_ratio_percent: settings
+                .remove("gc_compaction_ratio_percent")
+                .map(|x| x.parse::<u64>())
+                .transpose()
+                .context("Failed to parse 'gc_compaction_ratio_percent' as integer")?,
         };
         if !settings.is_empty() {
             bail!("Unrecognized tenant settings: {settings:?}")

docker-compose/compute_wrapper/Dockerfile

@@ -10,10 +10,7 @@ USER root
 RUN apt-get update &&       \
     apt-get install -y curl \
                        jq   \
-                       python3-pip \
                        netcat-openbsd
-#Faker is required for the pg_anon test
-RUN case $COMPUTE_IMAGE in compute-node-v17) OPT="--break-system-packages";; *) OPT= ;; esac && pip3 install $OPT Faker
 #This is required for the pg_hintplan test
 RUN mkdir -p /ext-src/pg_hint_plan-src && chown postgres /ext-src/pg_hint_plan-src 
 

docker-compose/docker_compose_test.sh

@@ -18,9 +18,6 @@ cd $(dirname $0)
 COMPUTE_CONTAINER_NAME=docker-compose-compute-1
 TEST_CONTAINER_NAME=docker-compose-neon-test-extensions-1
 PSQL_OPTION="-h localhost -U cloud_admin -p 55433 -d postgres"
-: ${http_proxy:=}
-: ${https_proxy:=}
-export http_proxy https_proxy
 
 cleanup() {
     echo "show container information"
@@ -35,12 +32,6 @@ for pg_version in ${TEST_VERSION_ONLY-14 15 16 17}; do
     echo "clean up containers if exists"
     cleanup
     PG_TEST_VERSION=$((pg_version < 16 ? 16 : pg_version))
-    # The support of pg_anon not yet added to PG17, so we have to add the corresponding option for other PG versions
-    if [ "${pg_version}" -ne 17 ]; then
-      SPEC_PATH="compute_wrapper/var/db/postgres/specs"
-      mv $SPEC_PATH/spec.json $SPEC_PATH/spec.bak
-      jq '.cluster.settings += [{"name": "session_preload_libraries","value": "anon","vartype": "string"}]' "${SPEC_PATH}/spec.bak" > "${SPEC_PATH}/spec.json"
-    fi
     PG_VERSION=$pg_version PG_TEST_VERSION=$PG_TEST_VERSION docker compose --profile test-extensions -f $COMPOSE_FILE up --build -d
 
     echo "wait until the compute is ready. timeout after 60s. "
@@ -62,27 +53,12 @@ for pg_version in ${TEST_VERSION_ONLY-14 15 16 17}; do
     done
 
     if [ $pg_version -ge 16 ]; then
-        echo Enabling trust connection
-        docker exec $COMPUTE_CONTAINER_NAME bash -c "sed -i '\$d' /var/db/postgres/compute/pg_hba.conf && echo -e 'host\t all\t all\t all\t trust' >> /var/db/postgres/compute/pg_hba.conf && psql $PSQL_OPTION -c 'select pg_reload_conf()' "
-        echo Adding postgres role
-        docker exec $COMPUTE_CONTAINER_NAME psql $PSQL_OPTION -c "CREATE ROLE postgres SUPERUSER LOGIN"
         # This is required for the pg_hint_plan test, to prevent flaky log message causing the test to fail
         # It cannot be moved to Dockerfile now because the database directory is created after the start of the container
         echo Adding dummy config
         docker exec $COMPUTE_CONTAINER_NAME touch /var/db/postgres/compute/compute_ctl_temp_override.conf
-        # This block is required for the pg_anon extension test.
-        # The test assumes that it is running on the same host with the postgres engine.
-        # In our case it's not true, that's why we are copying files to the compute node
+        # The following block copies the files for the pg_hintplan test to the compute node for the extension test in an isolated docker-compose environment
         TMPDIR=$(mktemp -d)
-        # Add support for pg_anon for pg_v16
-        if [ $pg_version -ne 17 ]; then
-          docker cp $TEST_CONTAINER_NAME:/ext-src/pg_anon-src/data $TMPDIR/data
-          echo -e '1\t too \t many \t tabs' > $TMPDIR/data/bad.csv
-          docker cp $TMPDIR/data $COMPUTE_CONTAINER_NAME:/tmp/tmp_anon_alternate_data
-        rm -rf $TMPDIR
-        fi
-        TMPDIR=$(mktemp -d)
-        # The following block does the same for the pg_hintplan test
         docker cp $TEST_CONTAINER_NAME:/ext-src/pg_hint_plan-src/data $TMPDIR/data
         docker cp $TMPDIR/data $COMPUTE_CONTAINER_NAME:/ext-src/pg_hint_plan-src/
         rm -rf $TMPDIR
@@ -106,8 +82,4 @@ for pg_version in ${TEST_VERSION_ONLY-14 15 16 17}; do
         fi
     fi
     cleanup
-    # Restore the original spec.json
-    if [ "$pg_version" -ne 17 ]; then
-      mv "$SPEC_PATH/spec.bak" "$SPEC_PATH/spec.json"
-    fi
 done

docs/rfcs/041-sharded-ingest.md

@@ -0,0 +1,255 @@
+# 
+Created on Aug 2024
+Implemented on Jan 2025
+
+## Summary
+
+Data in large tenants is split up between multiple pageservers according to key hashes, as
+introduced in the [sharding RFC](031-sharding-static.md) and [shard splitting RFC](032-shard-splitting.md).
+
+Whereas currently we send all WAL to all pageserver shards, and each shard filters out the data that it needs,
+in this RFC we add a mechanism to filter the WAL on the safekeeper, so that each shard receives
+only the data it needs.
+
+This will place some extra CPU load on the safekeepers, in exchange for reducing the network bandwidth
+for ingesting WAL back to scaling as O(1) with shard count, rather than O(N_shards).
+
+## Motivation
+
+1. Large databases require higher shard counts.  Whereas currently we run with up to 8 shards for tenants
+with a few TB of storage, the next order of magnitude capacity increase will require tens of shards, such
+that sending all WAL to all shards is impractical in terms of bandwidth.
+2. For contemporary database sizes (~2TB), the pageserver is the bottleneck for ingest: since each
+   shard has to decode and process the whole WAL, sharding doesn't fully relieve this bottleneck.  To achieve significantly higher ingest speeds, we need to filter the WAL earlier so that each pageserver
+   only has to process relevant parts.
+
+## Non Goals (if relevant)
+
+We do not seek to introduce multiple WALs per timeline, or to share the work of handling a timeline's
+WAL across safekeepers (beyond simple 3x replication).  This RFC may be thought of as an incremental
+move of the ingestion bottleneck up the stack: instead of high write rates bottlenecking on the
+pageserver, they will bottleneck on the safekeeper.
+
+## Impacted components (e.g. pageserver, safekeeper, console, etc)
+
+Safekeeper, pageserver.
+
+There will be no control plane or storage controller coordination needed, as pageservers will directly
+indicate their sharding parameters to the safekeeper when subscribing for WAL.
+
+## Proposed implementation
+
+Terminology:
+- "Data pages" refers to postgres relation blocks, and SLRU blocks.
+- "Metadata pages" refers to everything else the pageserver stores, such as relation sizes and
+  directories of relations.
+
+### Phase 1: Refactor ingest
+
+Currently, pageserver ingest code is structured approximately as follows:
+1. `handle_walreceiver_connection` reads a stream of binary WAL records off a network
+   socket
+2. `WalIngest::ingest_record` to translate the record into a series of page-level modifications
+3. `DatadirModification` accumulates page updates from several `ingest_record` calls, and when
+   its `commit()` method is called, flushes these into a Timeline's open `InMemoryLayer`.
+
+This process currently assumes access to a pageserver `Timeline` throughout `ingest_record` and
+from `DatadirModification`, which is used to do read-modify-write cycles on metadata pages
+such as relation sizes and the master DBDIR page.  It also assumes that records are ingested
+strictly one after the other: they cannot be ingested in parallel because each record assumes
+that earlier records' changes have already been applied to `Timeline`.
+
+This code will be refactored to disentangle the simple, fast decode of relation page writes
+from the more complex logic for updating internal metadata.  An intermediate representation
+called `InterpretedWalRecords` will be introduced.  This is similar to the internal state of
+a `DatadirModification`, but does not require access to a Timeline.  Instead of storing
+metadata updates as materialized writes to pages, it will accumulate these as abstract operations,
+for example rather than including a write to a relation size key, this structure will include
+an operation that indicates "Update relation _foo_'s size to the max of its current value and
+_bar_", such that these may be applied later to a real Timeline.
+
+The `DatadirModification` will be aware of the `EphemeralFile` format, so that as it accumulates
+simple page writes of relation blocks, it can write them directly into a buffer in the serialized
+format.  This will avoid the need to later deserialize/reserialize this data when passing the
+structure between safekeeper and pageserver.
+
+The new pipeline will be:
+1. `handle_walreceiver_connection` reads a stream of binary WAL records off a network
+2. A `InterpretedWalRecords` is generated from the incoming WAL records.  This does not
+   require a reference to a Timeline.
+3. The logic that is current spread between `WalIngest` and `DatadirModification` for updating
+   metadata will be refactored to consume the metadata operations from the `InterpretedWalRecords`
+   and turn them into literal writes to metadata pages.  This part must be done sequentially.
+4. The resulting buffer of metadata page writes is combined with the buffer of relation block
+   writes, and written into the `InMemoryLayer`.
+
+Implemented in:
+1. https://github.com/neondatabase/neon/pull/9472
+2. https://github.com/neondatabase/neon/pull/9504
+3. https://github.com/neondatabase/neon/pull/9524
+
+### Phase 2: Decode & filter on safekeeper
+
+In the previous phase, the ingest code was modified to be able to do most of its work without access to
+a Timeline: this first stage of ingest simply converts a series of binary wal records into
+a buffer of relation/SLRU page writes, and a buffer of abstract metadata writes.
+
+The modified ingest code may be transplanted from pageserver to safekeeper (probably via a
+shared crate).  The safekeeper->pageserver network protocol is modified to:
+ - in subscription requests, send the `ShardIdentity` from the pageserver to the safekeeper
+ - in responses, transmit a `InterpretedWalRecords` instead of a raw `WalRecord`.
+ - use the `ShardIdentity` to filter the `ProcessedWalIngest` to relevant content for
+   the subscribing shard before transmitting it.
+
+The overall behavior of the pageserver->safekeeper interaction remains the same, in terms of
+consistent LSN feedback, and connection management.  Only the payload of the subscriptions
+changes, to express an LSN range of WAL as a filtered `ProcessedWalIngest` instead of the
+raw data.
+
+The ingest code on the pageserver can now skip the part where it does the first phase of
+processing, as it will receive pre-processed, compressed data off the wire.
+
+Note that `InterpretedWalRecord` batches multiple `InterpretedWalRecord(s)` in the same network
+message. Safekeeper reads WAL in chunks of 16 blocks and then decodes as many Postgres WAL records
+as possible. Each Postgres WAL record maps to one `InterpretedWalRecord` for potentially multiple shards.
+Hence, the size of the batch is given by the number of Postgres WAL records that fit in 16 blocks.
+
+The protocol needs to support evolution. Protobuf was chosen here with the view that, in the future,
+we may migrate it to GRPC altogether
+
+Implemented in:
+1. https://github.com/neondatabase/neon/pull/9746
+2. https://github.com/neondatabase/neon/pull/9821
+
+### Phase 3: Fan out interpreted WAL
+
+In the previous phase, the initial processing of WAL was moved to the safekeeper, but it is still
+done once for each shard: this will generate O(N_shards) CPU work on the safekeeper (especially
+when considering converting to Protobuf format and compression).
+
+To avoid this, we fan-out WAL from one (tenant, timeline, shard) to all other shards subscribed on
+the same safekeeper. Under normal operation, the WAL will be read from disk, decoded and interpreted
+_only_ once per (safekeeper, timeline).
+
+When the first shard of a sharded timeline subscribes to a given safekeeper a task is spawned
+for the WAL reader (`InterpretedWalReader`). This task reads WAL, decodes, interprets it and sends
+it to the sender (`InterpretedWalSender`). The sender is a future that is polled from the connection
+task. When further shards subscribe on the safekeeper they will attach themselves to the existing WAL reader.
+There's two cases to consider:
+1. The shard's requested `start_lsn` is ahead of the current position of the WAL reader. In this case, the shard
+will start receiving data when the reader reaches that LSN. The intuition here is that there's little to gain
+by letting shards "front-run" since compute backpressure is based on the laggard LSN.
+2. The shard's requested `start_lsn` is below the current position of the WAL reader. In this case, the WAL reader
+gets reset to this requested position (same intuition). Special care is taken such that advanced shards do not receive
+interpreted WAL records below their current position.
+
+The approach above implies that there is at most one WAL reader per (tenant, timeline) on a given safekeeper at any point in time.
+If this turns out to be operationally problematic, there's a trick we can deploy: `--max-delta-for-fanout` is an optional safekeeper
+argument that controls the max absolute delta between a new shard and the current WAL position of the WAL reader. If the absolute
+delta is above that value, a new reader is spawned. Note that there's currently no concurrency control on the number of WAL readers,
+so it's recommended to use large values to avoid pushing CPU utilisation too high.
+
+Unsharded tenants do not spawn a separate task for the interpreted WAL reader since there's no benefit to it. Instead they poll
+the reader and sender concurrently from the connection task.
+
+Shard splits are interesting here because it is the only case when the same shard might have two subscriptions at the same time.
+This is handled by giving readers a unique identifier. Both shards will receive the same data while respecting their requested start
+position.
+
+Implemented in:
+1. https://github.com/neondatabase/neon/pull/10190
+
+## Deployment
+
+Each phase shall be deployed independently. Special care should be taken around protocol changes.
+
+## Observability Tips
+
+* The safekeeper logs the protocol requested by the pageserver
+along with the pageserver ID, tenant, timeline and shard: `starting streaming from`.
+* There's metrics for the number of wal readers:
+  * `safekeeper_wal_readers{kind="task", target=~"pageserver.*"}` gives the number of wal reader tasks for each SK
+  * `safekeeper_wal_readers{kind="future", target=~"pageserver.*"}` gives the numer of wal readers polled inline by each SK
+  * `safekeeper_interpreted_wal_reader_tasks` gives the number of wal reader tasks per tenant, timeline
+* Interesting log lines for the fan-out reader:
+  * `Spawning interpreted`: first shard creates the interpreted wal reader
+  * `Fanning out`: a subsequent shard attaches itself to an interpreted wal reader
+  * `Aborting interpreted`: all senders have finished and the reader task is being aborted
+
+## Future Optimizations
+
+This sections describes some improvement areas which may be revisited in the future.
+
+### Buffering of Interpreted WAL
+
+The interpreted WAL reader may buffer interpreted WAL records in user space to help with serving
+subscriptions that are lagging behind the current position of the reader.
+
+Counterpoints:
+* Safekeepers serve many thousands of timelines and allocating a buffer for each might be wasteful,
+especially given that it would go unused on the happy path.
+* WAL is buffered in the kernel page cache. Usually we'd only pay the CPU cost of decoding and interpreting.
+
+### Tweaking the Pagserver Safekeeper Selection Algorithm
+
+We could make the pageserver aware of which safekeeper's already host shards for the timeline along
+with their current WAL positions. The pageserver should then prefer safkeepers that are in the same
+AZ _and_ already have a shard with a position close to the desired start position.
+
+We currently run one safekeeper per AZ, so the point is mute until that changes.
+
+### Pipelining first ingest phase
+
+The first ingest phase is a stateless transformation of a binary WAL record into a pre-processed
+output per shard.  To put multiple CPUs to work, we may pipeline this processing up to some defined buffer
+depth.
+
+## Alternatives considered
+
+### Give safekeepers enough state to fully decode WAL
+
+In this RFC, we only do the first phase of ingest on the safekeeper, because this is
+the phase that is stateless.  Subsequent changes then happen on the pageserver, with
+access to the `Timeline` state.
+
+We could do more work on the safekeeper if we transmitted metadata state to the safekeeper
+when subscribing to the WAL: for example, by telling the safekeeper all the relation sizes,
+so that it could then generate all the metadata writes for relation sizes.
+
+We avoid doing this for several reasons:
+1. Complexity: it's a more invasive protocol change
+2. Decoupling: having the safekeeper understand the `ProcessedWalIngest` already somewhat
+   infects it with knowledge of the pageserver, but this is mainly an abstract structure
+   that describes postgres writes.  However, if we taught the safekeeper about the exact
+   way that pageserver deals with metadata keys, this would be a much tighter coupling.
+3. Load: once the WAL has been processed to the point that it can be split between shards,
+   it is preferable to share out work on the remaining shards rather than adding extra CPU
+   load to the safekeeper.
+
+### Do pre-processing on the compute instead of the safekeeper
+
+Since our first stage of ingest is stateless, it could be done at any stage in the pipeline,
+all the way up to the compute.
+
+We choose not to do this, because it is useful for the safekeeper to store the raw WAL rather
+than just the preprocessed WAL:
+- The safekeeper still needs to be able to serve raw WAL back to postgres for e.g. physical replication
+- It simplifies our paxos implementation to have the offset in the write log be literally
+  the same as the LSN
+- Raw WAL must have a stable protocol since we might have to re-ingest it at arbitrary points in the future.
+  Storing raw WAL give us more flexibility to evolve the pageserver, safekeeper protocol.
+
+### Do wal pre-processing on shard 0 or a separate service, send it to other shards from there
+
+If we wanted to keep the safekeepers as entirely pure stores of raw WAL bytes, then
+we could do the initial decode and shard-splitting in some other location:
+- Shard 0 could subscribe to the full WAL and then send writes to other shards
+- A new intermediate service between the safekeeper and pageserver could do the splitting.
+
+So why not?
+- Extra network hop from shard 0 to the final destination shard
+- Clearly there is more infrastructure involved here compared with doing it inline on the safekeeper.
+- Safekeepers already have very light CPU load: typical cloud instances shapes with appropriate
+  disks for the safekeepers effectively have "free" CPU resources.
+- Doing extra work on shard 0 would complicate scheduling of shards on pageservers, because
+  shard 0 would have significantly higher CPU load under write workloads than other shards.

libs/pageserver_api/src/config.rs

@@ -120,6 +120,7 @@ pub struct ConfigToml {
     pub no_sync: Option<bool>,
     pub wal_receiver_protocol: PostgresClientProtocol,
     pub page_service_pipelining: PageServicePipeliningConfig,
+    pub get_vectored_concurrent_io: GetVectoredConcurrentIo,
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -158,6 +159,25 @@ pub enum PageServiceProtocolPipelinedExecutionStrategy {
     Tasks,
 }
 
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(tag = "mode", rename_all = "kebab-case")]
+#[serde(deny_unknown_fields)]
+pub enum GetVectoredConcurrentIo {
+    /// The read path is fully sequential: layers are visited
+    /// one after the other and IOs are issued and waited upon
+    /// from the same task that traverses the layers.
+    Sequential,
+    /// The read path still traverses layers sequentially, and
+    /// index blocks will be read into the PS PageCache from
+    /// that task, with waiting.
+    /// But data IOs are dispatched and waited upon from a sidecar
+    /// task so that the traversing task can continue to traverse
+    /// layers while the IOs are in flight.
+    /// If the PS PageCache miss rate is low, this improves
+    /// throughput dramatically.
+    SidecarTask,
+}
+
 pub mod statvfs {
     pub mod mock {
         #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -301,6 +321,20 @@ pub struct TenantConfigToml {
     pub timeline_offloading: bool,
 
     pub wal_receiver_protocol_override: Option<PostgresClientProtocol>,
+
+    /// Enable rel_size_v2 for this tenant. Once enabled, the tenant will persist this information into
+    /// `index_part.json`, and it cannot be reversed.
+    pub rel_size_v2_enabled: Option<bool>,
+
+    // gc-compaction related configs
+    /// Enable automatic gc-compaction trigger on this tenant.
+    pub gc_compaction_enabled: bool,
+    /// The initial threshold for gc-compaction in KB. Once the total size of layers below the gc-horizon is above this threshold,
+    /// gc-compaction will be triggered.
+    pub gc_compaction_initial_threshold_kb: u64,
+    /// The ratio that triggers the auto gc-compaction. If (the total size of layers between L2 LSN and gc-horizon) / (size below the L2 LSN)
+    /// is above this ratio, gc-compaction will be triggered.
+    pub gc_compaction_ratio_percent: u64,
 }
 
 pub mod defaults {
@@ -450,6 +484,11 @@ impl Default for ConfigToml {
                     execution: PageServiceProtocolPipelinedExecutionStrategy::ConcurrentFutures,
                 })
             },
+            get_vectored_concurrent_io: if !cfg!(test) {
+                GetVectoredConcurrentIo::Sequential
+            } else {
+                GetVectoredConcurrentIo::SidecarTask
+            },
         }
     }
 }
@@ -494,6 +533,9 @@ pub mod tenant_conf_defaults {
     // By default ingest enough WAL for two new L0 layers before checking if new image
     // image layers should be created.
     pub const DEFAULT_IMAGE_LAYER_CREATION_CHECK_THRESHOLD: u8 = 2;
+    pub const DEFAULT_GC_COMPACTION_ENABLED: bool = false;
+    pub const DEFAULT_GC_COMPACTION_INITIAL_THRESHOLD_KB: u64 = 10240000;
+    pub const DEFAULT_GC_COMPACTION_RATIO_PERCENT: u64 = 100;
 }
 
 impl Default for TenantConfigToml {
@@ -538,6 +580,10 @@ impl Default for TenantConfigToml {
             lsn_lease_length_for_ts: LsnLease::DEFAULT_LENGTH_FOR_TS,
             timeline_offloading: false,
             wal_receiver_protocol_override: None,
+            rel_size_v2_enabled: None,
+            gc_compaction_enabled: DEFAULT_GC_COMPACTION_ENABLED,
+            gc_compaction_initial_threshold_kb: DEFAULT_GC_COMPACTION_INITIAL_THRESHOLD_KB,
+            gc_compaction_ratio_percent: DEFAULT_GC_COMPACTION_RATIO_PERCENT,
         }
     }
 }

libs/pageserver_api/src/models.rs

@@ -497,6 +497,14 @@ pub struct TenantConfigPatch {
     pub timeline_offloading: FieldPatch<bool>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub wal_receiver_protocol_override: FieldPatch<PostgresClientProtocol>,
+    #[serde(skip_serializing_if = "FieldPatch::is_noop")]
+    pub rel_size_v2_enabled: FieldPatch<bool>,
+    #[serde(skip_serializing_if = "FieldPatch::is_noop")]
+    pub gc_compaction_enabled: FieldPatch<bool>,
+    #[serde(skip_serializing_if = "FieldPatch::is_noop")]
+    pub gc_compaction_initial_threshold_kb: FieldPatch<u64>,
+    #[serde(skip_serializing_if = "FieldPatch::is_noop")]
+    pub gc_compaction_ratio_percent: FieldPatch<u64>,
 }
 
 /// An alternative representation of `pageserver::tenant::TenantConf` with
@@ -528,6 +536,10 @@ pub struct TenantConfig {
     pub lsn_lease_length_for_ts: Option<String>,
     pub timeline_offloading: Option<bool>,
     pub wal_receiver_protocol_override: Option<PostgresClientProtocol>,
+    pub rel_size_v2_enabled: Option<bool>,
+    pub gc_compaction_enabled: Option<bool>,
+    pub gc_compaction_initial_threshold_kb: Option<u64>,
+    pub gc_compaction_ratio_percent: Option<u64>,
 }
 
 impl TenantConfig {
@@ -557,6 +569,10 @@ impl TenantConfig {
             mut lsn_lease_length_for_ts,
             mut timeline_offloading,
             mut wal_receiver_protocol_override,
+            mut rel_size_v2_enabled,
+            mut gc_compaction_enabled,
+            mut gc_compaction_initial_threshold_kb,
+            mut gc_compaction_ratio_percent,
         } = self;
 
         patch.checkpoint_distance.apply(&mut checkpoint_distance);
@@ -601,6 +617,16 @@ impl TenantConfig {
         patch
             .wal_receiver_protocol_override
             .apply(&mut wal_receiver_protocol_override);
+        patch.rel_size_v2_enabled.apply(&mut rel_size_v2_enabled);
+        patch
+            .gc_compaction_enabled
+            .apply(&mut gc_compaction_enabled);
+        patch
+            .gc_compaction_initial_threshold_kb
+            .apply(&mut gc_compaction_initial_threshold_kb);
+        patch
+            .gc_compaction_ratio_percent
+            .apply(&mut gc_compaction_ratio_percent);
 
         Self {
             checkpoint_distance,
@@ -627,6 +653,10 @@ impl TenantConfig {
             lsn_lease_length_for_ts,
             timeline_offloading,
             wal_receiver_protocol_override,
+            rel_size_v2_enabled,
+            gc_compaction_enabled,
+            gc_compaction_initial_threshold_kb,
+            gc_compaction_ratio_percent,
         }
     }
 }

libs/safekeeper_api/src/models.rs

@@ -277,3 +277,8 @@ pub struct TimelineTermBumpResponse {
     pub previous_term: u64,
     pub current_term: u64,
 }
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct SafekeeperUtilization {
+    pub timeline_count: u64,
+}

libs/utils/src/env.rs

@@ -2,6 +2,7 @@
 
 use std::{fmt::Display, str::FromStr};
 
+/// For types `V` that implement [`FromStr`].
 pub fn var<V, E>(varname: &str) -> Option<V>
 where
     V: FromStr<Err = E>,
@@ -10,7 +11,9 @@ where
     match std::env::var(varname) {
         Ok(s) => Some(
             s.parse()
-                .map_err(|e| format!("failed to parse env var {varname}: {e:#}"))
+                .map_err(|e| {
+                    format!("failed to parse env var {varname} using FromStr::parse: {e:#}")
+                })
                 .unwrap(),
         ),
         Err(std::env::VarError::NotPresent) => None,
@@ -19,3 +22,24 @@ where
         }
     }
 }
+
+/// For types `V` that implement [`serde::de::DeserializeOwned`].
+pub fn var_serde_json_string<V>(varname: &str) -> Option<V>
+where
+    V: serde::de::DeserializeOwned,
+{
+    match std::env::var(varname) {
+        Ok(s) => Some({
+            let value = serde_json::Value::String(s);
+            serde_json::from_value(value)
+                .map_err(|e| {
+                    format!("failed to parse env var {varname} as a serde_json json string: {e:#}")
+                })
+                .unwrap()
+        }),
+        Err(std::env::VarError::NotPresent) => None,
+        Err(std::env::VarError::NotUnicode(_)) => {
+            panic!("env var {varname} is not unicode")
+        }
+    }
+}

libs/utils/src/failpoint_support.rs

@@ -11,31 +11,55 @@ use tracing::*;
 
 /// Declare a failpoint that can use to `pause` failpoint action.
 /// We don't want to block the executor thread, hence, spawn_blocking + await.
+///
+/// Optionally pass a cancellation token, and this failpoint will drop out of
+/// its pause when the cancellation token fires. This is useful for testing
+/// cases where we would like to block something, but test its clean shutdown behavior.
+/// The macro evaluates to a Result in that case, where Ok(()) is the case
+/// where the failpoint was not paused, and Err() is the case where cancellation
+/// token fired while evaluating the failpoint.
+///
+/// Remember to unpause the failpoint in the test; until that happens, one of the
+/// limited number of spawn_blocking thread pool threads is leaked.
 #[macro_export]
 macro_rules! pausable_failpoint {
-    ($name:literal) => {
+    ($name:literal) => {{
         if cfg!(feature = "testing") {
-            tokio::task::spawn_blocking({
-                let current = tracing::Span::current();
+            let cancel = ::tokio_util::sync::CancellationToken::new();
+            let _ = $crate::pausable_failpoint!($name, &cancel);
+        }
+    }};
+    ($name:literal, $cancel:expr) => {{
+        if cfg!(feature = "testing") {
+            let failpoint_fut = ::tokio::task::spawn_blocking({
+                let current = ::tracing::Span::current();
                 move || {
                     let _entered = current.entered();
-                    tracing::info!("at failpoint {}", $name);
-                    fail::fail_point!($name);
+                    ::tracing::info!("at failpoint {}", $name);
+                    ::fail::fail_point!($name);
+                }
+            });
+            let cancel_fut = async move {
+                $cancel.cancelled().await;
+            };
+            ::tokio::select! {
+                res = failpoint_fut => {
+                    res.expect("spawn_blocking");
+                    // continue with execution
+                    Ok(())
+                },
+                _ = cancel_fut => {
+                    Err(())
                 }
-            })
-            .await
-            .expect("spawn_blocking");
-        }
-    };
-    ($name:literal, $cond:expr) => {
-        if cfg!(feature = "testing") {
-            if $cond {
-                pausable_failpoint!($name)
             }
+        } else {
+            Ok(())
         }
-    };
+    }};
 }
 
+pub use pausable_failpoint;
+
 /// use with fail::cfg("$name", "return(2000)")
 ///
 /// The effect is similar to a "sleep(2000)" action, i.e. we sleep for the

libs/utils/src/sync/gate.rs

@@ -64,6 +64,12 @@ pub struct GateGuard {
     gate: Arc<GateInner>,
 }
 
+impl GateGuard {
+    pub fn try_clone(&self) -> Result<Self, GateError> {
+        Gate::enter_impl(self.gate.clone())
+    }
+}
+
 impl Drop for GateGuard {
     fn drop(&mut self) {
         if self.gate.closing.load(Ordering::Relaxed) {
@@ -107,11 +113,11 @@ impl Gate {
     /// to avoid blocking close() indefinitely: typically types that contain a Gate will
     /// also contain a CancellationToken.
     pub fn enter(&self) -> Result<GateGuard, GateError> {
-        let permit = self
-            .inner
-            .sem
-            .try_acquire()
-            .map_err(|_| GateError::GateClosed)?;
+        Self::enter_impl(self.inner.clone())
+    }
+
+    fn enter_impl(gate: Arc<GateInner>) -> Result<GateGuard, GateError> {
+        let permit = gate.sem.try_acquire().map_err(|_| GateError::GateClosed)?;
 
         // we now have the permit, let's disable the normal raii functionality and leave
         // "returning" the permit to our GateGuard::drop.
@@ -122,7 +128,7 @@ impl Gate {
 
         Ok(GateGuard {
             span_at_enter: tracing::Span::current(),
-            gate: self.inner.clone(),
+            gate,
         })
     }
 
@@ -252,4 +258,39 @@ mod tests {
         // Attempting to enter() is still forbidden
         gate.enter().expect_err("enter should fail finishing close");
     }
+
+    #[tokio::test(start_paused = true)]
+    async fn clone_gate_guard() {
+        let gate = Gate::default();
+        let forever = Duration::from_secs(24 * 7 * 365);
+
+        let guard1 = gate.enter().expect("gate isn't closed");
+
+        let guard2 = guard1.try_clone().expect("gate isn't clsoed");
+
+        let mut close_fut = std::pin::pin!(gate.close());
+
+        tokio::time::timeout(forever, &mut close_fut)
+            .await
+            .unwrap_err();
+
+        // we polled close_fut once, that should prevent all later enters and clones
+        gate.enter().unwrap_err();
+        guard1.try_clone().unwrap_err();
+        guard2.try_clone().unwrap_err();
+
+        // guard2 keeps gate open even if guard1 is closed
+        drop(guard1);
+        tokio::time::timeout(forever, &mut close_fut)
+            .await
+            .unwrap_err();
+
+        drop(guard2);
+
+        // now that the last guard is dropped, closing should complete
+        close_fut.await;
+
+        // entering is still forbidden
+        gate.enter().expect_err("enter should stilll fail");
+    }
 }

libs/wal_decoder/Cargo.toml

@@ -17,7 +17,6 @@ postgres_ffi.workspace = true
 serde.workspace = true
 thiserror.workspace = true
 tokio = { workspace = true, features = ["io-util"] }
-tonic.workspace = true
 tracing.workspace = true
 utils.workspace = true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }

libs/wal_decoder/src/models.rs

@@ -45,7 +45,7 @@ pub mod proto {
     #![allow(clippy::derive_partial_eq_without_eq)]
     // The generated ValueMeta has a `len` method generate for its `len` field.
     #![allow(clippy::len_without_is_empty)]
-    tonic::include_proto!("interpreted_wal");
+    include!(concat!(env!("OUT_DIR"), concat!("/interpreted_wal.rs")));
 }
 
 #[derive(Copy, Clone, Serialize, Deserialize)]

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -13,7 +13,7 @@ use rand::prelude::*;
 use tokio::task::JoinSet;
 use tracing::info;
 
-use std::collections::HashSet;
+use std::collections::{HashSet, VecDeque};
 use std::future::Future;
 use std::num::NonZeroUsize;
 use std::pin::Pin;
@@ -63,6 +63,10 @@ pub(crate) struct Args {
     #[clap(long)]
     set_io_mode: Option<pageserver_api::models::virtual_file::IoMode>,
 
+    /// Queue depth generated in each client.
+    #[clap(long, default_value = "1")]
+    queue_depth: NonZeroUsize,
+
     targets: Option<Vec<TenantTimelineId>>,
 }
 
@@ -298,6 +302,7 @@ async fn main_impl(
             start_work_barrier.wait().await;
             let client_start = Instant::now();
             let mut ticks_processed = 0;
+            let mut inflight = VecDeque::new();
             while !cancel.is_cancelled() {
                 // Detect if a request took longer than the RPS rate
                 if let Some(period) = &rps_period {
@@ -311,31 +316,37 @@ async fn main_impl(
                     ticks_processed = periods_passed_until_now;
                 }
 
-                let start = Instant::now();
-                let req = {
-                    let mut rng = rand::thread_rng();
-                    let r = &ranges[weights.sample(&mut rng)];
-                    let key: i128 = rng.gen_range(r.start..r.end);
-                    let key = Key::from_i128(key);
-                    assert!(key.is_rel_block_key());
-                    let (rel_tag, block_no) = key
-                        .to_rel_block()
-                        .expect("we filter non-rel-block keys out above");
-                    PagestreamGetPageRequest {
-                        hdr: PagestreamRequest {
-                            reqid: 0,
-                            request_lsn: if rng.gen_bool(args.req_latest_probability) {
-                                Lsn::MAX
-                            } else {
-                                r.timeline_lsn
+                while inflight.len() < args.queue_depth.get() {
+                    let start = Instant::now();
+                    let req = {
+                        let mut rng = rand::thread_rng();
+                        let r = &ranges[weights.sample(&mut rng)];
+                        let key: i128 = rng.gen_range(r.start..r.end);
+                        let key = Key::from_i128(key);
+                        assert!(key.is_rel_block_key());
+                        let (rel_tag, block_no) = key
+                            .to_rel_block()
+                            .expect("we filter non-rel-block keys out above");
+                        PagestreamGetPageRequest {
+                            hdr: PagestreamRequest {
+                                reqid: 0,
+                                request_lsn: if rng.gen_bool(args.req_latest_probability) {
+                                    Lsn::MAX
+                                } else {
+                                    r.timeline_lsn
+                                },
+                                not_modified_since: r.timeline_lsn,
                             },
-                            not_modified_since: r.timeline_lsn,
-                        },
-                        rel: rel_tag,
-                        blkno: block_no,
-                    }
-                };
-                client.getpage(req).await.unwrap();
+                            rel: rel_tag,
+                            blkno: block_no,
+                        }
+                    };
+                    client.getpage_send(req).await.unwrap();
+                    inflight.push_back(start);
+                }
+
+                let start = inflight.pop_front().unwrap();
+                client.getpage_recv().await.unwrap();
                 let end = Instant::now();
                 live_stats.request_done();
                 ticks_processed += 1;

pageserver/src/basebackup.rs

@@ -25,6 +25,7 @@ use tokio_tar::{Builder, EntryType, Header};
 
 use crate::context::RequestContext;
 use crate::pgdatadir_mapping::Version;
+use crate::tenant::storage_layer::IoConcurrency;
 use crate::tenant::Timeline;
 use pageserver_api::reltag::{RelTag, SlruKind};
 
@@ -123,6 +124,13 @@ where
         full_backup,
         replica,
         ctx,
+        io_concurrency: IoConcurrency::spawn_from_conf(
+            timeline.conf,
+            timeline
+                .gate
+                .enter()
+                .map_err(|e| BasebackupError::Server(e.into()))?,
+        ),
     };
     basebackup
         .send_tarball()
@@ -144,6 +152,7 @@ where
     full_backup: bool,
     replica: bool,
     ctx: &'a RequestContext,
+    io_concurrency: IoConcurrency,
 }
 
 /// A sink that accepts SLRU blocks ordered by key and forwards
@@ -303,7 +312,7 @@ where
             for part in slru_partitions.parts {
                 let blocks = self
                     .timeline
-                    .get_vectored(part, self.lsn, self.ctx)
+                    .get_vectored(part, self.lsn, self.io_concurrency.clone(), self.ctx)
                     .await
                     .map_err(|e| BasebackupError::Server(e.into()))?;
 
@@ -358,7 +367,7 @@ where
         let start_time = Instant::now();
         let aux_files = self
             .timeline
-            .list_aux_files(self.lsn, self.ctx)
+            .list_aux_files(self.lsn, self.ctx, self.io_concurrency.clone())
             .await
             .map_err(|e| BasebackupError::Server(e.into()))?;
         let aux_scan_time = start_time.elapsed();
@@ -422,7 +431,7 @@ where
         }
         let repl_origins = self
             .timeline
-            .get_replorigins(self.lsn, self.ctx)
+            .get_replorigins(self.lsn, self.ctx, self.io_concurrency.clone())
             .await
             .map_err(|e| BasebackupError::Server(e.into()))?;
         let n_origins = repl_origins.len();
@@ -489,7 +498,13 @@ where
             for blknum in startblk..endblk {
                 let img = self
                     .timeline
-                    .get_rel_page_at_lsn(src, blknum, Version::Lsn(self.lsn), self.ctx)
+                    .get_rel_page_at_lsn(
+                        src,
+                        blknum,
+                        Version::Lsn(self.lsn),
+                        self.ctx,
+                        self.io_concurrency.clone(),
+                    )
                     .await
                     .map_err(|e| BasebackupError::Server(e.into()))?;
                 segment_data.extend_from_slice(&img[..]);

pageserver/src/bin/pageserver.rs

@@ -135,6 +135,7 @@ fn main() -> anyhow::Result<()> {
     info!(?conf.virtual_file_io_mode, "starting with virtual_file IO mode");
     info!(?conf.wal_receiver_protocol, "starting with WAL receiver protocol");
     info!(?conf.page_service_pipelining, "starting with page service pipelining config");
+    info!(?conf.get_vectored_concurrent_io, "starting with get_vectored IO concurrency config");
 
     // The tenants directory contains all the pageserver local disk state.
     // Create if not exists and make sure all the contents are durable before proceeding.

pageserver/src/config.rs

@@ -191,6 +191,8 @@ pub struct PageServerConf {
     pub wal_receiver_protocol: PostgresClientProtocol,
 
     pub page_service_pipelining: pageserver_api::config::PageServicePipeliningConfig,
+
+    pub get_vectored_concurrent_io: pageserver_api::config::GetVectoredConcurrentIo,
 }
 
 /// Token for authentication to safekeepers
@@ -352,6 +354,7 @@ impl PageServerConf {
             no_sync,
             wal_receiver_protocol,
             page_service_pipelining,
+            get_vectored_concurrent_io,
         } = config_toml;
 
         let mut conf = PageServerConf {
@@ -396,6 +399,7 @@ impl PageServerConf {
             import_pgdata_aws_endpoint_url,
             wal_receiver_protocol,
             page_service_pipelining,
+            get_vectored_concurrent_io,
 
             // ------------------------------------------------------------
             // fields that require additional validation or custom handling

pageserver/src/http/routes.rs

@@ -84,6 +84,7 @@ use crate::tenant::remote_timeline_client::list_remote_tenant_shards;
 use crate::tenant::remote_timeline_client::list_remote_timelines;
 use crate::tenant::secondary::SecondaryController;
 use crate::tenant::size::ModelInputs;
+use crate::tenant::storage_layer::IoConcurrency;
 use crate::tenant::storage_layer::LayerAccessStatsReset;
 use crate::tenant::storage_layer::LayerName;
 use crate::tenant::timeline::import_pgdata;
@@ -2938,8 +2939,15 @@ async fn list_aux_files(
         active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
             .await?;
 
+    let io_concurrency = IoConcurrency::spawn_from_conf(
+        state.conf,
+        timeline.gate.enter().map_err(|_| ApiError::Cancelled)?,
+    );
+
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
-    let files = timeline.list_aux_files(body.lsn, &ctx).await?;
+    let files = timeline
+        .list_aux_files(body.lsn, &ctx, io_concurrency)
+        .await?;
     json_response(StatusCode::OK, files)
 }
 

pageserver/src/metrics.rs

@@ -100,71 +100,30 @@ pub(crate) static VEC_READ_NUM_LAYERS_VISITED: Lazy<Histogram> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-// Metrics collected on operations on the storage repository.
-#[derive(
-    Clone, Copy, enum_map::Enum, strum_macros::EnumString, strum_macros::Display, IntoStaticStr,
-)]
-pub(crate) enum GetKind {
-    Singular,
-    Vectored,
-}
-
-pub(crate) struct ReconstructTimeMetrics {
-    singular: Histogram,
-    vectored: Histogram,
-}
-
-pub(crate) static RECONSTRUCT_TIME: Lazy<ReconstructTimeMetrics> = Lazy::new(|| {
-    let inner = register_histogram_vec!(
-        "pageserver_getpage_reconstruct_seconds",
-        "Time spent in reconstruct_value (reconstruct a page from deltas)",
-        &["get_kind"],
-        CRITICAL_OP_BUCKETS.into(),
+pub(crate) static CONCURRENT_INITDBS: Lazy<UIntGauge> = Lazy::new(|| {
+    register_uint_gauge!(
+        "pageserver_concurrent_initdb",
+        "Number of initdb processes running"
     )
-    .expect("failed to define a metric");
-
-    ReconstructTimeMetrics {
-        singular: inner.with_label_values(&[GetKind::Singular.into()]),
-        vectored: inner.with_label_values(&[GetKind::Vectored.into()]),
-    }
+    .expect("failed to define a metric")
 });
 
-impl ReconstructTimeMetrics {
-    pub(crate) fn for_get_kind(&self, get_kind: GetKind) -> &Histogram {
-        match get_kind {
-            GetKind::Singular => &self.singular,
-            GetKind::Vectored => &self.vectored,
-        }
-    }
-}
-
-pub(crate) struct ReconstructDataTimeMetrics {
-    singular: Histogram,
-    vectored: Histogram,
-}
-
-impl ReconstructDataTimeMetrics {
-    pub(crate) fn for_get_kind(&self, get_kind: GetKind) -> &Histogram {
-        match get_kind {
-            GetKind::Singular => &self.singular,
-            GetKind::Vectored => &self.vectored,
-        }
-    }
-}
-
-pub(crate) static GET_RECONSTRUCT_DATA_TIME: Lazy<ReconstructDataTimeMetrics> = Lazy::new(|| {
-    let inner = register_histogram_vec!(
-        "pageserver_getpage_get_reconstruct_data_seconds",
-        "Time spent in get_reconstruct_value_data",
-        &["get_kind"],
-        CRITICAL_OP_BUCKETS.into(),
+pub(crate) static INITDB_SEMAPHORE_ACQUISITION_TIME: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
+        "pageserver_initdb_semaphore_seconds_global",
+        "Time spent getting a permit from the global initdb semaphore",
+        STORAGE_OP_BUCKETS.into()
     )
-    .expect("failed to define a metric");
+    .expect("failed to define metric")
+});
 
-    ReconstructDataTimeMetrics {
-        singular: inner.with_label_values(&[GetKind::Singular.into()]),
-        vectored: inner.with_label_values(&[GetKind::Vectored.into()]),
-    }
+pub(crate) static INITDB_RUN_TIME: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
+        "pageserver_initdb_seconds_global",
+        "Time spent performing initdb",
+        STORAGE_OP_BUCKETS.into()
+    )
+    .expect("failed to define metric")
 });
 
 pub(crate) struct GetVectoredLatency {
@@ -2524,12 +2483,19 @@ impl StorageTimeMetricsTimer {
         }
     }
 
-    /// Record the time from creation to now.
-    pub fn stop_and_record(self) {
-        let duration = self.start.elapsed().as_secs_f64();
-        self.metrics.timeline_sum.inc_by(duration);
+    /// Returns the elapsed duration of the timer.
+    pub fn elapsed(&self) -> Duration {
+        self.start.elapsed()
+    }
+
+    /// Record the time from creation to now and return it.
+    pub fn stop_and_record(self) -> Duration {
+        let duration = self.elapsed();
+        let seconds = duration.as_secs_f64();
+        self.metrics.timeline_sum.inc_by(seconds);
         self.metrics.timeline_count.inc();
-        self.metrics.global_histogram.observe(duration);
+        self.metrics.global_histogram.observe(seconds);
+        duration
     }
 
     /// Turns this timer into a timer, which will always record -- usually this means recording
@@ -2549,6 +2515,14 @@ impl Drop for AlwaysRecordingStorageTimeMetricsTimer {
     }
 }
 
+impl AlwaysRecordingStorageTimeMetricsTimer {
+    /// Returns the elapsed duration of the timer.
+    #[allow(unused)]
+    pub fn elapsed(&self) -> Duration {
+        self.0.as_ref().expect("not dropped yet").elapsed()
+    }
+}
+
 /// Timing facilities for an globally histogrammed metric, which is supported by per tenant and
 /// timeline total sum and count.
 #[derive(Clone, Debug)]
@@ -3893,7 +3867,6 @@ pub fn preinitialize_metrics(conf: &'static PageServerConf) {
     });
 
     // Custom
-    Lazy::force(&RECONSTRUCT_TIME);
     Lazy::force(&BASEBACKUP_QUERY_TIME);
     Lazy::force(&COMPUTE_COMMANDS_COUNTERS);
     Lazy::force(&tokio_epoll_uring::THREAD_LOCAL_METRICS_STORAGE);

pageserver/src/page_service.rs

@@ -39,6 +39,7 @@ use tokio::io::{AsyncWriteExt, BufWriter};
 use tokio::task::JoinHandle;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
+use utils::sync::gate::{Gate, GateGuard};
 use utils::sync::spsc_fold;
 use utils::{
     auth::{Claims, Scope, SwappableJwtAuth},
@@ -61,6 +62,7 @@ use crate::task_mgr::{self, COMPUTE_REQUEST_RUNTIME};
 use crate::tenant::mgr::ShardSelector;
 use crate::tenant::mgr::TenantManager;
 use crate::tenant::mgr::{GetActiveTenantError, GetTenantError, ShardResolveResult};
+use crate::tenant::storage_layer::IoConcurrency;
 use crate::tenant::timeline::{self, WaitLsnError};
 use crate::tenant::GetTimelineError;
 use crate::tenant::PageReconstructError;
@@ -90,6 +92,7 @@ pub struct Listener {
 pub struct Connections {
     cancel: CancellationToken,
     tasks: tokio::task::JoinSet<ConnectionHandlerResult>,
+    gate: Gate,
 }
 
 pub fn spawn(
@@ -110,6 +113,7 @@ pub fn spawn(
     let task = COMPUTE_REQUEST_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
         "libpq listener",
         libpq_listener_main(
+            conf,
             tenant_manager,
             pg_auth,
             tcp_listener,
@@ -134,11 +138,16 @@ impl Listener {
 }
 impl Connections {
     pub(crate) async fn shutdown(self) {
-        let Self { cancel, mut tasks } = self;
+        let Self {
+            cancel,
+            mut tasks,
+            gate,
+        } = self;
         cancel.cancel();
         while let Some(res) = tasks.join_next().await {
             Self::handle_connection_completion(res);
         }
+        gate.close().await;
     }
 
     fn handle_connection_completion(res: Result<anyhow::Result<()>, tokio::task::JoinError>) {
@@ -158,7 +167,9 @@ impl Connections {
 /// Returns Ok(()) upon cancellation via `cancel`, returning the set of
 /// open connections.
 ///
+#[allow(clippy::too_many_arguments)]
 pub async fn libpq_listener_main(
+    conf: &'static PageServerConf,
     tenant_manager: Arc<TenantManager>,
     auth: Option<Arc<SwappableJwtAuth>>,
     listener: tokio::net::TcpListener,
@@ -168,9 +179,15 @@ pub async fn libpq_listener_main(
     listener_cancel: CancellationToken,
 ) -> Connections {
     let connections_cancel = CancellationToken::new();
+    let connections_gate = Gate::default();
     let mut connection_handler_tasks = tokio::task::JoinSet::default();
 
     loop {
+        let gate_guard = match connections_gate.enter() {
+            Ok(guard) => guard,
+            Err(_) => break,
+        };
+
         let accepted = tokio::select! {
             biased;
             _ = listener_cancel.cancelled() => break,
@@ -190,6 +207,7 @@ pub async fn libpq_listener_main(
                 let connection_ctx = listener_ctx
                     .detached_child(TaskKind::PageRequestHandler, DownloadBehavior::Download);
                 connection_handler_tasks.spawn(page_service_conn_main(
+                    conf,
                     tenant_manager.clone(),
                     local_auth,
                     socket,
@@ -197,6 +215,7 @@ pub async fn libpq_listener_main(
                     pipelining_config.clone(),
                     connection_ctx,
                     connections_cancel.child_token(),
+                    gate_guard,
                 ));
             }
             Err(err) => {
@@ -211,13 +230,16 @@ pub async fn libpq_listener_main(
     Connections {
         cancel: connections_cancel,
         tasks: connection_handler_tasks,
+        gate: connections_gate,
     }
 }
 
 type ConnectionHandlerResult = anyhow::Result<()>;
 
 #[instrument(skip_all, fields(peer_addr))]
+#[allow(clippy::too_many_arguments)]
 async fn page_service_conn_main(
+    conf: &'static PageServerConf,
     tenant_manager: Arc<TenantManager>,
     auth: Option<Arc<SwappableJwtAuth>>,
     socket: tokio::net::TcpStream,
@@ -225,6 +247,7 @@ async fn page_service_conn_main(
     pipelining_config: PageServicePipeliningConfig,
     connection_ctx: RequestContext,
     cancel: CancellationToken,
+    gate_guard: GateGuard,
 ) -> ConnectionHandlerResult {
     let _guard = LIVE_CONNECTIONS
         .with_label_values(&["page_service"])
@@ -274,11 +297,13 @@ async fn page_service_conn_main(
     // But it's in a shared crate, so, we store connection_ctx inside PageServerHandler
     // and create the per-query context in process_query ourselves.
     let mut conn_handler = PageServerHandler::new(
+        conf,
         tenant_manager,
         auth,
         pipelining_config,
         connection_ctx,
         cancel.clone(),
+        gate_guard,
     );
     let pgbackend = PostgresBackend::new_from_io(socket, peer_addr, auth_type, None)?;
 
@@ -310,6 +335,7 @@ async fn page_service_conn_main(
 }
 
 struct PageServerHandler {
+    conf: &'static PageServerConf,
     auth: Option<Arc<SwappableJwtAuth>>,
     claims: Option<Claims>,
 
@@ -325,6 +351,8 @@ struct PageServerHandler {
     timeline_handles: Option<TimelineHandles>,
 
     pipelining_config: PageServicePipeliningConfig,
+
+    gate_guard: GateGuard,
 }
 
 struct TimelineHandles {
@@ -634,19 +662,23 @@ impl BatchedFeMessage {
 
 impl PageServerHandler {
     pub fn new(
+        conf: &'static PageServerConf,
         tenant_manager: Arc<TenantManager>,
         auth: Option<Arc<SwappableJwtAuth>>,
         pipelining_config: PageServicePipeliningConfig,
         connection_ctx: RequestContext,
         cancel: CancellationToken,
+        gate_guard: GateGuard,
     ) -> Self {
         PageServerHandler {
+            conf,
             auth,
             claims: None,
             connection_ctx,
             timeline_handles: Some(TimelineHandles::new(tenant_manager)),
             cancel,
             pipelining_config,
+            gate_guard,
         }
     }
 
@@ -1015,6 +1047,7 @@ impl PageServerHandler {
         &mut self,
         pgb_writer: &mut PostgresBackend<IO>,
         batch: BatchedFeMessage,
+        io_concurrency: IoConcurrency,
         cancel: &CancellationToken,
         protocol_version: PagestreamProtocolVersion,
         ctx: &RequestContext,
@@ -1084,6 +1117,7 @@ impl PageServerHandler {
                                 &*shard.upgrade()?,
                                 effective_request_lsn,
                                 pages,
+                                io_concurrency,
                                 ctx,
                             )
                             .instrument(span.clone())
@@ -1288,6 +1322,17 @@ impl PageServerHandler {
             }
         }
 
+        let io_concurrency = IoConcurrency::spawn_from_conf(
+            self.conf,
+            match self.gate_guard.try_clone() {
+                Ok(guard) => guard,
+                Err(_) => {
+                    info!("shutdown request received in page handler");
+                    return Err(QueryError::Shutdown);
+                }
+            },
+        );
+
         let pgb_reader = pgb
             .split()
             .context("implementation error: split pgb into reader and writer")?;
@@ -1309,6 +1354,7 @@ impl PageServerHandler {
                     request_span,
                     pipelining_config,
                     protocol_version,
+                    io_concurrency,
                     &ctx,
                 )
                 .await
@@ -1322,6 +1368,7 @@ impl PageServerHandler {
                     timeline_handles,
                     request_span,
                     protocol_version,
+                    io_concurrency,
                     &ctx,
                 )
                 .await
@@ -1349,6 +1396,7 @@ impl PageServerHandler {
         mut timeline_handles: TimelineHandles,
         request_span: Span,
         protocol_version: PagestreamProtocolVersion,
+        io_concurrency: IoConcurrency,
         ctx: &RequestContext,
     ) -> (
         (PostgresBackendReader<IO>, TimelineHandles),
@@ -1383,7 +1431,14 @@ impl PageServerHandler {
             };
 
             let err = self
-                .pagesteam_handle_batched_message(pgb_writer, msg, &cancel, protocol_version, ctx)
+                .pagesteam_handle_batched_message(
+                    pgb_writer,
+                    msg,
+                    io_concurrency.clone(),
+                    &cancel,
+                    protocol_version,
+                    ctx,
+                )
                 .await;
             match err {
                 Ok(()) => {}
@@ -1407,6 +1462,7 @@ impl PageServerHandler {
         request_span: Span,
         pipelining_config: PageServicePipeliningConfigPipelined,
         protocol_version: PagestreamProtocolVersion,
+        io_concurrency: IoConcurrency,
         ctx: &RequestContext,
     ) -> (
         (PostgresBackendReader<IO>, TimelineHandles),
@@ -1550,6 +1606,7 @@ impl PageServerHandler {
                     self.pagesteam_handle_batched_message(
                         pgb_writer,
                         batch,
+                        io_concurrency.clone(),
                         &cancel,
                         protocol_version,
                         &ctx,
@@ -1806,6 +1863,7 @@ impl PageServerHandler {
         timeline: &Timeline,
         effective_lsn: Lsn,
         requests: smallvec::SmallVec<[BatchedGetPageRequest; 1]>,
+        io_concurrency: IoConcurrency,
         ctx: &RequestContext,
     ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer), BatchedPageStreamError>> {
         debug_assert_current_span_has_tenant_and_timeline_id();
@@ -1832,6 +1890,7 @@ impl PageServerHandler {
             .get_rel_page_at_lsn_batched(
                 requests.iter().map(|p| (&p.req.rel, &p.req.blkno)),
                 effective_lsn,
+                io_concurrency,
                 ctx,
             )
             .await;

pageserver/src/pgdatadir_mapping.rs

@@ -17,6 +17,7 @@ use crate::span::{
     debug_assert_current_span_has_tenant_and_timeline_id,
     debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id,
 };
+use crate::tenant::storage_layer::IoConcurrency;
 use crate::tenant::timeline::GetVectoredError;
 use anyhow::{ensure, Context};
 use bytes::{Buf, Bytes, BytesMut};
@@ -200,6 +201,7 @@ impl Timeline {
         blknum: BlockNumber,
         version: Version<'_>,
         ctx: &RequestContext,
+        io_concurrency: IoConcurrency,
     ) -> Result<Bytes, PageReconstructError> {
         match version {
             Version::Lsn(effective_lsn) => {
@@ -208,6 +210,7 @@ impl Timeline {
                     .get_rel_page_at_lsn_batched(
                         pages.iter().map(|(tag, blknum)| (tag, blknum)),
                         effective_lsn,
+                        io_concurrency.clone(),
                         ctx,
                     )
                     .await;
@@ -246,6 +249,7 @@ impl Timeline {
         &self,
         pages: impl ExactSizeIterator<Item = (&RelTag, &BlockNumber)>,
         effective_lsn: Lsn,
+        io_concurrency: IoConcurrency,
         ctx: &RequestContext,
     ) -> Vec<Result<Bytes, PageReconstructError>> {
         debug_assert_current_span_has_tenant_and_timeline_id();
@@ -309,7 +313,10 @@ impl Timeline {
             acc.to_keyspace()
         };
 
-        match self.get_vectored(keyspace, effective_lsn, ctx).await {
+        match self
+            .get_vectored(keyspace, effective_lsn, io_concurrency, ctx)
+            .await
+        {
             Ok(results) => {
                 for (key, res) in results {
                     let mut key_slots = keys_slots.remove(&key).unwrap().into_iter();
@@ -889,9 +896,15 @@ impl Timeline {
         &self,
         lsn: Lsn,
         ctx: &RequestContext,
+        io_concurrency: IoConcurrency,
     ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
         let kv = self
-            .scan(KeySpace::single(Key::metadata_aux_key_range()), lsn, ctx)
+            .scan(
+                KeySpace::single(Key::metadata_aux_key_range()),
+                lsn,
+                ctx,
+                io_concurrency,
+            )
             .await?;
         let mut result = HashMap::new();
         let mut sz = 0;
@@ -914,8 +927,9 @@ impl Timeline {
         &self,
         lsn: Lsn,
         ctx: &RequestContext,
+        io_concurrency: IoConcurrency,
     ) -> Result<(), PageReconstructError> {
-        self.list_aux_files_v2(lsn, ctx).await?;
+        self.list_aux_files_v2(lsn, ctx, io_concurrency).await?;
         Ok(())
     }
 
@@ -923,17 +937,24 @@ impl Timeline {
         &self,
         lsn: Lsn,
         ctx: &RequestContext,
+        io_concurrency: IoConcurrency,
     ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
-        self.list_aux_files_v2(lsn, ctx).await
+        self.list_aux_files_v2(lsn, ctx, io_concurrency).await
     }
 
     pub(crate) async fn get_replorigins(
         &self,
         lsn: Lsn,
         ctx: &RequestContext,
+        io_concurrency: IoConcurrency,
     ) -> Result<HashMap<RepOriginId, Lsn>, PageReconstructError> {
         let kv = self
-            .scan(KeySpace::single(repl_origin_key_range()), lsn, ctx)
+            .scan(
+                KeySpace::single(repl_origin_key_range()),
+                lsn,
+                ctx,
+                io_concurrency,
+            )
             .await?;
         let mut result = HashMap::new();
         for (k, v) in kv {
@@ -2432,7 +2453,11 @@ mod tests {
             ("foo/bar2".to_string(), Bytes::from_static(b"content2")),
         ]);
 
-        let readback = tline.list_aux_files(Lsn(0x1008), &ctx).await?;
+        let io_concurrency = IoConcurrency::spawn_for_test();
+
+        let readback = tline
+            .list_aux_files(Lsn(0x1008), &ctx, io_concurrency.clone())
+            .await?;
         assert_eq!(readback, expect_1008);
 
         // Second modification: update one key, remove the other
@@ -2444,11 +2469,15 @@ mod tests {
         let expect_2008 =
             HashMap::from([("foo/bar1".to_string(), Bytes::from_static(b"content3"))]);
 
-        let readback = tline.list_aux_files(Lsn(0x2008), &ctx).await?;
+        let readback = tline
+            .list_aux_files(Lsn(0x2008), &ctx, io_concurrency.clone())
+            .await?;
         assert_eq!(readback, expect_2008);
 
         // Reading back in time works
-        let readback = tline.list_aux_files(Lsn(0x1008), &ctx).await?;
+        let readback = tline
+            .list_aux_files(Lsn(0x1008), &ctx, io_concurrency.clone())
+            .await?;
         assert_eq!(readback, expect_1008);
 
         Ok(())

pageserver/src/tenant.rs

@@ -95,6 +95,9 @@ use crate::deletion_queue::DeletionQueueError;
 use crate::import_datadir;
 use crate::is_uninit_mark;
 use crate::l0_flush::L0FlushGlobalState;
+use crate::metrics::CONCURRENT_INITDBS;
+use crate::metrics::INITDB_RUN_TIME;
+use crate::metrics::INITDB_SEMAPHORE_ACQUISITION_TIME;
 use crate::metrics::TENANT;
 use crate::metrics::{
     remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
@@ -5347,8 +5350,17 @@ async fn run_initdb(
         initdb_bin_path, initdb_target_dir, initdb_lib_dir,
     );
 
-    let _permit = INIT_DB_SEMAPHORE.acquire().await;
+    let _permit = {
+        let _timer = INITDB_SEMAPHORE_ACQUISITION_TIME.start_timer();
+        INIT_DB_SEMAPHORE.acquire().await
+    };
 
+    CONCURRENT_INITDBS.inc();
+    scopeguard::defer! {
+        CONCURRENT_INITDBS.dec();
+    }
+
+    let _timer = INITDB_RUN_TIME.start_timer();
     let res = postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
         superuser: &conf.superuser,
         locale: &conf.locale,
@@ -5463,6 +5475,12 @@ pub(crate) mod harness {
                 lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
                 timeline_offloading: Some(tenant_conf.timeline_offloading),
                 wal_receiver_protocol_override: tenant_conf.wal_receiver_protocol_override,
+                rel_size_v2_enabled: tenant_conf.rel_size_v2_enabled,
+                gc_compaction_enabled: Some(tenant_conf.gc_compaction_enabled),
+                gc_compaction_initial_threshold_kb: Some(
+                    tenant_conf.gc_compaction_initial_threshold_kb,
+                ),
+                gc_compaction_ratio_percent: Some(tenant_conf.gc_compaction_ratio_percent),
             }
         }
     }
@@ -5696,7 +5714,7 @@ mod tests {
     use pageserver_api::value::Value;
     use pageserver_compaction::helpers::overlaps_with;
     use rand::{thread_rng, Rng};
-    use storage_layer::PersistentLayerKey;
+    use storage_layer::{IoConcurrency, PersistentLayerKey};
     use tests::storage_layer::ValuesReconstructState;
     use tests::timeline::{GetVectoredError, ShutdownMode};
     use timeline::{CompactOptions, DeltaLayerTestDesc};
@@ -6477,6 +6495,7 @@ mod tests {
     async fn test_get_vectored() -> anyhow::Result<()> {
         let harness = TenantHarness::create("test_get_vectored").await?;
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
         let tline = tenant
             .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
             .await?;
@@ -6541,7 +6560,7 @@ mod tests {
                 .get_vectored_impl(
                     read.clone(),
                     reads_lsn,
-                    &mut ValuesReconstructState::new(),
+                    &mut ValuesReconstructState::new(io_concurrency.clone()),
                     &ctx,
                 )
                 .await;
@@ -6588,6 +6607,7 @@ mod tests {
         let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
 
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
         let tline = tenant
             .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
             .await?;
@@ -6622,7 +6642,7 @@ mod tests {
             .get_vectored_impl(
                 aux_keyspace.clone(),
                 read_lsn,
-                &mut ValuesReconstructState::new(),
+                &mut ValuesReconstructState::new(io_concurrency.clone()),
                 &ctx,
             )
             .await;
@@ -6670,6 +6690,7 @@ mod tests {
         )
         .await?;
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
 
         let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
         let gap_at_key = current_key.add(100);
@@ -6770,7 +6791,7 @@ mod tests {
             .get_vectored_impl(
                 read.clone(),
                 current_lsn,
-                &mut ValuesReconstructState::new(),
+                &mut ValuesReconstructState::new(io_concurrency.clone()),
                 &ctx,
             )
             .await?;
@@ -6813,6 +6834,7 @@ mod tests {
     async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
         let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
 
         let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
         let end_key = start_key.add(1000);
@@ -6905,7 +6927,7 @@ mod tests {
                         ranges: vec![child_gap_at_key..child_gap_at_key.next()],
                     },
                     query_lsn,
-                    &mut ValuesReconstructState::new(),
+                    &mut ValuesReconstructState::new(io_concurrency.clone()),
                     &ctx,
                 )
                 .await;
@@ -7351,6 +7373,7 @@ mod tests {
     async fn test_metadata_scan() -> anyhow::Result<()> {
         let harness = TenantHarness::create("test_metadata_scan").await?;
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
         let tline = tenant
             .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
             .await?;
@@ -7404,7 +7427,7 @@ mod tests {
                 .get_vectored_impl(
                     keyspace.clone(),
                     lsn,
-                    &mut ValuesReconstructState::default(),
+                    &mut ValuesReconstructState::new(io_concurrency.clone()),
                     &ctx,
                 )
                 .await?
@@ -7519,6 +7542,7 @@ mod tests {
         let harness = TenantHarness::create("test_aux_file_e2e").await.unwrap();
 
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
 
         let mut lsn = Lsn(0x08);
 
@@ -7538,7 +7562,10 @@ mod tests {
         }
 
         // we can read everything from the storage
-        let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
+        let files = tline
+            .list_aux_files(lsn, &ctx, io_concurrency.clone())
+            .await
+            .unwrap();
         assert_eq!(
             files.get("pg_logical/mappings/test1"),
             Some(&bytes::Bytes::from_static(b"first"))
@@ -7554,7 +7581,10 @@ mod tests {
             modification.commit(&ctx).await.unwrap();
         }
 
-        let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
+        let files = tline
+            .list_aux_files(lsn, &ctx, io_concurrency.clone())
+            .await
+            .unwrap();
         assert_eq!(
             files.get("pg_logical/mappings/test2"),
             Some(&bytes::Bytes::from_static(b"second"))
@@ -7565,7 +7595,10 @@ mod tests {
             .await
             .unwrap();
 
-        let files = child.list_aux_files(lsn, &ctx).await.unwrap();
+        let files = child
+            .list_aux_files(lsn, &ctx, io_concurrency.clone())
+            .await
+            .unwrap();
         assert_eq!(files.get("pg_logical/mappings/test1"), None);
         assert_eq!(files.get("pg_logical/mappings/test2"), None);
     }
@@ -7574,6 +7607,7 @@ mod tests {
     async fn test_metadata_image_creation() -> anyhow::Result<()> {
         let harness = TenantHarness::create("test_metadata_image_creation").await?;
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
         let tline = tenant
             .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
             .await?;
@@ -7593,8 +7627,9 @@ mod tests {
             keyspace: &KeySpace,
             lsn: Lsn,
             ctx: &RequestContext,
+            io_concurrency: IoConcurrency,
         ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
-            let mut reconstruct_state = ValuesReconstructState::default();
+            let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
             let res = tline
                 .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
                 .await?;
@@ -7642,7 +7677,8 @@ mod tests {
 
             if iter % 5 == 0 {
                 let (_, before_delta_file_accessed) =
-                    scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
+                    scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
+                        .await?;
                 tline
                     .compact(
                         &cancel,
@@ -7656,7 +7692,8 @@ mod tests {
                     )
                     .await?;
                 let (_, after_delta_file_accessed) =
-                    scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
+                    scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
+                        .await?;
                 assert!(after_delta_file_accessed < before_delta_file_accessed, "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}");
                 // Given that we already produced an image layer, there should be no delta layer needed for the scan, but still setting a low threshold there for unforeseen circumstances.
                 assert!(
@@ -7745,6 +7782,7 @@ mod tests {
     async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
         let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
 
         let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
         let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
@@ -7883,7 +7921,7 @@ mod tests {
         );
 
         // test vectored scan on parent timeline
-        let mut reconstruct_state = ValuesReconstructState::new();
+        let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
         let res = tline
             .get_vectored_impl(
                 KeySpace::single(Key::metadata_key_range()),
@@ -7909,7 +7947,7 @@ mod tests {
         );
 
         // test vectored scan on child timeline
-        let mut reconstruct_state = ValuesReconstructState::new();
+        let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
         let res = child
             .get_vectored_impl(
                 KeySpace::single(Key::metadata_key_range()),
@@ -7947,7 +7985,9 @@ mod tests {
         lsn: Lsn,
         ctx: &RequestContext,
     ) -> Result<Option<Bytes>, GetVectoredError> {
-        let mut reconstruct_state = ValuesReconstructState::new();
+        let io_concurrency =
+            IoConcurrency::spawn_from_conf(tline.conf, tline.gate.enter().unwrap());
+        let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
         let mut res = tline
             .get_vectored_impl(
                 KeySpace::single(key..key.next()),
@@ -8048,6 +8088,7 @@ mod tests {
             .await
             .unwrap();
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
 
         let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
         let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
@@ -8107,7 +8148,7 @@ mod tests {
 
         // Image layers are created at last_record_lsn
         let images = tline
-            .inspect_image_layers(Lsn(0x40), &ctx)
+            .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
             .await
             .unwrap()
             .into_iter()
@@ -8122,6 +8163,7 @@ mod tests {
             .await
             .unwrap();
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
 
         let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
         let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
@@ -8172,7 +8214,7 @@ mod tests {
 
         // Image layers are created at last_record_lsn
         let images = tline
-            .inspect_image_layers(Lsn(0x30), &ctx)
+            .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
             .await
             .unwrap()
             .into_iter()
@@ -8185,6 +8227,7 @@ mod tests {
     async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
         let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
         let (tenant, ctx) = harness.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
 
         fn get_key(id: u32) -> Key {
             // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
@@ -8326,7 +8369,7 @@ mod tests {
 
         // Check if the image layer at the GC horizon contains exactly what we want
         let image_at_gc_horizon = tline
-            .inspect_image_layers(Lsn(0x30), &ctx)
+            .inspect_image_layers(Lsn(0x30), &ctx, io_concurrency.clone())
             .await
             .unwrap()
             .into_iter()
@@ -10039,7 +10082,12 @@ mod tests {
 
         let keyspace = KeySpace::single(get_key(0)..get_key(10));
         let results = tline
-            .get_vectored(keyspace, delta_layer_end_lsn, &ctx)
+            .get_vectored(
+                keyspace,
+                delta_layer_end_lsn,
+                IoConcurrency::sequential(),
+                &ctx,
+            )
             .await
             .expect("No vectored errors");
         for (key, res) in results {

pageserver/src/tenant/config.rs

@@ -357,6 +357,18 @@ pub struct TenantConfOpt {
 
     #[serde(skip_serializing_if = "Option::is_none")]
     pub wal_receiver_protocol_override: Option<PostgresClientProtocol>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub rel_size_v2_enabled: Option<bool>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub gc_compaction_enabled: Option<bool>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub gc_compaction_initial_threshold_kb: Option<u64>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub gc_compaction_ratio_percent: Option<u64>,
 }
 
 impl TenantConfOpt {
@@ -425,6 +437,16 @@ impl TenantConfOpt {
             wal_receiver_protocol_override: self
                 .wal_receiver_protocol_override
                 .or(global_conf.wal_receiver_protocol_override),
+            rel_size_v2_enabled: self.rel_size_v2_enabled.or(global_conf.rel_size_v2_enabled),
+            gc_compaction_enabled: self
+                .gc_compaction_enabled
+                .unwrap_or(global_conf.gc_compaction_enabled),
+            gc_compaction_initial_threshold_kb: self
+                .gc_compaction_initial_threshold_kb
+                .unwrap_or(global_conf.gc_compaction_initial_threshold_kb),
+            gc_compaction_ratio_percent: self
+                .gc_compaction_ratio_percent
+                .unwrap_or(global_conf.gc_compaction_ratio_percent),
         }
     }
 
@@ -454,6 +476,10 @@ impl TenantConfOpt {
             mut lsn_lease_length_for_ts,
             mut timeline_offloading,
             mut wal_receiver_protocol_override,
+            mut rel_size_v2_enabled,
+            mut gc_compaction_enabled,
+            mut gc_compaction_initial_threshold_kb,
+            mut gc_compaction_ratio_percent,
         } = self;
 
         patch.checkpoint_distance.apply(&mut checkpoint_distance);
@@ -522,6 +548,16 @@ impl TenantConfOpt {
         patch
             .wal_receiver_protocol_override
             .apply(&mut wal_receiver_protocol_override);
+        patch.rel_size_v2_enabled.apply(&mut rel_size_v2_enabled);
+        patch
+            .gc_compaction_enabled
+            .apply(&mut gc_compaction_enabled);
+        patch
+            .gc_compaction_initial_threshold_kb
+            .apply(&mut gc_compaction_initial_threshold_kb);
+        patch
+            .gc_compaction_ratio_percent
+            .apply(&mut gc_compaction_ratio_percent);
 
         Ok(Self {
             checkpoint_distance,
@@ -548,6 +584,10 @@ impl TenantConfOpt {
             lsn_lease_length_for_ts,
             timeline_offloading,
             wal_receiver_protocol_override,
+            rel_size_v2_enabled,
+            gc_compaction_enabled,
+            gc_compaction_initial_threshold_kb,
+            gc_compaction_ratio_percent,
         })
     }
 }
@@ -603,6 +643,10 @@ impl From<TenantConfOpt> for models::TenantConfig {
             lsn_lease_length_for_ts: value.lsn_lease_length_for_ts.map(humantime),
             timeline_offloading: value.timeline_offloading,
             wal_receiver_protocol_override: value.wal_receiver_protocol_override,
+            rel_size_v2_enabled: value.rel_size_v2_enabled,
+            gc_compaction_enabled: value.gc_compaction_enabled,
+            gc_compaction_initial_threshold_kb: value.gc_compaction_initial_threshold_kb,
+            gc_compaction_ratio_percent: value.gc_compaction_ratio_percent,
         }
     }
 }

pageserver/src/tenant/layer_map.rs

@@ -57,6 +57,7 @@ use std::collections::{HashMap, VecDeque};
 use std::iter::Peekable;
 use std::ops::Range;
 use std::sync::Arc;
+use tokio::sync::watch;
 use utils::lsn::Lsn;
 
 use historic_layer_coverage::BufferedHistoricLayerCoverage;
@@ -67,7 +68,6 @@ use super::storage_layer::{LayerVisibilityHint, PersistentLayerDesc};
 ///
 /// LayerMap tracks what layers exist on a timeline.
 ///
-#[derive(Default)]
 pub struct LayerMap {
     //
     // 'open_layer' holds the current InMemoryLayer that is accepting new
@@ -93,7 +93,25 @@ pub struct LayerMap {
 
     /// L0 layers have key range Key::MIN..Key::MAX, and locating them using R-Tree search is very inefficient.
     /// So L0 layers are held in l0_delta_layers vector, in addition to the R-tree.
+    ///
+    /// NB: make sure to notify `watch_l0_deltas` on changes.
     l0_delta_layers: Vec<Arc<PersistentLayerDesc>>,
+
+    /// Notifies about L0 delta layer changes, sending the current number of L0 layers.
+    watch_l0_deltas: watch::Sender<usize>,
+}
+
+impl Default for LayerMap {
+    fn default() -> Self {
+        Self {
+            open_layer: Default::default(),
+            next_open_layer_at: Default::default(),
+            frozen_layers: Default::default(),
+            historic: Default::default(),
+            l0_delta_layers: Default::default(),
+            watch_l0_deltas: watch::channel(0).0,
+        }
+    }
 }
 
 /// The primary update API for the layer map.
@@ -466,6 +484,8 @@ impl LayerMap {
 
         if Self::is_l0(&layer_desc.key_range, layer_desc.is_delta) {
             self.l0_delta_layers.push(layer_desc.clone().into());
+            self.watch_l0_deltas
+                .send_replace(self.l0_delta_layers.len());
         }
 
         self.historic.insert(
@@ -488,6 +508,8 @@ impl LayerMap {
             let mut l0_delta_layers = std::mem::take(&mut self.l0_delta_layers);
             l0_delta_layers.retain(|other| other.key() != layer_key);
             self.l0_delta_layers = l0_delta_layers;
+            self.watch_l0_deltas
+                .send_replace(self.l0_delta_layers.len());
             // this assertion is related to use of Arc::ptr_eq in Self::compare_arced_layers,
             // there's a chance that the comparison fails at runtime due to it comparing (pointer,
             // vtable) pairs.
@@ -850,6 +872,11 @@ impl LayerMap {
         &self.l0_delta_layers
     }
 
+    /// Subscribes to L0 delta layer changes, sending the current number of L0 delta layers.
+    pub fn watch_level0_deltas(&self) -> watch::Receiver<usize> {
+        self.watch_l0_deltas.subscribe()
+    }
+
     /// debugging function to print out the contents of the layer map
     #[allow(unused)]
     pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {

pageserver/src/tenant/remote_timeline_client.rs

@@ -803,6 +803,12 @@ impl RemoteTimelineClient {
 
         upload_queue.dirty.metadata.apply(update);
 
+        // Defense in depth: if we somehow generated invalid metadata, do not persist it.
+        upload_queue
+            .dirty
+            .validate()
+            .map_err(|e| anyhow::anyhow!(e))?;
+
         self.schedule_index_upload(upload_queue);
 
         Ok(())

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -77,8 +77,32 @@ pub struct IndexPart {
     ///
     /// None means no aux files have been written to the storage before the point
     /// when this flag is introduced.
+    ///
+    /// This flag is not used any more as all tenants have been transitioned to the new aux file policy.
     #[serde(skip_serializing_if = "Option::is_none", default)]
     pub(crate) last_aux_file_policy: Option<AuxFilePolicy>,
+
+    #[serde(skip_serializing_if = "Option::is_none", default)]
+    pub(crate) rel_size_migration: Option<RelSizeMigration>,
+
+    /// The LSN of gc-compaction horizon. Once gc-compaction is finished for all layer files below an LSN, this LSN will be updated.
+    #[serde(skip_serializing_if = "Option::is_none", default)]
+    pub(crate) l2_lsn: Option<Lsn>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub enum RelSizeMigration {
+    /// The tenant is using the old rel_size format.
+    /// Note that this enum is persisted as `Option<RelSizeMigration>` in the index part, so
+    /// `None` is the same as `Some(RelSizeMigration::Legacy)`.
+    Legacy,
+    /// The tenant is migrating to the new rel_size format. Both old and new rel_size format are
+    /// persisted in the index part. The read path will read both formats and merge them.
+    Migrating,
+    /// The tenant has migrated to the new rel_size format. Only the new rel_size format is persisted
+    /// in the index part, and the read path will not read the old format.
+    Migrated,
 }
 
 impl IndexPart {
@@ -97,10 +121,12 @@ impl IndexPart {
     /// - 8: added `archived_at`
     /// - 9: +gc_blocking
     /// - 10: +import_pgdata
-    const LATEST_VERSION: usize = 10;
+    /// - 11: +rel_size_migration
+    /// - 12: +l2_lsn
+    const LATEST_VERSION: usize = 12;
 
     // Versions we may see when reading from a bucket.
-    pub const KNOWN_VERSIONS: &'static [usize] = &[1, 2, 3, 4, 5, 6, 7, 8, 9, 10];
+    pub const KNOWN_VERSIONS: &'static [usize] = &[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12];
 
     pub const FILE_NAME: &'static str = "index_part.json";
 
@@ -116,6 +142,8 @@ impl IndexPart {
             gc_blocking: None,
             last_aux_file_policy: None,
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         }
     }
 
@@ -152,6 +180,21 @@ impl IndexPart {
         };
         is_same_remote_layer_path(name, metadata, name, index_metadata)
     }
+
+    /// Check for invariants in the index: this is useful when uploading an index to ensure that if
+    /// we encounter a bug, we do not persist buggy metadata.
+    pub(crate) fn validate(&self) -> Result<(), String> {
+        if self.import_pgdata.is_none()
+            && self.metadata.ancestor_timeline().is_none()
+            && self.layer_metadata.is_empty()
+        {
+            // Unless we're in the middle of a raw pgdata import, or this is a child timeline,the index must
+            // always have at least one layer.
+            return Err("Index has no ancestor and no layers".to_string());
+        }
+
+        Ok(())
+    }
 }
 
 /// Metadata gathered for each of the layer files.
@@ -401,6 +444,8 @@ mod tests {
             gc_blocking: None,
             last_aux_file_policy: None,
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -446,6 +491,8 @@ mod tests {
             gc_blocking: None,
             last_aux_file_policy: None,
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -492,6 +539,8 @@ mod tests {
             gc_blocking: None,
             last_aux_file_policy: None,
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -541,6 +590,8 @@ mod tests {
             gc_blocking: None,
             last_aux_file_policy: None,
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let empty_layers_parsed = IndexPart::from_json_bytes(empty_layers_json.as_bytes()).unwrap();
@@ -585,6 +636,8 @@ mod tests {
             gc_blocking: None,
             last_aux_file_policy: None,
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -632,6 +685,8 @@ mod tests {
             gc_blocking: None,
             last_aux_file_policy: None,
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -684,6 +739,8 @@ mod tests {
             gc_blocking: None,
             last_aux_file_policy: Some(AuxFilePolicy::V2),
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -741,6 +798,8 @@ mod tests {
             gc_blocking: None,
             last_aux_file_policy: Default::default(),
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -799,6 +858,8 @@ mod tests {
             gc_blocking: None,
             last_aux_file_policy: Default::default(),
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -862,6 +923,8 @@ mod tests {
             last_aux_file_policy: Default::default(),
             archived_at: None,
             import_pgdata: None,
+            rel_size_migration: None,
+            l2_lsn: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -937,7 +1000,168 @@ mod tests {
                 started_at: parse_naive_datetime("2024-11-13T09:23:42.123000000"),
                 finished_at: parse_naive_datetime("2024-11-13T09:42:23.123000000"),
                 idempotency_key: import_pgdata::index_part_format::IdempotencyKey::new("specified-by-client-218a5213-5044-4562-a28d-d024c5f057f5".to_string()),
-            })))
+            }))),
+            rel_size_migration: None,
+            l2_lsn: None,
+        };
+
+        let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
+        assert_eq!(part, expected);
+    }
+
+    #[test]
+    fn v11_rel_size_migration_is_parsed() {
+        let example = r#"{
+            "version": 11,
+            "layer_metadata":{
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9": { "file_size": 25600000 },
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
+            },
+            "disk_consistent_lsn":"0/16960E8",
+            "metadata": {
+                "disk_consistent_lsn": "0/16960E8",
+                "prev_record_lsn": "0/1696070",
+                "ancestor_timeline": "e45a7f37d3ee2ff17dc14bf4f4e3f52e",
+                "ancestor_lsn": "0/0",
+                "latest_gc_cutoff_lsn": "0/1696070",
+                "initdb_lsn": "0/1696070",
+                "pg_version": 14
+            },
+            "gc_blocking": {
+                "started_at": "2024-07-19T09:00:00.123",
+                "reasons": ["DetachAncestor"]
+            },
+            "import_pgdata": {
+                "V1": {
+                    "Done": {
+                        "idempotency_key": "specified-by-client-218a5213-5044-4562-a28d-d024c5f057f5",
+                        "started_at": "2024-11-13T09:23:42.123",
+                        "finished_at": "2024-11-13T09:42:23.123"
+                    }
+                }
+            },
+            "rel_size_migration": "legacy"
+        }"#;
+
+        let expected = IndexPart {
+            version: 11,
+            layer_metadata: HashMap::from([
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), LayerFileMetadata {
+                    file_size: 25600000,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                }),
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), LayerFileMetadata {
+                    file_size: 9007199254741001,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                })
+            ]),
+            disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
+            metadata: TimelineMetadata::new(
+                Lsn::from_str("0/16960E8").unwrap(),
+                Some(Lsn::from_str("0/1696070").unwrap()),
+                Some(TimelineId::from_str("e45a7f37d3ee2ff17dc14bf4f4e3f52e").unwrap()),
+                Lsn::INVALID,
+                Lsn::from_str("0/1696070").unwrap(),
+                Lsn::from_str("0/1696070").unwrap(),
+                14,
+            ).with_recalculated_checksum().unwrap(),
+            deleted_at: None,
+            lineage: Default::default(),
+            gc_blocking: Some(GcBlocking {
+                started_at: parse_naive_datetime("2024-07-19T09:00:00.123000000"),
+                reasons: enumset::EnumSet::from_iter([GcBlockingReason::DetachAncestor]),
+            }),
+            last_aux_file_policy: Default::default(),
+            archived_at: None,
+            import_pgdata: Some(import_pgdata::index_part_format::Root::V1(import_pgdata::index_part_format::V1::Done(import_pgdata::index_part_format::Done{
+                started_at: parse_naive_datetime("2024-11-13T09:23:42.123000000"),
+                finished_at: parse_naive_datetime("2024-11-13T09:42:23.123000000"),
+                idempotency_key: import_pgdata::index_part_format::IdempotencyKey::new("specified-by-client-218a5213-5044-4562-a28d-d024c5f057f5".to_string()),
+            }))),
+            rel_size_migration: Some(RelSizeMigration::Legacy),
+            l2_lsn: None,
+        };
+
+        let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
+        assert_eq!(part, expected);
+    }
+
+    #[test]
+    fn v12_l2_lsn_is_parsed() {
+        let example = r#"{
+            "version": 12,
+            "layer_metadata":{
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9": { "file_size": 25600000 },
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
+            },
+            "disk_consistent_lsn":"0/16960E8",
+            "metadata": {
+                "disk_consistent_lsn": "0/16960E8",
+                "prev_record_lsn": "0/1696070",
+                "ancestor_timeline": "e45a7f37d3ee2ff17dc14bf4f4e3f52e",
+                "ancestor_lsn": "0/0",
+                "latest_gc_cutoff_lsn": "0/1696070",
+                "initdb_lsn": "0/1696070",
+                "pg_version": 14
+            },
+            "gc_blocking": {
+                "started_at": "2024-07-19T09:00:00.123",
+                "reasons": ["DetachAncestor"]
+            },
+            "import_pgdata": {
+                "V1": {
+                    "Done": {
+                        "idempotency_key": "specified-by-client-218a5213-5044-4562-a28d-d024c5f057f5",
+                        "started_at": "2024-11-13T09:23:42.123",
+                        "finished_at": "2024-11-13T09:42:23.123"
+                    }
+                }
+            },
+            "rel_size_migration": "legacy",
+            "l2_lsn": "0/16960E8"
+        }"#;
+
+        let expected = IndexPart {
+            version: 12,
+            layer_metadata: HashMap::from([
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), LayerFileMetadata {
+                    file_size: 25600000,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                }),
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), LayerFileMetadata {
+                    file_size: 9007199254741001,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                })
+            ]),
+            disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
+            metadata: TimelineMetadata::new(
+                Lsn::from_str("0/16960E8").unwrap(),
+                Some(Lsn::from_str("0/1696070").unwrap()),
+                Some(TimelineId::from_str("e45a7f37d3ee2ff17dc14bf4f4e3f52e").unwrap()),
+                Lsn::INVALID,
+                Lsn::from_str("0/1696070").unwrap(),
+                Lsn::from_str("0/1696070").unwrap(),
+                14,
+            ).with_recalculated_checksum().unwrap(),
+            deleted_at: None,
+            lineage: Default::default(),
+            gc_blocking: Some(GcBlocking {
+                started_at: parse_naive_datetime("2024-07-19T09:00:00.123000000"),
+                reasons: enumset::EnumSet::from_iter([GcBlockingReason::DetachAncestor]),
+            }),
+            last_aux_file_policy: Default::default(),
+            archived_at: None,
+            import_pgdata: Some(import_pgdata::index_part_format::Root::V1(import_pgdata::index_part_format::V1::Done(import_pgdata::index_part_format::Done{
+                started_at: parse_naive_datetime("2024-11-13T09:23:42.123000000"),
+                finished_at: parse_naive_datetime("2024-11-13T09:42:23.123000000"),
+                idempotency_key: import_pgdata::index_part_format::IdempotencyKey::new("specified-by-client-218a5213-5044-4562-a28d-d024c5f057f5".to_string()),
+            }))),
+            rel_size_migration: Some(RelSizeMigration::Legacy),
+            l2_lsn: Some("0/16960E8".parse::<Lsn>().unwrap()),
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -40,6 +40,10 @@ pub(crate) async fn upload_index_part(
     });
     pausable_failpoint!("before-upload-index-pausable");
 
+    // Safety: refuse to persist invalid index metadata, to mitigate the impact of any bug that produces this
+    // (this should never happen)
+    index_part.validate().map_err(|e| anyhow::anyhow!(e))?;
+
     // FIXME: this error comes too late
     let serialized = index_part.to_json_bytes()?;
     let serialized = Bytes::from(serialized);

pageserver/src/tenant/storage_layer.rs

@@ -10,18 +10,26 @@ mod layer_desc;
 mod layer_name;
 pub mod merge_iterator;
 
+use crate::config::PageServerConf;
 use crate::context::{AccessStatsBehavior, RequestContext};
 use bytes::Bytes;
+use futures::stream::FuturesUnordered;
+use futures::StreamExt;
 use pageserver_api::key::Key;
 use pageserver_api::keyspace::{KeySpace, KeySpaceRandomAccum};
 use pageserver_api::record::NeonWalRecord;
 use pageserver_api::value::Value;
-use std::cmp::{Ordering, Reverse};
+use std::cmp::Ordering;
 use std::collections::hash_map::Entry;
 use std::collections::{BinaryHeap, HashMap};
+use std::future::Future;
 use std::ops::Range;
+use std::pin::Pin;
+use std::sync::atomic::AtomicUsize;
 use std::sync::Arc;
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
+use tracing::{trace, Instrument};
+use utils::sync::gate::GateGuard;
 
 use utils::lsn::Lsn;
 
@@ -78,30 +86,151 @@ pub(crate) enum ValueReconstructSituation {
     Continue,
 }
 
-/// Reconstruct data accumulated for a single key during a vectored get
-#[derive(Debug, Default, Clone)]
-pub(crate) struct VectoredValueReconstructState {
-    pub(crate) records: Vec<(Lsn, NeonWalRecord)>,
-    pub(crate) img: Option<(Lsn, Bytes)>,
-
-    situation: ValueReconstructSituation,
+/// On disk representation of a value loaded in a buffer
+#[derive(Debug)]
+pub(crate) enum OnDiskValue {
+    /// Unencoded [`Value::Image`]
+    RawImage(Bytes),
+    /// Encoded [`Value`]. Can deserialize into an image or a WAL record
+    WalRecordOrImage(Bytes),
 }
 
-impl VectoredValueReconstructState {
-    fn get_cached_lsn(&self) -> Option<Lsn> {
-        self.img.as_ref().map(|img| img.0)
+/// Reconstruct data accumulated for a single key during a vectored get
+#[derive(Debug, Default)]
+pub(crate) struct VectoredValueReconstructState {
+    pub(crate) on_disk_values: Vec<(Lsn, OnDiskValueIoWaiter)>,
+
+    pub(crate) situation: ValueReconstructSituation,
+}
+
+#[derive(Debug)]
+pub(crate) struct OnDiskValueIoWaiter {
+    rx: tokio::sync::oneshot::Receiver<OnDiskValueIoResult>,
+}
+
+#[derive(Debug)]
+#[must_use]
+pub(crate) enum OnDiskValueIo {
+    /// Traversal identified this IO as required to complete the vectored get.
+    Required {
+        num_active_ios: Arc<AtomicUsize>,
+        tx: tokio::sync::oneshot::Sender<OnDiskValueIoResult>,
+    },
+    /// Sparse keyspace reads always read all the values for a given key,
+    /// even though only the first value is needed.
+    ///
+    /// This variant represents the unnecessary IOs for those values at lower LSNs
+    /// that aren't needed, but are currently still being done.
+    ///
+    /// The execution of unnecessary IOs was a pre-existing behavior before concurrent IO.
+    /// We added this explicit representation here so that we can drop
+    /// unnecessary IO results immediately, instead of buffering them in
+    /// `oneshot` channels inside [`VectoredValueReconstructState`] until
+    /// [`VectoredValueReconstructState::collect_pending_ios`] gets called.
+    Unnecessary,
+}
+
+type OnDiskValueIoResult = Result<OnDiskValue, std::io::Error>;
+
+impl OnDiskValueIo {
+    pub(crate) fn complete(self, res: OnDiskValueIoResult) {
+        match self {
+            OnDiskValueIo::Required { num_active_ios, tx } => {
+                num_active_ios.fetch_sub(1, std::sync::atomic::Ordering::Release);
+                let _ = tx.send(res);
+            }
+            OnDiskValueIo::Unnecessary => {
+                // Nobody cared, see variant doc comment.
+            }
+        }
     }
 }
 
-impl From<VectoredValueReconstructState> for ValueReconstructState {
-    fn from(mut state: VectoredValueReconstructState) -> Self {
-        // walredo expects the records to be descending in terms of Lsn
-        state.records.sort_by_key(|(lsn, _)| Reverse(*lsn));
+#[derive(Debug, thiserror::Error)]
+pub(crate) enum WaitCompletionError {
+    #[error("OnDiskValueIo was dropped without completing, likely the sidecar task panicked")]
+    IoDropped,
+}
 
-        ValueReconstructState {
-            records: state.records,
-            img: state.img,
+impl OnDiskValueIoWaiter {
+    pub(crate) async fn wait_completion(self) -> Result<OnDiskValueIoResult, WaitCompletionError> {
+        // NB: for Unnecessary IOs, this method never gets called because we don't add them to `on_disk_values`.
+        self.rx.await.map_err(|_| WaitCompletionError::IoDropped)
+    }
+}
+
+impl VectoredValueReconstructState {
+    /// # Cancel-Safety
+    ///
+    /// Technically fine to stop polling this future, but, the IOs will still
+    /// be executed to completion by the sidecar task and hold on to / consume resources.
+    /// Better not do it to make reasonsing about the system easier.
+    pub(crate) async fn collect_pending_ios(
+        self,
+    ) -> Result<ValueReconstructState, PageReconstructError> {
+        use utils::bin_ser::BeSer;
+
+        let mut res = Ok(ValueReconstructState::default());
+
+        // We should try hard not to bail early, so that by the time we return from this
+        // function, all IO for this value is done. It's not required -- we could totally
+        // stop polling the IO futures in the sidecar task, they need to support that,
+        // but just stopping to poll doesn't reduce the IO load on the disk. It's easier
+        // to reason about the system if we just wait for all IO to complete, even if
+        // we're no longer interested in the result.
+        //
+        // Revisit this when IO futures are replaced with a more sophisticated IO system
+        // and an IO scheduler, where we know which IOs were submitted and which ones
+        // just queued. Cf the comment on IoConcurrency::spawn_io.
+        for (lsn, waiter) in self.on_disk_values {
+            let value_recv_res = waiter
+                .wait_completion()
+                // we rely on the caller to poll us to completion, so this is not a bail point
+                .await;
+            // Force not bailing early by wrapping the code into a closure.
+            #[allow(clippy::redundant_closure_call)]
+            let _: () = (|| {
+                match (&mut res, value_recv_res) {
+                    (Err(_), _) => {
+                        // We've already failed, no need to process more.
+                    }
+                    (Ok(_), Err(wait_err)) => {
+                        // This shouldn't happen - likely the sidecar task panicked.
+                        res = Err(PageReconstructError::Other(wait_err.into()));
+                    }
+                    (Ok(_), Ok(Err(err))) => {
+                        let err: std::io::Error = err;
+                        // TODO: returning IO error here will fail a compute query.
+                        // Probably not what we want, we're not doing `maybe_fatal_err`
+                        // in the IO futures.
+                        // But it's been like that for a long time, not changing it
+                        // as part of concurrent IO.
+                        // => https://github.com/neondatabase/neon/issues/10454
+                        res = Err(PageReconstructError::Other(err.into()));
+                    }
+                    (Ok(ok), Ok(Ok(OnDiskValue::RawImage(img)))) => {
+                        assert!(ok.img.is_none());
+                        ok.img = Some((lsn, img));
+                    }
+                    (Ok(ok), Ok(Ok(OnDiskValue::WalRecordOrImage(buf)))) => {
+                        match Value::des(&buf) {
+                            Ok(Value::WalRecord(rec)) => {
+                                ok.records.push((lsn, rec));
+                            }
+                            Ok(Value::Image(img)) => {
+                                assert!(ok.img.is_none());
+                                ok.img = Some((lsn, img));
+                            }
+                            Err(err) => {
+                                res = Err(PageReconstructError::Other(err.into()));
+                            }
+                        }
+                    }
+                }
+            })();
         }
+
+        res
     }
 }
 
@@ -109,7 +238,7 @@ impl From<VectoredValueReconstructState> for ValueReconstructState {
 pub(crate) struct ValuesReconstructState {
     /// The keys will be removed after `get_vectored` completes. The caller outside `Timeline`
     /// should not expect to get anything from this hashmap.
-    pub(crate) keys: HashMap<Key, Result<VectoredValueReconstructState, PageReconstructError>>,
+    pub(crate) keys: HashMap<Key, VectoredValueReconstructState>,
     /// The keys which are already retrieved
     keys_done: KeySpaceRandomAccum,
 
@@ -119,27 +248,365 @@ pub(crate) struct ValuesReconstructState {
     // Statistics that are still accessible as a caller of `get_vectored_impl`.
     layers_visited: u32,
     delta_layers_visited: u32,
+
+    pub(crate) io_concurrency: IoConcurrency,
+    num_active_ios: Arc<AtomicUsize>,
+}
+
+/// The level of IO concurrency to be used on the read path
+///
+/// The desired end state is that we always do parallel IO.
+/// This struct and the dispatching in the impl will be removed once
+/// we've built enough confidence.
+pub(crate) enum IoConcurrency {
+    Sequential,
+    SidecarTask {
+        task_id: usize,
+        ios_tx: tokio::sync::mpsc::UnboundedSender<IoFuture>,
+    },
+}
+
+type IoFuture = Pin<Box<dyn Send + Future<Output = ()>>>;
+
+pub(crate) enum SelectedIoConcurrency {
+    Sequential,
+    SidecarTask(GateGuard),
+}
+
+impl std::fmt::Debug for IoConcurrency {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            IoConcurrency::Sequential => write!(f, "Sequential"),
+            IoConcurrency::SidecarTask { .. } => write!(f, "SidecarTask"),
+        }
+    }
+}
+
+impl std::fmt::Debug for SelectedIoConcurrency {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            SelectedIoConcurrency::Sequential => write!(f, "Sequential"),
+            SelectedIoConcurrency::SidecarTask(_) => write!(f, "SidecarTask"),
+        }
+    }
+}
+
+impl IoConcurrency {
+    /// Force sequential IO. This is a temporary workaround until we have
+    /// moved plumbing-through-the-call-stack
+    /// of IoConcurrency into `RequestContextq.
+    ///
+    /// DO NOT USE for new code.
+    ///
+    /// Tracking issue: <https://github.com/neondatabase/neon/issues/10460>.
+    pub(crate) fn sequential() -> Self {
+        Self::spawn(SelectedIoConcurrency::Sequential)
+    }
+
+    pub(crate) fn spawn_from_conf(
+        conf: &'static PageServerConf,
+        gate_guard: GateGuard,
+    ) -> IoConcurrency {
+        use pageserver_api::config::GetVectoredConcurrentIo;
+        let selected = match conf.get_vectored_concurrent_io {
+            GetVectoredConcurrentIo::Sequential => SelectedIoConcurrency::Sequential,
+            GetVectoredConcurrentIo::SidecarTask => SelectedIoConcurrency::SidecarTask(gate_guard),
+        };
+        Self::spawn(selected)
+    }
+
+    pub(crate) fn spawn(io_concurrency: SelectedIoConcurrency) -> Self {
+        match io_concurrency {
+            SelectedIoConcurrency::Sequential => IoConcurrency::Sequential,
+            SelectedIoConcurrency::SidecarTask(gate_guard) => {
+                let (ios_tx, ios_rx) = tokio::sync::mpsc::unbounded_channel();
+                static TASK_ID: AtomicUsize = AtomicUsize::new(0);
+                let task_id = TASK_ID.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+                // TODO: enrich the span with more context (tenant,shard,timeline) + (basebackup|pagestream|...)
+                let span =
+                    tracing::info_span!(parent: None, "IoConcurrency_sidecar", task_id = task_id);
+                trace!(task_id, "spawning sidecar task");
+                tokio::spawn(async move {
+                    trace!("start");
+                    scopeguard::defer!{ trace!("end") };
+                    type IosRx = tokio::sync::mpsc::UnboundedReceiver<IoFuture>;
+                    enum State {
+                        Waiting {
+                            // invariant: is_empty(), but we recycle the allocation
+                            empty_futures: FuturesUnordered<IoFuture>,
+                            ios_rx: IosRx,
+                        },
+                        Executing {
+                            futures: FuturesUnordered<IoFuture>,
+                            ios_rx: IosRx,
+                        },
+                        ShuttingDown {
+                            futures: FuturesUnordered<IoFuture>,
+                        },
+                    }
+                    let mut state = State::Waiting {
+                        empty_futures: FuturesUnordered::new(),
+                        ios_rx,
+                    };
+                    loop {
+                        match state {
+                            State::Waiting {
+                                empty_futures,
+                                mut ios_rx,
+                            } => {
+                                assert!(empty_futures.is_empty());
+                                tokio::select! {
+                                    fut = ios_rx.recv() => {
+                                        if let Some(fut) = fut {
+                                            trace!("received new io future");
+                                            empty_futures.push(fut);
+                                            state = State::Executing { futures: empty_futures, ios_rx };
+                                        } else {
+                                            state = State::ShuttingDown { futures: empty_futures }
+                                        }
+                                    }
+                                }
+                            }
+                            State::Executing {
+                                mut futures,
+                                mut ios_rx,
+                            } => {
+                                tokio::select! {
+                                    res = futures.next() => {
+                                        trace!("io future completed");
+                                        assert!(res.is_some());
+                                        if futures.is_empty() {
+                                            state = State::Waiting { empty_futures: futures, ios_rx};
+                                        } else {
+                                            state = State::Executing { futures, ios_rx };
+                                        }
+                                    }
+                                    fut = ios_rx.recv() => {
+                                        if let Some(fut) = fut {
+                                            trace!("received new io future");
+                                            futures.push(fut);
+                                            state =  State::Executing { futures, ios_rx};
+                                        } else {
+                                            state = State::ShuttingDown { futures };
+                                        }
+                                    }
+                                }
+                            }
+                            State::ShuttingDown {
+                                mut futures,
+                            } => {
+                                trace!("shutting down");
+                                while let Some(()) = futures.next().await {
+                                    trace!("io future completed (shutdown)");
+                                    // drain
+                                }
+                                trace!("shutdown complete");
+                                break;
+                            }
+                        }
+                    }
+                    drop(gate_guard); // drop it right before we exit
+                }.instrument(span));
+                IoConcurrency::SidecarTask { task_id, ios_tx }
+            }
+        }
+    }
+
+    pub(crate) fn clone(&self) -> Self {
+        match self {
+            IoConcurrency::Sequential => IoConcurrency::Sequential,
+            IoConcurrency::SidecarTask { task_id, ios_tx } => IoConcurrency::SidecarTask {
+                task_id: *task_id,
+                ios_tx: ios_tx.clone(),
+            },
+        }
+    }
+
+    /// Submit an IO to be executed in the background. DEADLOCK RISK, read the full doc string.
+    ///
+    /// The IO is represented as an opaque future.
+    /// IO completion must be handled inside the future, e.g., through a oneshot channel.
+    ///
+    /// The API seems simple but there are multiple **pitfalls** involving
+    /// DEADLOCK RISK.
+    ///
+    /// First, there are no guarantees about the exexecution of the IO.
+    /// It may be `await`ed in-place before this function returns.
+    /// It may be polled partially by this task and handed off to another task to be finished.
+    /// It may be polled and then dropped before returning ready.
+    ///
+    /// This means that submitted IOs must not be interedependent.
+    /// Interdependence may be through shared limited resources, e.g.,
+    /// - VirtualFile file descriptor cache slot acquisition
+    /// - tokio-epoll-uring slot
+    ///
+    /// # Why current usage is safe from deadlocks
+    ///
+    /// Textbook condition for a deadlock is that _all_ of the following be given
+    /// - Mutual exclusion
+    /// - Hold and wait
+    /// - No preemption
+    /// - Circular wait
+    ///
+    /// The current usage is safe because:
+    /// - Mutual exclusion: IO futures definitely use mutexes, no way around that for now
+    /// - Hold and wait: IO futures currently hold two kinds of locks/resources while waiting
+    ///   for acquisition of other resources:
+    ///    - VirtualFile file descriptor cache slot tokio mutex
+    ///    - tokio-epoll-uring slot (uses tokio notify => wait queue, much like mutex)
+    /// - No preemption: there's no taking-away of acquired locks/resources => given
+    /// - Circular wait: this is the part of the condition that isn't met: all IO futures
+    ///   first acquire VirtualFile mutex, then tokio-epoll-uring slot.
+    ///   There is no IO future that acquires slot before VirtualFile.
+    ///   Hence there can be no circular waiting.
+    ///   Hence there cannot be a deadlock.
+    ///
+    /// This is a very fragile situation and must be revisited whenver any code called from
+    /// inside the IO futures is changed.
+    ///
+    /// We will move away from opaque IO futures towards well-defined IOs at some point in
+    /// the future when we have shipped this first version of concurrent IO to production
+    /// and are ready to retire the Sequential mode which runs the futures in place.
+    /// Right now, while brittle, the opaque IO approach allows us to ship the feature
+    /// with minimal changes to the code and minimal changes to existing behavior in Sequential mode.
+    ///
+    /// Also read the comment in `collect_pending_ios`.
+    pub(crate) async fn spawn_io<F>(&mut self, fut: F)
+    where
+        F: std::future::Future<Output = ()> + Send + 'static,
+    {
+        match self {
+            IoConcurrency::Sequential => fut.await,
+            IoConcurrency::SidecarTask { ios_tx, .. } => {
+                let fut = Box::pin(fut);
+                // NB: experiments showed that doing an opportunistic poll of `fut` here was bad for throughput
+                // while insignificant for latency.
+                // It would make sense to revisit the tokio-epoll-uring API in the future such that we can try
+                // a submission here, but never poll the future. That way, io_uring can make proccess while
+                // the future sits in the ios_tx queue.
+                match ios_tx.send(fut) {
+                    Ok(()) => {}
+                    Err(_) => {
+                        unreachable!("the io task must have exited, likely it panicked")
+                    }
+                }
+            }
+        }
+    }
+
+    #[cfg(test)]
+    pub(crate) fn spawn_for_test() -> impl std::ops::DerefMut<Target = Self> {
+        use std::ops::{Deref, DerefMut};
+        use tracing::info;
+        use utils::sync::gate::Gate;
+
+        // Spawn needs a Gate, give it one.
+        struct Wrapper {
+            inner: IoConcurrency,
+            #[allow(dead_code)]
+            gate: Box<Gate>,
+        }
+        impl Deref for Wrapper {
+            type Target = IoConcurrency;
+
+            fn deref(&self) -> &Self::Target {
+                &self.inner
+            }
+        }
+        impl DerefMut for Wrapper {
+            fn deref_mut(&mut self) -> &mut Self::Target {
+                &mut self.inner
+            }
+        }
+        let gate = Box::new(Gate::default());
+
+        // The default behavior when running Rust unit tests without any further
+        // flags is to use the new behavior.
+        // The CI uses the following environment variable to unit test both old
+        // and new behavior.
+        // NB: the Python regression & perf tests take the `else` branch
+        // below and have their own defaults management.
+        let selected = {
+            // The pageserver_api::config type is unsuitable because it's internally tagged.
+            #[derive(serde::Deserialize)]
+            #[serde(rename_all = "kebab-case")]
+            enum TestOverride {
+                Sequential,
+                SidecarTask,
+            }
+            use once_cell::sync::Lazy;
+            static TEST_OVERRIDE: Lazy<TestOverride> = Lazy::new(|| {
+                utils::env::var_serde_json_string(
+                    "NEON_PAGESERVER_UNIT_TEST_GET_VECTORED_CONCURRENT_IO",
+                )
+                .unwrap_or(TestOverride::SidecarTask)
+            });
+
+            match *TEST_OVERRIDE {
+                TestOverride::Sequential => SelectedIoConcurrency::Sequential,
+                TestOverride::SidecarTask => {
+                    SelectedIoConcurrency::SidecarTask(gate.enter().expect("just created it"))
+                }
+            }
+        };
+
+        info!(?selected, "get_vectored_concurrent_io test");
+
+        Wrapper {
+            inner: Self::spawn(selected),
+            gate,
+        }
+    }
+}
+
+/// Make noise in case the [`ValuesReconstructState`] gets dropped while
+/// there are still IOs in flight.
+/// Refer to `collect_pending_ios` for why we prefer not to do that.
+//
+/// We log from here instead of from the sidecar task because the [`ValuesReconstructState`]
+/// gets dropped in a tracing span with more context.
+/// We repeat the sidecar tasks's `task_id` so we can correlate what we emit here with
+/// the logs / panic handler logs from the sidecar task, which also logs the `task_id`.
+impl Drop for ValuesReconstructState {
+    fn drop(&mut self) {
+        let num_active_ios = self
+            .num_active_ios
+            .load(std::sync::atomic::Ordering::Acquire);
+        if num_active_ios == 0 {
+            return;
+        }
+        let sidecar_task_id = match &self.io_concurrency {
+            IoConcurrency::Sequential => None,
+            IoConcurrency::SidecarTask { task_id, .. } => Some(*task_id),
+        };
+        tracing::warn!(
+            num_active_ios,
+            ?sidecar_task_id,
+            backtrace=%std::backtrace::Backtrace::force_capture(),
+            "dropping ValuesReconstructState while some IOs have not been completed",
+        );
+    }
 }
 
 impl ValuesReconstructState {
-    pub(crate) fn new() -> Self {
+    pub(crate) fn new(io_concurrency: IoConcurrency) -> Self {
         Self {
             keys: HashMap::new(),
             keys_done: KeySpaceRandomAccum::new(),
             keys_with_image_coverage: None,
             layers_visited: 0,
             delta_layers_visited: 0,
+            io_concurrency,
+            num_active_ios: Arc::new(AtomicUsize::new(0)),
         }
     }
 
-    /// Associate a key with the error which it encountered and mark it as done
-    pub(crate) fn on_key_error(&mut self, key: Key, err: PageReconstructError) {
-        let previous = self.keys.insert(key, Err(err));
-        if let Some(Ok(state)) = previous {
-            if state.situation == ValueReconstructSituation::Continue {
-                self.keys_done.add_key(key);
-            }
-        }
+    /// Absolutely read [`IoConcurrency::spawn_io`] to learn about assumptions & pitfalls.
+    pub(crate) async fn spawn_io<F>(&mut self, fut: F)
+    where
+        F: std::future::Future<Output = ()> + Send + 'static,
+    {
+        self.io_concurrency.spawn_io(fut).await;
     }
 
     pub(crate) fn on_layer_visited(&mut self, layer: &ReadableLayer) {
@@ -159,29 +626,6 @@ impl ValuesReconstructState {
         self.layers_visited
     }
 
-    /// This function is called after reading a keyspace from a layer.
-    /// It checks if the read path has now moved past the cached Lsn for any keys.
-    ///
-    /// Implementation note: We intentionally iterate over the keys for which we've
-    /// already collected some reconstruct data. This avoids scaling complexity with
-    /// the size of the search space.
-    pub(crate) fn on_lsn_advanced(&mut self, keyspace: &KeySpace, advanced_to: Lsn) {
-        for (key, value) in self.keys.iter_mut() {
-            if !keyspace.contains(key) {
-                continue;
-            }
-
-            if let Ok(state) = value {
-                if state.situation != ValueReconstructSituation::Complete
-                    && state.get_cached_lsn() >= Some(advanced_to)
-                {
-                    state.situation = ValueReconstructSituation::Complete;
-                    self.keys_done.add_key(*key);
-                }
-            }
-        }
-    }
-
     /// On hitting image layer, we can mark all keys in this range as done, because
     /// if the image layer does not contain a key, it is deleted/never added.
     pub(crate) fn on_image_layer_visited(&mut self, key_range: &Range<Key>) {
@@ -199,70 +643,42 @@ impl ValuesReconstructState {
     ///
     /// If the key is in the sparse keyspace (i.e., aux files), we do not track them in
     /// `key_done`.
-    pub(crate) fn update_key(
-        &mut self,
-        key: &Key,
-        lsn: Lsn,
-        value: Value,
-    ) -> ValueReconstructSituation {
-        let state = self
-            .keys
-            .entry(*key)
-            .or_insert(Ok(VectoredValueReconstructState::default()));
+    // TODO: rename this method & update description.
+    pub(crate) fn update_key(&mut self, key: &Key, lsn: Lsn, completes: bool) -> OnDiskValueIo {
+        let state = self.keys.entry(*key).or_default();
+
         let is_sparse_key = key.is_sparse();
-        if let Ok(state) = state {
-            let key_done = match state.situation {
-                ValueReconstructSituation::Complete => {
-                    if is_sparse_key {
-                        // Sparse keyspace might be visited multiple times because
-                        // we don't track unmapped keyspaces.
-                        return ValueReconstructSituation::Complete;
-                    } else {
-                        unreachable!()
-                    }
-                }
-                ValueReconstructSituation::Continue => match value {
-                    Value::Image(img) => {
-                        state.img = Some((lsn, img));
-                        true
-                    }
-                    Value::WalRecord(rec) => {
-                        debug_assert!(
-                            Some(lsn) > state.get_cached_lsn(),
-                            "Attempt to collect a record below cached LSN for walredo: {} < {}",
-                            lsn,
-                            state
-                                .get_cached_lsn()
-                                .expect("Assertion can only fire if a cached lsn is present")
-                        );
 
-                        let will_init = rec.will_init();
-                        state.records.push((lsn, rec));
-                        will_init
-                    }
-                },
-            };
-
-            if key_done && state.situation == ValueReconstructSituation::Continue {
-                state.situation = ValueReconstructSituation::Complete;
-                if !is_sparse_key {
-                    self.keys_done.add_key(*key);
+        let required_io = match state.situation {
+            ValueReconstructSituation::Complete => {
+                if is_sparse_key {
+                    // Sparse keyspace might be visited multiple times because
+                    // we don't track unmapped keyspaces.
+                    return OnDiskValueIo::Unnecessary;
+                } else {
+                    unreachable!()
                 }
             }
+            ValueReconstructSituation::Continue => {
+                self.num_active_ios
+                    .fetch_add(1, std::sync::atomic::Ordering::Release);
+                let (tx, rx) = tokio::sync::oneshot::channel();
+                state.on_disk_values.push((lsn, OnDiskValueIoWaiter { rx }));
+                OnDiskValueIo::Required {
+                    tx,
+                    num_active_ios: Arc::clone(&self.num_active_ios),
+                }
+            }
+        };
 
-            state.situation
-        } else {
-            ValueReconstructSituation::Complete
+        if completes && state.situation == ValueReconstructSituation::Continue {
+            state.situation = ValueReconstructSituation::Complete;
+            if !is_sparse_key {
+                self.keys_done.add_key(*key);
+            }
         }
-    }
 
-    /// Returns the Lsn at which this key is cached if one exists.
-    /// The read path should go no further than this Lsn for the given key.
-    pub(crate) fn get_cached_lsn(&self, key: &Key) -> Option<Lsn> {
-        self.keys
-            .get(key)
-            .and_then(|k| k.as_ref().ok())
-            .and_then(|state| state.get_cached_lsn())
+        required_io
     }
 
     /// Returns the key space describing the keys that have
@@ -276,12 +692,6 @@ impl ValuesReconstructState {
     }
 }
 
-impl Default for ValuesReconstructState {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
 /// A key that uniquely identifies a layer in a timeline
 #[derive(Debug, PartialEq, Eq, Clone, Hash)]
 pub(crate) enum LayerId {
@@ -720,3 +1130,78 @@ impl<T: std::fmt::Display> std::fmt::Debug for RangeDisplayDebug<'_, T> {
         write!(f, "{}..{}", self.0.start, self.0.end)
     }
 }
+
+#[cfg(test)]
+mod tests2 {
+    use pageserver_api::key::DBDIR_KEY;
+    use tracing::info;
+
+    use super::*;
+    use crate::tenant::storage_layer::IoConcurrency;
+
+    /// TODO: currently this test relies on manual visual inspection of the --no-capture output.
+    /// Should look like so:
+    /// ```text
+    /// RUST_LOG=trace cargo nextest run  --features testing  --no-capture test_io_concurrency_noise
+    /// running 1 test
+    /// 2025-01-21T17:42:01.335679Z  INFO get_vectored_concurrent_io test selected=SidecarTask
+    /// 2025-01-21T17:42:01.335680Z TRACE spawning sidecar task task_id=0
+    /// 2025-01-21T17:42:01.335937Z TRACE IoConcurrency_sidecar{task_id=0}: start
+    /// 2025-01-21T17:42:01.335972Z TRACE IoConcurrency_sidecar{task_id=0}: received new io future
+    /// 2025-01-21T17:42:01.335999Z  INFO IoConcurrency_sidecar{task_id=0}: waiting for signal to complete IO
+    /// 2025-01-21T17:42:01.336229Z  WARN dropping ValuesReconstructState while some IOs have not been completed num_active_ios=1 sidecar_task_id=Some(0) backtrace=   0: <pageserver::tenant::storage_layer::ValuesReconstructState as core::ops::drop::Drop>::drop
+    ///              at ./src/tenant/storage_layer.rs:553:24
+    ///    1: core::ptr::drop_in_place<pageserver::tenant::storage_layer::ValuesReconstructState>
+    ///              at /home/christian/.rustup/toolchains/1.84.0-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/core/src/ptr/mod.rs:521:1
+    ///    2: core::mem::drop
+    ///              at /home/christian/.rustup/toolchains/1.84.0-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/core/src/mem/mod.rs:942:24
+    ///    3: pageserver::tenant::storage_layer::tests2::test_io_concurrency_noise::{{closure}}
+    ///              at ./src/tenant/storage_layer.rs:1159:9
+    ///   ...
+    ///   49: <unknown>
+    /// 2025-01-21T17:42:01.452293Z  INFO IoConcurrency_sidecar{task_id=0}: completing IO
+    /// 2025-01-21T17:42:01.452357Z TRACE IoConcurrency_sidecar{task_id=0}: io future completed
+    /// 2025-01-21T17:42:01.452473Z TRACE IoConcurrency_sidecar{task_id=0}: end
+    /// test tenant::storage_layer::tests2::test_io_concurrency_noise ... ok
+    ///
+    /// ```
+    #[tokio::test]
+    async fn test_io_concurrency_noise() {
+        crate::tenant::harness::setup_logging();
+
+        let io_concurrency = IoConcurrency::spawn_for_test();
+        match *io_concurrency {
+            IoConcurrency::Sequential => {
+                // This test asserts behavior in sidecar mode, doesn't make sense in sequential mode.
+                return;
+            }
+            IoConcurrency::SidecarTask { .. } => {}
+        }
+        let mut reconstruct_state = ValuesReconstructState::new(io_concurrency.clone());
+
+        let (io_fut_is_waiting_tx, io_fut_is_waiting) = tokio::sync::oneshot::channel();
+        let (do_complete_io, should_complete_io) = tokio::sync::oneshot::channel();
+        let (io_fut_exiting_tx, io_fut_exiting) = tokio::sync::oneshot::channel();
+
+        let io = reconstruct_state.update_key(&DBDIR_KEY, Lsn(8), true);
+        reconstruct_state
+            .spawn_io(async move {
+                info!("waiting for signal to complete IO");
+                io_fut_is_waiting_tx.send(()).unwrap();
+                should_complete_io.await.unwrap();
+                info!("completing IO");
+                io.complete(Ok(OnDiskValue::RawImage(Bytes::new())));
+                io_fut_exiting_tx.send(()).unwrap();
+            })
+            .await;
+
+        io_fut_is_waiting.await.unwrap();
+
+        // this is what makes the noise
+        drop(reconstruct_state);
+
+        do_complete_io.send(()).unwrap();
+
+        io_fut_exiting.await.unwrap();
+    }
+}

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -41,13 +41,12 @@ use crate::tenant::vectored_blob_io::{
     BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
     VectoredReadPlanner,
 };
-use crate::tenant::PageReconstructError;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
 use crate::virtual_file::IoBufferMut;
 use crate::virtual_file::{self, MaybeFatalIo, VirtualFile};
 use crate::TEMP_FILE_SUFFIX;
 use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
-use anyhow::{anyhow, bail, ensure, Context, Result};
+use anyhow::{bail, ensure, Context, Result};
 use camino::{Utf8Path, Utf8PathBuf};
 use futures::StreamExt;
 use itertools::Itertools;
@@ -60,7 +59,7 @@ use pageserver_api::shard::TenantShardId;
 use pageserver_api::value::Value;
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
-use std::collections::VecDeque;
+use std::collections::{HashMap, VecDeque};
 use std::fs::File;
 use std::io::SeekFrom;
 use std::ops::Range;
@@ -77,7 +76,10 @@ use utils::{
     lsn::Lsn,
 };
 
-use super::{AsLayerDesc, LayerName, PersistentLayerDesc, ValuesReconstructState};
+use super::{
+    AsLayerDesc, LayerName, OnDiskValue, OnDiskValueIo, PersistentLayerDesc, ResidentLayer,
+    ValuesReconstructState,
+};
 
 ///
 /// Header stored in the beginning of the file
@@ -226,7 +228,7 @@ pub struct DeltaLayerInner {
     index_start_blk: u32,
     index_root_blk: u32,
 
-    file: VirtualFile,
+    file: Arc<VirtualFile>,
     file_id: FileId,
 
     layer_key_range: Range<Key>,
@@ -795,9 +797,11 @@ impl DeltaLayerInner {
         max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
         ctx: &RequestContext,
     ) -> anyhow::Result<Self> {
-        let file = VirtualFile::open_v2(path, ctx)
-            .await
-            .context("open layer file")?;
+        let file = Arc::new(
+            VirtualFile::open_v2(path, ctx)
+                .await
+                .context("open layer file")?,
+        );
 
         let file_id = page_cache::next_file_id();
 
@@ -842,12 +846,11 @@ impl DeltaLayerInner {
     // Look up the keys in the provided keyspace and update
     // the reconstruct state with whatever is found.
     //
-    // If the key is cached, go no further than the cached Lsn.
-    //
     // Currently, the index is visited for each range, but this
     // can be further optimised to visit the index only once.
     pub(super) async fn get_values_reconstruct_data(
         &self,
+        this: ResidentLayer,
         keyspace: KeySpace,
         lsn_range: Range<Lsn>,
         reconstruct_state: &mut ValuesReconstructState,
@@ -875,17 +878,14 @@ impl DeltaLayerInner {
             data_end_offset,
             index_reader,
             planner,
-            reconstruct_state,
             ctx,
         )
         .await
         .map_err(GetVectoredError::Other)?;
 
-        self.do_reads_and_update_state(reads, reconstruct_state, ctx)
+        self.do_reads_and_update_state(this, reads, reconstruct_state, ctx)
             .await;
 
-        reconstruct_state.on_lsn_advanced(&keyspace, lsn_range.start);
-
         Ok(())
     }
 
@@ -895,7 +895,6 @@ impl DeltaLayerInner {
         data_end_offset: u64,
         index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
         mut planner: VectoredReadPlanner,
-        reconstruct_state: &mut ValuesReconstructState,
         ctx: &RequestContext,
     ) -> anyhow::Result<Vec<VectoredRead>>
     where
@@ -922,10 +921,9 @@ impl DeltaLayerInner {
                 assert!(key >= range.start);
 
                 let outside_lsn_range = !lsn_range.contains(&lsn);
-                let below_cached_lsn = reconstruct_state.get_cached_lsn(&key) >= Some(lsn);
 
                 let flag = {
-                    if outside_lsn_range || below_cached_lsn {
+                    if outside_lsn_range {
                         BlobFlag::Ignore
                     } else if blob_ref.will_init() {
                         BlobFlag::ReplaceAll
@@ -994,98 +992,78 @@ impl DeltaLayerInner {
 
     async fn do_reads_and_update_state(
         &self,
+        this: ResidentLayer,
         reads: Vec<VectoredRead>,
         reconstruct_state: &mut ValuesReconstructState,
         ctx: &RequestContext,
     ) {
-        let vectored_blob_reader = VectoredBlobReader::new(&self.file);
-        let mut ignore_key_with_err = None;
-
         let max_vectored_read_bytes = self
             .max_vectored_read_bytes
             .expect("Layer is loaded with max vectored bytes config")
             .0
             .into();
         let buf_size = Self::get_min_read_buffer_size(&reads, max_vectored_read_bytes);
-        let mut buf = Some(IoBufferMut::with_capacity(buf_size));
 
         // Note that reads are processed in reverse order (from highest key+lsn).
         // This is the order that `ReconstructState` requires such that it can
         // track when a key is done.
         for read in reads.into_iter().rev() {
-            let res = vectored_blob_reader
-                .read_blobs(&read, buf.take().expect("Should have a buffer"), ctx)
-                .await;
-
-            let blobs_buf = match res {
-                Ok(blobs_buf) => blobs_buf,
-                Err(err) => {
-                    let kind = err.kind();
-                    for (_, blob_meta) in read.blobs_at.as_slice() {
-                        reconstruct_state.on_key_error(
-                            blob_meta.key,
-                            PageReconstructError::Other(anyhow!(
-                                "Failed to read blobs from virtual file {}: {}",
-                                self.file.path(),
-                                kind
-                            )),
-                        );
-                    }
-
-                    // We have "lost" the buffer since the lower level IO api
-                    // doesn't return the buffer on error. Allocate a new one.
-                    buf = Some(IoBufferMut::with_capacity(buf_size));
-
-                    continue;
-                }
-            };
-            let view = BufView::new_slice(&blobs_buf.buf);
-            for meta in blobs_buf.blobs.iter().rev() {
-                if Some(meta.meta.key) == ignore_key_with_err {
-                    continue;
-                }
-                let blob_read = meta.read(&view).await;
-                let blob_read = match blob_read {
-                    Ok(buf) => buf,
-                    Err(e) => {
-                        reconstruct_state.on_key_error(
-                            meta.meta.key,
-                            PageReconstructError::Other(anyhow!(e).context(format!(
-                                "Failed to decompress blob from virtual file {}",
-                                self.file.path(),
-                            ))),
-                        );
-
-                        ignore_key_with_err = Some(meta.meta.key);
-                        continue;
-                    }
-                };
-
-                let value = Value::des(&blob_read);
-
-                let value = match value {
-                    Ok(v) => v,
-                    Err(e) => {
-                        reconstruct_state.on_key_error(
-                            meta.meta.key,
-                            PageReconstructError::Other(anyhow!(e).context(format!(
-                                "Failed to deserialize blob from virtual file {}",
-                                self.file.path(),
-                            ))),
-                        );
-
-                        ignore_key_with_err = Some(meta.meta.key);
-                        continue;
-                    }
-                };
-
-                // Invariant: once a key reaches [`ValueReconstructSituation::Complete`]
-                // state, no further updates shall be made to it. The call below will
-                // panic if the invariant is violated.
-                reconstruct_state.update_key(&meta.meta.key, meta.meta.lsn, value);
+            let mut ios: HashMap<(Key, Lsn), OnDiskValueIo> = Default::default();
+            for (_, blob_meta) in read.blobs_at.as_slice().iter().rev() {
+                let io = reconstruct_state.update_key(
+                    &blob_meta.key,
+                    blob_meta.lsn,
+                    blob_meta.will_init,
+                );
+                ios.insert((blob_meta.key, blob_meta.lsn), io);
             }
 
-            buf = Some(blobs_buf.buf);
+            let read_extend_residency = this.clone();
+            let read_from = self.file.clone();
+            let read_ctx = ctx.attached_child();
+            reconstruct_state
+                .spawn_io(async move {
+                    let vectored_blob_reader = VectoredBlobReader::new(&read_from);
+                    let buf = IoBufferMut::with_capacity(buf_size);
+
+                    let res = vectored_blob_reader.read_blobs(&read, buf, &read_ctx).await;
+                    match res {
+                        Ok(blobs_buf) => {
+                            let view = BufView::new_slice(&blobs_buf.buf);
+                            for meta in blobs_buf.blobs.iter().rev() {
+                                let io = ios.remove(&(meta.meta.key, meta.meta.lsn)).unwrap();
+
+                                let blob_read = meta.read(&view).await;
+                                let blob_read = match blob_read {
+                                    Ok(buf) => buf,
+                                    Err(e) => {
+                                        io.complete(Err(e));
+                                        continue;
+                                    }
+                                };
+
+                                io.complete(Ok(OnDiskValue::WalRecordOrImage(
+                                    blob_read.into_bytes(),
+                                )));
+                            }
+
+                            assert!(ios.is_empty());
+                        }
+                        Err(err) => {
+                            for (_, sender) in ios {
+                                sender.complete(Err(std::io::Error::new(
+                                    err.kind(),
+                                    "vec read failed",
+                                )));
+                            }
+                        }
+                    }
+
+                    // keep layer resident until this IO is done; this spawned IO future generally outlives the
+                    // call to `self` / the `Arc<DownloadedLayer>` / the `ResidentLayer` that guarantees residency
+                    drop(read_extend_residency);
+                })
+                .await;
         }
     }
 
@@ -1224,7 +1202,14 @@ impl DeltaLayerInner {
             let actionable = if let Some((key, lsn, start_offset)) = prev.take() {
                 let end_offset = offset;
 
-                Some((BlobMeta { key, lsn }, start_offset..end_offset))
+                Some((
+                    BlobMeta {
+                        key,
+                        lsn,
+                        will_init: false,
+                    },
+                    start_offset..end_offset,
+                ))
             } else {
                 None
             };
@@ -1560,7 +1545,9 @@ impl DeltaLayerIterator<'_> {
                 let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
                 let blob_ref = BlobRef(value);
                 let offset = blob_ref.pos();
-                if let Some(batch_plan) = self.planner.handle(key, lsn, offset) {
+                if let Some(batch_plan) =
+                    self.planner.handle(key, lsn, offset, blob_ref.will_init())
+                {
                     break batch_plan;
                 }
             } else {
@@ -1673,7 +1660,6 @@ pub(crate) mod test {
             .expect("In memory disk finish should never fail");
         let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
         let planner = VectoredReadPlanner::new(100);
-        let mut reconstruct_state = ValuesReconstructState::new();
         let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
 
         let keyspace = KeySpace {
@@ -1691,7 +1677,6 @@ pub(crate) mod test {
             disk_offset,
             reader,
             planner,
-            &mut reconstruct_state,
             &ctx,
         )
         .await
@@ -1935,7 +1920,6 @@ pub(crate) mod test {
             );
 
             let planner = VectoredReadPlanner::new(constants::MAX_VECTORED_READ_BYTES);
-            let mut reconstruct_state = ValuesReconstructState::new();
             let keyspace = pick_random_keyspace(rng, &entries_meta.key_range);
             let data_end_offset = inner.index_start_blk as u64 * PAGE_SZ as u64;
 
@@ -1945,7 +1929,6 @@ pub(crate) mod test {
                 data_end_offset,
                 index_reader,
                 planner,
-                &mut reconstruct_state,
                 &ctx,
             )
             .await?;

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -38,12 +38,11 @@ use crate::tenant::vectored_blob_io::{
     BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
     VectoredReadPlanner,
 };
-use crate::tenant::PageReconstructError;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
 use crate::virtual_file::IoBufferMut;
 use crate::virtual_file::{self, MaybeFatalIo, VirtualFile};
 use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
-use anyhow::{anyhow, bail, ensure, Context, Result};
+use anyhow::{bail, ensure, Context, Result};
 use bytes::Bytes;
 use camino::{Utf8Path, Utf8PathBuf};
 use hex;
@@ -56,12 +55,13 @@ use pageserver_api::shard::{ShardIdentity, TenantShardId};
 use pageserver_api::value::Value;
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
-use std::collections::VecDeque;
+use std::collections::{HashMap, VecDeque};
 use std::fs::File;
 use std::io::SeekFrom;
 use std::ops::Range;
 use std::os::unix::prelude::FileExt;
 use std::str::FromStr;
+use std::sync::Arc;
 use tokio::sync::OnceCell;
 use tokio_stream::StreamExt;
 use tracing::*;
@@ -73,7 +73,10 @@ use utils::{
 };
 
 use super::layer_name::ImageLayerName;
-use super::{AsLayerDesc, LayerName, PersistentLayerDesc, ValuesReconstructState};
+use super::{
+    AsLayerDesc, LayerName, OnDiskValue, OnDiskValueIo, PersistentLayerDesc, ResidentLayer,
+    ValuesReconstructState,
+};
 
 ///
 /// Header stored in the beginning of the file
@@ -164,7 +167,7 @@ pub struct ImageLayerInner {
     key_range: Range<Key>,
     lsn: Lsn,
 
-    file: VirtualFile,
+    file: Arc<VirtualFile>,
     file_id: FileId,
 
     max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
@@ -391,9 +394,11 @@ impl ImageLayerInner {
         max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
         ctx: &RequestContext,
     ) -> anyhow::Result<Self> {
-        let file = VirtualFile::open_v2(path, ctx)
-            .await
-            .context("open layer file")?;
+        let file = Arc::new(
+            VirtualFile::open_v2(path, ctx)
+                .await
+                .context("open layer file")?,
+        );
         let file_id = page_cache::next_file_id();
         let block_reader = FileBlockReader::new(&file, file_id);
         let summary_blk = block_reader
@@ -439,6 +444,7 @@ impl ImageLayerInner {
     // the reconstruct state with whatever is found.
     pub(super) async fn get_values_reconstruct_data(
         &self,
+        this: ResidentLayer,
         keyspace: KeySpace,
         reconstruct_state: &mut ValuesReconstructState,
         ctx: &RequestContext,
@@ -448,7 +454,7 @@ impl ImageLayerInner {
             .await
             .map_err(GetVectoredError::Other)?;
 
-        self.do_reads_and_update_state(reads, reconstruct_state, ctx)
+        self.do_reads_and_update_state(this, reads, reconstruct_state, ctx)
             .await;
 
         reconstruct_state.on_image_layer_visited(&self.key_range);
@@ -570,6 +576,7 @@ impl ImageLayerInner {
 
     async fn do_reads_and_update_state(
         &self,
+        this: ResidentLayer,
         reads: Vec<VectoredRead>,
         reconstruct_state: &mut ValuesReconstructState,
         ctx: &RequestContext,
@@ -580,8 +587,13 @@ impl ImageLayerInner {
             .0
             .into();
 
-        let vectored_blob_reader = VectoredBlobReader::new(&self.file);
         for read in reads.into_iter() {
+            let mut ios: HashMap<(Key, Lsn), OnDiskValueIo> = Default::default();
+            for (_, blob_meta) in read.blobs_at.as_slice() {
+                let io = reconstruct_state.update_key(&blob_meta.key, blob_meta.lsn, true);
+                ios.insert((blob_meta.key, blob_meta.lsn), io);
+            }
+
             let buf_size = read.size();
 
             if buf_size > max_vectored_read_bytes {
@@ -611,50 +623,51 @@ impl ImageLayerInner {
                 }
             }
 
-            let buf = IoBufferMut::with_capacity(buf_size);
-            let res = vectored_blob_reader.read_blobs(&read, buf, ctx).await;
+            let read_extend_residency = this.clone();
+            let read_from = self.file.clone();
+            let read_ctx = ctx.attached_child();
+            reconstruct_state
+                .spawn_io(async move {
+                    let buf = IoBufferMut::with_capacity(buf_size);
+                    let vectored_blob_reader = VectoredBlobReader::new(&read_from);
+                    let res = vectored_blob_reader.read_blobs(&read, buf, &read_ctx).await;
 
-            match res {
-                Ok(blobs_buf) => {
-                    let view = BufView::new_slice(&blobs_buf.buf);
-                    for meta in blobs_buf.blobs.iter() {
-                        let img_buf = meta.read(&view).await;
+                    match res {
+                        Ok(blobs_buf) => {
+                            let view = BufView::new_slice(&blobs_buf.buf);
+                            for meta in blobs_buf.blobs.iter() {
+                                let io: OnDiskValueIo =
+                                    ios.remove(&(meta.meta.key, meta.meta.lsn)).unwrap();
+                                let img_buf = meta.read(&view).await;
 
-                        let img_buf = match img_buf {
-                            Ok(img_buf) => img_buf,
-                            Err(e) => {
-                                reconstruct_state.on_key_error(
-                                    meta.meta.key,
-                                    PageReconstructError::Other(anyhow!(e).context(format!(
-                                        "Failed to decompress blob from virtual file {}",
-                                        self.file.path(),
-                                    ))),
-                                );
+                                let img_buf = match img_buf {
+                                    Ok(img_buf) => img_buf,
+                                    Err(e) => {
+                                        io.complete(Err(e));
+                                        continue;
+                                    }
+                                };
 
-                                continue;
+                                io.complete(Ok(OnDiskValue::RawImage(img_buf.into_bytes())));
                             }
-                        };
-                        reconstruct_state.update_key(
-                            &meta.meta.key,
-                            self.lsn,
-                            Value::Image(img_buf.into_bytes()),
-                        );
+
+                            assert!(ios.is_empty());
+                        }
+                        Err(err) => {
+                            for (_, io) in ios {
+                                io.complete(Err(std::io::Error::new(
+                                    err.kind(),
+                                    "vec read failed",
+                                )));
+                            }
+                        }
                     }
-                }
-                Err(err) => {
-                    let kind = err.kind();
-                    for (_, blob_meta) in read.blobs_at.as_slice() {
-                        reconstruct_state.on_key_error(
-                            blob_meta.key,
-                            PageReconstructError::from(anyhow!(
-                                "Failed to read blobs from virtual file {}: {}",
-                                self.file.path(),
-                                kind
-                            )),
-                        );
-                    }
-                }
-            };
+
+                    // keep layer resident until this IO is done; this spawned IO future generally outlives the
+                    // call to `self` / the `Arc<DownloadedLayer>` / the `ResidentLayer` that guarantees residency
+                    drop(read_extend_residency);
+                })
+                .await;
         }
     }
 
@@ -1069,6 +1082,7 @@ impl ImageLayerIterator<'_> {
                     Key::from_slice(&raw_key[..KEY_SIZE]),
                     self.image_layer.lsn,
                     offset,
+                    true,
                 ) {
                     break batch_plan;
                 }

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -8,23 +8,22 @@ use crate::assert_u64_eq_usize::{u64_to_usize, U64IsUsize, UsizeIsU64};
 use crate::config::PageServerConf;
 use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
 use crate::tenant::ephemeral_file::EphemeralFile;
+use crate::tenant::storage_layer::{OnDiskValue, OnDiskValueIo};
 use crate::tenant::timeline::GetVectoredError;
-use crate::tenant::PageReconstructError;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
 use crate::{l0_flush, page_cache};
-use anyhow::{anyhow, Result};
+use anyhow::Result;
 use camino::Utf8PathBuf;
 use pageserver_api::key::CompactKey;
 use pageserver_api::key::Key;
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::InMemoryLayerInfo;
 use pageserver_api::shard::TenantShardId;
-use pageserver_api::value::Value;
 use std::collections::{BTreeMap, HashMap};
 use std::sync::{Arc, OnceLock};
 use std::time::Instant;
 use tracing::*;
-use utils::{bin_ser::BeSer, id::TimelineId, lsn::Lsn, vec_map::VecMap};
+use utils::{id::TimelineId, lsn::Lsn, vec_map::VecMap};
 use wal_decoder::serialized_batch::{SerializedValueBatch, SerializedValueMeta, ValueMeta};
 // avoid binding to Write (conflicts with std::io::Write)
 // while being able to use std::fmt::Write's methods
@@ -36,9 +35,7 @@ use std::sync::atomic::Ordering as AtomicOrdering;
 use std::sync::atomic::{AtomicU64, AtomicUsize};
 use tokio::sync::RwLock;
 
-use super::{
-    DeltaLayerWriter, PersistentLayerDesc, ValueReconstructSituation, ValuesReconstructState,
-};
+use super::{DeltaLayerWriter, PersistentLayerDesc, ValuesReconstructState};
 
 pub(crate) mod vectored_dio_read;
 
@@ -415,10 +412,8 @@ impl InMemoryLayer {
 
     // Look up the keys in the provided keyspace and update
     // the reconstruct state with whatever is found.
-    //
-    // If the key is cached, go no further than the cached Lsn.
     pub(crate) async fn get_values_reconstruct_data(
-        &self,
+        self: &Arc<InMemoryLayer>,
         keyspace: KeySpace,
         end_lsn: Lsn,
         reconstruct_state: &mut ValuesReconstructState,
@@ -435,6 +430,9 @@ impl InMemoryLayer {
             read: vectored_dio_read::LogicalRead<Vec<u8>>,
         }
         let mut reads: HashMap<Key, Vec<ValueRead>> = HashMap::new();
+        let mut ios: HashMap<(Key, Lsn), OnDiskValueIo> = Default::default();
+
+        let lsn_range = self.start_lsn..end_lsn;
 
         for range in keyspace.ranges.iter() {
             for (key, vec_map) in inner
@@ -442,12 +440,7 @@ impl InMemoryLayer {
                 .range(range.start.to_compact()..range.end.to_compact())
             {
                 let key = Key::from_compact(*key);
-                let lsn_range = match reconstruct_state.get_cached_lsn(&key) {
-                    Some(cached_lsn) => (cached_lsn + 1)..end_lsn,
-                    None => self.start_lsn..end_lsn,
-                };
-
-                let slice = vec_map.slice_range(lsn_range);
+                let slice = vec_map.slice_range(lsn_range.clone());
 
                 for (entry_lsn, index_entry) in slice.iter().rev() {
                     let IndexEntryUnpacked {
@@ -463,55 +456,59 @@ impl InMemoryLayer {
                             Vec::with_capacity(len as usize),
                         ),
                     });
+
+                    let io = reconstruct_state.update_key(&key, *entry_lsn, will_init);
+                    ios.insert((key, *entry_lsn), io);
+
                     if will_init {
                         break;
                     }
                 }
             }
         }
+        drop(inner); // release the lock before we spawn the IO; if it's serial-mode IO we will deadlock on the read().await below
+        let read_from = Arc::clone(self);
+        let read_ctx = ctx.attached_child();
+        reconstruct_state
+            .spawn_io(async move {
+                let inner = read_from.inner.read().await;
+                let f = vectored_dio_read::execute(
+                    &inner.file,
+                    reads
+                        .iter()
+                        .flat_map(|(_, value_reads)| value_reads.iter().map(|v| &v.read)),
+                    &read_ctx,
+                );
+                send_future::SendFuture::send(f) // https://github.com/rust-lang/rust/issues/96865
+                    .await;
 
-        // Execute the reads.
-
-        let f = vectored_dio_read::execute(
-            &inner.file,
-            reads
-                .iter()
-                .flat_map(|(_, value_reads)| value_reads.iter().map(|v| &v.read)),
-            &ctx,
-        );
-        send_future::SendFuture::send(f) // https://github.com/rust-lang/rust/issues/96865
-            .await;
-
-        // Process results into the reconstruct state
-        'next_key: for (key, value_reads) in reads {
-            for ValueRead { entry_lsn, read } in value_reads {
-                match read.into_result().expect("we run execute() above") {
-                    Err(e) => {
-                        reconstruct_state.on_key_error(key, PageReconstructError::from(anyhow!(e)));
-                        continue 'next_key;
-                    }
-                    Ok(value_buf) => {
-                        let value = Value::des(&value_buf);
-                        if let Err(e) = value {
-                            reconstruct_state
-                                .on_key_error(key, PageReconstructError::from(anyhow!(e)));
-                            continue 'next_key;
+                for (key, value_reads) in reads {
+                    for ValueRead { entry_lsn, read } in value_reads {
+                        let io = ios.remove(&(key, entry_lsn)).expect("sender must exist");
+                        match read.into_result().expect("we run execute() above") {
+                            Err(e) => {
+                                io.complete(Err(std::io::Error::new(
+                                    e.kind(),
+                                    "dio vec read failed",
+                                )));
+                            }
+                            Ok(value_buf) => {
+                                io.complete(Ok(OnDiskValue::WalRecordOrImage(value_buf.into())));
+                            }
                         }
-
-                        let key_situation =
-                            reconstruct_state.update_key(&key, entry_lsn, value.unwrap());
-                        if key_situation == ValueReconstructSituation::Complete {
-                            // TODO: metric to see if we fetched more values than necessary
-                            continue 'next_key;
-                        }
-
-                        // process the next value in the next iteration of the loop
                     }
                 }
-            }
-        }
 
-        reconstruct_state.on_lsn_advanced(&keyspace, self.start_lsn);
+                assert!(ios.is_empty());
+
+                // Keep layer existent until this IO is done;
+                // This is kinda forced for InMemoryLayer because we need to inner.read() anyway,
+                // but it's less obvious for DeltaLayer and ImageLayer. So, keep this explicit
+                // drop for consistency among all three layer types.
+                drop(inner);
+                drop(read_from);
+            })
+            .await;
 
         Ok(())
     }
@@ -606,6 +603,7 @@ impl InMemoryLayer {
         // Write the batch to the file
         inner.file.write_raw(&raw, ctx).await?;
         let new_size = inner.file.len();
+
         let expected_new_len = base_offset
             .checked_add(raw.len().into_u64())
             // write_raw would error if we were to overflow u64.

pageserver/src/tenant/storage_layer/layer.rs

@@ -308,7 +308,7 @@ impl Layer {
         reconstruct_data: &mut ValuesReconstructState,
         ctx: &RequestContext,
     ) -> Result<(), GetVectoredError> {
-        let layer = self
+        let downloaded = self
             .0
             .get_or_maybe_download(true, Some(ctx))
             .await
@@ -318,11 +318,15 @@ impl Layer {
                 }
                 other => GetVectoredError::Other(anyhow::anyhow!(other)),
             })?;
+        let this = ResidentLayer {
+            downloaded: downloaded.clone(),
+            owner: self.clone(),
+        };
 
         self.record_access(ctx);
 
-        layer
-            .get_values_reconstruct_data(keyspace, lsn_range, reconstruct_data, &self.0, ctx)
+        downloaded
+            .get_values_reconstruct_data(this, keyspace, lsn_range, reconstruct_data, ctx)
             .instrument(tracing::debug_span!("get_values_reconstruct_data", layer=%self))
             .await
             .map_err(|err| match err {
@@ -1768,25 +1772,25 @@ impl DownloadedLayer {
 
     async fn get_values_reconstruct_data(
         &self,
+        this: ResidentLayer,
         keyspace: KeySpace,
         lsn_range: Range<Lsn>,
         reconstruct_data: &mut ValuesReconstructState,
-        owner: &Arc<LayerInner>,
         ctx: &RequestContext,
     ) -> Result<(), GetVectoredError> {
         use LayerKind::*;
 
         match self
-            .get(owner, ctx)
+            .get(&this.owner.0, ctx)
             .await
             .map_err(GetVectoredError::Other)?
         {
             Delta(d) => {
-                d.get_values_reconstruct_data(keyspace, lsn_range, reconstruct_data, ctx)
+                d.get_values_reconstruct_data(this, keyspace, lsn_range, reconstruct_data, ctx)
                     .await
             }
             Image(i) => {
-                i.get_values_reconstruct_data(keyspace, reconstruct_data, ctx)
+                i.get_values_reconstruct_data(this, keyspace, reconstruct_data, ctx)
                     .await
             }
         }

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

@@ -1,6 +1,6 @@
 use std::time::UNIX_EPOCH;
 
-use pageserver_api::key::CONTROLFILE_KEY;
+use pageserver_api::key::{Key, CONTROLFILE_KEY};
 use tokio::task::JoinSet;
 use utils::{
     completion::{self, Completion},
@@ -9,7 +9,13 @@ use utils::{
 
 use super::failpoints::{Failpoint, FailpointKind};
 use super::*;
-use crate::{context::DownloadBehavior, tenant::storage_layer::LayerVisibilityHint};
+use crate::{
+    context::DownloadBehavior,
+    tenant::{
+        harness::test_img,
+        storage_layer::{IoConcurrency, LayerVisibilityHint},
+    },
+};
 use crate::{task_mgr::TaskKind, tenant::harness::TenantHarness};
 
 /// Used in tests to advance a future to wanted await point, and not futher.
@@ -28,23 +34,55 @@ async fn smoke_test() {
     let span = h.span();
     let download_span = span.in_scope(|| tracing::info_span!("downloading", timeline_id = 1));
     let (tenant, _) = h.load().await;
+    let io_concurrency = IoConcurrency::spawn_for_test();
 
     let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Download);
 
+    let image_layers = vec![(
+        Lsn(0x40),
+        vec![(
+            Key::from_hex("620000000033333333444444445500000000").unwrap(),
+            test_img("foo"),
+        )],
+    )];
+
+    // Create a test timeline with one real layer, and one synthetic test layer.  The synthetic
+    // one is only there so that we can GC the real one without leaving the timeline's metadata
+    // empty, which is an illegal state (see [`IndexPart::validate`]).
     let timeline = tenant
-        .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
+        .create_test_timeline_with_layers(
+            TimelineId::generate(),
+            Lsn(0x10),
+            14,
+            &ctx,
+            Default::default(),
+            image_layers,
+            Lsn(0x100),
+        )
         .await
         .unwrap();
 
-    let layer = {
+    // Grab one of the timeline's layers to exercise in the test, and the other layer that is just
+    // there to avoid the timeline being illegally empty
+    let (layer, dummy_layer) = {
         let mut layers = {
             let layers = timeline.layers.read().await;
             layers.likely_resident_layers().cloned().collect::<Vec<_>>()
         };
 
-        assert_eq!(layers.len(), 1);
+        assert_eq!(layers.len(), 2);
 
-        layers.swap_remove(0)
+        layers.sort_by_key(|l| l.layer_desc().get_key_range().start);
+        let synthetic_layer = layers.pop().unwrap();
+        let real_layer = layers.pop().unwrap();
+        tracing::info!(
+            "real_layer={:?} ({}), synthetic_layer={:?} ({})",
+            real_layer,
+            real_layer.layer_desc().file_size,
+            synthetic_layer,
+            synthetic_layer.layer_desc().file_size
+        );
+        (real_layer, synthetic_layer)
     };
 
     // all layers created at pageserver are like `layer`, initialized with strong
@@ -55,7 +93,7 @@ async fn smoke_test() {
     };
 
     let img_before = {
-        let mut data = ValuesReconstructState::default();
+        let mut data = ValuesReconstructState::new(io_concurrency.clone());
         layer
             .get_values_reconstruct_data(
                 controlfile_keyspace.clone(),
@@ -65,10 +103,13 @@ async fn smoke_test() {
             )
             .await
             .unwrap();
+
         data.keys
             .remove(&CONTROLFILE_KEY)
             .expect("must be present")
-            .expect("should not error")
+            .collect_pending_ios()
+            .await
+            .expect("must not error")
             .img
             .take()
             .expect("tenant harness writes the control file")
@@ -87,7 +128,7 @@ async fn smoke_test() {
 
     // on accesses when the layer is evicted, it will automatically be downloaded.
     let img_after = {
-        let mut data = ValuesReconstructState::default();
+        let mut data = ValuesReconstructState::new(io_concurrency.clone());
         layer
             .get_values_reconstruct_data(
                 controlfile_keyspace.clone(),
@@ -101,7 +142,9 @@ async fn smoke_test() {
         data.keys
             .remove(&CONTROLFILE_KEY)
             .expect("must be present")
-            .expect("should not error")
+            .collect_pending_ios()
+            .await
+            .expect("must not error")
             .img
             .take()
             .expect("tenant harness writes the control file")
@@ -173,10 +216,13 @@ async fn smoke_test() {
 
     let rtc = &timeline.remote_client;
 
+    // Simulate GC removing our test layer.
     {
-        let layers = &[layer];
         let mut g = timeline.layers.write().await;
+
+        let layers = &[layer];
         g.open_mut().unwrap().finish_gc_timeline(layers);
+
         // this just updates the remote_physical_size for demonstration purposes
         rtc.schedule_gc_update(layers).unwrap();
     }
@@ -191,7 +237,10 @@ async fn smoke_test() {
 
     rtc.wait_completion().await.unwrap();
 
-    assert_eq!(rtc.get_remote_physical_size(), 0);
+    assert_eq!(
+        rtc.get_remote_physical_size(),
+        dummy_layer.metadata().file_size
+    );
     assert_eq!(0, LAYER_IMPL_METRICS.inits_cancelled.get())
 }
 

pageserver/src/tenant/tasks.rs

@@ -67,10 +67,9 @@ pub(crate) async fn concurrent_background_tasks_rate_limit_permit(
 ) -> tokio::sync::SemaphorePermit<'static> {
     let _guard = crate::metrics::BACKGROUND_LOOP_SEMAPHORE.measure_acquisition(loop_kind);
 
-    pausable_failpoint!(
-        "initial-size-calculation-permit-pause",
-        loop_kind == BackgroundLoopKind::InitialLogicalSizeCalculation
-    );
+    if loop_kind == BackgroundLoopKind::InitialLogicalSizeCalculation {
+        pausable_failpoint!("initial-size-calculation-permit-pause");
+    }
 
     // TODO: assert that we run on BACKGROUND_RUNTIME; requires tokio_unstable Handle::id();
     match CONCURRENT_BACKGROUND_TASKS.acquire().await {

pageserver/src/tenant/timeline.rs

@@ -20,6 +20,7 @@ use camino::Utf8Path;
 use chrono::{DateTime, Utc};
 use enumset::EnumSet;
 use fail::fail_point;
+use futures::{stream::FuturesUnordered, StreamExt};
 use handle::ShardTimelineId;
 use offload::OffloadError;
 use once_cell::sync::Lazy;
@@ -60,6 +61,7 @@ use utils::{
 use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
 
 use std::sync::atomic::Ordering as AtomicOrdering;
+use std::sync::OnceLock;
 use std::sync::{Arc, Mutex, RwLock, Weak};
 use std::time::{Duration, Instant, SystemTime};
 use std::{
@@ -72,8 +74,8 @@ use std::{
     collections::btree_map::Entry,
     ops::{Deref, Range},
 };
-use std::{pin::pin, sync::OnceLock};
 
+use crate::l0_flush::{self, L0FlushGlobalState};
 use crate::{
     aux_file::AuxFileSizeEstimator,
     page_service::TenantManagerTypes,
@@ -81,7 +83,10 @@ use crate::{
         config::AttachmentMode,
         layer_map::{LayerMap, SearchResult},
         metadata::TimelineMetadata,
-        storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
+        storage_layer::{
+            inmemory_layer::IndexEntry, IoConcurrency, PersistentLayerDesc,
+            ValueReconstructSituation,
+        },
     },
     walingest::WalLagCooldown,
     walredo,
@@ -102,10 +107,6 @@ use crate::{
 use crate::{
     disk_usage_eviction_task::EvictionCandidate, tenant::storage_layer::delta_layer::DeltaEntry,
 };
-use crate::{
-    l0_flush::{self, L0FlushGlobalState},
-    metrics::GetKind,
-};
 use crate::{
     metrics::ScanLatencyOngoingRecording, tenant::timeline::logical_size::CurrentLogicalSize,
 };
@@ -1005,9 +1006,7 @@ impl Timeline {
             ranges: vec![key..key.next()],
         };
 
-        // Initialise the reconstruct state for the key with the cache
-        // entry returned above.
-        let mut reconstruct_state = ValuesReconstructState::new();
+        let mut reconstruct_state = ValuesReconstructState::new(IoConcurrency::sequential());
 
         let vectored_res = self
             .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
@@ -1050,6 +1049,7 @@ impl Timeline {
         &self,
         keyspace: KeySpace,
         lsn: Lsn,
+        io_concurrency: super::storage_layer::IoConcurrency,
         ctx: &RequestContext,
     ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
         if !lsn.is_valid() {
@@ -1084,7 +1084,7 @@ impl Timeline {
             .get_vectored_impl(
                 keyspace.clone(),
                 lsn,
-                &mut ValuesReconstructState::new(),
+                &mut ValuesReconstructState::new(io_concurrency),
                 ctx,
             )
             .await;
@@ -1109,6 +1109,7 @@ impl Timeline {
         keyspace: KeySpace,
         lsn: Lsn,
         ctx: &RequestContext,
+        io_concurrency: super::storage_layer::IoConcurrency,
     ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
         if !lsn.is_valid() {
             return Err(GetVectoredError::InvalidLsn(lsn));
@@ -1140,7 +1141,7 @@ impl Timeline {
             .get_vectored_impl(
                 keyspace.clone(),
                 lsn,
-                &mut ValuesReconstructState::default(),
+                &mut ValuesReconstructState::new(io_concurrency),
                 ctx,
             )
             .await;
@@ -1159,39 +1160,56 @@ impl Timeline {
         reconstruct_state: &mut ValuesReconstructState,
         ctx: &RequestContext,
     ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
-        let get_kind = if keyspace.total_raw_size() == 1 {
-            GetKind::Singular
-        } else {
-            GetKind::Vectored
+        let traversal_res: Result<(), _> = self
+            .get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
+            .await;
+        if let Err(err) = traversal_res {
+            // Wait for all the spawned IOs to complete.
+            // See comments on `spawn_io` inside `storage_layer` for more details.
+            let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
+                .into_values()
+                .map(|state| state.collect_pending_ios())
+                .collect::<FuturesUnordered<_>>();
+            while collect_futs.next().await.is_some() {}
+            return Err(err);
         };
 
-        let get_data_timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME
-            .for_get_kind(get_kind)
-            .start_timer();
-        self.get_vectored_reconstruct_data(keyspace.clone(), lsn, reconstruct_state, ctx)
-            .await?;
-        get_data_timer.stop_and_record();
-
-        let reconstruct_timer = crate::metrics::RECONSTRUCT_TIME
-            .for_get_kind(get_kind)
-            .start_timer();
-        let mut results: BTreeMap<Key, Result<Bytes, PageReconstructError>> = BTreeMap::new();
         let layers_visited = reconstruct_state.get_layers_visited();
 
-        for (key, res) in std::mem::take(&mut reconstruct_state.keys) {
-            match res {
-                Err(err) => {
-                    results.insert(key, Err(err));
-                }
-                Ok(state) => {
-                    let state = ValueReconstructState::from(state);
+        let futs = FuturesUnordered::new();
+        for (key, state) in std::mem::take(&mut reconstruct_state.keys) {
+            futs.push({
+                let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
+                async move {
+                    assert_eq!(state.situation, ValueReconstructSituation::Complete);
 
-                    let reconstruct_res = self.reconstruct_value(key, lsn, state).await;
-                    results.insert(key, reconstruct_res);
+                    let converted = match state.collect_pending_ios().await {
+                        Ok(ok) => ok,
+                        Err(err) => {
+                            return (key, Err(err));
+                        }
+                    };
+
+                    // The walredo module expects the records to be descending in terms of Lsn.
+                    // And we submit the IOs in that order, so, there shuold be no need to sort here.
+                    debug_assert!(
+                        converted
+                            .records
+                            .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
+                        "{converted:?}"
+                    );
+
+                    (
+                        key,
+                        walredo_self.reconstruct_value(key, lsn, converted).await,
+                    )
                 }
-            }
+            });
         }
-        reconstruct_timer.stop_and_record();
+
+        let results = futs
+            .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
+            .await;
 
         // For aux file keys (v1 or v2) the vectored read path does not return an error
         // when they're missing. Instead they are omitted from the resulting btree
@@ -2804,12 +2822,10 @@ impl Timeline {
             "initial size calculation",
             // NB: don't log errors here, task_mgr will do that.
             async move {
-                let cancel = task_mgr::shutdown_token();
                 self_clone
                     .initial_logical_size_calculation_task(
                         initial_part_end,
                         cancel_wait_for_background_loop_concurrency_limit_semaphore,
-                        cancel,
                         background_ctx,
                     )
                     .await;
@@ -2819,11 +2835,21 @@ impl Timeline {
         );
     }
 
+    /// # Cancellation
+    ///
+    /// This method is sensitive to `Timeline::cancel`.
+    ///
+    /// It is _not_ sensitive to task_mgr::shutdown_token().
+    ///
+    /// # Cancel-Safety
+    ///
+    /// It does Timeline IO, hence this should be polled to completion because
+    /// we could be leaving in-flight IOs behind, which is safe, but annoying
+    /// to reason about.
     async fn initial_logical_size_calculation_task(
         self: Arc<Self>,
         initial_part_end: Lsn,
         skip_concurrency_limiter: CancellationToken,
-        cancel: CancellationToken,
         background_ctx: RequestContext,
     ) {
         scopeguard::defer! {
@@ -2836,7 +2862,6 @@ impl Timeline {
             let self_ref = &self;
             let skip_concurrency_limiter = &skip_concurrency_limiter;
             async move {
-                let cancel = task_mgr::shutdown_token();
                 let wait_for_permit = super::tasks::concurrent_background_tasks_rate_limit_permit(
                     BackgroundLoopKind::InitialLogicalSizeCalculation,
                     background_ctx,
@@ -2850,9 +2875,6 @@ impl Timeline {
                     _ = self_ref.cancel.cancelled() => {
                         return Err(CalculateLogicalSizeError::Cancelled);
                     }
-                    _ = cancel.cancelled() => {
-                        return Err(CalculateLogicalSizeError::Cancelled);
-                    },
                     () = skip_concurrency_limiter.cancelled() => {
                         // Some action that is part of a end user interaction requested logical size
                         // => break out of the rate limit
@@ -2869,6 +2891,14 @@ impl Timeline {
                     crate::metrics::initial_logical_size::START_CALCULATION.retry(circumstances)
                 };
 
+                let io_concurrency = IoConcurrency::spawn_from_conf(
+                    self_ref.conf,
+                    self_ref
+                        .gate
+                        .enter()
+                        .map_err(|_| CalculateLogicalSizeError::Cancelled)?,
+                );
+
                 let calculated_size = self_ref
                     .logical_size_calculation_task(
                         initial_part_end,
@@ -2878,7 +2908,11 @@ impl Timeline {
                     .await?;
 
                 self_ref
-                    .trigger_aux_file_size_computation(initial_part_end, background_ctx)
+                    .trigger_aux_file_size_computation(
+                        initial_part_end,
+                        background_ctx,
+                        io_concurrency,
+                    )
                     .await?;
 
                 // TODO: add aux file size to logical size
@@ -2911,22 +2945,18 @@ impl Timeline {
                             )
                             .expect("10min < 1hour"),
                         );
-                        tokio::time::sleep(sleep_duration).await;
+                        tokio::select! {
+                            _ = tokio::time::sleep(sleep_duration) => {}
+                            _ = self.cancel.cancelled() => return ControlFlow::Break(()),
+                        }
                     }
                 }
             }
         };
 
-        let (calculated_size, metrics_guard) = tokio::select! {
-            res = retrying  => {
-                match res {
-                    ControlFlow::Continue(calculated_size) => calculated_size,
-                    ControlFlow::Break(()) => return,
-                }
-            }
-            _ = cancel.cancelled() => {
-                return;
-            }
+        let (calculated_size, metrics_guard) = match retrying.await {
+            ControlFlow::Continue(calculated_size) => calculated_size,
+            ControlFlow::Break(()) => return,
         };
 
         // we cannot query current_logical_size.current_size() to know the current
@@ -2982,9 +3012,6 @@ impl Timeline {
         receiver
     }
 
-    /// # Cancel-Safety
-    ///
-    /// This method is cancellation-safe.
     #[instrument(skip_all)]
     async fn logical_size_calculation_task(
         self: &Arc<Self>,
@@ -3002,32 +3029,13 @@ impl Timeline {
             .enter()
             .map_err(|_| CalculateLogicalSizeError::Cancelled)?;
 
-        let self_calculation = Arc::clone(self);
-
-        let mut calculation = pin!(async {
-            let ctx = ctx.attached_child();
-            self_calculation
-                .calculate_logical_size(lsn, cause, &guard, &ctx)
-                .await
-        });
-
-        tokio::select! {
-            res = &mut calculation => { res }
-            _ = self.cancel.cancelled() => {
-                debug!("cancelling logical size calculation for timeline shutdown");
-                calculation.await
-            }
-        }
+        self.calculate_logical_size(lsn, cause, &guard, ctx).await
     }
 
     /// Calculate the logical size of the database at the latest LSN.
     ///
     /// NOTE: counted incrementally, includes ancestors. This can be a slow operation,
     /// especially if we need to download remote layers.
-    ///
-    /// # Cancel-Safety
-    ///
-    /// This method is cancellation-safe.
     async fn calculate_logical_size(
         &self,
         up_to_lsn: Lsn,
@@ -3040,7 +3048,10 @@ impl Timeline {
             self.timeline_id, up_to_lsn
         );
 
-        pausable_failpoint!("timeline-calculate-logical-size-pause");
+        if let Err(()) = pausable_failpoint!("timeline-calculate-logical-size-pause", &self.cancel)
+        {
+            return Err(CalculateLogicalSizeError::Cancelled);
+        }
 
         // See if we've already done the work for initial size calculation.
         // This is a short-cut for timelines that are mostly unused.
@@ -3617,6 +3628,12 @@ impl Timeline {
                     return;
                 }
 
+                // Break to notify potential waiters as soon as we've flushed the requested LSN. If
+                // more requests have arrived in the meanwhile, we'll resume flushing afterwards.
+                if flushed_to_lsn >= frozen_to_lsn {
+                    break Ok(());
+                }
+
                 let timer = self.metrics.flush_time_histo.start_timer();
 
                 let num_frozen_layers;
@@ -4128,6 +4145,7 @@ impl Timeline {
 
     /// Create image layers for Postgres data. Assumes the caller passes a partition that is not too large,
     /// so that at most one image layer will be produced from this function.
+    #[allow(clippy::too_many_arguments)]
     async fn create_image_layer_for_rel_blocks(
         self: &Arc<Self>,
         partition: &KeySpace,
@@ -4136,6 +4154,7 @@ impl Timeline {
         ctx: &RequestContext,
         img_range: Range<Key>,
         start: Key,
+        io_concurrency: IoConcurrency,
     ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
         let mut wrote_keys = false;
 
@@ -4164,7 +4183,12 @@ impl Timeline {
                     || (last_key_in_range && key_request_accum.raw_size() > 0)
                 {
                     let results = self
-                        .get_vectored(key_request_accum.consume_keyspace(), lsn, ctx)
+                        .get_vectored(
+                            key_request_accum.consume_keyspace(),
+                            lsn,
+                            io_concurrency.clone(),
+                            ctx,
+                        )
                         .await?;
 
                     if self.cancel.is_cancelled() {
@@ -4243,9 +4267,10 @@ impl Timeline {
         img_range: Range<Key>,
         mode: ImageLayerCreationMode,
         start: Key,
+        io_concurrency: IoConcurrency,
     ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
         // Metadata keys image layer creation.
-        let mut reconstruct_state = ValuesReconstructState::default();
+        let mut reconstruct_state = ValuesReconstructState::new(io_concurrency);
         let begin = Instant::now();
         let data = self
             .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
@@ -4462,6 +4487,13 @@ impl Timeline {
                 )))
             });
 
+            let io_concurrency = IoConcurrency::spawn_from_conf(
+                self.conf,
+                self.gate
+                    .enter()
+                    .map_err(|_| CreateImageLayersError::Cancelled)?,
+            );
+
             if !compact_metadata {
                 let ImageLayerCreationOutcome {
                     image,
@@ -4474,6 +4506,7 @@ impl Timeline {
                         ctx,
                         img_range,
                         start,
+                        io_concurrency,
                     )
                     .await?;
 
@@ -4492,6 +4525,7 @@ impl Timeline {
                         img_range,
                         mode,
                         start,
+                        io_concurrency,
                     )
                     .await?;
                 start = next_start_key;
@@ -5678,9 +5712,17 @@ impl Timeline {
         info!("force created image layer {}", image_layer.local_path());
         {
             let mut guard = self.layers.write().await;
-            guard.open_mut().unwrap().force_insert_layer(image_layer);
+            guard
+                .open_mut()
+                .unwrap()
+                .force_insert_layer(image_layer.clone());
         }
 
+        // Update remote_timeline_client state to reflect existence of this layer
+        self.remote_client
+            .schedule_layer_file_upload(image_layer)
+            .unwrap();
+
         Ok(())
     }
 
@@ -5731,9 +5773,17 @@ impl Timeline {
         info!("force created delta layer {}", delta_layer.local_path());
         {
             let mut guard = self.layers.write().await;
-            guard.open_mut().unwrap().force_insert_layer(delta_layer);
+            guard
+                .open_mut()
+                .unwrap()
+                .force_insert_layer(delta_layer.clone());
         }
 
+        // Update remote_timeline_client state to reflect existence of this layer
+        self.remote_client
+            .schedule_layer_file_upload(delta_layer)
+            .unwrap();
+
         Ok(())
     }
 
@@ -5743,13 +5793,14 @@ impl Timeline {
         self: &Arc<Timeline>,
         lsn: Lsn,
         ctx: &RequestContext,
+        io_concurrency: IoConcurrency,
     ) -> anyhow::Result<Vec<(Key, Bytes)>> {
         let mut all_data = Vec::new();
         let guard = self.layers.read().await;
         for layer in guard.layer_map()?.iter_historic_layers() {
             if !layer.is_delta() && layer.image_layer_lsn() == lsn {
                 let layer = guard.get_from_desc(&layer);
-                let mut reconstruct_data = ValuesReconstructState::default();
+                let mut reconstruct_data = ValuesReconstructState::new(io_concurrency.clone());
                 layer
                     .get_values_reconstruct_data(
                         KeySpace::single(Key::MIN..Key::MAX),
@@ -5758,8 +5809,9 @@ impl Timeline {
                         ctx,
                     )
                     .await?;
-                for (k, v) in reconstruct_data.keys {
-                    all_data.push((k, v?.img.unwrap().1));
+                for (k, v) in std::mem::take(&mut reconstruct_data.keys) {
+                    let v = v.collect_pending_ios().await?;
+                    all_data.push((k, v.img.unwrap().1));
                 }
             }
         }

pageserver/src/tenant/timeline/compaction.rs

@@ -42,8 +42,8 @@ use crate::tenant::storage_layer::merge_iterator::MergeIterator;
 use crate::tenant::storage_layer::{
     AsLayerDesc, PersistentLayerDesc, PersistentLayerKey, ValueReconstructState,
 };
-use crate::tenant::timeline::ImageLayerCreationOutcome;
 use crate::tenant::timeline::{drop_rlock, DeltaLayerWriter, ImageLayerWriter};
+use crate::tenant::timeline::{ImageLayerCreationOutcome, IoConcurrency};
 use crate::tenant::timeline::{Layer, ResidentLayer};
 use crate::tenant::{gc_block, DeltaLayer, MaybeOffloaded};
 use crate::virtual_file::{MaybeFatalIo, VirtualFile};
@@ -3170,6 +3170,7 @@ impl TimelineAdaptor {
                 ctx,
                 key_range.clone(),
                 start,
+                IoConcurrency::sequential(),
             )
             .await?;
 

pageserver/src/tenant/timeline/delete.rs

@@ -112,7 +112,7 @@ pub(super) async fn delete_local_timeline_directory(
 }
 
 /// It is important that this gets called when DeletionGuard is being held.
-/// For more context see comments in [`DeleteTimelineFlow::prepare`]
+/// For more context see comments in [`make_timeline_delete_guard`]
 async fn remove_maybe_offloaded_timeline_from_tenant(
     tenant: &Tenant,
     timeline: &TimelineOrOffloaded,
@@ -193,10 +193,8 @@ impl DeleteTimelineFlow {
     ) -> Result<(), DeleteTimelineError> {
         super::debug_assert_current_span_has_tenant_and_timeline_id();
 
-        let allow_offloaded_children = false;
-        let set_stopping = true;
         let (timeline, mut guard) =
-            Self::prepare(tenant, timeline_id, allow_offloaded_children, set_stopping)?;
+            make_timeline_delete_guard(tenant, timeline_id, TimelineDeleteGuardKind::Delete)?;
 
         guard.mark_in_progress()?;
 
@@ -333,75 +331,6 @@ impl DeleteTimelineFlow {
         Ok(())
     }
 
-    pub(super) fn prepare(
-        tenant: &Tenant,
-        timeline_id: TimelineId,
-        allow_offloaded_children: bool,
-        set_stopping: bool,
-    ) -> Result<(TimelineOrOffloaded, DeletionGuard), DeleteTimelineError> {
-        // Note the interaction between this guard and deletion guard.
-        // Here we attempt to lock deletion guard when we're holding a lock on timelines.
-        // This is important because when you take into account `remove_timeline_from_tenant`
-        // we remove timeline from memory when we still hold the deletion guard.
-        // So here when timeline deletion is finished timeline wont be present in timelines map at all
-        // which makes the following sequence impossible:
-        // T1: get preempted right before the try_lock on `Timeline::delete_progress`
-        // T2: do a full deletion, acquire and drop `Timeline::delete_progress`
-        // T1: acquire deletion lock, do another `DeleteTimelineFlow::run`
-        // For more context see this discussion: `https://github.com/neondatabase/neon/pull/4552#discussion_r1253437346`
-        let timelines = tenant.timelines.lock().unwrap();
-        let timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
-
-        let timeline = match timelines.get(&timeline_id) {
-            Some(t) => TimelineOrOffloaded::Timeline(Arc::clone(t)),
-            None => match timelines_offloaded.get(&timeline_id) {
-                Some(t) => TimelineOrOffloaded::Offloaded(Arc::clone(t)),
-                None => return Err(DeleteTimelineError::NotFound),
-            },
-        };
-
-        // Ensure that there are no child timelines, because we are about to remove files,
-        // which will break child branches
-        let mut children = Vec::new();
-        if !allow_offloaded_children {
-            children.extend(timelines_offloaded.iter().filter_map(|(id, entry)| {
-                (entry.ancestor_timeline_id == Some(timeline_id)).then_some(*id)
-            }));
-        }
-        children.extend(timelines.iter().filter_map(|(id, entry)| {
-            (entry.get_ancestor_timeline_id() == Some(timeline_id)).then_some(*id)
-        }));
-
-        if !children.is_empty() {
-            return Err(DeleteTimelineError::HasChildren(children));
-        }
-
-        // Note that using try_lock here is important to avoid a deadlock.
-        // Here we take lock on timelines and then the deletion guard.
-        // At the end of the operation we're holding the guard and need to lock timelines map
-        // to remove the timeline from it.
-        // Always if you have two locks that are taken in different order this can result in a deadlock.
-
-        let delete_progress = Arc::clone(timeline.delete_progress());
-        let delete_lock_guard = match delete_progress.try_lock_owned() {
-            Ok(guard) => DeletionGuard(guard),
-            Err(_) => {
-                // Unfortunately if lock fails arc is consumed.
-                return Err(DeleteTimelineError::AlreadyInProgress(Arc::clone(
-                    timeline.delete_progress(),
-                )));
-            }
-        };
-
-        if set_stopping {
-            if let TimelineOrOffloaded::Timeline(timeline) = &timeline {
-                timeline.set_state(TimelineState::Stopping);
-            }
-        }
-
-        Ok((timeline, delete_lock_guard))
-    }
-
     fn schedule_background(
         guard: DeletionGuard,
         conf: &'static PageServerConf,
@@ -483,6 +412,80 @@ impl DeleteTimelineFlow {
     }
 }
 
+#[derive(Copy, Clone, PartialEq, Eq)]
+pub(super) enum TimelineDeleteGuardKind {
+    Offload,
+    Delete,
+}
+
+pub(super) fn make_timeline_delete_guard(
+    tenant: &Tenant,
+    timeline_id: TimelineId,
+    guard_kind: TimelineDeleteGuardKind,
+) -> Result<(TimelineOrOffloaded, DeletionGuard), DeleteTimelineError> {
+    // Note the interaction between this guard and deletion guard.
+    // Here we attempt to lock deletion guard when we're holding a lock on timelines.
+    // This is important because when you take into account `remove_timeline_from_tenant`
+    // we remove timeline from memory when we still hold the deletion guard.
+    // So here when timeline deletion is finished timeline wont be present in timelines map at all
+    // which makes the following sequence impossible:
+    // T1: get preempted right before the try_lock on `Timeline::delete_progress`
+    // T2: do a full deletion, acquire and drop `Timeline::delete_progress`
+    // T1: acquire deletion lock, do another `DeleteTimelineFlow::run`
+    // For more context see this discussion: `https://github.com/neondatabase/neon/pull/4552#discussion_r1253437346`
+    let timelines = tenant.timelines.lock().unwrap();
+    let timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
+
+    let timeline = match timelines.get(&timeline_id) {
+        Some(t) => TimelineOrOffloaded::Timeline(Arc::clone(t)),
+        None => match timelines_offloaded.get(&timeline_id) {
+            Some(t) => TimelineOrOffloaded::Offloaded(Arc::clone(t)),
+            None => return Err(DeleteTimelineError::NotFound),
+        },
+    };
+
+    // Ensure that there are no child timelines, because we are about to remove files,
+    // which will break child branches
+    let mut children = Vec::new();
+    if guard_kind == TimelineDeleteGuardKind::Delete {
+        children.extend(timelines_offloaded.iter().filter_map(|(id, entry)| {
+            (entry.ancestor_timeline_id == Some(timeline_id)).then_some(*id)
+        }));
+    }
+    children.extend(timelines.iter().filter_map(|(id, entry)| {
+        (entry.get_ancestor_timeline_id() == Some(timeline_id)).then_some(*id)
+    }));
+
+    if !children.is_empty() {
+        return Err(DeleteTimelineError::HasChildren(children));
+    }
+
+    // Note that using try_lock here is important to avoid a deadlock.
+    // Here we take lock on timelines and then the deletion guard.
+    // At the end of the operation we're holding the guard and need to lock timelines map
+    // to remove the timeline from it.
+    // Always if you have two locks that are taken in different order this can result in a deadlock.
+
+    let delete_progress = Arc::clone(timeline.delete_progress());
+    let delete_lock_guard = match delete_progress.try_lock_owned() {
+        Ok(guard) => DeletionGuard(guard),
+        Err(_) => {
+            // Unfortunately if lock fails arc is consumed.
+            return Err(DeleteTimelineError::AlreadyInProgress(Arc::clone(
+                timeline.delete_progress(),
+            )));
+        }
+    };
+
+    if guard_kind == TimelineDeleteGuardKind::Delete {
+        if let TimelineOrOffloaded::Timeline(timeline) = &timeline {
+            timeline.set_state(TimelineState::Stopping);
+        }
+    }
+
+    Ok((timeline, delete_lock_guard))
+}
+
 pub(super) struct DeletionGuard(OwnedMutexGuard<DeleteTimelineFlow>);
 
 impl Deref for DeletionGuard {

pageserver/src/tenant/timeline/handle.rs

@@ -588,32 +588,40 @@ impl<T: Types> Drop for Cache<T> {
             let Some(handle_inner_arc) = handle_inner_weak.upgrade() else {
                 continue;
             };
-            let handle_timeline = handle_inner_arc
+            let Some(handle_timeline) = handle_inner_arc
                 // locking rules: drop lock before acquiring other lock below
                 .lock()
                 .expect("poisoned")
-                .shutdown();
+                .shutdown()
+            else {
+                // Concurrent PerTimelineState::shutdown.
+                continue;
+            };
+            // Clean up per_timeline_state so the HandleInner allocation can be dropped.
             let per_timeline_state = handle_timeline.per_timeline_state();
             let mut handles_lock_guard = per_timeline_state.handles.lock().expect("mutex poisoned");
             let Some(handles) = &mut *handles_lock_guard else {
                 continue;
             };
             let Some(removed_handle_inner_arc) = handles.remove(&self.id) else {
-                // There could have been a shutdown inbetween us upgrading the weak and locking the mutex.
+                // Concurrent PerTimelineState::shutdown.
                 continue;
             };
-            drop(handles_lock_guard); // locking rules: remember them when!
-            assert!(Arc::ptr_eq(&removed_handle_inner_arc, &handle_inner_arc,));
+            drop(handles_lock_guard); // locking rules!
+            assert!(Arc::ptr_eq(&removed_handle_inner_arc, &handle_inner_arc));
         }
     }
 }
 
 impl<T: Types> HandleInner<T> {
-    fn shutdown(&mut self) -> Arc<T::Timeline> {
+    fn shutdown(&mut self) -> Option<Arc<T::Timeline>> {
         match std::mem::replace(self, HandleInner::ShutDown) {
-            HandleInner::KeepingTimelineGateOpen { timeline, .. } => timeline,
+            HandleInner::KeepingTimelineGateOpen { timeline, .. } => Some(timeline),
             HandleInner::ShutDown => {
-                unreachable!("handles are only shut down once in their lifetime");
+                // Duplicate shutdowns are possible because both Cache::drop and PerTimelineState::shutdown
+                // may do it concurrently, but locking rules disallow holding per-timeline-state lock and
+                // the handle lock at the same time.
+                None
             }
         }
     }
@@ -1132,4 +1140,101 @@ mod tests {
         // There should be no strong references to the timeline object except the one on "stack".
         assert_eq!(Arc::strong_count(&shard0), refcount_start);
     }
+
+    #[tokio::test(start_paused = true)]
+    async fn test_reference_cycle_broken_when_cache_is_dropped() {
+        crate::tenant::harness::setup_logging();
+        let timeline_id = TimelineId::generate();
+        let shard0 = Arc::new_cyclic(|myself| StubTimeline {
+            gate: Default::default(),
+            id: timeline_id,
+            shard: ShardIdentity::unsharded(),
+            per_timeline_state: PerTimelineState::default(),
+            myself: myself.clone(),
+        });
+        let mgr = StubManager {
+            shards: vec![shard0.clone()],
+        };
+        let key = DBDIR_KEY;
+
+        let mut cache = Cache::<TestTypes>::default();
+
+        // helper to check if a handle is referenced by per_timeline_state
+        let per_timeline_state_refs_handle = |handle_weak: &Weak<Mutex<HandleInner<_>>>| {
+            let per_timeline_state = shard0.per_timeline_state.handles.lock().unwrap();
+            let per_timeline_state = per_timeline_state.as_ref().unwrap();
+            per_timeline_state
+                .values()
+                .any(|v| Weak::ptr_eq(&Arc::downgrade(v), handle_weak))
+        };
+
+        // Fill the cache.
+        let handle = cache
+            .get(timeline_id, ShardSelector::Page(key), &mgr)
+            .await
+            .expect("we have the timeline");
+        assert!(Weak::ptr_eq(&handle.myself, &shard0.myself));
+        let handle_inner_weak = Arc::downgrade(&handle.inner);
+        assert!(
+            per_timeline_state_refs_handle(&handle_inner_weak),
+            "we still hold `handle` _and_ haven't dropped `cache` yet"
+        );
+
+        // Drop the cache.
+        drop(cache);
+
+        assert!(
+            !(per_timeline_state_refs_handle(&handle_inner_weak)),
+            "nothing should reference the handle allocation anymore"
+        );
+        assert!(
+            Weak::upgrade(&handle_inner_weak).is_some(),
+            "the local `handle` still keeps the allocation alive"
+        );
+        // but obviously the cache is gone so no new allocations can be handed out.
+
+        // Drop handle.
+        drop(handle);
+        assert!(
+            Weak::upgrade(&handle_inner_weak).is_none(),
+            "the local `handle` is dropped, so the allocation should be dropped by now"
+        );
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn test_reference_cycle_broken_when_per_timeline_state_shutdown() {
+        crate::tenant::harness::setup_logging();
+        let timeline_id = TimelineId::generate();
+        let shard0 = Arc::new_cyclic(|myself| StubTimeline {
+            gate: Default::default(),
+            id: timeline_id,
+            shard: ShardIdentity::unsharded(),
+            per_timeline_state: PerTimelineState::default(),
+            myself: myself.clone(),
+        });
+        let mgr = StubManager {
+            shards: vec![shard0.clone()],
+        };
+        let key = DBDIR_KEY;
+
+        let mut cache = Cache::<TestTypes>::default();
+        let handle = cache
+            .get(timeline_id, ShardSelector::Page(key), &mgr)
+            .await
+            .expect("we have the timeline");
+        // grab a weak reference to the inner so can later try to Weak::upgrade it and assert that fails
+        let handle_inner_weak = Arc::downgrade(&handle.inner);
+
+        // drop the handle, obviously the lifetime of `inner` is at least as long as each strong reference to it
+        drop(handle);
+        assert!(Weak::upgrade(&handle_inner_weak).is_some(), "can still");
+
+        // Shutdown the per_timeline_state.
+        shard0.per_timeline_state.shutdown();
+        assert!(Weak::upgrade(&handle_inner_weak).is_none(), "can no longer");
+
+        // cache only contains Weak's, so, it can outlive the per_timeline_state without
+        // Drop explicitly solely to make this point.
+        drop(cache);
+    }
 }

pageserver/src/tenant/timeline/offload.rs

@@ -2,10 +2,11 @@ use std::sync::Arc;
 
 use pageserver_api::models::{TenantState, TimelineState};
 
-use super::delete::{delete_local_timeline_directory, DeleteTimelineFlow, DeletionGuard};
+use super::delete::{delete_local_timeline_directory, DeletionGuard};
 use super::Timeline;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::remote_timeline_client::ShutdownIfArchivedError;
+use crate::tenant::timeline::delete::{make_timeline_delete_guard, TimelineDeleteGuardKind};
 use crate::tenant::{OffloadedTimeline, Tenant, TenantManifestError, TimelineOrOffloaded};
 
 #[derive(thiserror::Error, Debug)]
@@ -36,13 +37,10 @@ pub(crate) async fn offload_timeline(
     debug_assert_current_span_has_tenant_and_timeline_id();
     tracing::info!("offloading archived timeline");
 
-    let allow_offloaded_children = true;
-    let set_stopping = false;
-    let (timeline, guard) = DeleteTimelineFlow::prepare(
+    let (timeline, guard) = make_timeline_delete_guard(
         tenant,
         timeline.timeline_id,
-        allow_offloaded_children,
-        set_stopping,
+        TimelineDeleteGuardKind::Offload,
     )
     .map_err(|e| OffloadError::Other(anyhow::anyhow!(e)))?;
 
@@ -106,7 +104,7 @@ pub(crate) async fn offload_timeline(
 }
 
 /// It is important that this gets called when DeletionGuard is being held.
-/// For more context see comments in [`DeleteTimelineFlow::prepare`]
+/// For more context see comments in [`make_timeline_delete_guard`]
 ///
 /// Returns the strong count of the timeline `Arc`
 fn remove_timeline_from_tenant(

pageserver/src/tenant/vectored_blob_io.rs

@@ -35,6 +35,7 @@ use crate::virtual_file::{self, VirtualFile};
 pub struct BlobMeta {
     pub key: Key,
     pub lsn: Lsn,
+    pub will_init: bool,
 }
 
 /// A view into the vectored blobs read buffer.
@@ -310,7 +311,15 @@ pub enum BlobFlag {
 /// * Iterate over the collected blobs and coalesce them into reads at the end
 pub struct VectoredReadPlanner {
     // Track all the blob offsets. Start offsets must be ordered.
-    blobs: BTreeMap<Key, Vec<(Lsn, u64, u64)>>,
+    // Values in the value tuples are:
+    // (
+    //   lsn of the blob,
+    //   start offset of the blob in the underlying file,
+    //   end offset of the blob in the underlying file,
+    //   whether the blob initializes the page image or not
+    //   see [`pageserver_api::record::NeonWalRecord::will_init`]
+    // )
+    blobs: BTreeMap<Key, Vec<(Lsn, u64, u64, bool)>>,
     // Arguments for previous blob passed into [`VectoredReadPlanner::handle`]
     prev: Option<(Key, Lsn, u64, BlobFlag)>,
 
@@ -371,12 +380,12 @@ impl VectoredReadPlanner {
         match flag {
             BlobFlag::None => {
                 let blobs_for_key = self.blobs.entry(key).or_default();
-                blobs_for_key.push((lsn, start_offset, end_offset));
+                blobs_for_key.push((lsn, start_offset, end_offset, false));
             }
             BlobFlag::ReplaceAll => {
                 let blobs_for_key = self.blobs.entry(key).or_default();
                 blobs_for_key.clear();
-                blobs_for_key.push((lsn, start_offset, end_offset));
+                blobs_for_key.push((lsn, start_offset, end_offset, true));
             }
             BlobFlag::Ignore => {}
         }
@@ -387,11 +396,17 @@ impl VectoredReadPlanner {
         let mut reads = Vec::new();
 
         for (key, blobs_for_key) in self.blobs {
-            for (lsn, start_offset, end_offset) in blobs_for_key {
+            for (lsn, start_offset, end_offset, will_init) in blobs_for_key {
                 let extended = match &mut current_read_builder {
-                    Some(read_builder) => {
-                        read_builder.extend(start_offset, end_offset, BlobMeta { key, lsn })
-                    }
+                    Some(read_builder) => read_builder.extend(
+                        start_offset,
+                        end_offset,
+                        BlobMeta {
+                            key,
+                            lsn,
+                            will_init,
+                        },
+                    ),
                     None => VectoredReadExtended::No,
                 };
 
@@ -399,7 +414,11 @@ impl VectoredReadPlanner {
                     let next_read_builder = ChunkedVectoredReadBuilder::new(
                         start_offset,
                         end_offset,
-                        BlobMeta { key, lsn },
+                        BlobMeta {
+                            key,
+                            lsn,
+                            will_init,
+                        },
                         self.max_read_size,
                     );
 
@@ -527,7 +546,7 @@ impl<'a> VectoredBlobReader<'a> {
 pub struct StreamingVectoredReadPlanner {
     read_builder: Option<ChunkedVectoredReadBuilder>,
     // Arguments for previous blob passed into [`StreamingVectoredReadPlanner::handle`]
-    prev: Option<(Key, Lsn, u64)>,
+    prev: Option<(Key, Lsn, u64, bool)>,
     /// Max read size per batch. This is not a strict limit. If there are [0, 100) and [100, 200), while the `max_read_size` is 150,
     /// we will produce a single batch instead of split them.
     max_read_size: u64,
@@ -550,27 +569,47 @@ impl StreamingVectoredReadPlanner {
         }
     }
 
-    pub fn handle(&mut self, key: Key, lsn: Lsn, offset: u64) -> Option<VectoredRead> {
+    pub fn handle(
+        &mut self,
+        key: Key,
+        lsn: Lsn,
+        offset: u64,
+        will_init: bool,
+    ) -> Option<VectoredRead> {
         // Implementation note: internally lag behind by one blob such that
         // we have a start and end offset when initialising [`VectoredRead`]
-        let (prev_key, prev_lsn, prev_offset) = match self.prev {
+        let (prev_key, prev_lsn, prev_offset, prev_will_init) = match self.prev {
             None => {
-                self.prev = Some((key, lsn, offset));
+                self.prev = Some((key, lsn, offset, will_init));
                 return None;
             }
             Some(prev) => prev,
         };
 
-        let res = self.add_blob(prev_key, prev_lsn, prev_offset, offset, false);
+        let res = self.add_blob(
+            prev_key,
+            prev_lsn,
+            prev_offset,
+            offset,
+            false,
+            prev_will_init,
+        );
 
-        self.prev = Some((key, lsn, offset));
+        self.prev = Some((key, lsn, offset, will_init));
 
         res
     }
 
     pub fn handle_range_end(&mut self, offset: u64) -> Option<VectoredRead> {
-        let res = if let Some((prev_key, prev_lsn, prev_offset)) = self.prev {
-            self.add_blob(prev_key, prev_lsn, prev_offset, offset, true)
+        let res = if let Some((prev_key, prev_lsn, prev_offset, prev_will_init)) = self.prev {
+            self.add_blob(
+                prev_key,
+                prev_lsn,
+                prev_offset,
+                offset,
+                true,
+                prev_will_init,
+            )
         } else {
             None
         };
@@ -587,10 +626,19 @@ impl StreamingVectoredReadPlanner {
         start_offset: u64,
         end_offset: u64,
         is_last_blob_in_read: bool,
+        will_init: bool,
     ) -> Option<VectoredRead> {
         match &mut self.read_builder {
             Some(read_builder) => {
-                let extended = read_builder.extend(start_offset, end_offset, BlobMeta { key, lsn });
+                let extended = read_builder.extend(
+                    start_offset,
+                    end_offset,
+                    BlobMeta {
+                        key,
+                        lsn,
+                        will_init,
+                    },
+                );
                 assert_eq!(extended, VectoredReadExtended::Yes);
             }
             None => {
@@ -598,7 +646,11 @@ impl StreamingVectoredReadPlanner {
                     Some(ChunkedVectoredReadBuilder::new_streaming(
                         start_offset,
                         end_offset,
-                        BlobMeta { key, lsn },
+                        BlobMeta {
+                            key,
+                            lsn,
+                            will_init,
+                        },
                     ))
                 };
             }
@@ -812,7 +864,7 @@ mod tests {
         let mut planner = StreamingVectoredReadPlanner::new(max_read_size, 1000);
         let mut reads = Vec::new();
         for (key, lsn, offset, _) in blob_descriptions.clone() {
-            reads.extend(planner.handle(key, lsn, offset));
+            reads.extend(planner.handle(key, lsn, offset, false));
         }
         reads.extend(planner.handle_range_end(652 * 1024));
 
@@ -850,7 +902,7 @@ mod tests {
         let mut planner = StreamingVectoredReadPlanner::new(max_read_size, 2);
         let mut reads = Vec::new();
         for (key, lsn, offset, _) in blob_descriptions.clone() {
-            reads.extend(planner.handle(key, lsn, offset));
+            reads.extend(planner.handle(key, lsn, offset, false));
         }
         reads.extend(planner.handle_range_end(652 * 1024));
 
@@ -875,7 +927,7 @@ mod tests {
         {
             let mut planner = StreamingVectoredReadPlanner::new(max_read_size, 1);
             let mut reads = Vec::new();
-            reads.extend(planner.handle(key, lsn, 0));
+            reads.extend(planner.handle(key, lsn, 0, false));
             reads.extend(planner.handle_range_end(652 * 1024));
             assert_eq!(reads.len(), 1);
             validate_read(&reads[0], &[(key, lsn, 0, BlobFlag::None)]);
@@ -883,8 +935,8 @@ mod tests {
         {
             let mut planner = StreamingVectoredReadPlanner::new(max_read_size, 1);
             let mut reads = Vec::new();
-            reads.extend(planner.handle(key, lsn, 0));
-            reads.extend(planner.handle(key, lsn, 128 * 1024));
+            reads.extend(planner.handle(key, lsn, 0, false));
+            reads.extend(planner.handle(key, lsn, 128 * 1024, false));
             reads.extend(planner.handle_range_end(652 * 1024));
             assert_eq!(reads.len(), 2);
             validate_read(&reads[0], &[(key, lsn, 0, BlobFlag::None)]);
@@ -893,8 +945,8 @@ mod tests {
         {
             let mut planner = StreamingVectoredReadPlanner::new(max_read_size, 2);
             let mut reads = Vec::new();
-            reads.extend(planner.handle(key, lsn, 0));
-            reads.extend(planner.handle(key, lsn, 128 * 1024));
+            reads.extend(planner.handle(key, lsn, 0, false));
+            reads.extend(planner.handle(key, lsn, 128 * 1024, false));
             reads.extend(planner.handle_range_end(652 * 1024));
             assert_eq!(reads.len(), 1);
             validate_read(
@@ -923,6 +975,7 @@ mod tests {
         let meta = BlobMeta {
             key: Key::MIN,
             lsn: Lsn(0),
+            will_init: false,
         };
 
         for (idx, (blob, offset)) in blobs.iter().zip(offsets.iter()).enumerate() {

pageserver/src/walingest.rs

@@ -499,7 +499,13 @@ impl WalIngest {
 
                 let content = modification
                     .tline
-                    .get_rel_page_at_lsn(src_rel, blknum, Version::Modified(modification), ctx)
+                    .get_rel_page_at_lsn(
+                        src_rel,
+                        blknum,
+                        Version::Modified(modification),
+                        ctx,
+                        crate::tenant::storage_layer::IoConcurrency::sequential(),
+                    )
                     .await?;
                 modification.put_rel_page_image(dst_rel, blknum, content)?;
                 num_blocks_copied += 1;
@@ -1489,6 +1495,7 @@ mod tests {
     use super::*;
     use crate::tenant::harness::*;
     use crate::tenant::remote_timeline_client::{remote_initdb_archive_path, INITDB_PATH};
+    use crate::tenant::storage_layer::IoConcurrency;
     use postgres_ffi::RELSEG_SIZE;
 
     use crate::DEFAULT_PG_VERSION;
@@ -1532,6 +1539,7 @@ mod tests {
     #[tokio::test]
     async fn test_relsize() -> Result<()> {
         let (tenant, ctx) = TenantHarness::create("test_relsize").await?.load().await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
         let tline = tenant
             .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
             .await?;
@@ -1599,7 +1607,13 @@ mod tests {
         // Check page contents at each LSN
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x20)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    0,
+                    Version::Lsn(Lsn(0x20)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 2")
@@ -1607,7 +1621,13 @@ mod tests {
 
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x30)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    0,
+                    Version::Lsn(Lsn(0x30)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 3")
@@ -1615,14 +1635,26 @@ mod tests {
 
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x40)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    0,
+                    Version::Lsn(Lsn(0x40)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 3")
         );
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x40)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    1,
+                    Version::Lsn(Lsn(0x40)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1 at 4")
@@ -1630,21 +1662,39 @@ mod tests {
 
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x50)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    0,
+                    Version::Lsn(Lsn(0x50)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 3")
         );
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x50)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    1,
+                    Version::Lsn(Lsn(0x50)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1 at 4")
         );
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 2, Version::Lsn(Lsn(0x50)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    2,
+                    Version::Lsn(Lsn(0x50)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 2 at 5")
@@ -1667,14 +1717,26 @@ mod tests {
         );
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x60)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    0,
+                    Version::Lsn(Lsn(0x60)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 3")
         );
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x60)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    1,
+                    Version::Lsn(Lsn(0x60)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1 at 4")
@@ -1689,7 +1751,13 @@ mod tests {
         );
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 2, Version::Lsn(Lsn(0x50)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    2,
+                    Version::Lsn(Lsn(0x50)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 2 at 5")
@@ -1722,14 +1790,26 @@ mod tests {
         );
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x70)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    0,
+                    Version::Lsn(Lsn(0x70)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             ZERO_PAGE
         );
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x70)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    1,
+                    Version::Lsn(Lsn(0x70)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1")
@@ -1750,7 +1830,13 @@ mod tests {
         for blk in 2..1500 {
             assert_eq!(
                 tline
-                    .get_rel_page_at_lsn(TESTREL_A, blk, Version::Lsn(Lsn(0x80)), &ctx)
+                    .get_rel_page_at_lsn(
+                        TESTREL_A,
+                        blk,
+                        Version::Lsn(Lsn(0x80)),
+                        &ctx,
+                        io_concurrency.clone()
+                    )
                     .instrument(test_span.clone())
                     .await?,
                 ZERO_PAGE
@@ -1758,7 +1844,13 @@ mod tests {
         }
         assert_eq!(
             tline
-                .get_rel_page_at_lsn(TESTREL_A, 1500, Version::Lsn(Lsn(0x80)), &ctx)
+                .get_rel_page_at_lsn(
+                    TESTREL_A,
+                    1500,
+                    Version::Lsn(Lsn(0x80)),
+                    &ctx,
+                    io_concurrency.clone()
+                )
                 .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1500")
@@ -1851,6 +1943,7 @@ mod tests {
             .await?
             .load()
             .await;
+        let io_concurrency = IoConcurrency::spawn_for_test();
         let tline = tenant
             .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
             .await?;
@@ -1903,7 +1996,13 @@ mod tests {
             let data = format!("foo blk {} at {}", blkno, lsn);
             assert_eq!(
                 tline
-                    .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(lsn), &ctx)
+                    .get_rel_page_at_lsn(
+                        TESTREL_A,
+                        blkno,
+                        Version::Lsn(lsn),
+                        &ctx,
+                        io_concurrency.clone()
+                    )
                     .instrument(test_span.clone())
                     .await?,
                 test_img(&data)
@@ -1931,7 +2030,13 @@ mod tests {
             let data = format!("foo blk {} at {}", blkno, lsn);
             assert_eq!(
                 tline
-                    .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(Lsn(0x60)), &ctx)
+                    .get_rel_page_at_lsn(
+                        TESTREL_A,
+                        blkno,
+                        Version::Lsn(Lsn(0x60)),
+                        &ctx,
+                        io_concurrency.clone()
+                    )
                     .instrument(test_span.clone())
                     .await?,
                 test_img(&data)
@@ -1950,7 +2055,13 @@ mod tests {
             let data = format!("foo blk {} at {}", blkno, lsn);
             assert_eq!(
                 tline
-                    .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(Lsn(0x50)), &ctx)
+                    .get_rel_page_at_lsn(
+                        TESTREL_A,
+                        blkno,
+                        Version::Lsn(Lsn(0x50)),
+                        &ctx,
+                        io_concurrency.clone()
+                    )
                     .instrument(test_span.clone())
                     .await?,
                 test_img(&data)
@@ -1987,7 +2098,13 @@ mod tests {
             let data = format!("foo blk {} at {}", blkno, lsn);
             assert_eq!(
                 tline
-                    .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(Lsn(0x80)), &ctx)
+                    .get_rel_page_at_lsn(
+                        TESTREL_A,
+                        blkno,
+                        Version::Lsn(Lsn(0x80)),
+                        &ctx,
+                        io_concurrency.clone()
+                    )
                     .instrument(test_span.clone())
                     .await?,
                 test_img(&data)

pgxn/neon/file_cache.c

@@ -911,57 +911,85 @@ lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 			if (entry->access_count++ == 0)
 				dlist_delete(&entry->list_node);
 		}
-		else
+		/*-----------
+		 * If the chunk wasn't already in the LFC then we have these
+		 * options, in order of preference:
+		 *
+		 * Unless there is no space available, we can:
+		 *  1. Use an entry from the `holes` list, and
+		 *  2. Create a new entry.
+		 * We can always, regardless of space in the LFC:
+		 *  3. evict an entry from LRU, and
+		 *  4. ignore the write operation (the least favorite option)
+		 */
+		else if (lfc_ctl->used < lfc_ctl->limit)
 		{
-			/*
-			 * We have two choices if all cache pages are pinned (i.e. used in IO
-			 * operations):
-			 *
-			 * 1) Wait until some of this operation is completed and pages is
-			 * unpinned.
-			 *
-			 * 2) Allocate one more chunk, so that specified cache size is more
-			 * recommendation than hard limit.
-			 *
-			 * As far as probability of such event (that all pages are pinned) is
-			 * considered to be very very small: there are should be very large
-			 * number of concurrent IO operations and them are limited by
-			 * max_connections, we prefer not to complicate code and use second
-			 * approach.
-			 */
-			if (lfc_ctl->used >= lfc_ctl->limit && !dlist_is_empty(&lfc_ctl->lru))
-			{
-				/* Cache overflow: evict least recently used chunk */
-				FileCacheEntry *victim = dlist_container(FileCacheEntry, list_node, dlist_pop_head_node(&lfc_ctl->lru));
-	
-				for (int i = 0; i < BLOCKS_PER_CHUNK; i++)
-				{
-					lfc_ctl->used_pages -= (victim->bitmap[i >> 5] >> (i & 31)) & 1;
-				}
-				CriticalAssert(victim->access_count == 0);
-				entry->offset = victim->offset; /* grab victim's chunk */
-				hash_search_with_hash_value(lfc_hash, &victim->key, victim->hash, HASH_REMOVE, NULL);
-				neon_log(DEBUG2, "Swap file cache page");
-			}
-			else if (!dlist_is_empty(&lfc_ctl->holes))
+			if (!dlist_is_empty(&lfc_ctl->holes))
 			{
 				/* We can reuse a hole that was left behind when the LFC was shrunk previously */
-				FileCacheEntry *hole = dlist_container(FileCacheEntry, list_node, dlist_pop_head_node(&lfc_ctl->holes));
-				uint32		offset = hole->offset;
-				bool		hole_found;
-	
-				hash_search_with_hash_value(lfc_hash, &hole->key, hole->hash, HASH_REMOVE, &hole_found);
+				FileCacheEntry *hole = dlist_container(FileCacheEntry, list_node,
+													   dlist_pop_head_node(&lfc_ctl->holes));
+				uint32 offset = hole->offset;
+				bool hole_found;
+
+				hash_search_with_hash_value(lfc_hash, &hole->key,
+											hole->hash, HASH_REMOVE, &hole_found);
 				CriticalAssert(hole_found);
-	
+
 				lfc_ctl->used += 1;
-				entry->offset = offset;	/* reuse the hole */
+				entry->offset = offset;			/* reuse the hole */
 			}
 			else
 			{
 				lfc_ctl->used += 1;
-				entry->offset = lfc_ctl->size++;	/* allocate new chunk at end
-													 * of file */
+				entry->offset = lfc_ctl->size++;/* allocate new chunk at end
+												 * of file */
 			}
+		}
+		/*
+		 * We've already used up all allocated LFC entries.
+		 *
+		 * If we can clear an entry from the LRU, do that.
+		 * If we can't (e.g. because all other slots are being accessed)
+		 * then we will remove this entry from the hash and continue
+		 * on to the next chunk, as we may not exceed the limit.
+		 */
+		else if (!dlist_is_empty(&lfc_ctl->lru))
+		{
+			/* Cache overflow: evict least recently used chunk */
+			FileCacheEntry *victim = dlist_container(FileCacheEntry, list_node,
+													 dlist_pop_head_node(&lfc_ctl->lru));
+
+			for (int i = 0; i < BLOCKS_PER_CHUNK; i++)
+			{
+				lfc_ctl->used_pages -= (victim->bitmap[i >> 5] >> (i & 31)) & 1;
+			}
+
+			CriticalAssert(victim->access_count == 0);
+			entry->offset = victim->offset; /* grab victim's chunk */
+			hash_search_with_hash_value(lfc_hash, &victim->key,
+										victim->hash, HASH_REMOVE, NULL);
+			neon_log(DEBUG2, "Swap file cache page");
+		}
+		else
+		{
+			/* Can't add this chunk - we don't have the space for it */
+			hash_search_with_hash_value(lfc_hash, &entry->key, hash,
+										HASH_REMOVE, NULL);
+
+			/*
+			 * We can't process this chunk due to lack of space in LFC,
+			 * so skip to the next one
+			 */
+			LWLockRelease(lfc_lock);
+			blkno += blocks_in_chunk;
+			buf_offset += blocks_in_chunk;
+			nblocks -= blocks_in_chunk;
+			continue;
+		}
+
+		if (!found)
+		{
 			entry->access_count = 1;
 			entry->hash = hash;
 			memset(entry->bitmap, 0, sizeof entry->bitmap);

poetry.lock

@@ -395,6 +395,7 @@ files = [
 
 [package.dependencies]
 botocore-stubs = "*"
+mypy-boto3-kms = {version = ">=1.26.0,<1.27.0", optional = true, markers = "extra == \"kms\""}
 mypy-boto3-s3 = {version = ">=1.26.0,<1.27.0", optional = true, markers = "extra == \"s3\""}
 types-s3transfer = "*"
 typing-extensions = ">=4.1.0"
@@ -1948,6 +1949,17 @@ install-types = ["pip"]
 mypyc = ["setuptools (>=50)"]
 reports = ["lxml"]
 
+[[package]]
+name = "mypy-boto3-kms"
+version = "1.26.147"
+description = "Type annotations for boto3.KMS 1.26.147 service generated with mypy-boto3-builder 7.14.5"
+optional = false
+python-versions = ">=3.7"
+files = [
+    {file = "mypy-boto3-kms-1.26.147.tar.gz", hash = "sha256:816a4d1bb0585e1b9620a3f96c1d69a06f53b7b5621858579dd77c60dbb5fa5c"},
+    {file = "mypy_boto3_kms-1.26.147-py3-none-any.whl", hash = "sha256:493f0db674a25c88769f5cb8ab8ac00d3dda5dfc903d5cda34c990ee64689f79"},
+]
+
 [[package]]
 name = "mypy-boto3-s3"
 version = "1.26.0.post1"
@@ -2286,6 +2298,7 @@ files = [
     {file = "psycopg2_binary-2.9.10-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:bb89f0a835bcfc1d42ccd5f41f04870c1b936d8507c6df12b7737febc40f0909"},
     {file = "psycopg2_binary-2.9.10-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:f0c2d907a1e102526dd2986df638343388b94c33860ff3bbe1384130828714b1"},
     {file = "psycopg2_binary-2.9.10-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:f8157bed2f51db683f31306aa497311b560f2265998122abe1dce6428bd86567"},
+    {file = "psycopg2_binary-2.9.10-cp313-cp313-win_amd64.whl", hash = "sha256:27422aa5f11fbcd9b18da48373eb67081243662f9b46e6fd07c3eb46e4535142"},
     {file = "psycopg2_binary-2.9.10-cp38-cp38-macosx_12_0_x86_64.whl", hash = "sha256:eb09aa7f9cecb45027683bb55aebaaf45a0df8bf6de68801a6afdc7947bb09d4"},
     {file = "psycopg2_binary-2.9.10-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b73d6d7f0ccdad7bc43e6d34273f70d587ef62f824d7261c4ae9b8b1b6af90e8"},
     {file = "psycopg2_binary-2.9.10-cp38-cp38-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ce5ab4bf46a211a8e924d307c1b1fcda82368586a19d0a24f8ae166f5c784864"},
@@ -3386,6 +3399,16 @@ files = [
     {file = "wrapt-1.14.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:8ad85f7f4e20964db4daadcab70b47ab05c7c1cf2a7c1e51087bfaa83831854c"},
     {file = "wrapt-1.14.1-cp310-cp310-win32.whl", hash = "sha256:a9a52172be0b5aae932bef82a79ec0a0ce87288c7d132946d645eba03f0ad8a8"},
     {file = "wrapt-1.14.1-cp310-cp310-win_amd64.whl", hash = "sha256:6d323e1554b3d22cfc03cd3243b5bb815a51f5249fdcbb86fda4bf62bab9e164"},
+    {file = "wrapt-1.14.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ecee4132c6cd2ce5308e21672015ddfed1ff975ad0ac8d27168ea82e71413f55"},
+    {file = "wrapt-1.14.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:2020f391008ef874c6d9e208b24f28e31bcb85ccff4f335f15a3251d222b92d9"},
+    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2feecf86e1f7a86517cab34ae6c2f081fd2d0dac860cb0c0ded96d799d20b335"},
+    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:240b1686f38ae665d1b15475966fe0472f78e71b1b4903c143a842659c8e4cb9"},
+    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a9008dad07d71f68487c91e96579c8567c98ca4c3881b9b113bc7b33e9fd78b8"},
+    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:6447e9f3ba72f8e2b985a1da758767698efa72723d5b59accefd716e9e8272bf"},
+    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:acae32e13a4153809db37405f5eba5bac5fbe2e2ba61ab227926a22901051c0a"},
+    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:49ef582b7a1152ae2766557f0550a9fcbf7bbd76f43fbdc94dd3bf07cc7168be"},
+    {file = "wrapt-1.14.1-cp311-cp311-win32.whl", hash = "sha256:358fe87cc899c6bb0ddc185bf3dbfa4ba646f05b1b0b9b5a27c2cb92c2cea204"},
+    {file = "wrapt-1.14.1-cp311-cp311-win_amd64.whl", hash = "sha256:26046cd03936ae745a502abf44dac702a5e6880b2b01c29aea8ddf3353b68224"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux1_i686.whl", hash = "sha256:43ca3bbbe97af00f49efb06e352eae40434ca9d915906f77def219b88e85d907"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux1_x86_64.whl", hash = "sha256:6b1a564e6cb69922c7fe3a678b9f9a3c54e72b469875aa8018f18b4d1dd1adf3"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux2010_i686.whl", hash = "sha256:00b6d4ea20a906c0ca56d84f93065b398ab74b927a7a3dbd470f6fc503f95dc3"},
@@ -3655,4 +3678,4 @@ cffi = ["cffi (>=1.11)"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.11"
-content-hash = "e6904aca09abc6c805604b21a5702a97e0056406f9ec7469b091d35ee10a6b16"
+content-hash = "b25702660d965fa9f21db0b20fa07b545d199d8ccb303c67eece77b338b08494"

pyproject.toml

@@ -17,7 +17,7 @@ Jinja2 = "^3.1.5"
 types-requests = "^2.31.0.0"
 types-psycopg2 = "^2.9.21.20241019"
 boto3 = "^1.34.11"
-boto3-stubs = {extras = ["s3"], version = "^1.26.16"}
+boto3-stubs = {extras = ["s3", "kms"], version = "^1.26.16"}
 moto = {extras = ["server"], version = "^5.0.6"}
 backoff = "^2.2.1"
 pytest-lazy-fixture = "^0.6.3"

safekeeper/client/src/mgmt_api.rs

@@ -102,6 +102,11 @@ impl Client {
         self.get(&uri).await
     }
 
+    pub async fn utilization(&self) -> Result<reqwest::Response> {
+        let uri = format!("{}/v1/utilization/", self.mgmt_api_endpoint);
+        self.get(&uri).await
+    }
+
     async fn get<U: IntoUrl>(&self, uri: U) -> Result<reqwest::Response> {
         self.request(Method::GET, uri, ()).await
     }

safekeeper/src/http/routes.rs

@@ -127,6 +127,13 @@ async fn timeline_create_handler(mut request: Request<Body>) -> Result<Response<
     json_response(StatusCode::OK, ())
 }
 
+async fn utilization_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
+    check_permission(&request, None)?;
+    let global_timelines = get_global_timelines(&request);
+    let utilization = global_timelines.get_timeline_counts();
+    json_response(StatusCode::OK, utilization)
+}
+
 /// List all (not deleted) timelines.
 /// Note: it is possible to do the same with debug_dump.
 async fn timeline_list_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
@@ -620,6 +627,7 @@ pub fn make_router(
                 failpoints_handler(r, cancel).await
             })
         })
+        .get("/v1/uzilization", |r| request_span(r, utilization_handler))
         .delete("/v1/tenant/:tenant_id", |r| {
             request_span(r, tenant_delete_handler)
         })

safekeeper/src/timelines_global_map.rs

@@ -13,6 +13,7 @@ use anyhow::{bail, Context, Result};
 use camino::Utf8PathBuf;
 use camino_tempfile::Utf8TempDir;
 use safekeeper_api::membership::Configuration;
+use safekeeper_api::models::SafekeeperUtilization;
 use safekeeper_api::ServerInfo;
 use serde::Serialize;
 use std::collections::HashMap;
@@ -416,6 +417,20 @@ impl GlobalTimelines {
             .collect()
     }
 
+    /// Returns statistics about timeline counts
+    pub fn get_timeline_counts(&self) -> SafekeeperUtilization {
+        let global_lock = self.state.lock().unwrap();
+        let timeline_count = global_lock
+            .timelines
+            .values()
+            .filter(|t| match t {
+                GlobalMapTimeline::CreationInProgress => false,
+                GlobalMapTimeline::Timeline(t) => !t.is_cancelled(),
+            })
+            .count() as u64;
+        SafekeeperUtilization { timeline_count }
+    }
+
     /// Returns all timelines belonging to a given tenant. Used for deleting all timelines of a tenant,
     /// and that's why it can return cancelled timelines, to retry deleting them.
     fn get_all_for_tenant(&self, tenant_id: TenantId) -> Vec<Arc<Timeline>> {

storage_broker/src/bin/storage_broker.rs

@@ -32,7 +32,6 @@ use tokio::sync::broadcast::error::RecvError;
 use tokio::time;
 use tonic::body::{self, empty_body, BoxBody};
 use tonic::codegen::Service;
-use tonic::transport::server::Connected;
 use tonic::Code;
 use tonic::{Request, Response, Status};
 use tracing::*;
@@ -459,9 +458,10 @@ impl BrokerService for Broker {
         &self,
         request: Request<tonic::Streaming<SafekeeperTimelineInfo>>,
     ) -> Result<Response<()>, Status> {
-        let remote_addr = request
-            .remote_addr()
-            .expect("TCPConnectInfo inserted by handler");
+        let &RemoteAddr(remote_addr) = request
+            .extensions()
+            .get()
+            .expect("RemoteAddr inserted by handler");
         let mut publisher = self.registry.register_publisher(remote_addr);
 
         let mut stream = request.into_inner();
@@ -484,9 +484,10 @@ impl BrokerService for Broker {
         &self,
         request: Request<SubscribeSafekeeperInfoRequest>,
     ) -> Result<Response<Self::SubscribeSafekeeperInfoStream>, Status> {
-        let remote_addr = request
-            .remote_addr()
-            .expect("TCPConnectInfo inserted by handler");
+        let &RemoteAddr(remote_addr) = request
+            .extensions()
+            .get()
+            .expect("RemoteAddr inserted by handler");
         let proto_key = request
             .into_inner()
             .subscription_key
@@ -537,9 +538,10 @@ impl BrokerService for Broker {
         &self,
         request: Request<SubscribeByFilterRequest>,
     ) -> std::result::Result<Response<Self::SubscribeByFilterStream>, Status> {
-        let remote_addr = request
-            .remote_addr()
-            .expect("TCPConnectInfo inserted by handler");
+        let &RemoteAddr(remote_addr) = request
+            .extensions()
+            .get()
+            .expect("RemoteAddr inserted by handler");
         let proto_filter = request.into_inner();
         let ttid_filter = proto_filter.tenant_timeline_id.as_ref();
 
@@ -628,6 +630,9 @@ async fn http1_handler(
     Ok(resp)
 }
 
+#[derive(Clone, Copy)]
+struct RemoteAddr(SocketAddr);
+
 #[tokio::main]
 async fn main() -> Result<(), Box<dyn std::error::Error>> {
     let args = Args::parse();
@@ -687,13 +692,13 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
             .max_concurrent_streams(None);
 
         let storage_broker_server_cloned = storage_broker_server.clone();
-        let connect_info = stream.connect_info();
+        let remote_addr = RemoteAddr(addr);
         let service_fn_ = async move {
             service_fn(move |mut req| {
                 // That's what tonic's MakeSvc.call does to pass conninfo to
                 // the request handler (and where its request.remote_addr()
                 // expects it to find).
-                req.extensions_mut().insert(connect_info.clone());
+                req.extensions_mut().insert(remote_addr);
 
                 // Technically this second clone is not needed, but consume
                 // by async block is apparently unavoidable. BTW, error

storage_controller/src/service/chaos_injector.rs

@@ -1,11 +1,17 @@
-use std::{sync::Arc, time::Duration};
+use std::{
+    collections::{BTreeMap, HashMap},
+    sync::Arc,
+    time::Duration,
+};
 
 use pageserver_api::controller_api::ShardSchedulingPolicy;
 use rand::seq::SliceRandom;
 use rand::thread_rng;
 use tokio_util::sync::CancellationToken;
+use utils::id::NodeId;
+use utils::shard::TenantShardId;
 
-use super::Service;
+use super::{Node, Scheduler, Service, TenantShard};
 
 pub struct ChaosInjector {
     service: Arc<Service>,
@@ -35,50 +41,86 @@ impl ChaosInjector {
         }
     }
 
+    /// If a shard has a secondary and attached location, then re-assign the secondary to be
+    /// attached and the attached to be secondary.
+    ///
+    /// Only modifies tenants if they're in Active scheduling policy.
+    fn maybe_migrate_to_secondary(
+        &self,
+        tenant_shard_id: TenantShardId,
+        nodes: &Arc<HashMap<NodeId, Node>>,
+        tenants: &mut BTreeMap<TenantShardId, TenantShard>,
+        scheduler: &mut Scheduler,
+    ) {
+        let shard = tenants
+            .get_mut(&tenant_shard_id)
+            .expect("Held lock between choosing ID and this get");
+
+        if !matches!(shard.get_scheduling_policy(), ShardSchedulingPolicy::Active) {
+            // Skip non-active scheduling policies, so that a shard with a policy like Pause can
+            // be pinned without being disrupted by us.
+            tracing::info!(
+                "Skipping shard {tenant_shard_id}: scheduling policy is {:?}",
+                shard.get_scheduling_policy()
+            );
+            return;
+        }
+
+        // Pick a secondary to promote
+        let Some(new_location) = shard
+            .intent
+            .get_secondary()
+            .choose(&mut thread_rng())
+            .cloned()
+        else {
+            tracing::info!(
+                "Skipping shard {tenant_shard_id}: no secondary location, can't migrate"
+            );
+            return;
+        };
+
+        let Some(old_location) = *shard.intent.get_attached() else {
+            tracing::info!("Skipping shard {tenant_shard_id}: currently has no attached location");
+            return;
+        };
+
+        tracing::info!("Injecting chaos: migrate {tenant_shard_id} {old_location}->{new_location}");
+
+        shard.intent.demote_attached(scheduler, old_location);
+        shard.intent.promote_attached(scheduler, new_location);
+        self.service.maybe_reconcile_shard(shard, nodes);
+    }
+
     async fn inject_chaos(&mut self) {
         // Pick some shards to interfere with
         let batch_size = 128;
         let mut inner = self.service.inner.write().unwrap();
         let (nodes, tenants, scheduler) = inner.parts_mut();
         let tenant_ids = tenants.keys().cloned().collect::<Vec<_>>();
-        let victims = tenant_ids.choose_multiple(&mut thread_rng(), batch_size);
 
-        for victim in victims {
-            let shard = tenants
-                .get_mut(victim)
-                .expect("Held lock between choosing ID and this get");
-
-            if !matches!(shard.get_scheduling_policy(), ShardSchedulingPolicy::Active) {
-                // Skip non-active scheduling policies, so that a shard with a policy like Pause can
-                // be pinned without being disrupted by us.
-                tracing::info!(
-                    "Skipping shard {victim}: scheduling policy is {:?}",
-                    shard.get_scheduling_policy()
-                );
-                continue;
+        // Prefer to migrate tenants that are currently outside their home AZ.  This avoids the chaos injector
+        // continuously pushing tenants outside their home AZ: instead, we'll tend to cycle between picking some
+        // random tenants to move, and then on next chaos iteration moving them back, then picking some new
+        // random tenants on the next iteration.
+        let mut victims = Vec::with_capacity(batch_size);
+        for shard in tenants.values() {
+            if shard.is_attached_outside_preferred_az(nodes) {
+                victims.push(shard.tenant_shard_id);
             }
 
-            // Pick a secondary to promote
-            let Some(new_location) = shard
-                .intent
-                .get_secondary()
-                .choose(&mut thread_rng())
-                .cloned()
-            else {
-                tracing::info!("Skipping shard {victim}: no secondary location, can't migrate");
-                continue;
-            };
+            if victims.len() >= batch_size {
+                break;
+            }
+        }
 
-            let Some(old_location) = *shard.intent.get_attached() else {
-                tracing::info!("Skipping shard {victim}: currently has no attached location");
-                continue;
-            };
+        let choose_random = batch_size.saturating_sub(victims.len());
+        tracing::info!("Injecting chaos: found {} shards to migrate back to home AZ, picking {choose_random} random shards to migrate", victims.len());
 
-            tracing::info!("Injecting chaos: migrate {victim} {old_location}->{new_location}");
+        let random_victims = tenant_ids.choose_multiple(&mut thread_rng(), choose_random);
+        victims.extend(random_victims);
 
-            shard.intent.demote_attached(scheduler, old_location);
-            shard.intent.promote_attached(scheduler, new_location);
-            self.service.maybe_reconcile_shard(shard, nodes);
+        for victim in victims {
+            self.maybe_migrate_to_secondary(victim, nodes, tenants, scheduler);
         }
     }
 }

storage_controller/src/tenant_shard.rs

@@ -1793,6 +1793,23 @@ impl TenantShard {
             }
         }
     }
+
+    /// Returns true if the tenant shard is attached to a node that is outside the preferred AZ.
+    ///
+    /// If the shard does not have a preferred AZ, returns false.
+    pub(crate) fn is_attached_outside_preferred_az(&self, nodes: &HashMap<NodeId, Node>) -> bool {
+        self.intent
+            .get_attached()
+            .map(|node_id| {
+                Some(
+                    nodes
+                        .get(&node_id)
+                        .expect("referenced node exists")
+                        .get_availability_zone_id(),
+                ) == self.intent.preferred_az_id.as_ref()
+            })
+            .unwrap_or(false)
+    }
 }
 
 impl Drop for TenantShard {

storage_scrubber/src/pageserver_physical_gc.rs

@@ -8,6 +8,8 @@ use crate::checks::{
 };
 use crate::metadata_stream::{stream_tenant_timelines, stream_tenants};
 use crate::{init_remote, BucketConfig, NodeKind, RootTarget, TenantShardTimelineId, MAX_RETRIES};
+use async_stream::try_stream;
+use futures::future::Either;
 use futures_util::{StreamExt, TryStreamExt};
 use pageserver::tenant::remote_timeline_client::index::LayerFileMetadata;
 use pageserver::tenant::remote_timeline_client::manifest::OffloadedTimelineManifest;
@@ -578,7 +580,7 @@ async fn gc_timeline(
     target: &RootTarget,
     mode: GcMode,
     ttid: TenantShardTimelineId,
-    accumulator: &Arc<std::sync::Mutex<TenantRefAccumulator>>,
+    accumulator: &std::sync::Mutex<TenantRefAccumulator>,
     tenant_manifest_info: Arc<Option<RemoteTenantManifestInfo>>,
 ) -> anyhow::Result<GcSummary> {
     let mut summary = GcSummary::default();
@@ -721,9 +723,9 @@ pub async fn pageserver_physical_gc(
 
     let remote_client = Arc::new(remote_client);
     let tenants = if tenant_shard_ids.is_empty() {
-        futures::future::Either::Left(stream_tenants(&remote_client, &target))
+        Either::Left(stream_tenants(&remote_client, &target))
     } else {
-        futures::future::Either::Right(futures::stream::iter(tenant_shard_ids.into_iter().map(Ok)))
+        Either::Right(futures::stream::iter(tenant_shard_ids.into_iter().map(Ok)))
     };
 
     // How many tenants to process in parallel.  We need to be mindful of pageservers
@@ -731,16 +733,16 @@ pub async fn pageserver_physical_gc(
     const CONCURRENCY: usize = 32;
 
     // Accumulate information about each tenant for cross-shard GC step we'll do at the end
-    let accumulator = Arc::new(std::sync::Mutex::new(TenantRefAccumulator::default()));
+    let accumulator = std::sync::Mutex::new(TenantRefAccumulator::default());
+
+    // Accumulate information about how many manifests we have GCd
+    let manifest_gc_summary = std::sync::Mutex::new(GcSummary::default());
 
     // Generate a stream of TenantTimelineId
-    enum GcSummaryOrContent<T> {
-        Content(T),
-        GcSummary(GcSummary),
-    }
     let timelines = tenants.map_ok(|tenant_shard_id| {
         let target_ref = &target;
         let remote_client_ref = &remote_client;
+        let manifest_gc_summary_ref = &manifest_gc_summary;
         async move {
             let gc_manifest_result = gc_tenant_manifests(
                 remote_client_ref,
@@ -757,55 +759,48 @@ pub async fn pageserver_physical_gc(
                     (GcSummary::default(), None)
                 }
             };
+            manifest_gc_summary_ref
+                .lock()
+                .unwrap()
+                .merge(summary_from_manifest);
             let tenant_manifest_arc = Arc::new(tenant_manifest_opt);
-            let summary_from_manifest = Ok(GcSummaryOrContent::<(_, _)>::GcSummary(
-                summary_from_manifest,
-            ));
-            stream_tenant_timelines(remote_client_ref, target_ref, tenant_shard_id)
-                .await
-                .map(|stream| {
-                    stream
-                        .zip(futures::stream::iter(std::iter::repeat(
-                            tenant_manifest_arc,
-                        )))
-                        .map(|(ttid_res, tenant_manifest_arc)| {
-                            ttid_res.map(move |ttid| {
-                                GcSummaryOrContent::Content((ttid, tenant_manifest_arc))
-                            })
-                        })
-                        .chain(futures::stream::iter([summary_from_manifest].into_iter()))
-                })
+            let mut timelines = Box::pin(
+                stream_tenant_timelines(remote_client_ref, target_ref, tenant_shard_id).await?,
+            );
+            Ok(try_stream! {
+                while let Some(ttid_res) = timelines.next().await {
+                    let ttid = ttid_res?;
+                    yield (ttid, tenant_manifest_arc.clone());
+                }
+            })
         }
     });
-    let timelines = std::pin::pin!(timelines.try_buffered(CONCURRENCY));
-    let timelines = timelines.try_flatten();
 
     let mut summary = GcSummary::default();
-
-    // Drain futures for per-shard GC, populating accumulator as a side effect
     {
-        let timelines = timelines.map_ok(|summary_or_ttid| match summary_or_ttid {
-            GcSummaryOrContent::Content((ttid, tenant_manifest_arc)) => {
-                futures::future::Either::Left(gc_timeline(
-                    &remote_client,
-                    &min_age,
-                    &target,
-                    mode,
-                    ttid,
-                    &accumulator,
-                    tenant_manifest_arc,
-                ))
-            }
-            GcSummaryOrContent::GcSummary(gc_summary) => {
-                futures::future::Either::Right(futures::future::ok(gc_summary))
-            }
+        let timelines = std::pin::pin!(timelines.try_buffered(CONCURRENCY));
+        let timelines = timelines.try_flatten();
+
+        let timelines = timelines.map_ok(|(ttid, tenant_manifest_arc)| {
+            gc_timeline(
+                &remote_client,
+                &min_age,
+                &target,
+                mode,
+                ttid,
+                &accumulator,
+                tenant_manifest_arc,
+            )
         });
         let mut timelines = std::pin::pin!(timelines.try_buffered(CONCURRENCY));
 
+        // Drain futures for per-shard GC, populating accumulator as a side effect
         while let Some(i) = timelines.next().await {
             summary.merge(i?);
         }
     }
+    // Streams are lazily evaluated, so only now do we have access to the inner object
+    summary.merge(manifest_gc_summary.into_inner().unwrap());
 
     // Execute cross-shard GC, using the accumulator's full view of all the shards built in the per-shard GC
     let Some(client) = controller_client else {
@@ -813,8 +808,7 @@ pub async fn pageserver_physical_gc(
         return Ok(summary);
     };
 
-    let (ancestor_shards, ancestor_refs) = Arc::into_inner(accumulator)
-        .unwrap()
+    let (ancestor_shards, ancestor_refs) = accumulator
         .into_inner()
         .unwrap()
         .into_gc_ancestors(client, &mut summary)

test_runner/conftest.py

@@ -15,4 +15,5 @@ pytest_plugins = (
     "fixtures.compare_fixtures",
     "fixtures.slow",
     "fixtures.reruns",
+    "fixtures.fast_import",
 )

test_runner/fixtures/common_types.py

@@ -208,6 +208,10 @@ class ShardIndex:
             shard_count=int(input[2:4], 16),
         )
 
+    @property
+    def is_sharded(self) -> bool:
+        return self.shard_count != 0
+
 
 class TenantShardId:
     def __init__(self, tenant_id: TenantId, shard_number: int, shard_count: int):

test_runner/fixtures/fast_import.py

@@ -0,0 +1,108 @@
+import os
+import shutil
+import subprocess
+import tempfile
+from collections.abc import Iterator
+from pathlib import Path
+
+import pytest
+
+from fixtures.log_helper import log
+from fixtures.neon_cli import AbstractNeonCli
+from fixtures.pg_version import PgVersion
+
+
+class FastImport(AbstractNeonCli):
+    COMMAND = "fast_import"
+    cmd: subprocess.CompletedProcess[str] | None = None
+
+    def __init__(
+        self,
+        extra_env: dict[str, str] | None,
+        binpath: Path,
+        pg_distrib_dir: Path,
+        pg_version: PgVersion,
+        workdir: Path,
+    ):
+        if extra_env is None:
+            env_vars = {}
+        else:
+            env_vars = extra_env.copy()
+
+        if not (binpath / self.COMMAND).exists():
+            raise Exception(f"{self.COMMAND} binary not found at '{binpath}'")
+        super().__init__(env_vars, binpath)
+
+        pg_dir = pg_distrib_dir / pg_version.v_prefixed
+        self.pg_distrib_dir = pg_distrib_dir
+        self.pg_version = pg_version
+        self.pg_bin = pg_dir / "bin"
+        if not (self.pg_bin / "postgres").exists():
+            raise Exception(f"postgres binary was not found at '{self.pg_bin}'")
+        self.pg_lib = pg_dir / "lib"
+        if env_vars.get("LD_LIBRARY_PATH") is not None:
+            self.pg_lib = Path(env_vars["LD_LIBRARY_PATH"])
+        elif os.getenv("LD_LIBRARY_PATH") is not None:
+            self.pg_lib = Path(str(os.getenv("LD_LIBRARY_PATH")))
+        if not workdir.exists():
+            raise Exception(f"Working directory '{workdir}' does not exist")
+        self.workdir = workdir
+
+    def run(
+        self,
+        pg_port: int | None = None,
+        source_connection_string: str | None = None,
+        restore_connection_string: str | None = None,
+        s3prefix: str | None = None,
+        interactive: bool = False,
+    ) -> subprocess.CompletedProcess[str]:
+        if self.cmd is not None:
+            raise Exception("Command already executed")
+        args = [
+            f"--pg-bin-dir={self.pg_bin}",
+            f"--pg-lib-dir={self.pg_lib}",
+            f"--working-directory={self.workdir}",
+        ]
+        if pg_port is not None:
+            args.append(f"--pg-port={pg_port}")
+        if source_connection_string is not None:
+            args.append(f"--source-connection-string={source_connection_string}")
+        if restore_connection_string is not None:
+            args.append(f"--restore-connection-string={restore_connection_string}")
+        if s3prefix is not None:
+            args.append(f"--s3-prefix={s3prefix}")
+        if interactive:
+            args.append("--interactive")
+
+        self.cmd = self.raw_cli(args)
+        return self.cmd
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *args):
+        if self.workdir.exists():
+            shutil.rmtree(self.workdir)
+
+
+@pytest.fixture(scope="function")
+def fast_import(
+    pg_version: PgVersion,
+    test_output_dir: Path,
+    neon_binpath: Path,
+    pg_distrib_dir: Path,
+) -> Iterator[FastImport]:
+    workdir = Path(tempfile.mkdtemp())
+    with FastImport(None, neon_binpath, pg_distrib_dir, pg_version, workdir) as fi:
+        yield fi
+
+        if fi.cmd is None:
+            return
+
+        # dump stdout & stderr into test log dir
+        with open(test_output_dir / "fast_import.stdout", "w") as f:
+            f.write(fi.cmd.stdout)
+        with open(test_output_dir / "fast_import.stderr", "w") as f:
+            f.write(fi.cmd.stderr)
+
+        log.info("Written logs to %s", test_output_dir)

test_runner/fixtures/metrics.py

@@ -126,12 +126,8 @@ PAGESERVER_GLOBAL_METRICS: tuple[str, ...] = (
     "pageserver_page_cache_read_accesses_total",
     "pageserver_page_cache_size_current_bytes",
     "pageserver_page_cache_size_max_bytes",
-    "pageserver_getpage_reconstruct_seconds_bucket",
-    "pageserver_getpage_reconstruct_seconds_count",
-    "pageserver_getpage_reconstruct_seconds_sum",
     *[f"pageserver_basebackup_query_seconds_{x}" for x in ["bucket", "count", "sum"]],
     *histogram("pageserver_smgr_query_seconds_global"),
-    *histogram("pageserver_getpage_get_reconstruct_data_seconds"),
     *histogram("pageserver_wait_lsn_seconds"),
     *histogram("pageserver_remote_operation_seconds"),
     *histogram("pageserver_io_operations_seconds"),

test_runner/fixtures/neon_fixtures.py

@@ -26,6 +26,7 @@ from urllib.parse import quote, urlparse
 
 import asyncpg
 import backoff
+import boto3
 import httpx
 import psycopg2
 import psycopg2.sql
@@ -36,6 +37,8 @@ from _pytest.config import Config
 from _pytest.config.argparsing import Parser
 from _pytest.fixtures import FixtureRequest
 from jwcrypto import jwk
+from mypy_boto3_kms import KMSClient
+from mypy_boto3_s3 import S3Client
 
 # Type-related stuff
 from psycopg2.extensions import connection as PgConnection
@@ -198,6 +201,30 @@ def mock_s3_server(port_distributor: PortDistributor) -> Iterator[MockS3Server]:
     mock_s3_server.kill()
 
 
+@pytest.fixture(scope="session")
+def mock_kms(mock_s3_server: MockS3Server) -> Iterator[KMSClient]:
+    yield boto3.client(
+        "kms",
+        endpoint_url=mock_s3_server.endpoint(),
+        region_name=mock_s3_server.region(),
+        aws_access_key_id=mock_s3_server.access_key(),
+        aws_secret_access_key=mock_s3_server.secret_key(),
+        aws_session_token=mock_s3_server.session_token(),
+    )
+
+
+@pytest.fixture(scope="session")
+def mock_s3_client(mock_s3_server: MockS3Server) -> Iterator[S3Client]:
+    yield boto3.client(
+        "s3",
+        endpoint_url=mock_s3_server.endpoint(),
+        region_name=mock_s3_server.region(),
+        aws_access_key_id=mock_s3_server.access_key(),
+        aws_secret_access_key=mock_s3_server.secret_key(),
+        aws_session_token=mock_s3_server.session_token(),
+    )
+
+
 class PgProtocol:
     """Reusable connection logic"""
 
@@ -313,6 +340,10 @@ class PgProtocol:
         """
         return self.safe_psql(query, log_query=log_query)[0][0]
 
+    def show_timeline_id(self) -> TimelineId:
+        """SHOW neon.timeline_id"""
+        return TimelineId(cast("str", self.safe_psql("show neon.timeline_id")[0][0]))
+
 
 class PageserverWalReceiverProtocol(StrEnum):
     VANILLA = "vanilla"
@@ -387,6 +418,7 @@ class NeonEnvBuilder:
         storage_controller_port_override: int | None = None,
         pageserver_virtual_file_io_mode: str | None = None,
         pageserver_wal_receiver_protocol: PageserverWalReceiverProtocol | None = None,
+        pageserver_get_vectored_concurrent_io: str | None = None,
     ):
         self.repo_dir = repo_dir
         self.rust_log_override = rust_log_override
@@ -426,6 +458,9 @@ class NeonEnvBuilder:
         self.storage_controller_config: dict[Any, Any] | None = None
 
         self.pageserver_virtual_file_io_engine: str | None = pageserver_virtual_file_io_engine
+        self.pageserver_get_vectored_concurrent_io: str | None = (
+            pageserver_get_vectored_concurrent_io
+        )
 
         self.pageserver_default_tenant_config_compaction_algorithm: dict[str, Any] | None = (
             pageserver_default_tenant_config_compaction_algorithm
@@ -452,6 +487,7 @@ class NeonEnvBuilder:
         self.test_name = test_name
         self.compatibility_neon_binpath = compatibility_neon_binpath
         self.compatibility_pg_distrib_dir = compatibility_pg_distrib_dir
+        self.test_may_use_compatibility_snapshot_binaries = False
         self.version_combination = combination
         self.mixdir = self.test_output_dir / "mixdir_neon"
         if self.version_combination is not None:
@@ -463,6 +499,7 @@ class NeonEnvBuilder:
             ), "the environment variable COMPATIBILITY_POSTGRES_DISTRIB_DIR is required when using mixed versions"
             self.mixdir.mkdir(mode=0o755, exist_ok=True)
             self._mix_versions()
+            self.test_may_use_compatibility_snapshot_binaries = True
 
     def init_configs(self, default_remote_storage_if_missing: bool = True) -> NeonEnv:
         # Cannot create more than one environment from one builder
@@ -1062,6 +1099,7 @@ class NeonEnv:
         self.pageserver_virtual_file_io_engine = config.pageserver_virtual_file_io_engine
         self.pageserver_virtual_file_io_mode = config.pageserver_virtual_file_io_mode
         self.pageserver_wal_receiver_protocol = config.pageserver_wal_receiver_protocol
+        self.pageserver_get_vectored_concurrent_io = config.pageserver_get_vectored_concurrent_io
 
         # Create the neon_local's `NeonLocalInitConf`
         cfg: dict[str, Any] = {
@@ -1115,12 +1153,24 @@ class NeonEnv:
 
             # Batching (https://github.com/neondatabase/neon/issues/9377):
             # enable batching by default in tests and benchmarks.
+            ps_cfg["page_service_pipelining"] = {
+                "mode": "pipelined",
+                "execution": "concurrent-futures",
+                "max_batch_size": 32,
+            }
+
+            # Concurrent IO (https://github.com/neondatabase/neon/issues/9378):
+            # enable concurrent IO by default in tests and benchmarks.
             # Compat tests are exempt because old versions fail to parse the new config.
-            if not config.compatibility_neon_binpath:
-                ps_cfg["page_service_pipelining"] = {
-                    "mode": "pipelined",
-                    "execution": "concurrent-futures",
-                    "max_batch_size": 32,
+            get_vectored_concurrent_io = self.pageserver_get_vectored_concurrent_io
+            if config.test_may_use_compatibility_snapshot_binaries:
+                log.info(
+                    "Forcing use of binary-built-in default to avoid forward-compatibility related test failures"
+                )
+                get_vectored_concurrent_io = None
+            if get_vectored_concurrent_io is not None:
+                ps_cfg["get_vectored_concurrent_io"] = {
+                    "mode": self.pageserver_get_vectored_concurrent_io,
                 }
 
             if self.pageserver_virtual_file_io_engine is not None:
@@ -1457,6 +1507,7 @@ def neon_simple_env(
     pageserver_virtual_file_io_engine: str,
     pageserver_default_tenant_config_compaction_algorithm: dict[str, Any] | None,
     pageserver_virtual_file_io_mode: str | None,
+    pageserver_get_vectored_concurrent_io: str | None,
 ) -> Iterator[NeonEnv]:
     """
     Simple Neon environment, with 1 safekeeper and 1 pageserver. No authentication, no fsync.
@@ -1489,6 +1540,7 @@ def neon_simple_env(
         pageserver_virtual_file_io_engine=pageserver_virtual_file_io_engine,
         pageserver_default_tenant_config_compaction_algorithm=pageserver_default_tenant_config_compaction_algorithm,
         pageserver_virtual_file_io_mode=pageserver_virtual_file_io_mode,
+        pageserver_get_vectored_concurrent_io=pageserver_get_vectored_concurrent_io,
         combination=combination,
     ) as builder:
         env = builder.init_start()
@@ -1515,6 +1567,7 @@ def neon_env_builder(
     pageserver_default_tenant_config_compaction_algorithm: dict[str, Any] | None,
     record_property: Callable[[str, object], None],
     pageserver_virtual_file_io_mode: str | None,
+    pageserver_get_vectored_concurrent_io: str | None,
 ) -> Iterator[NeonEnvBuilder]:
     """
     Fixture to create a Neon environment for test.
@@ -1557,6 +1610,7 @@ def neon_env_builder(
         test_overlay_dir=test_overlay_dir,
         pageserver_default_tenant_config_compaction_algorithm=pageserver_default_tenant_config_compaction_algorithm,
         pageserver_virtual_file_io_mode=pageserver_virtual_file_io_mode,
+        pageserver_get_vectored_concurrent_io=pageserver_get_vectored_concurrent_io,
     ) as builder:
         yield builder
         # Propogate `preserve_database_files` to make it possible to use in other fixtures,

test_runner/fixtures/parametrize.py

@@ -44,6 +44,11 @@ def pageserver_virtual_file_io_mode() -> str | None:
     return os.getenv("PAGESERVER_VIRTUAL_FILE_IO_MODE")
 
 
+@pytest.fixture(scope="function", autouse=True)
+def pageserver_get_vectored_concurrent_io() -> str | None:
+    return os.getenv("PAGESERVER_GET_VECTORED_CONCURRENT_IO")
+
+
 def get_pageserver_default_tenant_config_compaction_algorithm() -> dict[str, Any] | None:
     toml_table = os.getenv("PAGESERVER_DEFAULT_TENANT_CONFIG_COMPACTION_ALGORITHM")
     if toml_table is None:

test_runner/regress/test_attach_tenant_config.py

@@ -176,6 +176,10 @@ def test_fully_custom_config(positive_env: NeonEnv):
             "type": "interpreted",
             "args": {"format": "bincode", "compression": {"zstd": {"level": 1}}},
         },
+        "rel_size_v2_enabled": True,
+        "gc_compaction_enabled": True,
+        "gc_compaction_initial_threshold_kb": 1024000,
+        "gc_compaction_ratio_percent": 200,
     }
 
     vps_http = env.storage_controller.pageserver_api()

test_runner/regress/test_compatibility.py

@@ -251,6 +251,8 @@ def test_forward_compatibility(
         os.environ.get("ALLOW_FORWARD_COMPATIBILITY_BREAKAGE", "false").lower() == "true"
     )
 
+    neon_env_builder.test_may_use_compatibility_snapshot_binaries = True
+
     try:
         neon_env_builder.num_safekeepers = 3
 

test_runner/regress/test_import_pgdata.py

@@ -1,21 +1,28 @@
+import base64
 import json
 import re
 import time
 from enum import Enum
+from pathlib import Path
 
 import psycopg2
 import psycopg2.errors
 import pytest
 from fixtures.common_types import Lsn, TenantId, TenantShardId, TimelineId
+from fixtures.fast_import import FastImport
 from fixtures.log_helper import log
-from fixtures.neon_fixtures import NeonEnvBuilder, VanillaPostgres
+from fixtures.neon_fixtures import NeonEnvBuilder, PgBin, PgProtocol, VanillaPostgres
 from fixtures.pageserver.http import (
     ImportPgdataIdemptencyKey,
     PageserverApiException,
 )
 from fixtures.pg_version import PgVersion
-from fixtures.remote_storage import RemoteStorageKind
+from fixtures.port_distributor import PortDistributor
+from fixtures.remote_storage import MockS3Server, RemoteStorageKind
 from fixtures.utils import run_only_on_postgres
+from mypy_boto3_kms import KMSClient
+from mypy_boto3_kms.type_defs import EncryptResponseTypeDef
+from mypy_boto3_s3 import S3Client
 from pytest_httpserver import HTTPServer
 from werkzeug.wrappers.request import Request
 from werkzeug.wrappers.response import Response
@@ -98,13 +105,15 @@ def test_pgdata_import_smoke(
     while True:
         relblock_size = vanilla_pg.safe_psql_scalar("select pg_relation_size('t')")
         log.info(
-            f"relblock size: {relblock_size/8192} pages (target: {target_relblock_size//8192}) pages"
+            f"relblock size: {relblock_size / 8192} pages (target: {target_relblock_size // 8192}) pages"
         )
         if relblock_size >= target_relblock_size:
             break
         addrows = int((target_relblock_size - relblock_size) // 8192)
         assert addrows >= 1, "forward progress"
-        vanilla_pg.safe_psql(f"insert into t select generate_series({nrows+1}, {nrows + addrows})")
+        vanilla_pg.safe_psql(
+            f"insert into t select generate_series({nrows + 1}, {nrows + addrows})"
+        )
         nrows += addrows
     expect_nrows = nrows
     expect_sum = (
@@ -313,3 +322,135 @@ def test_pgdata_import_smoke(
     validate_vanilla_equivalence(br_initdb_endpoint)
     with pytest.raises(psycopg2.errors.UndefinedTable):
         br_initdb_endpoint.safe_psql("select * from othertable")
+
+
+def test_fast_import_binary(
+    test_output_dir,
+    vanilla_pg: VanillaPostgres,
+    port_distributor: PortDistributor,
+    fast_import: FastImport,
+):
+    vanilla_pg.start()
+    vanilla_pg.safe_psql("CREATE TABLE foo (a int); INSERT INTO foo SELECT generate_series(1, 10);")
+
+    pg_port = port_distributor.get_port()
+    fast_import.run(pg_port, vanilla_pg.connstr())
+    vanilla_pg.stop()
+
+    pgbin = PgBin(test_output_dir, fast_import.pg_distrib_dir, fast_import.pg_version)
+    with VanillaPostgres(
+        fast_import.workdir / "pgdata", pgbin, pg_port, False
+    ) as new_pgdata_vanilla_pg:
+        new_pgdata_vanilla_pg.start()
+
+        # database name and user are hardcoded in fast_import binary, and they are different from normal vanilla postgres
+        conn = PgProtocol(dsn=f"postgresql://cloud_admin@localhost:{pg_port}/neondb")
+        res = conn.safe_psql("SELECT count(*) FROM foo;")
+        log.info(f"Result: {res}")
+        assert res[0][0] == 10
+
+
+def test_fast_import_restore_to_connstring(
+    test_output_dir,
+    vanilla_pg: VanillaPostgres,
+    port_distributor: PortDistributor,
+    fast_import: FastImport,
+    pg_distrib_dir: Path,
+    pg_version: PgVersion,
+):
+    vanilla_pg.start()
+    vanilla_pg.safe_psql("CREATE TABLE foo (a int); INSERT INTO foo SELECT generate_series(1, 10);")
+
+    pgdatadir = test_output_dir / "restore-pgdata"
+    pg_bin = PgBin(test_output_dir, pg_distrib_dir, pg_version)
+    port = port_distributor.get_port()
+    with VanillaPostgres(pgdatadir, pg_bin, port) as restore_vanilla_pg:
+        restore_vanilla_pg.configure(["shared_preload_libraries='neon_rmgr'"])
+        restore_vanilla_pg.start()
+
+        fast_import.run(
+            source_connection_string=vanilla_pg.connstr(),
+            restore_connection_string=restore_vanilla_pg.connstr(),
+        )
+        vanilla_pg.stop()
+
+        res = restore_vanilla_pg.safe_psql("SELECT count(*) FROM foo;")
+        log.info(f"Result: {res}")
+        assert res[0][0] == 10
+
+
+def test_fast_import_restore_to_connstring_from_s3_spec(
+    test_output_dir,
+    vanilla_pg: VanillaPostgres,
+    port_distributor: PortDistributor,
+    fast_import: FastImport,
+    pg_distrib_dir: Path,
+    pg_version: PgVersion,
+    mock_s3_server: MockS3Server,
+    mock_kms: KMSClient,
+    mock_s3_client: S3Client,
+):
+    # Prepare KMS and S3
+    key_response = mock_kms.create_key(
+        Description="Test key",
+        KeyUsage="ENCRYPT_DECRYPT",
+        Origin="AWS_KMS",
+    )
+    key_id = key_response["KeyMetadata"]["KeyId"]
+
+    def encrypt(x: str) -> EncryptResponseTypeDef:
+        return mock_kms.encrypt(KeyId=key_id, Plaintext=x)
+
+    # Start source postgres and ingest data
+    vanilla_pg.start()
+    vanilla_pg.safe_psql("CREATE TABLE foo (a int); INSERT INTO foo SELECT generate_series(1, 10);")
+
+    # Start target postgres
+    pgdatadir = test_output_dir / "restore-pgdata"
+    pg_bin = PgBin(test_output_dir, pg_distrib_dir, pg_version)
+    port = port_distributor.get_port()
+    with VanillaPostgres(pgdatadir, pg_bin, port) as restore_vanilla_pg:
+        restore_vanilla_pg.configure(["shared_preload_libraries='neon_rmgr'"])
+        restore_vanilla_pg.start()
+
+        # Encrypt connstrings and put spec into S3
+        source_connstring_encrypted = encrypt(vanilla_pg.connstr())
+        restore_connstring_encrypted = encrypt(restore_vanilla_pg.connstr())
+        spec = {
+            "encryption_secret": {"KMS": {"key_id": key_id}},
+            "source_connstring_ciphertext_base64": base64.b64encode(
+                source_connstring_encrypted["CiphertextBlob"]
+            ).decode("utf-8"),
+            "restore_connstring_ciphertext_base64": base64.b64encode(
+                restore_connstring_encrypted["CiphertextBlob"]
+            ).decode("utf-8"),
+        }
+
+        mock_s3_client.create_bucket(Bucket="test-bucket")
+        mock_s3_client.put_object(
+            Bucket="test-bucket", Key="test-prefix/spec.json", Body=json.dumps(spec)
+        )
+
+        # Run fast_import
+        if fast_import.extra_env is None:
+            fast_import.extra_env = {}
+        fast_import.extra_env["AWS_ACCESS_KEY_ID"] = mock_s3_server.access_key()
+        fast_import.extra_env["AWS_SECRET_ACCESS_KEY"] = mock_s3_server.secret_key()
+        fast_import.extra_env["AWS_SESSION_TOKEN"] = mock_s3_server.session_token()
+        fast_import.extra_env["AWS_REGION"] = mock_s3_server.region()
+        fast_import.extra_env["AWS_ENDPOINT_URL"] = mock_s3_server.endpoint()
+        fast_import.extra_env["RUST_LOG"] = "aws_config=debug,aws_sdk_kms=debug"
+        fast_import.run(s3prefix="s3://test-bucket/test-prefix")
+        vanilla_pg.stop()
+
+        res = restore_vanilla_pg.safe_psql("SELECT count(*) FROM foo;")
+        log.info(f"Result: {res}")
+        assert res[0][0] == 10
+
+
+# TODO: Maybe test with pageserver?
+# 1. run whole neon env
+# 2. create timeline with some s3 path???
+# 3. run fast_import with s3 prefix
+# 4. ??? mock http where pageserver will report progress
+# 5. run compute on this timeline and check if data is there

test_runner/regress/test_local_file_cache.py

@@ -7,9 +7,78 @@ import threading
 import time
 
 import pytest
-from fixtures.neon_fixtures import NeonEnvBuilder
+from fixtures.neon_fixtures import NeonEnv, NeonEnvBuilder
 from fixtures.utils import USE_LFC, query_scalar
 
+"""
+Test whether LFC doesn't error out when the LRU is empty, but the LFC is
+already at its maximum size.
+
+If we don't handle this safely, we might allocate more hash entries than
+otherwise considered safe, thus causing ERRORs in hash_search(HASH_ENTER) once
+we hit lfc->used >= lfc->limit.
+"""
+
+
+@pytest.mark.skipif(not USE_LFC, reason="LFC is disabled, skipping")
+def test_local_file_cache_all_pinned(neon_simple_env: NeonEnv):
+    env = neon_simple_env
+    endpoint = env.endpoints.create_start(
+        "main",
+        config_lines=[
+            "neon.max_file_cache_size='1MB'",
+            "neon.file_cache_size_limit='1MB'",
+        ],
+    )
+    top_cur = endpoint.connect().cursor()
+
+    stop = threading.Event()
+    n_rows = 10000
+    n_threads = 5
+    n_updates_per_connection = 1000
+
+    top_cur.execute("CREATE TABLE lfctest (id int4 PRIMARY KEY, n int) WITH (fillfactor=10)")
+    top_cur.execute(f"INSERT INTO lfctest SELECT g, 1 FROM generate_series(1, {n_rows}) g")
+
+    # Start threads that will perform random UPDATEs. Each UPDATE
+    # increments the counter on the row, so that we can check at the
+    # end that the sum of all the counters match the number of updates
+    # performed (plus the initial 1 on each row).
+    #
+    # Furthermore, each thread will reconnect between every 1000 updates.
+    def run_updates(n_updates_performed_q: queue.Queue[int]):
+        n_updates_performed = 0
+        conn = endpoint.connect()
+        cur = conn.cursor()
+        while not stop.is_set():
+            id = random.randint(1, n_rows)
+            cur.execute(f"UPDATE lfctest SET n = n + 1 WHERE id = {id}")
+            n_updates_performed += 1
+            if n_updates_performed % n_updates_per_connection == 0:
+                cur.close()
+                conn.close()
+                conn = endpoint.connect()
+                cur = conn.cursor()
+        n_updates_performed_q.put(n_updates_performed)
+
+    n_updates_performed_q: queue.Queue[int] = queue.Queue()
+    threads: list[threading.Thread] = []
+    for _i in range(n_threads):
+        thread = threading.Thread(target=run_updates, args=(n_updates_performed_q,), daemon=True)
+        thread.start()
+        threads.append(thread)
+
+    time.sleep(15)
+
+    stop.set()
+
+    n_updates_performed = 0
+    for thread in threads:
+        thread.join()
+        n_updates_performed += n_updates_performed_q.get()
+
+    assert query_scalar(top_cur, "SELECT SUM(n) FROM lfctest") == n_rows + n_updates_performed
+
 
 @pytest.mark.skipif(not USE_LFC, reason="LFC is disabled, skipping")
 def test_local_file_cache_unlink(neon_env_builder: NeonEnvBuilder):

test_runner/regress/test_storage_scrubber.py

@@ -227,7 +227,9 @@ def test_scrubber_physical_gc_ancestors(neon_env_builder: NeonEnvBuilder, shard_
     new_shard_count = 4
     assert shard_count is None or new_shard_count > shard_count
     shards = env.storage_controller.tenant_shard_split(tenant_id, shard_count=new_shard_count)
-    env.storage_controller.reconcile_until_idle()  # Move shards to their final locations immediately
+    env.storage_controller.reconcile_until_idle(
+        timeout_secs=120
+    )  # Move shards to their final locations immediately
 
     # Create a timeline after split, to ensure scrubber can handle timelines that exist in child shards but not ancestors
     env.storage_controller.pageserver_api().timeline_create(
@@ -269,6 +271,8 @@ def test_scrubber_physical_gc_ancestors(neon_env_builder: NeonEnvBuilder, shard_
         ps.http_client().timeline_compact(
             shard, timeline_id, force_image_layer_creation=True, wait_until_uploaded=True
         )
+        # Add some WAL so that we don't gc at the latest remote consistent lsn
+        workload.churn_rows(1)
         ps.http_client().timeline_gc(shard, timeline_id, 0)
 
     # We will use a min_age_secs=1 threshold for deletion, let it pass

workspace_hack/Cargo.toml

@@ -17,8 +17,6 @@ license.workspace = true
 [dependencies]
 ahash = { version = "0.8" }
 anyhow = { version = "1", features = ["backtrace"] }
-axum = { version = "0.7", features = ["ws"] }
-axum-core = { version = "0.4", default-features = false, features = ["tracing"] }
 base64-594e8ee84c453af0 = { package = "base64", version = "0.13", features = ["alloc"] }
 base64-647d43efb71741da = { package = "base64", version = "0.21", features = ["alloc"] }
 base64ct = { version = "1", default-features = false, features = ["std"] }
@@ -46,7 +44,7 @@ hex = { version = "0.4", features = ["serde"] }
 hmac = { version = "0.12", default-features = false, features = ["reset"] }
 hyper-582f2526e08bb6a0 = { package = "hyper", version = "0.14", features = ["full"] }
 hyper-dff4ba8e3ae991db = { package = "hyper", version = "1", features = ["full"] }
-hyper-util = { version = "0.1", features = ["client-legacy", "server-auto", "service"] }
+hyper-util = { version = "0.1", features = ["client-legacy", "http1", "http2", "server", "service"] }
 indexmap-dff4ba8e3ae991db = { package = "indexmap", version = "1", default-features = false, features = ["std"] }
 indexmap-f595c2ba2a3f28df = { package = "indexmap", version = "2", features = ["serde"] }
 itertools = { version = "0.12" }
@@ -87,12 +85,11 @@ tikv-jemalloc-sys = { version = "0.6", features = ["profiling", "stats", "unpref
 time = { version = "0.3", features = ["macros", "serde-well-known"] }
 tokio = { version = "1", features = ["full", "test-util"] }
 tokio-rustls = { version = "0.26", default-features = false, features = ["logging", "ring", "tls12"] }
-tokio-stream = { version = "0.1", features = ["net"] }
+tokio-stream = { version = "0.1" }
 tokio-util = { version = "0.7", features = ["codec", "compat", "io", "rt"] }
 toml_edit = { version = "0.22", features = ["serde"] }
-tonic = { version = "0.12", features = ["tls-roots"] }
-tower-9fbad63c4bcf4a8f = { package = "tower", version = "0.4", default-features = false, features = ["balance", "buffer", "limit", "util"] }
-tower-d8f496e17d97b5cb = { package = "tower", version = "0.5", default-features = false, features = ["log", "make", "util"] }
+tonic = { version = "0.12", default-features = false, features = ["codegen", "prost", "tls-roots"] }
+tower = { version = "0.4", default-features = false, features = ["balance", "buffer", "limit", "util"] }
 tracing = { version = "0.1", features = ["log"] }
 tracing-core = { version = "0.1" }
 url = { version = "2", features = ["serde"] }