Merge pull request #10416 from neondatabase/rc/release-proxy/2025-01-16

Proxy release 2025-01-16

.github/actionlint.yml

@@ -25,3 +25,4 @@ config-variables:
   - PGREGRESS_PG17_PROJECT_ID
   - SLACK_ON_CALL_QA_STAGING_STREAM
   - DEV_AWS_OIDC_ROLE_MANAGE_BENCHMARK_EC2_VMS_ARN
+  - SLACK_ON_CALL_STORAGE_STAGING_STREAM

.github/workflows/build_and_test.yml

@@ -346,25 +346,22 @@ jobs:
       # XXX: no coverage data handling here, since benchmarks are run on release builds,
       # while coverage is currently collected for the debug ones
 
-  report-benchmarks-failures:
+  report-benchmarks-results-to-slack:
     needs: [ benchmarks, create-test-report ]
-    if: github.ref_name == 'main' && failure() && needs.benchmarks.result == 'failure'
-    permissions:
-      id-token: write # aws-actions/configure-aws-credentials
-      statuses: write
-      contents: write
-      pull-requests: write
+    if: github.ref_name == 'main' && !cancelled() && contains(fromJSON('["success", "failure"]'), needs.benchmarks.result)
     runs-on: ubuntu-22.04
 
     steps:
-    - uses: slackapi/slack-github-action@v1
+    - uses: slackapi/slack-github-action@v2
       with:
-        channel-id: C060CNA47S9 # on-call-staging-storage-stream
-        slack-message: |
-          Benchmarks failed on main <${{ github.event.head_commit.url }}|${{ github.sha }}>
-          <${{ needs.create-test-report.outputs.report-url }}|Allure report>
-      env:
-        SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}
+        method: chat.postMessage
+        token: ${{ secrets.SLACK_BOT_TOKEN }}
+        payload: |
+          channel: "${{ vars.SLACK_ON_CALL_STORAGE_STAGING_STREAM }}"
+          text: |
+            Benchmarks on main: *${{ needs.benchmarks.result }}*
+            - <${{ needs.create-test-report.outputs.report-url }}|Allure report>
+            - <${{ github.event.head_commit.url }}|${{ github.sha }}>
 
   create-test-report:
     needs: [ check-permissions, build-and-test-locally, coverage-report, build-build-tools-image, benchmarks ]

Cargo.lock

@@ -3981,9 +3981,11 @@ name = "pagectl"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "bincode",
  "camino",
  "clap",
  "humantime",
+ "itertools 0.10.5",
  "pageserver",
  "pageserver_api",
  "postgres_ffi",
@@ -4005,6 +4007,7 @@ dependencies = [
  "arc-swap",
  "async-compression",
  "async-stream",
+ "bincode",
  "bit_field",
  "byteorder",
  "bytes",
@@ -5655,6 +5658,7 @@ dependencies = [
  "crc32c",
  "criterion",
  "desim",
+ "env_logger 0.10.2",
  "fail",
  "futures",
  "hex",
@@ -5683,6 +5687,7 @@ dependencies = [
  "serde",
  "serde_json",
  "sha2",
+ "smallvec",
  "storage_broker",
  "strum",
  "strum_macros",
@@ -5707,10 +5712,13 @@ dependencies = [
 name = "safekeeper_api"
 version = "0.1.0"
 dependencies = [
+ "anyhow",
  "const_format",
+ "pageserver_api",
  "postgres_ffi",
  "pq_proto",
  "serde",
+ "serde_json",
  "tokio",
  "utils",
 ]
@@ -7558,12 +7566,21 @@ dependencies = [
  "anyhow",
  "async-compression",
  "bytes",
+ "camino",
+ "camino-tempfile",
+ "criterion",
+ "futures",
  "pageserver_api",
  "postgres_ffi",
+ "pprof",
  "prost",
+ "remote_storage",
  "serde",
+ "serde_json",
  "thiserror",
+ "tikv-jemallocator",
  "tokio",
+ "tokio-util",
  "tonic",
  "tonic-build",
  "tracing",

compute/compute-node.Dockerfile

@@ -66,6 +66,7 @@ RUN cd postgres && \
     make MAKELEVEL=0 -j $(getconf _NPROCESSORS_ONLN) -s -C src/interfaces/libpq install && \
     # Enable some of contrib extensions
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/autoinc.control && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/dblink.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/bloom.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/earthdistance.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/insert_username.control && \
@@ -871,7 +872,7 @@ RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux
     chmod +x rustup-init && \
     ./rustup-init -y --no-modify-path --profile minimal --default-toolchain stable && \
     rm rustup-init && \
-    cargo install --locked --version 0.12.6 cargo-pgrx && \
+    cargo install --locked --version 0.12.9 cargo-pgrx && \
     /bin/bash -c 'cargo pgrx init --pg${PG_VERSION:1}=/usr/local/pgsql/bin/pg_config'
 
 USER root
@@ -908,19 +909,19 @@ RUN apt update && apt install --no-install-recommends --no-install-suggests -y p
     mkdir pgrag-src && cd pgrag-src && tar xzf ../pgrag.tar.gz --strip-components=1 -C . && \
     \
     cd exts/rag && \
-    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
     cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/rag.control && \
     \
     cd ../rag_bge_small_en_v15 && \
-    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
     ORT_LIB_LOCATION=/home/nonroot/onnxruntime-src/build/Linux \
         REMOTE_ONNX_URL=http://pg-ext-s3-gateway/pgrag-data/bge_small_en_v15.onnx \
         cargo pgrx install --release --features remote_onnx && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/rag_bge_small_en_v15.control && \
     \
     cd ../rag_jina_reranker_v1_tiny_en && \
-    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
     ORT_LIB_LOCATION=/home/nonroot/onnxruntime-src/build/Linux \
         REMOTE_ONNX_URL=http://pg-ext-s3-gateway/pgrag-data/jina_reranker_v1_tiny_en.onnx \
         cargo pgrx install --release --features remote_onnx && \
@@ -945,7 +946,8 @@ RUN wget https://github.com/supabase/pg_jsonschema/archive/refs/tags/v0.3.3.tar.
     # against postgres forks that decided to change their ABI name (like us).
     # With that we can build extensions without forking them and using stock
     # pgx. As this feature is new few manual version bumps were required.
-    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx-tests = "0.12.6"/pgrx-tests = "0.12.9"/g' Cargo.toml && \
     cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_jsonschema.control
 
@@ -963,7 +965,8 @@ ARG PG_VERSION
 RUN wget https://github.com/supabase/pg_graphql/archive/refs/tags/v1.5.9.tar.gz -O pg_graphql.tar.gz && \
     echo "cf768385a41278be1333472204fc0328118644ae443182cf52f7b9b23277e497 pg_graphql.tar.gz" | sha256sum --check && \
     mkdir pg_graphql-src && cd pg_graphql-src && tar xzf ../pg_graphql.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgrx = "=0.12.6"/pgrx = { version = "0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx = "=0.12.6"/pgrx = { version = "=0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx-tests = "=0.12.6"/pgrx-tests = "=0.12.9"/g' Cargo.toml && \
     cargo pgrx install --release && \
     # it's needed to enable extension because it uses untrusted C language
     sed -i 's/superuser = false/superuser = true/g' /usr/local/pgsql/share/extension/pg_graphql.control && \
@@ -984,9 +987,8 @@ ARG PG_VERSION
 RUN wget https://github.com/kelvich/pg_tiktoken/archive/9118dd4549b7d8c0bbc98e04322499f7bf2fa6f7.tar.gz -O pg_tiktoken.tar.gz && \
     echo "a5bc447e7920ee149d3c064b8b9f0086c0e83939499753178f7d35788416f628 pg_tiktoken.tar.gz" | sha256sum --check && \
     mkdir pg_tiktoken-src && cd pg_tiktoken-src && tar xzf ../pg_tiktoken.tar.gz --strip-components=1 -C . && \
-    # TODO update pgrx version in the pg_tiktoken repo and remove this line
-    sed -i 's/pgrx = { version = "=0.10.2",/pgrx = { version = "0.11.3",/g' Cargo.toml && \
-    sed -i 's/pgrx-tests = "=0.10.2"/pgrx-tests = "0.11.3"/g' Cargo.toml && \
+    sed -i 's/pgrx = { version = "=0.12.6",/pgrx = { version = "0.12.9",/g' Cargo.toml && \
+    sed -i 's/pgrx-tests = "=0.12.6"/pgrx-tests = "0.12.9"/g' Cargo.toml && \
     cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_tiktoken.control
 
@@ -1028,7 +1030,11 @@ ARG PG_VERSION
 RUN wget https://github.com/neondatabase/pg_session_jwt/archive/refs/tags/v0.2.0.tar.gz -O pg_session_jwt.tar.gz && \
     echo "5ace028e591f2e000ca10afa5b1ca62203ebff014c2907c0ec3b29c36f28a1bb pg_session_jwt.tar.gz" | sha256sum --check && \
     mkdir pg_session_jwt-src && cd pg_session_jwt-src && tar xzf ../pg_session_jwt.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "=0.12.6", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/pgrx = "0.12.6"/pgrx = { version = "0.12.9", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    sed -i 's/version = "0.12.6"/version = "0.12.9"/g' pgrx-tests/Cargo.toml && \
+    sed -i 's/pgrx = "=0.12.6"/pgrx = { version = "=0.12.9", features = [ "unsafe-postgres" ] }/g' pgrx-tests/Cargo.toml && \
+    sed -i 's/pgrx-macros = "=0.12.6"/pgrx-macros = "=0.12.9"/g' pgrx-tests/Cargo.toml && \
+    sed -i 's/pgrx-pg-config = "=0.12.6"/pgrx-pg-config = "=0.12.9"/g' pgrx-tests/Cargo.toml && \
     cargo pgrx install --release
 
 #########################################################################################

compute_tools/src/bin/fast_import.rs

@@ -31,7 +31,7 @@ use camino::{Utf8Path, Utf8PathBuf};
 use clap::Parser;
 use compute_tools::extension_server::{get_pg_version, PostgresMajorVersion};
 use nix::unistd::Pid;
-use tracing::{info, info_span, warn, Instrument};
+use tracing::{error, info, info_span, warn, Instrument};
 use utils::fs_ext::is_directory_empty;
 
 #[path = "fast_import/aws_s3_sync.rs"]
@@ -41,12 +41,19 @@ mod child_stdio_to_log;
 #[path = "fast_import/s3_uri.rs"]
 mod s3_uri;
 
+const PG_WAIT_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(600);
+const PG_WAIT_RETRY_INTERVAL: std::time::Duration = std::time::Duration::from_millis(300);
+
 #[derive(clap::Parser)]
 struct Args {
     #[clap(long)]
     working_directory: Utf8PathBuf,
     #[clap(long, env = "NEON_IMPORTER_S3_PREFIX")]
-    s3_prefix: s3_uri::S3Uri,
+    s3_prefix: Option<s3_uri::S3Uri>,
+    #[clap(long)]
+    source_connection_string: Option<String>,
+    #[clap(short, long)]
+    interactive: bool,
     #[clap(long)]
     pg_bin_dir: Utf8PathBuf,
     #[clap(long)]
@@ -77,30 +84,70 @@ pub(crate) async fn main() -> anyhow::Result<()> {
 
     info!("starting");
 
-    let Args {
-        working_directory,
-        s3_prefix,
-        pg_bin_dir,
-        pg_lib_dir,
-    } = Args::parse();
+    let args = Args::parse();
 
-    let aws_config = aws_config::load_defaults(BehaviorVersion::v2024_03_28()).await;
+    // Validate arguments
+    if args.s3_prefix.is_none() && args.source_connection_string.is_none() {
+        anyhow::bail!("either s3_prefix or source_connection_string must be specified");
+    }
+    if args.s3_prefix.is_some() && args.source_connection_string.is_some() {
+        anyhow::bail!("only one of s3_prefix or source_connection_string can be specified");
+    }
 
-    let spec: Spec = {
-        let spec_key = s3_prefix.append("/spec.json");
-        let s3_client = aws_sdk_s3::Client::new(&aws_config);
-        let object = s3_client
-            .get_object()
-            .bucket(&spec_key.bucket)
-            .key(spec_key.key)
-            .send()
-            .await
-            .context("get spec from s3")?
-            .body
-            .collect()
-            .await
-            .context("download spec body")?;
-        serde_json::from_slice(&object.into_bytes()).context("parse spec as json")?
+    let working_directory = args.working_directory;
+    let pg_bin_dir = args.pg_bin_dir;
+    let pg_lib_dir = args.pg_lib_dir;
+
+    // Initialize AWS clients only if s3_prefix is specified
+    let (aws_config, kms_client) = if args.s3_prefix.is_some() {
+        let config = aws_config::load_defaults(BehaviorVersion::v2024_03_28()).await;
+        let kms = aws_sdk_kms::Client::new(&config);
+        (Some(config), Some(kms))
+    } else {
+        (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 spec: Spec = {
+            let spec_key = s3_prefix.append("/spec.json");
+            let s3_client = aws_sdk_s3::Client::new(aws_config.as_ref().unwrap());
+            let object = s3_client
+                .get_object()
+                .bucket(&spec_key.bucket)
+                .key(spec_key.key)
+                .send()
+                .await
+                .context("get spec from s3")?
+                .body
+                .collect()
+                .await
+                .context("download spec body")?;
+            serde_json::from_slice(&object.into_bytes()).context("parse spec as json")?
+        };
+
+        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")?
+            }
+        }
+    } else {
+        args.source_connection_string.unwrap()
     };
 
     match tokio::fs::create_dir(&working_directory).await {
@@ -123,15 +170,6 @@ pub(crate) async fn main() -> anyhow::Result<()> {
         .await
         .context("create pgdata directory")?;
 
-    //
-    // Setup clients
-    //
-    let aws_config = aws_config::load_defaults(BehaviorVersion::v2024_03_28()).await;
-    let kms_client = aws_sdk_kms::Client::new(&aws_config);
-
-    //
-    //  Initialize pgdata
-    //
     let pgbin = pg_bin_dir.join("postgres");
     let pg_version = match get_pg_version(pgbin.as_ref()) {
         PostgresMajorVersion::V14 => 14,
@@ -170,7 +208,13 @@ pub(crate) async fn main() -> anyhow::Result<()> {
         .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"])
+        .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())
@@ -185,44 +229,58 @@ pub(crate) async fn main() -> anyhow::Result<()> {
         )
         .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 {
-        let res = tokio_postgres::connect(&restore_pg_connstring, tokio_postgres::NoTls).await;
-        if res.is_ok() {
-            info!("postgres is ready, could connect to it");
-            break;
+        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;
+            }
         }
     }
 
-    //
-    // Decrypt connection string
-    //
-    let source_connection_string = {
-        match spec.encryption_secret {
-            EncryptionSecret::KMS { key_id } => {
-                let mut output = kms_client
-                    .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")?
-            }
-        }
-    };
-
-    //
-    // Start the work
-    //
+    let restore_pg_connstring = restore_pg_connstring.replace("dbname=postgres", "dbname=neondb");
 
     let dumpdir = working_directory.join("dumpdir");
 
@@ -310,6 +368,12 @@ pub(crate) async fn main() -> anyhow::Result<()> {
         }
     }
 
+    // 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(
@@ -325,21 +389,24 @@ pub(crate) async fn main() -> anyhow::Result<()> {
             .context("wait for postgres to shut down")?;
     }
 
-    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")?;
+        }
     }
 
     Ok(())

control_plane/storcon_cli/src/main.rs

@@ -9,8 +9,9 @@ use clap::{Parser, Subcommand};
 use pageserver_api::{
     controller_api::{
         AvailabilityZone, NodeAvailabilityWrapper, NodeDescribeResponse, NodeShardResponse,
-        SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest,
-        TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
+        SafekeeperDescribeResponse, SafekeeperSchedulingPolicyRequest, ShardSchedulingPolicy,
+        ShardsPreferredAzsRequest, SkSchedulingPolicy, TenantCreateRequest, TenantDescribeResponse,
+        TenantPolicyRequest,
     },
     models::{
         EvictionPolicy, EvictionPolicyLayerAccessThreshold, LocationConfigSecondary,
@@ -231,6 +232,13 @@ enum Command {
     },
     /// List safekeepers known to the storage controller
     Safekeepers {},
+    /// Set the scheduling policy of the specified safekeeper
+    SafekeeperScheduling {
+        #[arg(long)]
+        node_id: NodeId,
+        #[arg(long)]
+        scheduling_policy: SkSchedulingPolicyArg,
+    },
 }
 
 #[derive(Parser)]
@@ -283,6 +291,24 @@ impl FromStr for PlacementPolicyArg {
     }
 }
 
+#[derive(Debug, Clone)]
+struct SkSchedulingPolicyArg(SkSchedulingPolicy);
+
+impl FromStr for SkSchedulingPolicyArg {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "active" => Ok(Self(SkSchedulingPolicy::Active)),
+            "disabled" => Ok(Self(SkSchedulingPolicy::Disabled)),
+            "decomissioned" => Ok(Self(SkSchedulingPolicy::Decomissioned)),
+            _ => Err(anyhow::anyhow!(
+                "Unknown scheduling policy '{s}', try active,disabled,decomissioned"
+            )),
+        }
+    }
+}
+
 #[derive(Debug, Clone)]
 struct ShardSchedulingPolicyArg(ShardSchedulingPolicy);
 
@@ -477,16 +503,7 @@ async fn main() -> anyhow::Result<()> {
             println!("{table}");
         }
         Command::Tenants { node_id: None } => {
-            let mut resp = storcon_client
-                .dispatch::<(), Vec<TenantDescribeResponse>>(
-                    Method::GET,
-                    "control/v1/tenant".to_string(),
-                    None,
-                )
-                .await?;
-
-            resp.sort_by(|a, b| a.tenant_id.cmp(&b.tenant_id));
-
+            // Set up output formatting
             let mut table = comfy_table::Table::new();
             table.set_header([
                 "TenantId",
@@ -496,20 +513,55 @@ async fn main() -> anyhow::Result<()> {
                 "Placement",
                 "Scheduling",
             ]);
-            for tenant in resp {
-                let shard_zero = tenant.shards.into_iter().next().unwrap();
-                table.add_row([
-                    format!("{}", tenant.tenant_id),
-                    shard_zero
-                        .preferred_az_id
-                        .as_ref()
-                        .cloned()
-                        .unwrap_or("".to_string()),
-                    format!("{}", shard_zero.tenant_shard_id.shard_count.literal()),
-                    format!("{:?}", tenant.stripe_size),
-                    format!("{:?}", tenant.policy),
-                    format!("{:?}", shard_zero.scheduling_policy),
-                ]);
+
+            // Pagination loop over listing API
+            let mut start_after = None;
+            const LIMIT: usize = 1000;
+            loop {
+                let path = match start_after {
+                    None => format!("control/v1/tenant?limit={LIMIT}"),
+                    Some(start_after) => {
+                        format!("control/v1/tenant?limit={LIMIT}&start_after={start_after}")
+                    }
+                };
+
+                let resp = storcon_client
+                    .dispatch::<(), Vec<TenantDescribeResponse>>(Method::GET, path, None)
+                    .await?;
+
+                if resp.is_empty() {
+                    // End of data reached
+                    break;
+                }
+
+                // Give some visual feedback while we're building up the table (comfy_table doesn't have
+                // streaming output)
+                if resp.len() >= LIMIT {
+                    eprint!(".");
+                }
+
+                start_after = Some(resp.last().unwrap().tenant_id);
+
+                for tenant in resp {
+                    let shard_zero = tenant.shards.into_iter().next().unwrap();
+                    table.add_row([
+                        format!("{}", tenant.tenant_id),
+                        shard_zero
+                            .preferred_az_id
+                            .as_ref()
+                            .cloned()
+                            .unwrap_or("".to_string()),
+                        format!("{}", shard_zero.tenant_shard_id.shard_count.literal()),
+                        format!("{:?}", tenant.stripe_size),
+                        format!("{:?}", tenant.policy),
+                        format!("{:?}", shard_zero.scheduling_policy),
+                    ]);
+                }
+            }
+
+            // Terminate progress dots
+            if table.row_count() > LIMIT {
+                eprint!("");
             }
 
             println!("{table}");
@@ -1176,6 +1228,23 @@ async fn main() -> anyhow::Result<()> {
             }
             println!("{table}");
         }
+        Command::SafekeeperScheduling {
+            node_id,
+            scheduling_policy,
+        } => {
+            let scheduling_policy = scheduling_policy.0;
+            storcon_client
+                .dispatch::<SafekeeperSchedulingPolicyRequest, ()>(
+                    Method::POST,
+                    format!("control/v1/safekeeper/{node_id}/scheduling_policy"),
+                    Some(SafekeeperSchedulingPolicyRequest { scheduling_policy }),
+                )
+                .await?;
+            println!(
+                "Scheduling policy of {node_id} set to {}",
+                String::from(scheduling_policy)
+            );
+        }
     }
 
     Ok(())

docs/rfcs/035-safekeeper-dynamic-membership-change.md

@@ -81,7 +81,7 @@ configuration generation in them is less than its current one. Namely, it
 refuses to vote, to truncate WAL in `handle_elected` and to accept WAL. In
 response it sends its current configuration generation to let walproposer know.
 
-Safekeeper gets `PUT /v1/tenants/{tenant_id}/timelines/{timeline_id}/configuration` 
+Safekeeper gets `PUT /v1/tenants/{tenant_id}/timelines/{timeline_id}/configuration`
 accepting `Configuration`. Safekeeper switches to the given conf it is higher than its
 current one and ignores it otherwise. In any case it replies with
 ```
@@ -103,7 +103,7 @@ currently and tries to communicate with all of them. However, the list does not
 define consensus members. Instead, on start walproposer tracks highest
 configuration it receives from `AcceptorGreeting`s. Once it assembles greetings
 from majority of `sk_set` and majority of `new_sk_set` (if it is present), it
-establishes this configuration as its own and moves to voting. 
+establishes this configuration as its own and moves to voting.
 
 It should stop talking to safekeepers not listed in the configuration at this
 point, though it is not unsafe to continue doing so.
@@ -119,7 +119,7 @@ refusal to accept due to configuration change) it simply restarts.
 The following algorithm can be executed anywhere having access to configuration
 storage and safekeepers. It is safe to interrupt / restart it and run multiple
 instances of it concurrently, though likely one of them won't make
-progress then. It accepts `desired_set: Vec<NodeId>` as input. 
+progress then. It accepts `desired_set: Vec<NodeId>` as input.
 
 Algorithm will refuse to make the change if it encounters previous interrupted
 change attempt, but in this case it will try to finish it.
@@ -140,7 +140,7 @@ storage are reachable.
    safe. Failed CAS aborts the procedure.
 4) Call `PUT` `configuration` on safekeepers from the current set,
    delivering them `joint_conf`. Collecting responses from majority is required
-   to proceed. If any response returned generation higher than 
+   to proceed. If any response returned generation higher than
    `joint_conf.generation`, abort (another switch raced us). Otherwise, choose
    max `<last_log_term, flush_lsn>` among responses and establish it as
    (in memory) `sync_position`. Also choose max `term` and establish it as (in
@@ -149,49 +149,49 @@ storage are reachable.
    without ack from the new set. Similarly, we'll bump term on new majority
    to `sync_term` so that two computes with the same term are never elected.
 4) Initialize timeline on safekeeper(s) from `new_sk_set` where it
-   doesn't exist yet by doing `pull_timeline` from the majority of the 
+   doesn't exist yet by doing `pull_timeline` from the majority of the
    current set. Doing that on majority of `new_sk_set` is enough to
    proceed, but it is reasonable to ensure that all `new_sk_set` members
    are initialized -- if some of them are down why are we migrating there?
-5) Call `POST` `bump_term(sync_term)` on safekeepers from the new set. 
+5) Call `POST` `bump_term(sync_term)` on safekeepers from the new set.
    Success on majority is enough.
 6) Repeatedly call `PUT` `configuration` on safekeepers from the new set,
    delivering them `joint_conf` and collecting their positions. This will
-   switch them to the `joint_conf` which generally won't be needed 
+   switch them to the `joint_conf` which generally won't be needed
    because `pull_timeline` already includes it and plus additionally would be
    broadcast by compute. More importantly, we may proceed to the next step
-   only when `<last_log_term, flush_lsn>` on the majority of the new set reached 
-   `sync_position`. Similarly, on the happy path no waiting is not needed because 
+   only when `<last_log_term, flush_lsn>` on the majority of the new set reached
+   `sync_position`. Similarly, on the happy path no waiting is not needed because
    `pull_timeline` already includes it. However, we should double
     check to be safe. For example, timeline could have been created earlier e.g.
-    manually or after try-to-migrate, abort, try-to-migrate-again sequence. 
-7) Create `new_conf: Configuration` incrementing `join_conf` generation and having new 
-   safekeeper set as `sk_set` and None `new_sk_set`. Write it to configuration 
+    manually or after try-to-migrate, abort, try-to-migrate-again sequence.
+7) Create `new_conf: Configuration` incrementing `join_conf` generation and having new
+   safekeeper set as `sk_set` and None `new_sk_set`. Write it to configuration
    storage under one more CAS.
 8) Call `PUT` `configuration` on safekeepers from the new set,
-   delivering them `new_conf`. It is enough to deliver it to the majority 
+   delivering them `new_conf`. It is enough to deliver it to the majority
    of the new set; the rest can be updated by compute.
 
 I haven't put huge effort to make the description above very precise, because it
 is natural language prone to interpretations anyway. Instead I'd like to make TLA+
 spec of it.
 
-Description above focuses on safety. To make the flow practical and live, here a few more 
+Description above focuses on safety. To make the flow practical and live, here a few more
 considerations.
-1) It makes sense to ping new set to ensure it we are migrating to live node(s) before 
+1) It makes sense to ping new set to ensure it we are migrating to live node(s) before
   step 3.
-2) If e.g. accidentally wrong new sk set has been specified, before CAS in step `6` is completed 
+2) If e.g. accidentally wrong new sk set has been specified, before CAS in step `6` is completed
    it is safe to rollback to the old conf with one more CAS.
-3) On step 4 timeline might be already created on members of the new set for various reasons; 
+3) On step 4 timeline might be already created on members of the new set for various reasons;
    the simplest is the procedure restart. There are more complicated scenarious like mentioned
-   in step 5. Deleting and re-doing `pull_timeline` is generally unsafe without involving 
-   generations, so seems simpler to treat existing timeline as success. However, this also 
+   in step 5. Deleting and re-doing `pull_timeline` is generally unsafe without involving
+   generations, so seems simpler to treat existing timeline as success. However, this also
    has a disadvantage: you might imagine an surpassingly unlikely schedule where condition in
    the step 5 is never reached until compute is (re)awaken up to synchronize new member(s).
    I don't think we'll observe this in practice, but can add waking up compute if needed.
 4) In the end timeline should be locally deleted on the safekeeper(s) which are
    in the old set but not in the new one, unless they are unreachable. To be
-   safe this also should be done under generation number (deletion proceeds only if 
+   safe this also should be done under generation number (deletion proceeds only if
    current configuration is <= than one in request and safekeeper is not memeber of it).
 5) If current conf fetched on step 1 is already not joint and members equal to `desired_set`,
    jump to step 7, using it as `new_conf`.
@@ -202,47 +202,87 @@ The procedure ought to be driven from somewhere. Obvious candidates are control
 plane and storage_controller; and as each of them already has db we don't want
 yet another storage. I propose to manage safekeepers in storage_controller
 because 1) since it is in rust it simplifies simulation testing (more on this
-below) 2) it already manages pageservers. 
+below) 2) it already manages pageservers.
 
 This assumes that migration will be fully usable only after we migrate all
 tenants/timelines to storage_controller. It is discussible whether we want also
 to manage pageserver attachments for all of these, but likely we do.
 
-This requires us to define storcon <-> cplane interface.
+This requires us to define storcon <-> cplane interface and changes.
 
-### storage_controller <-> control plane interface
+### storage_controller <-> control plane interface and changes
 
 First of all, control plane should
 [change](https://neondb.slack.com/archives/C03438W3FLZ/p1719226543199829)
 storing safekeepers per timeline instead of per tenant because we can't migrate
-tenants atomically. 
+tenants atomically.
 
 The important question is how updated configuration is delivered from
 storage_controller to control plane to provide it to computes. As always, there
 are two options, pull and push. Let's do it the same push as with pageserver
 `/notify-attach` because 1) it keeps storage_controller out of critical compute
-start path 2) provides easier upgrade: there won't be such a thing as 'timeline
-managed by control plane / storcon', cplane just takes the value out of its db
-when needed 3) uniformity. It makes storage_controller responsible for retrying notifying
-control plane until it succeeds.
+start path 2) uniformity. It makes storage_controller responsible for retrying
+notifying control plane until it succeeds.
 
-So, cplane `/notify-safekeepers` for the timeline accepts `Configuration` and
-updates it in the db if the provided conf generation is higher (the cplane db
-should also store generations for this). Similarly to [`/notify-attach`](https://www.notion.so/neondatabase/Storage-Controller-Control-Plane-interface-6de56dd310a043bfa5c2f5564fa98365), it
-should update db which makes the call successful, and then try to schedule
-`apply_config` if possible, it is ok if not. storage_controller 
-should rate limit calling the endpoint, but likely this won't be needed, as migration
+It is not needed for the control plane to fully know the `Configuration`. It is
+enough for it to only to be aware of the list of safekeepers in the latest
+configuration to supply it to compute, plus associated generation number to
+protect from stale update requests and to also pass it to compute.
+
+So, cplane `/notify-safekeepers` for the timeline can accept JSON like
+```
+{
+   tenant_id: String,
+   timeline_id: String,
+   generation: u32,
+   safekeepers: Vec<SafekeeperId>,
+}
+```
+where `SafekeeperId` is
+```
+{
+   node_id: u64,
+   host: String
+}
+```
+In principle `host` is redundant, but may be useful for observability.
+
+The request updates list of safekeepers in the db if the provided conf
+generation is higher (the cplane db should also store generations for this).
+Similarly to
+[`/notify-attach`](https://www.notion.so/neondatabase/Storage-Controller-Control-Plane-interface-6de56dd310a043bfa5c2f5564fa98365),
+it should update db which makes the call successful, and then try to schedule
+`apply_config` if possible, it is ok if not. storage_controller should rate
+limit calling the endpoint, but likely this won't be needed, as migration
 throughput is limited by `pull_timeline`.
 
 Timeline (branch) creation in cplane should call storage_controller POST
 `tenant/:tenant_id/timeline` like it currently does for sharded tenants.
-Response should be augmented with `safekeeper_conf: Configuration`. The call
-should be retried until succeeds.
+Response should be augmented with `safekeepers_generation` and `safekeepers`
+fields like described in `/notify-safekeepers` above. Initially (currently)
+these fields may be absent; in this case cplane chooses safekeepers on its own
+like it currently does. The call should be retried until succeeds.
 
 Timeline deletion and tenant deletion in cplane should call appropriate
 storage_controller endpoints like it currently does for sharded tenants. The
 calls should be retried until they succeed.
 
+When compute receives safekeepers list from control plane it needs to know the
+generation to checked whether it should be updated (note that compute may get
+safekeeper list from either cplane or safekeepers). Currently `neon.safekeepers`
+GUC is just a comma separates list of `host:port`. Let's prefix it with
+`g#<generation>:` to this end, so it will look like
+```
+g#42:safekeeper-0.eu-central-1.aws.neon.tech:6401,safekeeper-2.eu-central-1.aws.neon.tech:6401,safekeeper-1.eu-central-1.aws.neon.tech:6401
+```
+
+To summarize, list of cplane changes:
+- per tenant -> per timeline safekeepers management and addition of int `safekeeper_generation` field.
+- `/notify-safekeepers` endpoint.
+- Branch creation call may return list of safekeepers and when it is
+  present cplane should adopt it instead of choosing on its own like it does currently.
+- `neon.safekeepers` GUC should be prefixed with `g#<generation>:`.
+
 ### storage_controller implementation
 
 Current 'load everything on startup and keep in memory' easy design is fine.
@@ -360,10 +400,10 @@ source safekeeper might fail, which is not a problem if we are going to
 decomission the node but leaves garbage otherwise. I'd propose in the first version
 1) Don't attempt deletion at all if node status is `offline`.
 2) If it failed, just issue warning.
-And add PUT `/control/v1/safekeepers/:node_id/scrub` endpoint which would find and 
-remove garbage timelines for manual use. It will 1) list all timelines on the 
-safekeeper 2) compare each one against configuration storage: if timeline 
-doesn't exist at all (had been deleted), it can be deleted. Otherwise, it can 
+And add PUT `/control/v1/safekeepers/:node_id/scrub` endpoint which would find and
+remove garbage timelines for manual use. It will 1) list all timelines on the
+safekeeper 2) compare each one against configuration storage: if timeline
+doesn't exist at all (had been deleted), it can be deleted. Otherwise, it can
 be deleted under generation number if node is not member of current generation.
 
 Automating this is untrivial; we'd need to register all potential missing
@@ -412,8 +452,8 @@ There should be following layers of tests:
 3) Since simulation testing injects at relatively high level points (not
    syscalls), it omits some code, in particular `pull_timeline`. Thus it is
    better to have basic tests covering whole system as well. Extended version of
-   `test_restarts_under_load` would do: start background load and do migration 
-   under it, then restart endpoint and check that no reported commits 
+   `test_restarts_under_load` would do: start background load and do migration
+   under it, then restart endpoint and check that no reported commits
    had been lost. I'd also add one more creating classic network split scenario, with
    one compute talking to AC and another to BD while migration from nodes ABC to ABD
    happens.
@@ -422,35 +462,51 @@ There should be following layers of tests:
 
 ## Order of implementation and rollout
 
-Note that 
+Note that
 - Control plane parts and integration with it is fully independent from everything else
   (tests would use simulation and neon_local).
+- It is reasonable to make compute <-> safekeepers protocol change
+  independent of enabling generations.
 - There is a lot of infra work making storage_controller aware of timelines and safekeepers
   and its impl/rollout should be separate from migration itself.
-- Initially walproposer can just stop working while it observers joint configuration.
+- Initially walproposer can just stop working while it observes joint configuration.
   Such window would be typically very short anyway.
+- Obviously we want to test the whole thing thoroughly on staging and only then
+  gradually enable in prod.
 
-To rollout smoothly, both walproposer and safekeeper should have flag
-`configurations_enabled`; when set to false, they would work as currently, i.e.
-walproposer is able to commit on whatever safekeeper set it is provided. Until
-all timelines are managed by storcon we'd need to use current script to migrate
-and update/drop entries in the storage_controller database if it has any.
+Let's have the following implementation bits for gradual rollout:
+- compute gets `neon.safekeepers_proto_version` flag.
+  Initially both compute and safekeepers will be able to talk both
+  versions so that we can delay force restart of them and for
+  simplicity of rollback in case it is needed.
+- storcon gets `-set-safekeepers` config option disabled by
+  default. Timeline creation request chooses safekeepers
+  (and returns them in response to cplane) only when it is set to
+  true.
+- control_plane [see above](storage_controller-<->-control-plane interface-and-changes)
+  prefixes `neon.safekeepers` GUC with generation number. When it is 0
+  (or prefix not present at all), walproposer behaves as currently, committing on
+  the provided safekeeper list -- generations are disabled.
+  If it is non 0 it follows this RFC rules.
+- We provide a script for manual migration to storage controller.
+  It selects timeline(s) from control plane (specified or all of them) db
+  and calls special import endpoint on storage controller which is very
+  similar to timeline creation: it inserts into the db, sets
+  configuration to initial on the safekeepers, calls cplane
+  `notify-safekeepers`.
 
-Safekeepers would need to be able to talk both current and new protocol version
-with compute to reduce number of computes restarted in prod once v2 protocol is
-deployed (though before completely switching we'd need to force this).
-
-Let's have the following rollout order:
-- storage_controller becomes aware of safekeepers;
-- storage_controller gets timeline creation for new timelines and deletion requests, but
-  doesn't manage all timelines yet. Migration can be tested on these new timelines.
-  To keep control plane and storage_controller databases in sync while control 
-  plane still chooses the safekeepers initially (until all timelines are imported
-  it can choose better), `TimelineCreateRequest` can get optional safekeepers
-  field with safekeepers chosen by cplane.
-- Then we can import all existing timelines from control plane to
-  storage_controller and gradually enable configurations region by region.
+Then the rollout for a region would be:
+- Current situation: safekeepers are choosen by control_plane.
+- We manually migrate some timelines, test moving them around.
+- Then we enable `--set-safekeepers` so that all new timelines
+  are on storage controller.
+- Finally migrate all existing timelines using the script (no
+  compute should be speaking old proto version at this point).
 
+Until all timelines are managed by storcon we'd need to use current ad hoc
+script to migrate if needed. To keep state clean, all storage controller managed
+timelines must be migrated before that, or controller db and configurations
+state of safekeepers dropped manually.
 
 Very rough implementation order:
 - Add concept of configurations to safekeepers (including control file),
@@ -458,10 +514,10 @@ Very rough implementation order:
 - Implement walproposer changes, including protocol.
 - Implement storconn part. Use it in neon_local (and pytest).
 - Make cplane store safekeepers per timeline instead of per tenant.
-- Implement cplane/storcon integration. Route branch creation/deletion 
+- Implement cplane/storcon integration. Route branch creation/deletion
   through storcon. Then we can test migration of new branches.
-- Finally import existing branches. Then we can drop cplane 
-  safekeeper selection code. Gradually enable configurations at 
+- Finally import existing branches. Then we can drop cplane
+  safekeeper selection code. Gradually enable configurations at
   computes and safekeepers. Before that, all computes must talk only
   v3 protocol version.
 

docs/rfcs/040-profiling.md

@@ -0,0 +1,247 @@
+# CPU and Memory Profiling
+
+Created 2025-01-12 by Erik Grinaker.
+
+See also [internal user guide](https://www.notion.so/neondatabase/Storage-CPU-Memory-Profiling-14bf189e004780228ec7d04442742324?pvs=4).
+
+## Summary
+
+This document proposes a standard cross-team pattern for CPU and memory profiling across
+applications and languages, using the [pprof](https://github.com/google/pprof) profile format.
+
+It enables both ad hoc profiles via HTTP endpoints, and continuous profiling across the fleet via
+[Grafana Cloud Profiles](https://grafana.com/docs/grafana-cloud/monitor-applications/profiles/).
+Continuous profiling incurs an overhead of about 0.1% CPU usage and 3% slower heap allocations.
+
+## Motivation
+
+CPU and memory profiles are crucial observability tools for understanding performance issues,
+resource exhaustion, and resource costs. They allow answering questions like:
+
+* Why is this process using 100% CPU?
+* How do I make this go faster?
+* Why did this process run out of memory?
+* Why are we paying for all these CPU cores and memory chips?
+
+Go has [first-class support](https://pkg.go.dev/net/http/pprof) for profiling included in its
+standard library, using the [pprof](https://github.com/google/pprof) profile format and associated
+tooling.
+
+This is not the case for Rust and C, where obtaining profiles can be rather cumbersome. It requires
+installing and running additional tools like `perf` as root on production nodes, with analysis tools
+that can be hard to use and often don't give good results. This is not only annoying, but can also
+significantly affect the resolution time of production incidents.
+
+This proposal will:
+
+* Provide CPU and heap profiles in pprof format via HTTP API.
+* Record continuous profiles in Grafana for aggregate historical analysis.
+* Make it easy for anyone to see a flamegraph in less than one minute.
+* Be reasonably consistent across teams and services (Rust, Go, C).
+
+## Non Goals (For Now)
+
+* [Additional profile types](https://grafana.com/docs/pyroscope/next/configure-client/profile-types/)
+  like mutexes, locks, goroutines, etc.
+* [Runtime trace integration](https://grafana.com/docs/pyroscope/next/configure-client/trace-span-profiles/).
+* [Profile-guided optimization](https://en.wikipedia.org/wiki/Profile-guided_optimization).
+
+## Using Profiles
+
+Ready-to-use profiles can be obtained using e.g. `curl`. For Rust services:
+
+```
+$ curl localhost:9898/profile/cpu >profile.pb.gz
+```
+
+pprof profiles can be explored using the [`pprof`](https://github.com/google/pprof) web UI, which
+provides flamegraphs, call graphs, plain text listings, and more:
+
+```
+$ pprof -http :6060 <profile>
+```
+
+Some endpoints (e.g. Rust-based ones) can also generate flamegraph SVGs directly:
+
+```
+$ curl localhost:9898/profile/cpu?format=svg >profile.svg
+$ open profile.svg
+```
+
+Continuous profiles are available in Grafana under Explore → Profiles → Explore Profiles
+(currently only in [staging](https://neonstaging.grafana.net/a/grafana-pyroscope-app/profiles-explorer)).
+
+## API Requirements
+
+* HTTP endpoints that return a profile in pprof format (with symbols).
+  * CPU: records a profile over the request time interval (`seconds` query parameter).
+  * Memory: returns the current in-use heap allocations.
+* Unauthenticated, as it should not expose user data or pose a denial-of-service risk.
+* Default sample frequency should not impact service (maximum 5% CPU overhead).
+* Linux-compatibility.
+
+Nice to have:
+
+* Return flamegraph SVG directly from the HTTP endpoint if requested.
+* Configurable sample frequency for CPU profiles.
+* Historical heap allocations, by count and bytes.
+* macOS-compatiblity.
+
+## Rust Profiling
+
+[`libs/utils/src/http/endpoint.rs`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs)
+contains ready-to-use HTTP endpoints for CPU and memory profiling:
+[`profile_cpu_handler`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs#L338) and [`profile_heap_handler`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs#L416).
+
+### CPU
+
+CPU profiles are provided by [pprof-rs](https://github.com/tikv/pprof-rs) via
+[`profile_cpu_handler`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs#L338).
+Expose it unauthenticated at `/profile/cpu`.
+
+Parameters:
+
+* `format`: profile output format (`pprof` or `svg`; default `pprof`).
+* `seconds`: duration to collect profile over, in seconds (default `5`).
+* `frequency`: how often to sample thread stacks, in Hz (default `99`).
+* `force`: if `true`, cancel a running profile and start a new one (default `false`).
+
+Works on Linux and macOS.
+
+### Memory
+
+Use the jemalloc allocator via [`tikv-jemallocator`](https://github.com/tikv/jemallocator),
+and enable profiling with samples every 2 MB allocated:
+
+```rust
+#[global_allocator]
+static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
+
+#[allow(non_upper_case_globals)]
+#[export_name = "malloc_conf"]
+pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:21\0";
+```
+
+pprof profiles are generated by
+[`jemalloc-pprof`](https://github.com/polarsignals/rust-jemalloc-pprof) via
+[`profile_heap_handler`](https://github.com/neondatabase/neon/blob/8327f68043e692c77f70d6a6dafa463636c01578/libs/utils/src/http/endpoint.rs#L416).
+Expose it unauthenticated at `/profile/heap`.
+
+Parameters:
+
+* `format`: profile output format (`pprof`, `svg`, or `jemalloc`; default `pprof`).
+
+Works on Linux only, due to [jemalloc limitations](https://github.com/jemalloc/jemalloc/issues/26).
+
+## Go Profiling
+
+The Go standard library includes pprof profiling via HTTP API in
+[`net/http/pprof`](https://pkg.go.dev/net/http/pprof). Expose it unauthenticated at
+`/debug/pprof`.
+
+Works on Linux and macOS.
+
+### CPU 
+
+Via `/debug/pprof/profile`. Parameters:
+
+* `debug`: profile output format (`0` is pprof, `1` or above is plaintext; default `0`).
+* `seconds`: duration to collect profile over, in seconds (default `30`).
+
+Does not support a frequency parameter (see [#57488](https://github.com/golang/go/issues/57488)),
+and defaults to 100 Hz. A lower frequency can be hardcoded via `SetCPUProfileRate`, but the default
+is likely ok (estimated 1% overhead).
+
+### Memory
+
+Via `/debug/pprof/heap`. Parameters:
+
+* `seconds`: take a delta profile over the given duration, in seconds (default `0`).
+* `gc`: if `1`, garbage collect before taking profile.
+
+## C Profiling
+
+[gperftools](https://github.com/gperftools/gperftools) provides in-process CPU and heap profiling
+with pprof output.
+
+However, continuous profiling of PostgreSQL is expensive (many computes), and has limited value
+since we don't own the internals anyway.
+
+Ad hoc profiling might still be useful, but the compute team considers existing tooling sufficient,
+so this is not a priority at the moment.
+
+## Grafana Continuous Profiling
+
+[Grafana Alloy](https://grafana.com/docs/alloy/latest/) continually scrapes CPU and memory profiles
+across the fleet, and archives them as time series. This can be used to analyze resource usage over
+time, either in aggregate or zoomed in to specific events and nodes.
+
+Profiles are retained for 30 days. Profile ingestion volume for CPU+heap at 60-second intervals
+is about 0.5 GB/node/day, or about $0.25/node/day = $7.5/node/month ($0.50/GB).
+
+It is currently enabled in [staging](https://neonstaging.grafana.net/a/grafana-pyroscope-app/profiles-explorer)
+for Pageserver and Safekeeper.
+
+### Scraping
+
+* CPU profiling: 59 seconds at 19 Hz every 60 seconds.
+* Heap profiling: heap snapshot with 2 MB frequency every 60 seconds.
+
+There are two main approaches that can be taken for CPU profiles:
+
+* Continuous low-frequency profiles (e.g. 19 Hz for 60 seconds every 60 seconds).
+* Occasional high-frequency profiles (e.g. 99 Hz for 5 seconds every 60 seconds).
+
+We choose continuous low-frequency profiles where possible. This has a fixed low overhead, instead
+of a spiky high overhead. It likely also gives a more representative view of resource usage.
+However, a 19 Hz rate gives a minimum resolution of 52.6 ms per sample, which may be larger than the
+actual runtime of small functions. Note that Go does not support a frequency parameter, so we must
+use a fixed frequency for all profiles via `SetCPUProfileRate()` (default 100 Hz).
+
+Only one CPU profile can be taken at a time. With continuous profiling, one will always be running.
+To allow also taking an ad hoc CPU profile, the Rust endpoint supports a `force` query parameter to
+cancel a running profile and start a new one.
+
+### Overhead
+
+With Rust:
+
+* CPU profiles at 19 Hz frequency: 0.1% overhead.
+* Heap profiles at 2 MB frequency: 3% allocation overhead.
+* Profile call/encoding/symbolization: 20 ms every 60 seconds, or 0.03% of 1 CPU (for Pageserver).
+* Profile symbolization caches: 125 MB memory, or 0.4% of 32 GB (for Pageserver).
+
+Benchmarks with pprof-rs showed that the CPU time for taking a stack trace of a 40-frame stack was
+11 µs using the `frame-pointer` feature, and 1.4 µs using `libunwind` with DWARF. `libunwind` saw
+frequent seg faults, so we use `frame-pointer` and build binaries with frame pointers (negligible
+overhead).
+
+CPU profiles work by installing an `ITIMER_PROF` for the process, which triggers a `SIGPROF` signal
+after a given amount of cumulative CPU time across all CPUs. The signal handler will run for one
+of the currently executing threads and take a stack trace. Thus, a 19 Hz profile will take 1 stack
+trace every 52.6 ms CPU time -- assuming 11 µs for a stack trace, this is 0.02% overhead, but
+likely 0.1% in practice (given e.g. context switches).
+
+Heap profiles work by probabilistically taking a stack trace on allocations, adjusted for the
+allocation size. A 1 MB allocation takes about 15 µs in benchmarks, and a stack trace about 1 µs,
+so we can estimate that a 2 MB sampling frequency has about 3% allocation overhead -- this is 
+consistent with benchmarks. This is significantly larger than CPU profiles, but mitigated by the
+fact that performance-sensitive code will avoid allocations as far as possible.
+
+Profile symbolization uses in-memory caches for symbol lookups. These take about 125 MB for
+Pageserver.
+
+## Alternatives Considered
+
+* eBPF profiles.
+  * Don't require instrumenting the binary.
+  * Use less resources.
+  * Can profile in kernel space too.
+  * Supported by Grafana.
+  * Less information about stack frames and spans.
+  * Limited tooling for local analysis.
+  * Does not support heap profiles.
+  * Does not work on macOS.
+
+* [Polar Signals](https://www.polarsignals.com) instead of Grafana.
+  * We already use Grafana for everything else. Appears good enough.

libs/pageserver_api/src/controller_api.rs

@@ -416,8 +416,6 @@ pub struct MetadataHealthListOutdatedResponse {
 }
 
 /// Publicly exposed safekeeper description
-///
-/// The `active` flag which we have in the DB is not included on purpose: it is deprecated.
 #[derive(Serialize, Deserialize, Clone)]
 pub struct SafekeeperDescribeResponse {
     pub id: NodeId,
@@ -433,6 +431,11 @@ pub struct SafekeeperDescribeResponse {
     pub scheduling_policy: SkSchedulingPolicy,
 }
 
+#[derive(Serialize, Deserialize, Clone)]
+pub struct SafekeeperSchedulingPolicyRequest {
+    pub scheduling_policy: SkSchedulingPolicy,
+}
+
 #[cfg(test)]
 mod test {
     use super::*;

libs/pageserver_api/src/key.rs

@@ -24,7 +24,9 @@ pub struct Key {
 
 /// When working with large numbers of Keys in-memory, it is more efficient to handle them as i128 than as
 /// a struct of fields.
-#[derive(Clone, Copy, Hash, PartialEq, Eq, Ord, PartialOrd, Serialize, Deserialize, Debug)]
+#[derive(
+    Clone, Copy, Default, Hash, PartialEq, Eq, Ord, PartialOrd, Serialize, Deserialize, Debug,
+)]
 pub struct CompactKey(i128);
 
 /// The storage key size.

libs/pageserver_api/src/models.rs

@@ -29,7 +29,7 @@ use utils::{
 };
 
 use crate::{
-    key::Key,
+    key::{CompactKey, Key},
     reltag::RelTag,
     shard::{ShardCount, ShardStripeSize, TenantShardId},
 };
@@ -1400,8 +1400,6 @@ pub enum PagestreamFeMessage {
     GetPage(PagestreamGetPageRequest),
     DbSize(PagestreamDbSizeRequest),
     GetSlruSegment(PagestreamGetSlruSegmentRequest),
-    #[cfg(feature = "testing")]
-    Test(PagestreamTestRequest),
 }
 
 // Wrapped in libpq CopyData
@@ -1413,8 +1411,6 @@ pub enum PagestreamBeMessage {
     Error(PagestreamErrorResponse),
     DbSize(PagestreamDbSizeResponse),
     GetSlruSegment(PagestreamGetSlruSegmentResponse),
-    #[cfg(feature = "testing")]
-    Test(PagestreamTestResponse),
 }
 
 // Keep in sync with `pagestore_client.h`
@@ -1426,9 +1422,6 @@ enum PagestreamBeMessageTag {
     Error = 103,
     DbSize = 104,
     GetSlruSegment = 105,
-    /// Test message discrimimant is unstable
-    #[cfg(feature = "testing")]
-    Test = 106,
 }
 impl TryFrom<u8> for PagestreamBeMessageTag {
     type Error = u8;
@@ -1440,8 +1433,6 @@ impl TryFrom<u8> for PagestreamBeMessageTag {
             103 => Ok(PagestreamBeMessageTag::Error),
             104 => Ok(PagestreamBeMessageTag::DbSize),
             105 => Ok(PagestreamBeMessageTag::GetSlruSegment),
-            #[cfg(feature = "testing")]
-            106 => Ok(PagestreamBeMessageTag::Test),
             _ => Err(value),
         }
     }
@@ -1559,20 +1550,6 @@ pub struct PagestreamDbSizeResponse {
     pub db_size: i64,
 }
 
-#[cfg(feature = "testing")]
-#[derive(Debug, PartialEq, Eq, Clone)]
-pub struct PagestreamTestRequest {
-    pub hdr: PagestreamRequest,
-    pub batch_key: u64,
-    pub message: String,
-}
-
-#[cfg(feature = "testing")]
-#[derive(Debug)]
-pub struct PagestreamTestResponse {
-    pub req: PagestreamTestRequest,
-}
-
 // This is a cut-down version of TenantHistorySize from the pageserver crate, omitting fields
 // that require pageserver-internal types.  It is sufficient to get the total size.
 #[derive(Serialize, Deserialize, Debug)]
@@ -1641,17 +1618,6 @@ impl PagestreamFeMessage {
                 bytes.put_u8(req.kind);
                 bytes.put_u32(req.segno);
             }
-            #[cfg(feature = "testing")]
-            Self::Test(req) => {
-                bytes.put_u8(5);
-                bytes.put_u64(req.hdr.reqid);
-                bytes.put_u64(req.hdr.request_lsn.0);
-                bytes.put_u64(req.hdr.not_modified_since.0);
-                bytes.put_u64(req.batch_key);
-                let message = req.message.as_bytes();
-                bytes.put_u64(message.len() as u64);
-                bytes.put_slice(message);
-            }
         }
 
         bytes.into()
@@ -1739,21 +1705,6 @@ impl PagestreamFeMessage {
                     segno: body.read_u32::<BigEndian>()?,
                 },
             )),
-            #[cfg(feature = "testing")]
-            5 => Ok(PagestreamFeMessage::Test(PagestreamTestRequest {
-                hdr: PagestreamRequest {
-                    reqid,
-                    request_lsn,
-                    not_modified_since,
-                },
-                batch_key: body.read_u64::<BigEndian>()?,
-                message: {
-                    let len = body.read_u64::<BigEndian>()?;
-                    let mut buf = vec![0; len as usize];
-                    body.read_exact(&mut buf)?;
-                    String::from_utf8(buf)?
-                },
-            })),
             _ => bail!("unknown smgr message tag: {:?}", msg_tag),
         }
     }
@@ -1797,15 +1748,6 @@ impl PagestreamBeMessage {
                         bytes.put_u32((resp.segment.len() / BLCKSZ as usize) as u32);
                         bytes.put(&resp.segment[..]);
                     }
-
-                    #[cfg(feature = "testing")]
-                    Self::Test(resp) => {
-                        bytes.put_u8(Tag::Test as u8);
-                        bytes.put_u64(resp.req.batch_key);
-                        let message = resp.req.message.as_bytes();
-                        bytes.put_u64(message.len() as u64);
-                        bytes.put_slice(message);
-                    }
                 }
             }
             PagestreamProtocolVersion::V3 => {
@@ -1874,18 +1816,6 @@ impl PagestreamBeMessage {
                         bytes.put_u32((resp.segment.len() / BLCKSZ as usize) as u32);
                         bytes.put(&resp.segment[..]);
                     }
-
-                    #[cfg(feature = "testing")]
-                    Self::Test(resp) => {
-                        bytes.put_u8(Tag::Test as u8);
-                        bytes.put_u64(resp.req.hdr.reqid);
-                        bytes.put_u64(resp.req.hdr.request_lsn.0);
-                        bytes.put_u64(resp.req.hdr.not_modified_since.0);
-                        bytes.put_u64(resp.req.batch_key);
-                        let message = resp.req.message.as_bytes();
-                        bytes.put_u64(message.len() as u64);
-                        bytes.put_slice(message);
-                    }
                 }
             }
         }
@@ -2028,28 +1958,6 @@ impl PagestreamBeMessage {
                         segment: segment.into(),
                     })
                 }
-                #[cfg(feature = "testing")]
-                Tag::Test => {
-                    let reqid = buf.read_u64::<BigEndian>()?;
-                    let request_lsn = Lsn(buf.read_u64::<BigEndian>()?);
-                    let not_modified_since = Lsn(buf.read_u64::<BigEndian>()?);
-                    let batch_key = buf.read_u64::<BigEndian>()?;
-                    let len = buf.read_u64::<BigEndian>()?;
-                    let mut msg = vec![0; len as usize];
-                    buf.read_exact(&mut msg)?;
-                    let message = String::from_utf8(msg)?;
-                    Self::Test(PagestreamTestResponse {
-                        req: PagestreamTestRequest {
-                            hdr: PagestreamRequest {
-                                reqid,
-                                request_lsn,
-                                not_modified_since,
-                            },
-                            batch_key,
-                            message,
-                        },
-                    })
-                }
             };
         let remaining = buf.into_inner();
         if !remaining.is_empty() {
@@ -2069,8 +1977,23 @@ impl PagestreamBeMessage {
             Self::Error(_) => "Error",
             Self::DbSize(_) => "DbSize",
             Self::GetSlruSegment(_) => "GetSlruSegment",
-            #[cfg(feature = "testing")]
-            Self::Test(_) => "Test",
+        }
+    }
+}
+
+#[derive(Debug, Serialize, Deserialize)]
+pub struct PageTraceEvent {
+    pub key: CompactKey,
+    pub effective_lsn: Lsn,
+    pub time: SystemTime,
+}
+
+impl Default for PageTraceEvent {
+    fn default() -> Self {
+        Self {
+            key: Default::default(),
+            effective_lsn: Default::default(),
+            time: std::time::UNIX_EPOCH,
         }
     }
 }

libs/pageserver_api/src/shard.rs

@@ -31,6 +31,8 @@
 //! - In a tenant with 4 shards, each shard has ShardCount(N), ShardNumber(i) where i in 0..N-1 (inclusive),
 //!   and their slugs are 0004, 0104, 0204, and 0304.
 
+use std::hash::{Hash, Hasher};
+
 use crate::{key::Key, models::ShardParameters};
 use postgres_ffi::relfile_utils::INIT_FORKNUM;
 use serde::{Deserialize, Serialize};
@@ -48,6 +50,23 @@ pub struct ShardIdentity {
     layout: ShardLayout,
 }
 
+/// Hash implementation
+///
+/// The stripe size cannot change dynamically, so it can be ignored for efficiency reasons.
+impl Hash for ShardIdentity {
+    fn hash<H: Hasher>(&self, state: &mut H) {
+        let ShardIdentity {
+            number,
+            count,
+            stripe_size: _,
+            layout: _,
+        } = self;
+
+        number.0.hash(state);
+        count.0.hash(state);
+    }
+}
+
 /// Stripe size in number of pages
 #[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Debug)]
 pub struct ShardStripeSize(pub u32);
@@ -59,7 +78,7 @@ impl Default for ShardStripeSize {
 }
 
 /// Layout version: for future upgrades where we might change how the key->shard mapping works
-#[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Debug)]
+#[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Hash, Debug)]
 pub struct ShardLayout(u8);
 
 const LAYOUT_V1: ShardLayout = ShardLayout(1);

libs/postgres_ffi/src/walrecord.rs

@@ -16,7 +16,7 @@ use utils::bin_ser::DeserializeError;
 use utils::lsn::Lsn;
 
 #[repr(C)]
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct XlMultiXactCreate {
     pub mid: MultiXactId,
     /* new MultiXact's ID */
@@ -46,7 +46,7 @@ impl XlMultiXactCreate {
 }
 
 #[repr(C)]
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct XlMultiXactTruncate {
     pub oldest_multi_db: Oid,
     /* to-be-truncated range of multixact offsets */
@@ -72,7 +72,7 @@ impl XlMultiXactTruncate {
 }
 
 #[repr(C)]
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct XlRelmapUpdate {
     pub dbid: Oid,   /* database ID, or 0 for shared map */
     pub tsid: Oid,   /* database's tablespace, or pg_global */
@@ -90,7 +90,7 @@ impl XlRelmapUpdate {
 }
 
 #[repr(C)]
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct XlReploriginDrop {
     pub node_id: RepOriginId,
 }
@@ -104,7 +104,7 @@ impl XlReploriginDrop {
 }
 
 #[repr(C)]
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct XlReploriginSet {
     pub remote_lsn: Lsn,
     pub node_id: RepOriginId,
@@ -911,7 +911,7 @@ impl XlSmgrCreate {
 }
 
 #[repr(C)]
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct XlSmgrTruncate {
     pub blkno: BlockNumber,
     pub rnode: RelFileNode,
@@ -984,7 +984,7 @@ impl XlDropDatabase {
 /// xl_xact_parsed_abort structs in PostgreSQL, but we use the same
 /// struct for commits and aborts.
 ///
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct XlXactParsedRecord {
     pub xid: TransactionId,
     pub info: u8,

libs/safekeeper_api/Cargo.toml

@@ -5,9 +5,12 @@ edition.workspace = true
 license.workspace = true
 
 [dependencies]
+anyhow.workspace = true
 const_format.workspace = true
 serde.workspace = true
+serde_json.workspace = true
 postgres_ffi.workspace = true
 pq_proto.workspace = true
 tokio.workspace = true
 utils.workspace = true
+pageserver_api.workspace = true

libs/safekeeper_api/src/lib.rs

@@ -4,12 +4,15 @@ use const_format::formatcp;
 use pq_proto::SystemId;
 use serde::{Deserialize, Serialize};
 
+pub mod membership;
 /// Public API types
 pub mod models;
 
 /// Consensus logical timestamp. Note: it is a part of sk control file.
 pub type Term = u64;
-pub const INVALID_TERM: Term = 0;
+/// With this term timeline is created initially. It
+/// is a normal term except wp is never elected with it.
+pub const INITIAL_TERM: Term = 0;
 
 /// Information about Postgres. Safekeeper gets it once and then verifies all
 /// further connections from computes match. Note: it is a part of sk control

libs/safekeeper_api/src/membership.rs

@@ -0,0 +1,166 @@
+//! Types defining safekeeper membership, see
+//! rfcs/035-safekeeper-dynamic-membership-change.md
+//! for details.
+
+use std::{collections::HashSet, fmt::Display};
+
+use anyhow;
+use anyhow::bail;
+use serde::{Deserialize, Serialize};
+use utils::id::NodeId;
+
+/// Number uniquely identifying safekeeper configuration.
+/// Note: it is a part of sk control file.
+pub type Generation = u32;
+/// 1 is the first valid generation, 0 is used as
+/// a placeholder before we fully migrate to generations.
+pub const INVALID_GENERATION: Generation = 0;
+pub const INITIAL_GENERATION: Generation = 1;
+
+/// Membership is defined by ids so e.g. walproposer uses them to figure out
+/// quorums, but we also carry host and port to give wp idea where to connect.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct SafekeeperId {
+    pub id: NodeId,
+    pub host: String,
+    /// We include here only port for computes -- that is, pg protocol tenant
+    /// only port, or wide pg protocol port if the former is not configured.
+    pub pg_port: u16,
+}
+
+impl Display for SafekeeperId {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "[id={}, ep={}:{}]", self.id, self.host, self.pg_port)
+    }
+}
+
+/// Set of safekeepers.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(transparent)]
+pub struct MemberSet {
+    pub members: Vec<SafekeeperId>,
+}
+
+impl MemberSet {
+    pub fn empty() -> Self {
+        MemberSet {
+            members: Vec::new(),
+        }
+    }
+
+    pub fn new(members: Vec<SafekeeperId>) -> anyhow::Result<Self> {
+        let hs: HashSet<NodeId> = HashSet::from_iter(members.iter().map(|sk| sk.id));
+        if hs.len() != members.len() {
+            bail!("duplicate safekeeper id in the set {:?}", members);
+        }
+        Ok(MemberSet { members })
+    }
+
+    pub fn contains(&self, sk: &SafekeeperId) -> bool {
+        self.members.iter().any(|m| m.id == sk.id)
+    }
+
+    pub fn add(&mut self, sk: SafekeeperId) -> anyhow::Result<()> {
+        if self.contains(&sk) {
+            bail!(format!(
+                "sk {} is already member of the set {}",
+                sk.id, self
+            ));
+        }
+        self.members.push(sk);
+        Ok(())
+    }
+}
+
+impl Display for MemberSet {
+    /// Display as a comma separated list of members.
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let sks_str = self
+            .members
+            .iter()
+            .map(|m| m.to_string())
+            .collect::<Vec<_>>();
+        write!(f, "({})", sks_str.join(", "))
+    }
+}
+
+/// Safekeeper membership configuration.
+/// Note: it is a part of both control file and http API.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct Configuration {
+    /// Unique id.
+    pub generation: Generation,
+    /// Current members of the configuration.
+    pub members: MemberSet,
+    /// Some means it is a joint conf.
+    pub new_members: Option<MemberSet>,
+}
+
+impl Configuration {
+    /// Used for pre-generations timelines, will be removed eventually.
+    pub fn empty() -> Self {
+        Configuration {
+            generation: INVALID_GENERATION,
+            members: MemberSet::empty(),
+            new_members: None,
+        }
+    }
+}
+
+impl Display for Configuration {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "gen={}, members={}, new_members={}",
+            self.generation,
+            self.members,
+            self.new_members
+                .as_ref()
+                .map(ToString::to_string)
+                .unwrap_or(String::from("none"))
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{MemberSet, SafekeeperId};
+    use utils::id::NodeId;
+
+    #[test]
+    fn test_member_set() {
+        let mut members = MemberSet::empty();
+        members
+            .add(SafekeeperId {
+                id: NodeId(42),
+                host: String::from("lala.org"),
+                pg_port: 5432,
+            })
+            .unwrap();
+
+        members
+            .add(SafekeeperId {
+                id: NodeId(42),
+                host: String::from("lala.org"),
+                pg_port: 5432,
+            })
+            .expect_err("duplicate must not be allowed");
+
+        members
+            .add(SafekeeperId {
+                id: NodeId(43),
+                host: String::from("bubu.org"),
+                pg_port: 5432,
+            })
+            .unwrap();
+
+        println!("members: {}", members);
+
+        let j = serde_json::to_string(&members).expect("failed to serialize");
+        println!("members json: {}", j);
+        assert_eq!(
+            j,
+            r#"[{"id":42,"host":"lala.org","pg_port":5432},{"id":43,"host":"bubu.org","pg_port":5432}]"#
+        );
+    }
+}

libs/safekeeper_api/src/models.rs

@@ -1,5 +1,6 @@
 //! Types used in safekeeper http API. Many of them are also reused internally.
 
+use pageserver_api::shard::ShardIdentity;
 use postgres_ffi::TimestampTz;
 use serde::{Deserialize, Serialize};
 use std::net::SocketAddr;
@@ -11,7 +12,7 @@ use utils::{
     pageserver_feedback::PageserverFeedback,
 };
 
-use crate::{ServerInfo, Term};
+use crate::{membership::Configuration, ServerInfo, Term};
 
 #[derive(Debug, Serialize)]
 pub struct SafekeeperStatus {
@@ -22,13 +23,16 @@ pub struct SafekeeperStatus {
 pub struct TimelineCreateRequest {
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
-    pub peer_ids: Option<Vec<NodeId>>,
+    pub mconf: Configuration,
     pub pg_version: u32,
     pub system_id: Option<u64>,
+    // By default WAL_SEGMENT_SIZE
     pub wal_seg_size: Option<u32>,
-    pub commit_lsn: Lsn,
-    // If not passed, it is assigned to the beginning of commit_lsn segment.
-    pub local_start_lsn: Option<Lsn>,
+    pub start_lsn: Lsn,
+    // Normal creation should omit this field (start_lsn initializes all LSNs).
+    // However, we allow specifying custom value higher than start_lsn for
+    // manual recovery case, see test_s3_wal_replay.
+    pub commit_lsn: Option<Lsn>,
 }
 
 /// Same as TermLsn, but serializes LSN using display serializer
@@ -143,7 +147,13 @@ pub type ConnectionId = u32;
 
 /// Serialize is used only for json'ing in API response. Also used internally.
 #[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct WalSenderState {
+pub enum WalSenderState {
+    Vanilla(VanillaWalSenderState),
+    Interpreted(InterpretedWalSenderState),
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct VanillaWalSenderState {
     pub ttid: TenantTimelineId,
     pub addr: SocketAddr,
     pub conn_id: ConnectionId,
@@ -152,6 +162,17 @@ pub struct WalSenderState {
     pub feedback: ReplicationFeedback,
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct InterpretedWalSenderState {
+    pub ttid: TenantTimelineId,
+    pub shard: ShardIdentity,
+    pub addr: SocketAddr,
+    pub conn_id: ConnectionId,
+    // postgres application_name
+    pub appname: Option<String>,
+    pub feedback: ReplicationFeedback,
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct WalReceiverState {
     /// None means it is recovery initiated by us (this safekeeper).
@@ -172,6 +193,7 @@ pub enum WalReceiverStatus {
 pub struct TimelineStatus {
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
+    pub mconf: Configuration,
     pub acceptor_state: AcceptorStateStatus,
     pub pg_info: ServerInfo,
     pub flush_lsn: Lsn,
@@ -186,6 +208,20 @@ pub struct TimelineStatus {
     pub walreceivers: Vec<WalReceiverState>,
 }
 
+/// Request to switch membership configuration.
+#[derive(Serialize, Deserialize)]
+#[serde(transparent)]
+pub struct TimelineMembershipSwitchRequest {
+    pub mconf: Configuration,
+}
+
+/// In response both previous and current configuration are sent.
+#[derive(Serialize, Deserialize)]
+pub struct TimelineMembershipSwitchResponse {
+    pub previous_conf: Configuration,
+    pub current_conf: Configuration,
+}
+
 fn lsn_invalid() -> Lsn {
     Lsn::INVALID
 }

libs/wal_decoder/Cargo.toml

@@ -24,3 +24,18 @@ workspace_hack = { version = "0.1", path = "../../workspace_hack" }
 
 [build-dependencies]
 tonic-build.workspace = true
+
+[dev-dependencies]
+criterion.workspace = true
+camino.workspace = true
+camino-tempfile.workspace = true
+remote_storage.workspace = true
+tokio-util.workspace = true
+serde_json.workspace = true
+futures.workspace = true
+tikv-jemallocator.workspace = true
+pprof.workspace = true
+
+[[bench]]
+name = "bench_interpret_wal"
+harness = false

libs/wal_decoder/benches/README.md

@@ -0,0 +1,34 @@
+## WAL Decoding and Interpretation Benchmarks
+
+Note that these benchmarks pull WAL from a public bucket in S3
+as a preparation step. Hence, you need a way to auth with AWS.
+You can achieve this by copying the `~/.aws/config` file from
+the AWS SSO notion page and exporting `AWS_PROFILE=dev` when invoking
+the benchmarks.
+
+To run benchmarks:
+
+```sh
+aws sso login --profile dev
+
+# All benchmarks.
+AWS_PROFILE=dev cargo bench --package wal_decoder
+
+# Specific file.
+AWS_PROFILE=dev cargo bench --package wal_decoder --bench bench_interpret_wal
+
+# Specific benchmark.
+AWS_PROFILE=dev cargo bench --package wal_decoder --bench bench_interpret_wal unsharded
+
+# List available benchmarks.
+cargo bench --package wal_decoder --benches -- --list
+
+# Generate flamegraph profiles using pprof-rs, profiling for 10 seconds.
+# Output in target/criterion/*/profile/flamegraph.svg.
+AWS_PROFILE=dev cargo bench --package wal_decoder --bench bench_interpret_wal unsharded -- --profile-time 10
+```
+
+Additional charts and statistics are available in `target/criterion/report/index.html`.
+
+Benchmarks are automatically compared against the previous run. To compare against other runs, see
+`--baseline` and `--save-baseline`.

libs/wal_decoder/benches/bench_interpret_wal.rs

@@ -0,0 +1,250 @@
+use anyhow::Context;
+use criterion::{criterion_group, criterion_main, Criterion};
+use futures::{stream::FuturesUnordered, StreamExt};
+use pageserver_api::shard::{ShardIdentity, ShardStripeSize};
+use postgres_ffi::{waldecoder::WalStreamDecoder, MAX_SEND_SIZE, WAL_SEGMENT_SIZE};
+use pprof::criterion::{Output, PProfProfiler};
+use serde::Deserialize;
+use std::{env, num::NonZeroUsize, sync::Arc};
+
+use camino::{Utf8Path, Utf8PathBuf};
+use camino_tempfile::Utf8TempDir;
+use remote_storage::{
+    DownloadOpts, GenericRemoteStorage, ListingMode, RemoteStorageConfig, RemoteStorageKind,
+    S3Config,
+};
+use tokio_util::sync::CancellationToken;
+use utils::{
+    lsn::Lsn,
+    shard::{ShardCount, ShardNumber},
+};
+use wal_decoder::models::InterpretedWalRecord;
+
+const S3_BUCKET: &str = "neon-github-public-dev";
+const S3_REGION: &str = "eu-central-1";
+const BUCKET_PREFIX: &str = "wal-snapshots/bulk-insert/";
+const METADATA_FILENAME: &str = "metadata.json";
+
+/// Use jemalloc, and configure it to sample allocations for profiles every 1 MB.
+/// This mirrors the configuration in bin/safekeeper.rs.
+#[global_allocator]
+static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
+
+#[allow(non_upper_case_globals)]
+#[export_name = "malloc_conf"]
+pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:20\0";
+
+async fn create_s3_client() -> anyhow::Result<Arc<GenericRemoteStorage>> {
+    let remote_storage_config = RemoteStorageConfig {
+        storage: RemoteStorageKind::AwsS3(S3Config {
+            bucket_name: S3_BUCKET.to_string(),
+            bucket_region: S3_REGION.to_string(),
+            prefix_in_bucket: Some(BUCKET_PREFIX.to_string()),
+            endpoint: None,
+            concurrency_limit: NonZeroUsize::new(100).unwrap(),
+            max_keys_per_list_response: None,
+            upload_storage_class: None,
+        }),
+        timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
+        small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
+    };
+    Ok(Arc::new(
+        GenericRemoteStorage::from_config(&remote_storage_config)
+            .await
+            .context("remote storage init")?,
+    ))
+}
+
+async fn download_bench_data(
+    client: Arc<GenericRemoteStorage>,
+    cancel: &CancellationToken,
+) -> anyhow::Result<Utf8TempDir> {
+    let temp_dir_parent: Utf8PathBuf = env::current_dir().unwrap().try_into()?;
+    let temp_dir = camino_tempfile::tempdir_in(temp_dir_parent)?;
+
+    eprintln!("Downloading benchmark data to {:?}", temp_dir);
+
+    let listing = client
+        .list(None, ListingMode::NoDelimiter, None, cancel)
+        .await?;
+
+    let mut downloads = listing
+        .keys
+        .into_iter()
+        .map(|obj| {
+            let client = client.clone();
+            let temp_dir_path = temp_dir.path().to_owned();
+
+            async move {
+                let remote_path = obj.key;
+                let download = client
+                    .download(&remote_path, &DownloadOpts::default(), cancel)
+                    .await?;
+                let mut body = tokio_util::io::StreamReader::new(download.download_stream);
+
+                let file_name = remote_path.object_name().unwrap();
+                let file_path = temp_dir_path.join(file_name);
+                let file = tokio::fs::OpenOptions::new()
+                    .create(true)
+                    .truncate(true)
+                    .write(true)
+                    .open(&file_path)
+                    .await?;
+
+                let mut writer = tokio::io::BufWriter::new(file);
+                tokio::io::copy_buf(&mut body, &mut writer).await?;
+
+                Ok::<(), anyhow::Error>(())
+            }
+        })
+        .collect::<FuturesUnordered<_>>();
+
+    while let Some(download) = downloads.next().await {
+        download?;
+    }
+
+    Ok(temp_dir)
+}
+
+struct BenchmarkData {
+    wal: Vec<u8>,
+    meta: BenchmarkMetadata,
+}
+
+#[derive(Deserialize)]
+struct BenchmarkMetadata {
+    pg_version: u32,
+    start_lsn: Lsn,
+}
+
+async fn load_bench_data(path: &Utf8Path, input_size: usize) -> anyhow::Result<BenchmarkData> {
+    eprintln!("Loading benchmark data from {:?}", path);
+
+    let mut entries = tokio::fs::read_dir(path).await?;
+    let mut ordered_segment_paths = Vec::new();
+    let mut metadata = None;
+
+    while let Some(entry) = entries.next_entry().await? {
+        if entry.file_name() == METADATA_FILENAME {
+            let bytes = tokio::fs::read(entry.path()).await?;
+            metadata = Some(
+                serde_json::from_slice::<BenchmarkMetadata>(&bytes)
+                    .context("failed to deserialize metadata.json")?,
+            );
+        } else {
+            ordered_segment_paths.push(entry.path());
+        }
+    }
+
+    ordered_segment_paths.sort();
+
+    let mut buffer = Vec::new();
+    for path in ordered_segment_paths {
+        if buffer.len() >= input_size {
+            break;
+        }
+
+        use async_compression::tokio::bufread::ZstdDecoder;
+        let file = tokio::fs::File::open(path).await?;
+        let reader = tokio::io::BufReader::new(file);
+        let decoder = ZstdDecoder::new(reader);
+        let mut reader = tokio::io::BufReader::new(decoder);
+        tokio::io::copy_buf(&mut reader, &mut buffer).await?;
+    }
+
+    buffer.truncate(input_size);
+
+    Ok(BenchmarkData {
+        wal: buffer,
+        meta: metadata.unwrap(),
+    })
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    const INPUT_SIZE: usize = 128 * 1024 * 1024;
+
+    let setup_runtime = tokio::runtime::Builder::new_current_thread()
+        .enable_all()
+        .build()
+        .unwrap();
+
+    let (_temp_dir, bench_data) = setup_runtime.block_on(async move {
+        let cancel = CancellationToken::new();
+        let client = create_s3_client().await.unwrap();
+        let temp_dir = download_bench_data(client, &cancel).await.unwrap();
+        let bench_data = load_bench_data(temp_dir.path(), INPUT_SIZE).await.unwrap();
+
+        (temp_dir, bench_data)
+    });
+
+    eprintln!(
+        "Benchmarking against {} MiB of WAL",
+        INPUT_SIZE / 1024 / 1024
+    );
+
+    let mut group = c.benchmark_group("decode-interpret-wal");
+    group.throughput(criterion::Throughput::Bytes(bench_data.wal.len() as u64));
+    group.sample_size(10);
+
+    group.bench_function("unsharded", |b| {
+        b.iter(|| decode_interpret_main(&bench_data, &[ShardIdentity::unsharded()]))
+    });
+
+    let eight_shards = (0..8)
+        .map(|i| ShardIdentity::new(ShardNumber(i), ShardCount(8), ShardStripeSize(8)).unwrap())
+        .collect::<Vec<_>>();
+
+    group.bench_function("8/8-shards", |b| {
+        b.iter(|| decode_interpret_main(&bench_data, &eight_shards))
+    });
+
+    let four_shards = eight_shards
+        .into_iter()
+        .filter(|s| s.number.0 % 2 == 0)
+        .collect::<Vec<_>>();
+    group.bench_function("4/8-shards", |b| {
+        b.iter(|| decode_interpret_main(&bench_data, &four_shards))
+    });
+
+    let two_shards = four_shards
+        .into_iter()
+        .filter(|s| s.number.0 % 4 == 0)
+        .collect::<Vec<_>>();
+    group.bench_function("2/8-shards", |b| {
+        b.iter(|| decode_interpret_main(&bench_data, &two_shards))
+    });
+}
+
+fn decode_interpret_main(bench: &BenchmarkData, shards: &[ShardIdentity]) {
+    let r = decode_interpret(bench, shards);
+    if let Err(e) = r {
+        panic!("{e:?}");
+    }
+}
+
+fn decode_interpret(bench: &BenchmarkData, shard: &[ShardIdentity]) -> anyhow::Result<()> {
+    let mut decoder = WalStreamDecoder::new(bench.meta.start_lsn, bench.meta.pg_version);
+    let xlogoff: usize = bench.meta.start_lsn.segment_offset(WAL_SEGMENT_SIZE);
+
+    for chunk in bench.wal[xlogoff..].chunks(MAX_SEND_SIZE) {
+        decoder.feed_bytes(chunk);
+        while let Some((lsn, recdata)) = decoder.poll_decode().unwrap() {
+            assert!(lsn.is_aligned());
+            let _ = InterpretedWalRecord::from_bytes_filtered(
+                recdata,
+                shard,
+                lsn,
+                bench.meta.pg_version,
+            )
+            .unwrap();
+        }
+    }
+
+    Ok(())
+}
+criterion_group!(
+    name=benches;
+    config=Criterion::default().with_profiler(PProfProfiler::new(100, Output::Flamegraph(None)));
+    targets=criterion_benchmark
+);
+criterion_main!(benches);

libs/wal_decoder/src/decoder.rs

@@ -1,6 +1,8 @@
 //! This module contains logic for decoding and interpreting
 //! raw bytes which represent a raw Postgres WAL record.
 
+use std::collections::HashMap;
+
 use crate::models::*;
 use crate::serialized_batch::SerializedValueBatch;
 use bytes::{Buf, Bytes};
@@ -14,15 +16,15 @@ use utils::lsn::Lsn;
 
 impl InterpretedWalRecord {
     /// Decode and interpreted raw bytes which represent one Postgres WAL record.
-    /// Data blocks which do not match the provided shard identity are filtered out.
+    /// Data blocks which do not match any of the provided shard identities are filtered out.
     /// Shard 0 is a special case since it tracks all relation sizes. We only give it
     /// the keys that are being written as that is enough for updating relation sizes.
     pub fn from_bytes_filtered(
         buf: Bytes,
-        shard: &ShardIdentity,
+        shards: &[ShardIdentity],
         next_record_lsn: Lsn,
         pg_version: u32,
-    ) -> anyhow::Result<InterpretedWalRecord> {
+    ) -> anyhow::Result<HashMap<ShardIdentity, InterpretedWalRecord>> {
         let mut decoded = DecodedWALRecord::default();
         decode_wal_record(buf, &mut decoded, pg_version)?;
         let xid = decoded.xl_xid;
@@ -33,43 +35,57 @@ impl InterpretedWalRecord {
             FlushUncommittedRecords::No
         };
 
-        let metadata_record =
-            MetadataRecord::from_decoded_filtered(&decoded, shard, next_record_lsn, pg_version)?;
-        let batch = SerializedValueBatch::from_decoded_filtered(
+        let mut shard_records: HashMap<ShardIdentity, InterpretedWalRecord> =
+            HashMap::with_capacity(shards.len());
+        for shard in shards {
+            shard_records.insert(
+                *shard,
+                InterpretedWalRecord {
+                    metadata_record: None,
+                    batch: SerializedValueBatch::default(),
+                    next_record_lsn,
+                    flush_uncommitted,
+                    xid,
+                },
+            );
+        }
+
+        MetadataRecord::from_decoded_filtered(
+            &decoded,
+            &mut shard_records,
+            next_record_lsn,
+            pg_version,
+        )?;
+        SerializedValueBatch::from_decoded_filtered(
             decoded,
-            shard,
+            &mut shard_records,
             next_record_lsn,
             pg_version,
         )?;
 
-        Ok(InterpretedWalRecord {
-            metadata_record,
-            batch,
-            next_record_lsn,
-            flush_uncommitted,
-            xid,
-        })
+        Ok(shard_records)
     }
 }
 
 impl MetadataRecord {
-    /// Builds a metadata record for this WAL record, if any.
+    /// Populates the given `shard_records` with metadata records from this WAL record, if any,
+    /// discarding those belonging to other shards.
     ///
-    /// Only metadata records relevant for the given shard are emitted. Currently, most metadata
+    /// Only metadata records relevant for the given shards is emitted. Currently, most metadata
     /// records are broadcast to all shards for simplicity, but this should be improved.
     fn from_decoded_filtered(
         decoded: &DecodedWALRecord,
-        shard: &ShardIdentity,
+        shard_records: &mut HashMap<ShardIdentity, InterpretedWalRecord>,
         next_record_lsn: Lsn,
         pg_version: u32,
-    ) -> anyhow::Result<Option<MetadataRecord>> {
+    ) -> anyhow::Result<()> {
         // Note: this doesn't actually copy the bytes since
         // the [`Bytes`] type implements it via a level of indirection.
         let mut buf = decoded.record.clone();
         buf.advance(decoded.main_data_offset);
 
         // First, generate metadata records from the decoded WAL record.
-        let mut metadata_record = match decoded.xl_rmid {
+        let metadata_record = match decoded.xl_rmid {
             pg_constants::RM_HEAP_ID | pg_constants::RM_HEAP2_ID => {
                 Self::decode_heapam_record(&mut buf, decoded, pg_version)?
             }
@@ -112,41 +128,65 @@ impl MetadataRecord {
         };
 
         // Next, filter the metadata record by shard.
-        match metadata_record {
-            Some(
-                MetadataRecord::Heapam(HeapamRecord::ClearVmBits(ref mut clear_vm_bits))
-                | MetadataRecord::Neonrmgr(NeonrmgrRecord::ClearVmBits(ref mut clear_vm_bits)),
-            ) => {
-                // Route VM page updates to the shards that own them. VM pages are stored in the VM fork
-                // of the main relation. These are sharded and managed just like regular relation pages.
-                // See: https://github.com/neondatabase/neon/issues/9855
-                let is_local_vm_page = |heap_blk| {
-                    let vm_blk = pg_constants::HEAPBLK_TO_MAPBLOCK(heap_blk);
-                    shard.is_key_local(&rel_block_to_key(clear_vm_bits.vm_rel, vm_blk))
-                };
-                // Send the old and new VM page updates to their respective shards.
-                clear_vm_bits.old_heap_blkno = clear_vm_bits
-                    .old_heap_blkno
-                    .filter(|&blkno| is_local_vm_page(blkno));
-                clear_vm_bits.new_heap_blkno = clear_vm_bits
-                    .new_heap_blkno
-                    .filter(|&blkno| is_local_vm_page(blkno));
-                // If neither VM page belongs to this shard, discard the record.
-                if clear_vm_bits.old_heap_blkno.is_none() && clear_vm_bits.new_heap_blkno.is_none()
-                {
-                    metadata_record = None
+        for (shard, record) in shard_records.iter_mut() {
+            match metadata_record {
+                Some(
+                    MetadataRecord::Heapam(HeapamRecord::ClearVmBits(ref clear_vm_bits))
+                    | MetadataRecord::Neonrmgr(NeonrmgrRecord::ClearVmBits(ref clear_vm_bits)),
+                ) => {
+                    // Route VM page updates to the shards that own them. VM pages are stored in the VM fork
+                    // of the main relation. These are sharded and managed just like regular relation pages.
+                    // See: https://github.com/neondatabase/neon/issues/9855
+                    let is_local_vm_page = |heap_blk| {
+                        let vm_blk = pg_constants::HEAPBLK_TO_MAPBLOCK(heap_blk);
+                        shard.is_key_local(&rel_block_to_key(clear_vm_bits.vm_rel, vm_blk))
+                    };
+                    // Send the old and new VM page updates to their respective shards.
+                    let updated_old_heap_blkno = clear_vm_bits
+                        .old_heap_blkno
+                        .filter(|&blkno| is_local_vm_page(blkno));
+                    let updated_new_heap_blkno = clear_vm_bits
+                        .new_heap_blkno
+                        .filter(|&blkno| is_local_vm_page(blkno));
+                    // If neither VM page belongs to this shard, discard the record.
+                    if updated_old_heap_blkno.is_some() || updated_new_heap_blkno.is_some() {
+                        // Clone the record and update it for the current shard.
+                        let mut for_shard = metadata_record.clone();
+                        match for_shard {
+                            Some(
+                                MetadataRecord::Heapam(HeapamRecord::ClearVmBits(
+                                    ref mut clear_vm_bits,
+                                ))
+                                | MetadataRecord::Neonrmgr(NeonrmgrRecord::ClearVmBits(
+                                    ref mut clear_vm_bits,
+                                )),
+                            ) => {
+                                clear_vm_bits.old_heap_blkno = updated_old_heap_blkno;
+                                clear_vm_bits.new_heap_blkno = updated_new_heap_blkno;
+                                record.metadata_record = for_shard;
+                            }
+                            _ => {
+                                unreachable!("for_shard is a clone of what we checked above")
+                            }
+                        }
+                    }
+                }
+                Some(MetadataRecord::LogicalMessage(LogicalMessageRecord::Put(_))) => {
+                    // Filter LogicalMessage records (AUX files) to only be stored on shard zero
+                    if shard.is_shard_zero() {
+                        record.metadata_record = metadata_record;
+                        // No other shards should receive this record, so we stop traversing shards early.
+                        break;
+                    }
+                }
+                _ => {
+                    // All other metadata records are sent to all shards.
+                    record.metadata_record = metadata_record.clone();
                 }
             }
-            Some(MetadataRecord::LogicalMessage(LogicalMessageRecord::Put(_))) => {
-                // Filter LogicalMessage records (AUX files) to only be stored on shard zero
-                if !shard.is_shard_zero() {
-                    metadata_record = None;
-                }
-            }
-            _ => {}
         }
 
-        Ok(metadata_record)
+        Ok(())
     }
 
     fn decode_heapam_record(

libs/wal_decoder/src/models.rs

@@ -48,7 +48,7 @@ pub mod proto {
     tonic::include_proto!("interpreted_wal");
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Copy, Clone, Serialize, Deserialize)]
 pub enum FlushUncommittedRecords {
     Yes,
     No,
@@ -64,7 +64,7 @@ pub struct InterpretedWalRecords {
 }
 
 /// An interpreted Postgres WAL record, ready to be handled by the pageserver
-#[derive(Serialize, Deserialize)]
+#[derive(Serialize, Deserialize, Clone)]
 pub struct InterpretedWalRecord {
     /// Optional metadata record - may cause writes to metadata keys
     /// in the storage engine
@@ -107,7 +107,7 @@ impl InterpretedWalRecord {
 
 /// The interpreted part of the Postgres WAL record which requires metadata
 /// writes to the underlying storage engine.
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum MetadataRecord {
     Heapam(HeapamRecord),
     Neonrmgr(NeonrmgrRecord),
@@ -123,12 +123,12 @@ pub enum MetadataRecord {
     Replorigin(ReploriginRecord),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum HeapamRecord {
     ClearVmBits(ClearVmBits),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct ClearVmBits {
     pub new_heap_blkno: Option<u32>,
     pub old_heap_blkno: Option<u32>,
@@ -136,29 +136,29 @@ pub struct ClearVmBits {
     pub flags: u8,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum NeonrmgrRecord {
     ClearVmBits(ClearVmBits),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum SmgrRecord {
     Create(SmgrCreate),
     Truncate(XlSmgrTruncate),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct SmgrCreate {
     pub rel: RelTag,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum DbaseRecord {
     Create(DbaseCreate),
     Drop(DbaseDrop),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct DbaseCreate {
     pub db_id: Oid,
     pub tablespace_id: Oid,
@@ -166,32 +166,32 @@ pub struct DbaseCreate {
     pub src_tablespace_id: Oid,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct DbaseDrop {
     pub db_id: Oid,
     pub tablespace_ids: Vec<Oid>,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum ClogRecord {
     ZeroPage(ClogZeroPage),
     Truncate(ClogTruncate),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct ClogZeroPage {
     pub segno: u32,
     pub rpageno: u32,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct ClogTruncate {
     pub pageno: u32,
     pub oldest_xid: TransactionId,
     pub oldest_xid_db: Oid,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum XactRecord {
     Commit(XactCommon),
     Abort(XactCommon),
@@ -200,7 +200,7 @@ pub enum XactRecord {
     Prepare(XactPrepare),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct XactCommon {
     pub parsed: XlXactParsedRecord,
     pub origin_id: u16,
@@ -209,73 +209,73 @@ pub struct XactCommon {
     pub lsn: Lsn,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct XactPrepare {
     pub xl_xid: TransactionId,
     pub data: Bytes,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum MultiXactRecord {
     ZeroPage(MultiXactZeroPage),
     Create(XlMultiXactCreate),
     Truncate(XlMultiXactTruncate),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct MultiXactZeroPage {
     pub slru_kind: SlruKind,
     pub segno: u32,
     pub rpageno: u32,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum RelmapRecord {
     Update(RelmapUpdate),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct RelmapUpdate {
     pub update: XlRelmapUpdate,
     pub buf: Bytes,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum XlogRecord {
     Raw(RawXlogRecord),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct RawXlogRecord {
     pub info: u8,
     pub lsn: Lsn,
     pub buf: Bytes,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum LogicalMessageRecord {
     Put(PutLogicalMessage),
     #[cfg(feature = "testing")]
     Failpoint,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct PutLogicalMessage {
     pub path: String,
     pub buf: Bytes,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum StandbyRecord {
     RunningXacts(StandbyRunningXacts),
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct StandbyRunningXacts {
     pub oldest_running_xid: TransactionId,
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub enum ReploriginRecord {
     Set(XlReploriginSet),
     Drop(XlReploriginDrop),

libs/wal_decoder/src/serialized_batch.rs

@@ -5,7 +5,7 @@
 //! Such batches are created from decoded PG wal records and ingested
 //! by the pageserver by writing directly to the ephemeral file.
 
-use std::collections::BTreeSet;
+use std::collections::{BTreeSet, HashMap};
 
 use bytes::{Bytes, BytesMut};
 use pageserver_api::key::rel_block_to_key;
@@ -22,6 +22,8 @@ use utils::lsn::Lsn;
 
 use pageserver_api::key::Key;
 
+use crate::models::InterpretedWalRecord;
+
 static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 
 /// Accompanying metadata for the batch
@@ -30,7 +32,7 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 /// relation sizes. In the case of "observed" values, we only need to know
 /// the key and LSN, so two types of metadata are supported to save on network
 /// bandwidth.
-#[derive(Serialize, Deserialize)]
+#[derive(Serialize, Deserialize, Clone)]
 pub enum ValueMeta {
     Serialized(SerializedValueMeta),
     Observed(ObservedValueMeta),
@@ -77,7 +79,7 @@ impl PartialEq for OrderedValueMeta {
 impl Eq for OrderedValueMeta {}
 
 /// Metadata for a [`Value`] serialized into the batch.
-#[derive(Serialize, Deserialize)]
+#[derive(Serialize, Deserialize, Clone)]
 pub struct SerializedValueMeta {
     pub key: CompactKey,
     pub lsn: Lsn,
@@ -89,14 +91,14 @@ pub struct SerializedValueMeta {
 }
 
 /// Metadata for a [`Value`] observed by the batch
-#[derive(Serialize, Deserialize)]
+#[derive(Serialize, Deserialize, Clone)]
 pub struct ObservedValueMeta {
     pub key: CompactKey,
     pub lsn: Lsn,
 }
 
 /// Batch of serialized [`Value`]s.
-#[derive(Serialize, Deserialize)]
+#[derive(Serialize, Deserialize, Clone)]
 pub struct SerializedValueBatch {
     /// [`Value`]s serialized in EphemeralFile's native format,
     /// ready for disk write by the pageserver
@@ -128,7 +130,8 @@ impl Default for SerializedValueBatch {
 }
 
 impl SerializedValueBatch {
-    /// Build a batch of serialized values from a decoded PG WAL record
+    /// Populates the given `shard_records` with value batches from this WAL record, if any,
+    /// discarding those belonging to other shards.
     ///
     /// The batch will only contain values for keys targeting the specifiec
     /// shard. Shard 0 is a special case, where any keys that don't belong to
@@ -136,21 +139,20 @@ impl SerializedValueBatch {
     /// but absent from the raw buffer [`SerializedValueBatch::raw`]).
     pub(crate) fn from_decoded_filtered(
         decoded: DecodedWALRecord,
-        shard: &ShardIdentity,
+        shard_records: &mut HashMap<ShardIdentity, InterpretedWalRecord>,
         next_record_lsn: Lsn,
         pg_version: u32,
-    ) -> anyhow::Result<SerializedValueBatch> {
-        // First determine how big the buffer needs to be and allocate it up-front.
+    ) -> anyhow::Result<()> {
+        // First determine how big the buffers need to be and allocate it up-front.
         // This duplicates some of the work below, but it's empirically much faster.
-        let estimated_buffer_size = Self::estimate_buffer_size(&decoded, shard, pg_version);
-        let mut buf = Vec::<u8>::with_capacity(estimated_buffer_size);
+        for (shard, record) in shard_records.iter_mut() {
+            assert!(record.batch.is_empty());
+
+            let estimate = Self::estimate_buffer_size(&decoded, shard, pg_version);
+            record.batch.raw = Vec::with_capacity(estimate);
+        }
 
-        let mut metadata: Vec<ValueMeta> = Vec::with_capacity(decoded.blocks.len());
-        let mut max_lsn: Lsn = Lsn(0);
-        let mut len: usize = 0;
         for blk in decoded.blocks.iter() {
-            let relative_off = buf.len() as u64;
-
             let rel = RelTag {
                 spcnode: blk.rnode_spcnode,
                 dbnode: blk.rnode_dbnode,
@@ -168,99 +170,98 @@ impl SerializedValueBatch {
                 );
             }
 
-            let key_is_local = shard.is_key_local(&key);
+            for (shard, record) in shard_records.iter_mut() {
+                let key_is_local = shard.is_key_local(&key);
 
-            tracing::debug!(
-                lsn=%next_record_lsn,
-                key=%key,
-                "ingest: shard decision {}",
-                if !key_is_local { "drop" } else { "keep" },
-            );
+                tracing::debug!(
+                    lsn=%next_record_lsn,
+                    key=%key,
+                    "ingest: shard decision {}",
+                    if !key_is_local { "drop" } else { "keep" },
+                );
 
-            if !key_is_local {
-                if shard.is_shard_zero() {
-                    // Shard 0 tracks relation sizes.  Although we will not store this block, we will observe
-                    // its blkno in case it implicitly extends a relation.
-                    metadata.push(ValueMeta::Observed(ObservedValueMeta {
+                if !key_is_local {
+                    if shard.is_shard_zero() {
+                        // Shard 0 tracks relation sizes.  Although we will not store this block, we will observe
+                        // its blkno in case it implicitly extends a relation.
+                        record
+                            .batch
+                            .metadata
+                            .push(ValueMeta::Observed(ObservedValueMeta {
+                                key: key.to_compact(),
+                                lsn: next_record_lsn,
+                            }))
+                    }
+
+                    continue;
+                }
+
+                // Instead of storing full-page-image WAL record,
+                // it is better to store extracted image: we can skip wal-redo
+                // in this case. Also some FPI records may contain multiple (up to 32) pages,
+                // so them have to be copied multiple times.
+                //
+                let val = if Self::block_is_image(&decoded, blk, pg_version) {
+                    // Extract page image from FPI record
+                    let img_len = blk.bimg_len as usize;
+                    let img_offs = blk.bimg_offset as usize;
+                    let mut image = BytesMut::with_capacity(BLCKSZ as usize);
+                    // TODO(vlad): skip the copy
+                    image.extend_from_slice(&decoded.record[img_offs..img_offs + img_len]);
+
+                    if blk.hole_length != 0 {
+                        let tail = image.split_off(blk.hole_offset as usize);
+                        image.resize(image.len() + blk.hole_length as usize, 0u8);
+                        image.unsplit(tail);
+                    }
+                    //
+                    // Match the logic of XLogReadBufferForRedoExtended:
+                    // The page may be uninitialized. If so, we can't set the LSN because
+                    // that would corrupt the page.
+                    //
+                    if !page_is_new(&image) {
+                        page_set_lsn(&mut image, next_record_lsn)
+                    }
+                    assert_eq!(image.len(), BLCKSZ as usize);
+
+                    Value::Image(image.freeze())
+                } else {
+                    Value::WalRecord(NeonWalRecord::Postgres {
+                        will_init: blk.will_init || blk.apply_image,
+                        rec: decoded.record.clone(),
+                    })
+                };
+
+                let relative_off = record.batch.raw.len() as u64;
+
+                val.ser_into(&mut record.batch.raw)
+                    .expect("Writing into in-memory buffer is infallible");
+
+                let val_ser_size = record.batch.raw.len() - relative_off as usize;
+
+                record
+                    .batch
+                    .metadata
+                    .push(ValueMeta::Serialized(SerializedValueMeta {
                         key: key.to_compact(),
                         lsn: next_record_lsn,
-                    }))
-                }
-
-                continue;
+                        batch_offset: relative_off,
+                        len: val_ser_size,
+                        will_init: val.will_init(),
+                    }));
+                record.batch.max_lsn = std::cmp::max(record.batch.max_lsn, next_record_lsn);
+                record.batch.len += 1;
             }
-
-            // Instead of storing full-page-image WAL record,
-            // it is better to store extracted image: we can skip wal-redo
-            // in this case. Also some FPI records may contain multiple (up to 32) pages,
-            // so them have to be copied multiple times.
-            //
-            let val = if Self::block_is_image(&decoded, blk, pg_version) {
-                // Extract page image from FPI record
-                let img_len = blk.bimg_len as usize;
-                let img_offs = blk.bimg_offset as usize;
-                let mut image = BytesMut::with_capacity(BLCKSZ as usize);
-                // TODO(vlad): skip the copy
-                image.extend_from_slice(&decoded.record[img_offs..img_offs + img_len]);
-
-                if blk.hole_length != 0 {
-                    let tail = image.split_off(blk.hole_offset as usize);
-                    image.resize(image.len() + blk.hole_length as usize, 0u8);
-                    image.unsplit(tail);
-                }
-                //
-                // Match the logic of XLogReadBufferForRedoExtended:
-                // The page may be uninitialized. If so, we can't set the LSN because
-                // that would corrupt the page.
-                //
-                if !page_is_new(&image) {
-                    page_set_lsn(&mut image, next_record_lsn)
-                }
-                assert_eq!(image.len(), BLCKSZ as usize);
-
-                Value::Image(image.freeze())
-            } else {
-                Value::WalRecord(NeonWalRecord::Postgres {
-                    will_init: blk.will_init || blk.apply_image,
-                    rec: decoded.record.clone(),
-                })
-            };
-
-            val.ser_into(&mut buf)
-                .expect("Writing into in-memory buffer is infallible");
-
-            let val_ser_size = buf.len() - relative_off as usize;
-
-            metadata.push(ValueMeta::Serialized(SerializedValueMeta {
-                key: key.to_compact(),
-                lsn: next_record_lsn,
-                batch_offset: relative_off,
-                len: val_ser_size,
-                will_init: val.will_init(),
-            }));
-            max_lsn = std::cmp::max(max_lsn, next_record_lsn);
-            len += 1;
         }
 
         if cfg!(any(debug_assertions, test)) {
-            let batch = Self {
-                raw: buf,
-                metadata,
-                max_lsn,
-                len,
-            };
-
-            batch.validate_lsn_order();
-
-            return Ok(batch);
+            // Validate that the batches are correct
+            for record in shard_records.values() {
+                record.batch.validate_lsn_order();
+            }
         }
 
-        Ok(Self {
-            raw: buf,
-            metadata,
-            max_lsn,
-            len,
-        })
+        Ok(())
     }
 
     /// Look into the decoded PG WAL record and determine

pageserver/Cargo.toml

@@ -8,7 +8,7 @@ license.workspace = true
 default = []
 # Enables test-only APIs, incuding failpoints. In particular, enables the `fail_point!` macro,
 # which adds some runtime cost to run tests on outage conditions
-testing = ["fail/failpoints", "pageserver_api/testing", "wal_decoder/testing", "pageserver_client/testing"]
+testing = ["fail/failpoints", "pageserver_api/testing", "wal_decoder/testing"]
 
 [dependencies]
 anyhow.workspace = true
@@ -16,6 +16,7 @@ arc-swap.workspace = true
 async-compression.workspace = true
 async-stream.workspace = true
 bit_field.workspace = true
+bincode.workspace = true
 byteorder.workspace = true
 bytes.workspace = true
 camino.workspace = true
@@ -113,7 +114,3 @@ harness = false
 [[bench]]
 name = "upload_queue"
 harness = false
-
-[[bin]]
-name = "test_helper_slow_client_reads"
-required-features = [ "testing" ]

pageserver/client/Cargo.toml

@@ -4,9 +4,6 @@ version = "0.1.0"
 edition.workspace = true
 license.workspace = true
 
-[features]
-testing = [ "pageserver_api/testing" ]
-
 [dependencies]
 pageserver_api.workspace = true
 thiserror.workspace = true

pageserver/client/src/page_service.rs

@@ -1,9 +1,6 @@
-use std::sync::{Arc, Mutex};
+use std::pin::Pin;
 
-use futures::{
-    stream::{SplitSink, SplitStream},
-    SinkExt, StreamExt,
-};
+use futures::SinkExt;
 use pageserver_api::{
     models::{
         PagestreamBeMessage, PagestreamFeMessage, PagestreamGetPageRequest,
@@ -13,6 +10,7 @@ use pageserver_api::{
 };
 use tokio::task::JoinHandle;
 use tokio_postgres::CopyOutStream;
+use tokio_stream::StreamExt;
 use tokio_util::sync::CancellationToken;
 use utils::{
     id::{TenantId, TimelineId},
@@ -64,28 +62,15 @@ impl Client {
             .client
             .copy_both_simple(&format!("pagestream_v3 {tenant_id} {timeline_id}"))
             .await?;
-        let (sink, stream) = copy_both.split(); // TODO: actually support splitting of the CopyBothDuplex so the lock inside this split adaptor goes away.
         let Client {
             cancel_on_client_drop,
             conn_task,
             client: _,
         } = self;
-        let shared = Arc::new(Mutex::new(PagestreamShared::ConnTaskRunning(
-            ConnTaskRunning {
-                cancel_on_client_drop,
-                conn_task,
-            },
-        )));
         Ok(PagestreamClient {
-            sink: PagestreamSender {
-                shared: shared.clone(),
-                sink,
-            },
-            stream: PagestreamReceiver {
-                shared: shared.clone(),
-                stream,
-            },
-            shared,
+            copy_both: Box::pin(copy_both),
+            conn_task,
+            cancel_on_client_drop,
         })
     }
 
@@ -112,28 +97,7 @@ impl Client {
 
 /// Create using [`Client::pagestream`].
 pub struct PagestreamClient {
-    shared: Arc<Mutex<PagestreamShared>>,
-    sink: PagestreamSender,
-    stream: PagestreamReceiver,
-}
-
-pub struct PagestreamSender {
-    #[allow(dead_code)]
-    shared: Arc<Mutex<PagestreamShared>>,
-    sink: SplitSink<tokio_postgres::CopyBothDuplex<bytes::Bytes>, bytes::Bytes>,
-}
-
-pub struct PagestreamReceiver {
-    #[allow(dead_code)]
-    shared: Arc<Mutex<PagestreamShared>>,
-    stream: SplitStream<tokio_postgres::CopyBothDuplex<bytes::Bytes>>,
-}
-
-enum PagestreamShared {
-    ConnTaskRunning(ConnTaskRunning),
-    ConnTaskCancelledJoinHandleReturnedOrDropped,
-}
-struct ConnTaskRunning {
+    copy_both: Pin<Box<tokio_postgres::CopyBothDuplex<bytes::Bytes>>>,
     cancel_on_client_drop: Option<tokio_util::sync::DropGuard>,
     conn_task: JoinHandle<()>,
 }
@@ -146,11 +110,11 @@ pub struct RelTagBlockNo {
 impl PagestreamClient {
     pub async fn shutdown(self) {
         let Self {
-            shared,
-            sink,
-            stream,
-        } = { self };
-        // The `copy_both` split into `sink` and `stream` contains internal channel sender, the receiver of which is polled by `conn_task`.
+            copy_both,
+            cancel_on_client_drop: cancel_conn_task,
+            conn_task,
+        } = self;
+        // The `copy_both` contains internal channel sender, the receiver of which is polled by `conn_task`.
         // When `conn_task` observes the sender has been dropped, it sends a `FeMessage::CopyFail` into the connection.
         // (see https://github.com/neondatabase/rust-postgres/blob/2005bf79573b8add5cf205b52a2b208e356cc8b0/tokio-postgres/src/copy_both.rs#L56).
         //
@@ -167,77 +131,27 @@ impl PagestreamClient {
         //
         // NB: page_service doesn't have a use case to exit the `pagestream` mode currently.
         // => https://github.com/neondatabase/neon/issues/6390
-        let ConnTaskRunning {
-            cancel_on_client_drop,
-            conn_task,
-        } = {
-            let mut guard = shared.lock().unwrap();
-            match std::mem::replace(
-                &mut *guard,
-                PagestreamShared::ConnTaskCancelledJoinHandleReturnedOrDropped,
-            ) {
-                PagestreamShared::ConnTaskRunning(conn_task_running) => conn_task_running,
-                PagestreamShared::ConnTaskCancelledJoinHandleReturnedOrDropped => unreachable!(),
-            }
-        };
-        let _ = cancel_on_client_drop.unwrap();
+        let _ = cancel_conn_task.unwrap();
         conn_task.await.unwrap();
-
-        // Now drop the split copy_both.
-        drop(sink);
-        drop(stream);
-    }
-
-    pub fn split(self) -> (PagestreamSender, PagestreamReceiver) {
-        let Self {
-            shared: _,
-            sink,
-            stream,
-        } = self;
-        (sink, stream)
+        drop(copy_both);
     }
 
     pub async fn getpage(
         &mut self,
         req: PagestreamGetPageRequest,
     ) -> anyhow::Result<PagestreamGetPageResponse> {
-        self.getpage_send(req).await?;
-        self.getpage_recv().await
-    }
+        let req = PagestreamFeMessage::GetPage(req);
+        let req: bytes::Bytes = req.serialize();
+        // let mut req = tokio_util::io::ReaderStream::new(&req);
+        let mut req = tokio_stream::once(Ok(req));
 
-    pub async fn getpage_send(&mut self, req: PagestreamGetPageRequest) -> anyhow::Result<()> {
-        self.sink.getpage_send(req).await
-    }
+        self.copy_both.send_all(&mut req).await?;
 
-    pub async fn getpage_recv(&mut self) -> anyhow::Result<PagestreamGetPageResponse> {
-        self.stream.getpage_recv().await
-    }
-}
-
-impl PagestreamSender {
-    // TODO: maybe make this impl Sink instead for better composability?
-    pub async fn send(&mut self, msg: PagestreamFeMessage) -> anyhow::Result<()> {
-        let msg = msg.serialize();
-        self.sink.send_all(&mut tokio_stream::once(Ok(msg))).await?;
-        Ok(())
-    }
-
-    pub async fn getpage_send(&mut self, req: PagestreamGetPageRequest) -> anyhow::Result<()> {
-        self.send(PagestreamFeMessage::GetPage(req)).await
-    }
-}
-
-impl PagestreamReceiver {
-    // TODO: maybe make this impl Stream instead for better composability?
-    pub async fn recv(&mut self) -> anyhow::Result<PagestreamBeMessage> {
-        let next: Option<Result<bytes::Bytes, _>> = self.stream.next().await;
+        let next: Option<Result<bytes::Bytes, _>> = self.copy_both.next().await;
         let next: bytes::Bytes = next.unwrap()?;
-        PagestreamBeMessage::deserialize(next)
-    }
 
-    pub async fn getpage_recv(&mut self) -> anyhow::Result<PagestreamGetPageResponse> {
-        let next: PagestreamBeMessage = self.recv().await?;
-        match next {
+        let msg = PagestreamBeMessage::deserialize(next)?;
+        match msg {
             PagestreamBeMessage::GetPage(p) => Ok(p),
             PagestreamBeMessage::Error(e) => anyhow::bail!("Error: {:?}", e),
             PagestreamBeMessage::Exists(_)
@@ -246,14 +160,7 @@ impl PagestreamReceiver {
             | PagestreamBeMessage::GetSlruSegment(_) => {
                 anyhow::bail!(
                     "unexpected be message kind in response to getpage request: {}",
-                    next.kind()
-                )
-            }
-            #[cfg(feature = "testing")]
-            PagestreamBeMessage::Test(_) => {
-                anyhow::bail!(
-                    "unexpected be message kind in response to getpage request: {}",
-                    next.kind()
+                    msg.kind()
                 )
             }
         }

pageserver/ctl/Cargo.toml

@@ -8,9 +8,11 @@ license.workspace = true
 
 [dependencies]
 anyhow.workspace = true
+bincode.workspace = true
 camino.workspace = true
 clap = { workspace = true, features = ["string"] }
 humantime.workspace = true
+itertools.workspace = true
 pageserver = { path = ".." }
 pageserver_api.workspace = true
 remote_storage = { path = "../../libs/remote_storage" }

pageserver/ctl/src/main.rs

@@ -9,7 +9,9 @@ mod index_part;
 mod key;
 mod layer_map_analyzer;
 mod layers;
+mod page_trace;
 
+use page_trace::PageTraceCmd;
 use std::{
     str::FromStr,
     time::{Duration, SystemTime},
@@ -64,6 +66,7 @@ enum Commands {
     Layer(LayerCmd),
     /// Debug print a hex key found from logs
     Key(key::DescribeKeyCommand),
+    PageTrace(PageTraceCmd),
 }
 
 /// Read and update pageserver metadata file
@@ -183,6 +186,7 @@ async fn main() -> anyhow::Result<()> {
                 .await?;
         }
         Commands::Key(dkc) => dkc.execute(),
+        Commands::PageTrace(cmd) => page_trace::main(&cmd)?,
     };
     Ok(())
 }

pageserver/ctl/src/page_trace.rs

@@ -0,0 +1,73 @@
+use std::collections::HashMap;
+use std::io::BufReader;
+
+use camino::Utf8PathBuf;
+use clap::Parser;
+use itertools::Itertools as _;
+use pageserver_api::key::{CompactKey, Key};
+use pageserver_api::models::PageTraceEvent;
+use pageserver_api::reltag::RelTag;
+
+/// Parses a page trace (as emitted by the `page_trace` timeline API), and outputs stats.
+#[derive(Parser)]
+pub(crate) struct PageTraceCmd {
+    /// Trace input file.
+    path: Utf8PathBuf,
+}
+
+pub(crate) fn main(cmd: &PageTraceCmd) -> anyhow::Result<()> {
+    let mut file = BufReader::new(std::fs::OpenOptions::new().read(true).open(&cmd.path)?);
+    let mut events: Vec<PageTraceEvent> = Vec::new();
+    loop {
+        match bincode::deserialize_from(&mut file) {
+            Ok(event) => events.push(event),
+            Err(err) => {
+                if let bincode::ErrorKind::Io(ref err) = *err {
+                    if err.kind() == std::io::ErrorKind::UnexpectedEof {
+                        break;
+                    }
+                }
+                return Err(err.into());
+            }
+        }
+    }
+
+    let mut reads_by_relation: HashMap<RelTag, i64> = HashMap::new();
+    let mut reads_by_key: HashMap<CompactKey, i64> = HashMap::new();
+
+    for event in events {
+        let key = Key::from_compact(event.key);
+        let reltag = RelTag {
+            spcnode: key.field2,
+            dbnode: key.field3,
+            relnode: key.field4,
+            forknum: key.field5,
+        };
+
+        *reads_by_relation.entry(reltag).or_default() += 1;
+        *reads_by_key.entry(event.key).or_default() += 1;
+    }
+
+    let multi_read_keys = reads_by_key
+        .into_iter()
+        .filter(|(_, count)| *count > 1)
+        .sorted_by_key(|(key, count)| (-*count, *key))
+        .collect_vec();
+
+    println!("Multi-read keys: {}", multi_read_keys.len());
+    for (key, count) in multi_read_keys {
+        println!("  {key}: {count}");
+    }
+
+    let reads_by_relation = reads_by_relation
+        .into_iter()
+        .sorted_by_key(|(rel, count)| (-*count, *rel))
+        .collect_vec();
+
+    println!("Reads by relation:");
+    for (reltag, count) in reads_by_relation {
+        println!("  {reltag}: {count}");
+    }
+
+    Ok(())
+}

pageserver/src/bin/test_helper_slow_client_reads.rs

@@ -1,65 +0,0 @@
-use std::{
-    io::{stdin, stdout, Read, Write},
-    time::Duration,
-};
-
-use clap::Parser;
-use pageserver_api::models::{PagestreamRequest, PagestreamTestRequest};
-use utils::{
-    id::{TenantId, TimelineId},
-    lsn::Lsn,
-};
-
-#[derive(clap::Parser)]
-struct Args {
-    connstr: String,
-    tenant_id: TenantId,
-    timeline_id: TimelineId,
-}
-
-#[tokio::main]
-async fn main() -> anyhow::Result<()> {
-    let Args {
-        connstr,
-        tenant_id,
-        timeline_id,
-    } = Args::parse();
-    let client = pageserver_client::page_service::Client::new(connstr).await?;
-    let client = client.pagestream(tenant_id, timeline_id).await?;
-    let (mut sender, _receiver) = client.split();
-
-    eprintln!("filling the pipe");
-    let mut msg = 0;
-    loop {
-        msg += 1;
-        let fut = sender.send(pageserver_api::models::PagestreamFeMessage::Test(
-            PagestreamTestRequest {
-                hdr: PagestreamRequest {
-                    reqid: 0,
-                    request_lsn: Lsn(23),
-                    not_modified_since: Lsn(23),
-                },
-                batch_key: 42,
-                message: format!("message {}", msg),
-            },
-        ));
-        let Ok(res) = tokio::time::timeout(Duration::from_secs(10), fut).await else {
-            eprintln!("pipe seems full");
-            break;
-        };
-        let _: () = res?;
-    }
-
-    let n = stdout().write(b"R")?;
-    assert_eq!(n, 1);
-    stdout().flush()?;
-
-    eprintln!("waiting for signal to tell us to exit");
-
-    let mut buf = [0u8; 1];
-    stdin().read_exact(&mut buf)?;
-
-    eprintln!("termination signal received, exiting");
-
-    anyhow::Ok(())
-}

pageserver/src/http/routes.rs

@@ -27,6 +27,7 @@ use pageserver_api::models::LocationConfigMode;
 use pageserver_api::models::LsnLease;
 use pageserver_api::models::LsnLeaseRequest;
 use pageserver_api::models::OffloadedTimelineInfo;
+use pageserver_api::models::PageTraceEvent;
 use pageserver_api::models::ShardParameters;
 use pageserver_api::models::TenantConfigPatchRequest;
 use pageserver_api::models::TenantDetails;
@@ -51,7 +52,9 @@ use pageserver_api::shard::TenantShardId;
 use remote_storage::DownloadError;
 use remote_storage::GenericRemoteStorage;
 use remote_storage::TimeTravelError;
+use scopeguard::defer;
 use tenant_size_model::{svg::SvgBranchKind, SizeResult, StorageModel};
+use tokio::time::Instant;
 use tokio_util::io::StreamReader;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -1521,6 +1524,71 @@ async fn timeline_gc_unblocking_handler(
     block_or_unblock_gc(request, false).await
 }
 
+/// Traces GetPage@LSN requests for a timeline, and emits metadata in an efficient binary encoding.
+/// Use the `pagectl page-trace` command to decode and analyze the output.
+async fn timeline_page_trace_handler(
+    request: Request<Body>,
+    cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    let state = get_state(&request);
+    check_permission(&request, None)?;
+
+    let size_limit: usize = parse_query_param(&request, "size_limit_bytes")?.unwrap_or(1024 * 1024);
+    let time_limit_secs: u64 = parse_query_param(&request, "time_limit_secs")?.unwrap_or(5);
+
+    // Convert size limit to event limit based on the serialized size of an event. The event size is
+    // fixed, as the default bincode serializer uses fixed-width integer encoding.
+    let event_size = bincode::serialize(&PageTraceEvent::default())
+        .map_err(|err| ApiError::InternalServerError(err.into()))?
+        .len();
+    let event_limit = size_limit / event_size;
+
+    let timeline =
+        active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
+            .await?;
+
+    // Install a page trace, unless one is already in progress. We just use a buffered channel,
+    // which may 2x the memory usage in the worst case, but it's still bounded.
+    let (trace_tx, mut trace_rx) = tokio::sync::mpsc::channel(event_limit);
+    let cur = timeline.page_trace.load();
+    let installed = cur.is_none()
+        && timeline
+            .page_trace
+            .compare_and_swap(cur, Some(Arc::new(trace_tx)))
+            .is_none();
+    if !installed {
+        return Err(ApiError::Conflict("page trace already active".to_string()));
+    }
+    defer!(timeline.page_trace.store(None)); // uninstall on return
+
+    // Collect the trace and return it to the client. We could stream the response, but this is
+    // simple and fine.
+    let mut body = Vec::with_capacity(size_limit);
+    let deadline = Instant::now() + Duration::from_secs(time_limit_secs);
+
+    while body.len() < size_limit {
+        tokio::select! {
+            event = trace_rx.recv() => {
+                let Some(event) = event else {
+                    break; // shouldn't happen (sender doesn't close, unless timeline dropped)
+                };
+                bincode::serialize_into(&mut body, &event)
+                    .map_err(|err| ApiError::InternalServerError(err.into()))?;
+            }
+            _ = tokio::time::sleep_until(deadline) => break, // time limit reached
+            _ = cancel.cancelled() => return Err(ApiError::Cancelled),
+        }
+    }
+
+    Ok(Response::builder()
+        .status(StatusCode::OK)
+        .header(header::CONTENT_TYPE, "application/octet-stream")
+        .body(hyper::Body::from(body))
+        .unwrap())
+}
+
 /// Adding a block is `POST ../block_gc`, removing a block is `POST ../unblock_gc`.
 ///
 /// Both are technically unsafe because they might fire off index uploads, thus they are POST.
@@ -3479,6 +3547,10 @@ pub fn make_router(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/unblock_gc",
             |r| api_handler(r, timeline_gc_unblocking_handler),
         )
+        .get(
+            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/page_trace",
+            |r| api_handler(r, timeline_page_trace_handler),
+        )
         .post("/v1/tenant/:tenant_shard_id/heatmap_upload", |r| {
             api_handler(r, secondary_upload_handler)
         })

pageserver/src/import_datadir.rs

@@ -278,6 +278,8 @@ async fn import_wal(
 
     let mut walingest = WalIngest::new(tline, startpoint, ctx).await?;
 
+    let shard = vec![*tline.get_shard_identity()];
+
     while last_lsn <= endpoint {
         // FIXME: assume postgresql tli 1 for now
         let filename = XLogFileName(1, segno, WAL_SEGMENT_SIZE);
@@ -314,10 +316,12 @@ async fn import_wal(
             if let Some((lsn, recdata)) = waldecoder.poll_decode()? {
                 let interpreted = InterpretedWalRecord::from_bytes_filtered(
                     recdata,
-                    tline.get_shard_identity(),
+                    &shard,
                     lsn,
                     tline.pg_version,
-                )?;
+                )?
+                .remove(tline.get_shard_identity())
+                .unwrap();
 
                 walingest
                     .ingest_record(interpreted, &mut modification, ctx)
@@ -411,6 +415,7 @@ pub async fn import_wal_from_tar(
     let mut offset = start_lsn.segment_offset(WAL_SEGMENT_SIZE);
     let mut last_lsn = start_lsn;
     let mut walingest = WalIngest::new(tline, start_lsn, ctx).await?;
+    let shard = vec![*tline.get_shard_identity()];
 
     // Ingest wal until end_lsn
     info!("importing wal until {}", end_lsn);
@@ -459,10 +464,12 @@ pub async fn import_wal_from_tar(
             if let Some((lsn, recdata)) = waldecoder.poll_decode()? {
                 let interpreted = InterpretedWalRecord::from_bytes_filtered(
                     recdata,
-                    tline.get_shard_identity(),
+                    &shard,
                     lsn,
                     tline.pg_version,
-                )?;
+                )?
+                .remove(tline.get_shard_identity())
+                .unwrap();
 
                 walingest
                     .ingest_record(interpreted, &mut modification, ctx)

pageserver/src/metrics.rs

@@ -1463,8 +1463,6 @@ pub enum SmgrQueryType {
     GetPageAtLsn,
     GetDbSize,
     GetSlruSegment,
-    #[cfg(feature = "testing")]
-    Test,
 }
 
 pub(crate) struct SmgrQueryTimePerTimeline {

pageserver/src/page_service.rs

@@ -67,6 +67,7 @@ use crate::tenant::PageReconstructError;
 use crate::tenant::Timeline;
 use crate::{basebackup, timed_after_cancellation};
 use pageserver_api::key::rel_block_to_key;
+use pageserver_api::models::PageTraceEvent;
 use pageserver_api::reltag::SlruKind;
 use postgres_ffi::pg_constants::DEFAULTTABLESPACE_OID;
 use postgres_ffi::BLCKSZ;
@@ -554,12 +555,6 @@ struct BatchedGetPageRequest {
     timer: SmgrOpTimer,
 }
 
-#[cfg(feature = "testing")]
-struct BatchedTestRequest {
-    req: models::PagestreamTestRequest,
-    timer: SmgrOpTimer,
-}
-
 enum BatchedFeMessage {
     Exists {
         span: Span,
@@ -591,12 +586,6 @@ enum BatchedFeMessage {
         shard: timeline::handle::Handle<TenantManagerTypes>,
         req: models::PagestreamGetSlruSegmentRequest,
     },
-    #[cfg(feature = "testing")]
-    Test {
-        span: Span,
-        shard: timeline::handle::Handle<TenantManagerTypes>,
-        requests: Vec<BatchedTestRequest>,
-    },
     RespondError {
         span: Span,
         error: BatchedPageStreamError,
@@ -617,12 +606,6 @@ impl BatchedFeMessage {
                     page.timer.observe_execution_start(at);
                 }
             }
-            #[cfg(feature = "testing")]
-            BatchedFeMessage::Test { requests, .. } => {
-                for req in requests {
-                    req.timer.observe_execution_start(at);
-                }
-            }
             BatchedFeMessage::RespondError { .. } => {}
         }
     }
@@ -883,22 +866,6 @@ impl PageServerHandler {
                     pages: smallvec::smallvec![BatchedGetPageRequest { req, timer }],
                 }
             }
-            #[cfg(feature = "testing")]
-            PagestreamFeMessage::Test(req) => {
-                let span = tracing::info_span!(parent: parent_span, "handle_test_request");
-                let shard = timeline_handles
-                    .get(tenant_id, timeline_id, ShardSelector::Zero)
-                    .instrument(span.clone()) // sets `shard_id` field
-                    .await?;
-                let timer =
-                    record_op_start_and_throttle(&shard, metrics::SmgrQueryType::Test, received_at)
-                        .await?;
-                BatchedFeMessage::Test {
-                    span,
-                    shard,
-                    requests: vec![BatchedTestRequest { req, timer }],
-                }
-            }
         };
         Ok(Some(batched_msg))
     }
@@ -961,46 +928,6 @@ impl PageServerHandler {
                 accum_pages.extend(this_pages);
                 Ok(())
             }
-            #[cfg(feature = "testing")]
-            (
-                Ok(BatchedFeMessage::Test {
-                    shard: accum_shard,
-                    requests: accum_requests,
-                    ..
-                }),
-                BatchedFeMessage::Test {
-                    shard: this_shard,
-                    requests: this_requests,
-                    ..
-                },
-            ) if (|| {
-                assert!(this_requests.len() == 1);
-                if accum_requests.len() >= max_batch_size.get() {
-                    trace!(%max_batch_size, "stopping batching because of batch size");
-                    assert_eq!(accum_requests.len(), max_batch_size.get());
-                    return false;
-                }
-                if (accum_shard.tenant_shard_id, accum_shard.timeline_id)
-                    != (this_shard.tenant_shard_id, this_shard.timeline_id)
-                {
-                    trace!("stopping batching because timeline object mismatch");
-                    // TODO: we _could_ batch & execute each shard seperately (and in parallel).
-                    // But the current logic for keeping responses in order does not support that.
-                    return false;
-                }
-                let this_batch_key = this_requests[0].req.batch_key;
-                let accum_batch_key = accum_requests[0].req.batch_key;
-                if this_requests[0].req.batch_key != accum_requests[0].req.batch_key {
-                    trace!(%accum_batch_key, %this_batch_key, "stopping batching because batch key changed");
-                    return false;
-                }
-                true
-            })() =>
-            {
-                // ok to batch
-                accum_requests.extend(this_requests);
-                Ok(())
-            }
             // something batched already but this message is unbatchable
             (_, this_msg) => {
                 // by default, don't continue batching
@@ -1127,27 +1054,6 @@ impl PageServerHandler {
                     span,
                 )
             }
-            #[cfg(feature = "testing")]
-            BatchedFeMessage::Test {
-                span,
-                shard,
-                requests,
-            } => {
-                fail::fail_point!("ps::handle-pagerequest-message::test");
-                (
-                    {
-                        let npages = requests.len();
-                        trace!(npages, "handling getpage request");
-                        let res = self
-                            .handle_test_request_batch(&shard, requests, ctx)
-                            .instrument(span.clone())
-                            .await;
-                        assert_eq!(res.len(), npages);
-                        res
-                    },
-                    span,
-                )
-            }
             BatchedFeMessage::RespondError { span, error } => {
                 // We've already decided to respond with an error, so we don't need to
                 // call the handler.
@@ -1813,6 +1719,20 @@ impl PageServerHandler {
             .query_metrics
             .observe_getpage_batch_start(requests.len());
 
+        // If a page trace is running, submit an event for this request.
+        if let Some(page_trace) = timeline.page_trace.load().as_ref() {
+            let time = SystemTime::now();
+            for batch in &requests {
+                let key = rel_block_to_key(batch.req.rel, batch.req.blkno).to_compact();
+                // Ignore error (trace buffer may be full or tracer may have disconnected).
+                _ = page_trace.try_send(PageTraceEvent {
+                    key,
+                    effective_lsn,
+                    time,
+                });
+            }
+        }
+
         let results = timeline
             .get_rel_page_at_lsn_batched(
                 requests.iter().map(|p| (&p.req.rel, &p.req.blkno)),
@@ -1871,51 +1791,6 @@ impl PageServerHandler {
         ))
     }
 
-    // NB: this impl mimics what we do for batched getpage requests.
-    #[cfg(feature = "testing")]
-    #[instrument(skip_all, fields(shard_id))]
-    async fn handle_test_request_batch(
-        &mut self,
-        timeline: &Timeline,
-        requests: Vec<BatchedTestRequest>,
-        _ctx: &RequestContext,
-    ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer), BatchedPageStreamError>> {
-        // real requests would do something with the timeline
-        let mut results = Vec::with_capacity(requests.len());
-        for _req in requests.iter() {
-            tokio::task::yield_now().await;
-
-            results.push({
-                if timeline.cancel.is_cancelled() {
-                    Err(PageReconstructError::Cancelled)
-                } else {
-                    Ok(())
-                }
-            });
-        }
-
-        // TODO: avoid creating the new Vec here
-        Vec::from_iter(
-            requests
-                .into_iter()
-                .zip(results.into_iter())
-                .map(|(req, res)| {
-                    res.map(|()| {
-                        (
-                            PagestreamBeMessage::Test(models::PagestreamTestResponse {
-                                req: req.req.clone(),
-                            }),
-                            req.timer,
-                        )
-                    })
-                    .map_err(|e| BatchedPageStreamError {
-                        err: PageStreamError::from(e),
-                        req: req.req.hdr,
-                    })
-                }),
-        )
-    }
-
     /// Note on "fullbackup":
     /// Full basebackups should only be used for debugging purposes.
     /// Originally, it was introduced to enable breaking storage format changes,

pageserver/src/tenant/timeline.rs

@@ -14,7 +14,7 @@ pub mod uninit;
 mod walreceiver;
 
 use anyhow::{anyhow, bail, ensure, Context, Result};
-use arc_swap::ArcSwap;
+use arc_swap::{ArcSwap, ArcSwapOption};
 use bytes::Bytes;
 use camino::Utf8Path;
 use chrono::{DateTime, Utc};
@@ -23,6 +23,7 @@ use fail::fail_point;
 use handle::ShardTimelineId;
 use offload::OffloadError;
 use once_cell::sync::Lazy;
+use pageserver_api::models::PageTraceEvent;
 use pageserver_api::{
     config::tenant_conf_defaults::DEFAULT_COMPACTION_THRESHOLD,
     key::{
@@ -42,6 +43,7 @@ use rand::Rng;
 use remote_storage::DownloadError;
 use serde_with::serde_as;
 use storage_broker::BrokerClientChannel;
+use tokio::sync::mpsc::Sender;
 use tokio::{
     runtime::Handle,
     sync::{oneshot, watch},
@@ -433,6 +435,9 @@ pub struct Timeline {
 
     /// Cf. [`crate::tenant::CreateTimelineIdempotency`].
     pub(crate) create_idempotency: crate::tenant::CreateTimelineIdempotency,
+
+    /// If Some, collects GetPage metadata for an ongoing PageTrace.
+    pub(crate) page_trace: ArcSwapOption<Sender<PageTraceEvent>>,
 }
 
 pub type TimelineDeleteProgress = Arc<tokio::sync::Mutex<DeleteTimelineFlow>>;
@@ -2380,6 +2385,8 @@ impl Timeline {
                 attach_wal_lag_cooldown,
 
                 create_idempotency,
+
+                page_trace: Default::default(),
             };
 
             result.repartition_threshold =

pageserver/src/tenant/timeline/compaction.rs

@@ -1776,7 +1776,10 @@ impl Timeline {
         base_img_from_ancestor: Option<(Key, Lsn, Bytes)>,
     ) -> anyhow::Result<KeyHistoryRetention> {
         // Pre-checks for the invariants
-        if cfg!(debug_assertions) {
+
+        let debug_mode = cfg!(debug_assertions) || cfg!(feature = "testing");
+
+        if debug_mode {
             for (log_key, _, _) in full_history {
                 assert_eq!(log_key, &key, "mismatched key");
             }
@@ -1922,15 +1925,19 @@ impl Timeline {
             output
         }
 
+        let mut key_exists = false;
         for (i, split_for_lsn) in split_history.into_iter().enumerate() {
             // TODO: there could be image keys inside the splits, and we can compute records_since_last_image accordingly.
             records_since_last_image += split_for_lsn.len();
-            let generate_image = if i == 0 && !has_ancestor {
+            // Whether to produce an image into the final layer files
+            let produce_image = if i == 0 && !has_ancestor {
                 // We always generate images for the first batch (below horizon / lowest retain_lsn)
                 true
             } else if i == batch_cnt - 1 {
                 // Do not generate images for the last batch (above horizon)
                 false
+            } else if records_since_last_image == 0 {
+                false
             } else if records_since_last_image >= delta_threshold_cnt {
                 // Generate images when there are too many records
                 true
@@ -1945,29 +1952,45 @@ impl Timeline {
                     break;
                 }
             }
-            if let Some((_, _, val)) = replay_history.first() {
-                if !val.will_init() {
-                    return Err(anyhow::anyhow!("invalid history, no base image")).with_context(
-                        || {
-                            generate_debug_trace(
-                                Some(&replay_history),
-                                full_history,
-                                retain_lsn_below_horizon,
-                                horizon,
-                            )
-                        },
-                    );
-                }
+            if replay_history.is_empty() && !key_exists {
+                // The key does not exist at earlier LSN, we can skip this iteration.
+                retention.push(Vec::new());
+                continue;
+            } else {
+                key_exists = true;
             }
-            if generate_image && records_since_last_image > 0 {
+            let Some((_, _, val)) = replay_history.first() else {
+                unreachable!("replay history should not be empty once it exists")
+            };
+            if !val.will_init() {
+                return Err(anyhow::anyhow!("invalid history, no base image")).with_context(|| {
+                    generate_debug_trace(
+                        Some(&replay_history),
+                        full_history,
+                        retain_lsn_below_horizon,
+                        horizon,
+                    )
+                });
+            }
+            // Whether to reconstruct the image. In debug mode, we will generate an image
+            // at every retain_lsn to ensure data is not corrupted, but we won't put the
+            // image into the final layer.
+            let generate_image = produce_image || debug_mode;
+            if produce_image {
                 records_since_last_image = 0;
-                let replay_history_for_debug = if cfg!(debug_assertions) {
+            }
+            let img_and_lsn = if generate_image {
+                let replay_history_for_debug = if debug_mode {
                     Some(replay_history.clone())
                 } else {
                     None
                 };
                 let replay_history_for_debug_ref = replay_history_for_debug.as_deref();
-                let history = std::mem::take(&mut replay_history);
+                let history = if produce_image {
+                    std::mem::take(&mut replay_history)
+                } else {
+                    replay_history.clone()
+                };
                 let mut img = None;
                 let mut records = Vec::with_capacity(history.len());
                 if let (_, lsn, Value::Image(val)) = history.first().as_ref().unwrap() {
@@ -2004,8 +2027,20 @@ impl Timeline {
                 }
                 records.reverse();
                 let state = ValueReconstructState { img, records };
-                let request_lsn = lsn_split_points[i]; // last batch does not generate image so i is always in range
+                // last batch does not generate image so i is always in range, unless we force generate
+                // an image during testing
+                let request_lsn = if i >= lsn_split_points.len() {
+                    Lsn::MAX
+                } else {
+                    lsn_split_points[i]
+                };
                 let img = self.reconstruct_value(key, request_lsn, state).await?;
+                Some((request_lsn, img))
+            } else {
+                None
+            };
+            if produce_image {
+                let (request_lsn, img) = img_and_lsn.unwrap();
                 replay_history.push((key, request_lsn, Value::Image(img.clone())));
                 retention.push(vec![(request_lsn, Value::Image(img))]);
             } else {
@@ -2273,6 +2308,8 @@ impl Timeline {
         let compact_key_range = job.compact_key_range;
         let compact_lsn_range = job.compact_lsn_range;
 
+        let debug_mode = cfg!(debug_assertions) || cfg!(feature = "testing");
+
         info!("running enhanced gc bottom-most compaction, dry_run={dry_run}, compact_key_range={}..{}, compact_lsn_range={}..{}", compact_key_range.start, compact_key_range.end, compact_lsn_range.start, compact_lsn_range.end);
 
         scopeguard::defer! {
@@ -2398,7 +2435,7 @@ impl Timeline {
                 .first()
                 .copied()
                 .unwrap_or(job_desc.gc_cutoff);
-            if cfg!(debug_assertions) {
+            if debug_mode {
                 assert_eq!(
                     res,
                     job_desc

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

@@ -264,6 +264,8 @@ pub(super) async fn handle_walreceiver_connection(
 
     let mut walingest = WalIngest::new(timeline.as_ref(), startpoint, &ctx).await?;
 
+    let shard = vec![*timeline.get_shard_identity()];
+
     let interpreted_proto_config = match protocol {
         PostgresClientProtocol::Vanilla => None,
         PostgresClientProtocol::Interpreted {
@@ -476,10 +478,12 @@ pub(super) async fn handle_walreceiver_connection(
                         // Deserialize and interpret WAL record
                         let interpreted = InterpretedWalRecord::from_bytes_filtered(
                             recdata,
-                            modification.tline.get_shard_identity(),
+                            &shard,
                             next_record_lsn,
                             modification.tline.pg_version,
-                        )?;
+                        )?
+                        .remove(timeline.get_shard_identity())
+                        .unwrap();
 
                         if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes)
                             && uncommitted_records > 0

pageserver/src/walingest.rs

@@ -2163,10 +2163,12 @@ mod tests {
             while let Some((lsn, recdata)) = decoder.poll_decode().unwrap() {
                 let interpreted = InterpretedWalRecord::from_bytes_filtered(
                     recdata,
-                    modification.tline.get_shard_identity(),
+                    &[*modification.tline.get_shard_identity()],
                     lsn,
                     modification.tline.pg_version,
                 )
+                .unwrap()
+                .remove(modification.tline.get_shard_identity())
                 .unwrap();
 
                 walingest

pgxn/neon/libpagestore.c

@@ -911,7 +911,74 @@ pageserver_receive(shardno_t shard_no)
 		}
 		PG_CATCH();
 		{
-			neon_shard_log(shard_no, LOG, "pageserver_receive: disconnect due malformatted response");
+			neon_shard_log(shard_no, LOG, "pageserver_receive: disconnect due to failure while parsing response");
+			pageserver_disconnect(shard_no);
+			PG_RE_THROW();
+		}
+		PG_END_TRY();
+
+		if (message_level_is_interesting(PageStoreTrace))
+		{
+			char	   *msg = nm_to_string((NeonMessage *) resp);
+
+			neon_shard_log(shard_no, PageStoreTrace, "got response: %s", msg);
+			pfree(msg);
+		}
+	}
+	else if (rc == -1)
+	{
+		neon_shard_log(shard_no, LOG, "pageserver_receive disconnect: psql end of copy data: %s", pchomp(PQerrorMessage(pageserver_conn)));
+		pageserver_disconnect(shard_no);
+		resp = NULL;
+	}
+	else if (rc == -2)
+	{
+		char	   *msg = pchomp(PQerrorMessage(pageserver_conn));
+
+		pageserver_disconnect(shard_no);
+		neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: could not read COPY data: %s", msg);
+	}
+	else
+	{
+		pageserver_disconnect(shard_no);
+		neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: unexpected PQgetCopyData return value: %d", rc);
+	}
+
+	shard->nresponses_received++;
+	return (NeonResponse *) resp;
+}
+
+static NeonResponse *
+pageserver_try_receive(shardno_t shard_no)
+{
+	StringInfoData resp_buff;
+	NeonResponse *resp;
+	PageServer *shard = &page_servers[shard_no];
+	PGconn	   *pageserver_conn = shard->conn;
+	/* read response */
+	int			rc;
+
+	if (shard->state != PS_Connected)
+		return NULL;
+
+	Assert(pageserver_conn);
+
+	rc = PQgetCopyData(shard->conn, &resp_buff.data, 1 /* async = true */);
+
+	if (rc == 0)
+		return NULL;
+	else if (rc > 0)
+	{
+		PG_TRY();
+		{
+			resp_buff.len = rc;
+			resp_buff.cursor = 0;
+			resp = nm_unpack_response(&resp_buff);
+			PQfreemem(resp_buff.data);
+		}
+		PG_CATCH();
+		{
+			neon_shard_log(shard_no, LOG, "pageserver_receive: disconnect due to failure while parsing response");
 			pageserver_disconnect(shard_no);
 			PG_RE_THROW();
 		}
@@ -980,6 +1047,7 @@ page_server_api api =
 	.send = pageserver_send,
 	.flush = pageserver_flush,
 	.receive = pageserver_receive,
+	.try_receive = pageserver_try_receive,
 	.disconnect = pageserver_disconnect_shard
 };
 

pgxn/neon/pagestore_client.h

@@ -192,9 +192,29 @@ typedef uint16 shardno_t;
 
 typedef struct
 {
+	/*
+	 * Send this request to the PageServer associated with this shard.
+	 */
 	bool		(*send) (shardno_t  shard_no, NeonRequest * request);
+	/*
+	 * Blocking read for the next response of this shard.
+	 *
+	 * When a CANCEL signal is handled, the connection state will be
+	 * unmodified.
+	 */
 	NeonResponse *(*receive) (shardno_t shard_no);
+	/*
+	 * Try get the next response from the TCP buffers, if any.
+	 * Returns NULL when the data is not yet available. 
+	 */
+	NeonResponse *(*try_receive) (shardno_t shard_no);
+	/*
+	 * Make sure all requests are sent to PageServer.
+	 */
 	bool		(*flush) (shardno_t shard_no);
+	/*
+	 * Disconnect from this pageserver shard.
+	 */
 	void        (*disconnect) (shardno_t shard_no);
 } page_server_api;
 

pgxn/neon/pagestore_smgr.c

@@ -405,6 +405,56 @@ compact_prefetch_buffers(void)
 	return false;
 }
 
+/*
+ * If there might be responses still in the TCP buffer, then
+ * we should try to use those, so as to reduce any TCP backpressure
+ * on the OS/PS side.
+ *
+ * This procedure handles that.
+ *
+ * Note that this is only valid as long as the only pipelined
+ * operations in the TCP buffer are getPage@Lsn requests.
+ */
+static void
+prefetch_pump_state(void)
+{
+	while (MyPState->ring_receive != MyPState->ring_flush)
+	{
+		NeonResponse   *response;
+		PrefetchRequest *slot;
+		MemoryContext	old;
+
+		slot = GetPrfSlot(MyPState->ring_receive);
+
+		old = MemoryContextSwitchTo(MyPState->errctx);
+		response = page_server->try_receive(slot->shard_no);
+		MemoryContextSwitchTo(old);
+
+		if (response == NULL)
+			break;
+
+		/* The slot should still be valid */
+		if (slot->status != PRFS_REQUESTED ||
+			slot->response != NULL ||
+			slot->my_ring_index != MyPState->ring_receive)
+			neon_shard_log(slot->shard_no, ERROR,
+						   "Incorrect prefetch slot state after receive: status=%d response=%p my=%lu receive=%lu",
+						   slot->status, slot->response,
+						   (long) slot->my_ring_index, (long) MyPState->ring_receive);
+
+		/* update prefetch state */
+		MyPState->n_responses_buffered += 1;
+		MyPState->n_requests_inflight -= 1;
+		MyPState->ring_receive += 1;
+		MyNeonCounters->getpage_prefetches_buffered =
+			MyPState->n_responses_buffered;
+
+		/* update slot state */
+		slot->status = PRFS_RECEIVED;
+		slot->response = response;
+	}
+}
+
 void
 readahead_buffer_resize(int newsize, void *extra)
 {
@@ -2808,6 +2858,8 @@ neon_prefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 			   MyPState->ring_last <= ring_index);
 	}
 
+	prefetch_pump_state();
+
 	return false;
 }
 
@@ -2849,6 +2901,8 @@ neon_prefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum)
 	Assert(ring_index < MyPState->ring_unused &&
 		   MyPState->ring_last <= ring_index);
 
+	prefetch_pump_state();
+
 	return false;
 }
 #endif /* PG_MAJORVERSION_NUM < 17 */
@@ -2891,6 +2945,8 @@ neon_writeback(SMgrRelation reln, ForkNumber forknum,
 	 */
 	neon_log(SmgrTrace, "writeback noop");
 
+	prefetch_pump_state();
+
 #ifdef DEBUG_COMPARE_LOCAL
 	if (IS_LOCAL_REL(reln))
 		mdwriteback(reln, forknum, blocknum, nblocks);
@@ -3145,6 +3201,8 @@ neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer
 	neon_get_request_lsns(InfoFromSMgrRel(reln), forkNum, blkno, &request_lsns, 1, NULL);
 	neon_read_at_lsn(InfoFromSMgrRel(reln), forkNum, blkno, request_lsns, buffer);
 
+	prefetch_pump_state();
+
 #ifdef DEBUG_COMPARE_LOCAL
 	if (forkNum == MAIN_FORKNUM && IS_LOCAL_REL(reln))
 	{
@@ -3282,6 +3340,8 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 	neon_read_at_lsnv(InfoFromSMgrRel(reln), forknum, blocknum, request_lsns,
 					  buffers, nblocks, read);
 
+	prefetch_pump_state();
+
 #ifdef DEBUG_COMPARE_LOCAL
 	if (forkNum == MAIN_FORKNUM && IS_LOCAL_REL(reln))
 	{
@@ -3450,6 +3510,8 @@ neon_write(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, const vo
 
 	lfc_write(InfoFromSMgrRel(reln), forknum, blocknum, buffer);
 
+	prefetch_pump_state();
+
 #ifdef DEBUG_COMPARE_LOCAL
 	if (IS_LOCAL_REL(reln))
 		#if PG_MAJORVERSION_NUM >= 17
@@ -3503,6 +3565,8 @@ neon_writev(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
 
 	lfc_writev(InfoFromSMgrRel(reln), forknum, blkno, buffers, nblocks);
 
+	prefetch_pump_state();
+
 #ifdef DEBUG_COMPARE_LOCAL
 	if (IS_LOCAL_REL(reln))
 		mdwritev(reln, forknum, blocknum, &buffer, 1, skipFsync);
@@ -3792,6 +3856,8 @@ neon_immedsync(SMgrRelation reln, ForkNumber forknum)
 
 	neon_log(SmgrTrace, "[NEON_SMGR] immedsync noop");
 
+	prefetch_pump_state();
+
 #ifdef DEBUG_COMPARE_LOCAL
 	if (IS_LOCAL_REL(reln))
 		mdimmedsync(reln, forknum);

safekeeper/Cargo.toml

@@ -26,6 +26,7 @@ hex.workspace = true
 humantime.workspace = true
 http.workspace = true
 hyper0.workspace = true
+itertools.workspace = true
 futures.workspace = true
 once_cell.workspace = true
 parking_lot.workspace = true
@@ -39,6 +40,7 @@ scopeguard.workspace = true
 reqwest = { workspace = true, features = ["json"] }
 serde.workspace = true
 serde_json.workspace = true
+smallvec.workspace = true
 strum.workspace = true
 strum_macros.workspace = true
 thiserror.workspace = true
@@ -63,6 +65,7 @@ storage_broker.workspace = true
 tokio-stream.workspace = true
 utils.workspace = true
 wal_decoder.workspace = true
+env_logger.workspace = true
 
 workspace_hack.workspace = true
 

safekeeper/src/bin/safekeeper.rs

@@ -207,6 +207,13 @@ struct Args {
     /// Also defines interval for eviction retries.
     #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_EVICTION_MIN_RESIDENT)]
     eviction_min_resident: Duration,
+    /// Enable fanning out WAL to different shards from the same reader
+    #[arg(long)]
+    wal_reader_fanout: bool,
+    /// Only fan out the WAL reader if the absoulte delta between the new requested position
+    /// and the current position of the reader is smaller than this value.
+    #[arg(long)]
+    max_delta_for_fanout: Option<u64>,
 }
 
 // Like PathBufValueParser, but allows empty string.
@@ -370,6 +377,8 @@ async fn main() -> anyhow::Result<()> {
         control_file_save_interval: args.control_file_save_interval,
         partial_backup_concurrency: args.partial_backup_concurrency,
         eviction_min_resident: args.eviction_min_resident,
+        wal_reader_fanout: args.wal_reader_fanout,
+        max_delta_for_fanout: args.max_delta_for_fanout,
     });
 
     // initialize sentry if SENTRY_DSN is provided

safekeeper/src/control_file.rs

@@ -3,6 +3,7 @@
 use anyhow::{bail, ensure, Context, Result};
 use byteorder::{LittleEndian, ReadBytesExt, WriteBytesExt};
 use camino::{Utf8Path, Utf8PathBuf};
+use safekeeper_api::membership::INVALID_GENERATION;
 use tokio::fs::File;
 use tokio::io::AsyncWriteExt;
 use utils::crashsafe::durable_rename;
@@ -13,14 +14,14 @@ use std::ops::Deref;
 use std::path::Path;
 use std::time::Instant;
 
-use crate::control_file_upgrade::downgrade_v9_to_v8;
+use crate::control_file_upgrade::downgrade_v10_to_v9;
 use crate::control_file_upgrade::upgrade_control_file;
 use crate::metrics::PERSIST_CONTROL_FILE_SECONDS;
 use crate::state::{EvictionState, TimelinePersistentState};
 use utils::bin_ser::LeSer;
 
 pub const SK_MAGIC: u32 = 0xcafeceefu32;
-pub const SK_FORMAT_VERSION: u32 = 9;
+pub const SK_FORMAT_VERSION: u32 = 10;
 
 // contains persistent metadata for safekeeper
 pub const CONTROL_FILE_NAME: &str = "safekeeper.control";
@@ -169,10 +170,11 @@ impl TimelinePersistentState {
         let mut buf: Vec<u8> = Vec::new();
         WriteBytesExt::write_u32::<LittleEndian>(&mut buf, SK_MAGIC)?;
 
-        if self.eviction_state == EvictionState::Present {
-            // temp hack for forward compatibility
-            const PREV_FORMAT_VERSION: u32 = 8;
-            let prev = downgrade_v9_to_v8(self);
+        if self.mconf.generation == INVALID_GENERATION {
+            // Temp hack for forward compatibility test: in case of none
+            // configuration save cfile in previous v9 format.
+            const PREV_FORMAT_VERSION: u32 = 9;
+            let prev = downgrade_v10_to_v9(self);
             WriteBytesExt::write_u32::<LittleEndian>(&mut buf, PREV_FORMAT_VERSION)?;
             prev.ser_into(&mut buf)?;
         } else {
@@ -233,6 +235,7 @@ impl Storage for FileStorage {
 #[cfg(test)]
 mod test {
     use super::*;
+    use safekeeper_api::membership::{Configuration, MemberSet};
     use tokio::fs;
     use utils::lsn::Lsn;
 
@@ -242,6 +245,11 @@ mod test {
     async fn test_read_write_safekeeper_state() -> anyhow::Result<()> {
         let tempdir = camino_tempfile::tempdir()?;
         let mut state = TimelinePersistentState::empty();
+        state.mconf = Configuration {
+            generation: 42,
+            members: MemberSet::empty(),
+            new_members: None,
+        };
         let mut storage = FileStorage::create_new(tempdir.path(), state.clone(), NO_SYNC).await?;
 
         // Make a change.

safekeeper/src/control_file_upgrade.rs

@@ -1,17 +1,22 @@
 //! Code to deal with safekeeper control file upgrades
+use std::vec;
+
 use crate::{
     safekeeper::{AcceptorState, PgUuid, TermHistory, TermLsn},
-    state::{EvictionState, PersistedPeers, TimelinePersistentState},
+    state::{EvictionState, TimelinePersistentState},
     wal_backup_partial,
 };
 use anyhow::{bail, Result};
 use pq_proto::SystemId;
-use safekeeper_api::{ServerInfo, Term};
+use safekeeper_api::{
+    membership::{Configuration, INVALID_GENERATION},
+    ServerInfo, Term,
+};
 use serde::{Deserialize, Serialize};
 use tracing::*;
 use utils::{
     bin_ser::LeSer,
-    id::{TenantId, TimelineId},
+    id::{NodeId, TenantId, TimelineId},
     lsn::Lsn,
 };
 
@@ -233,6 +238,90 @@ pub struct SafeKeeperStateV8 {
     pub partial_backup: wal_backup_partial::State,
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct PersistedPeers(pub Vec<(NodeId, PersistedPeerInfo)>);
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct PersistedPeerInfo {
+    /// LSN up to which safekeeper offloaded WAL to s3.
+    pub backup_lsn: Lsn,
+    /// Term of the last entry.
+    pub term: Term,
+    /// LSN of the last record.
+    pub flush_lsn: Lsn,
+    /// Up to which LSN safekeeper regards its WAL as committed.
+    pub commit_lsn: Lsn,
+}
+
+impl PersistedPeerInfo {
+    pub fn new() -> Self {
+        Self {
+            backup_lsn: Lsn::INVALID,
+            term: safekeeper_api::INITIAL_TERM,
+            flush_lsn: Lsn(0),
+            commit_lsn: Lsn(0),
+        }
+    }
+}
+
+// make clippy happy
+impl Default for PersistedPeerInfo {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Note: SafekeeperStateVn is old name for TimelinePersistentStateVn.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct TimelinePersistentStateV9 {
+    #[serde(with = "hex")]
+    pub tenant_id: TenantId,
+    #[serde(with = "hex")]
+    pub timeline_id: TimelineId,
+    /// persistent acceptor state
+    pub acceptor_state: AcceptorState,
+    /// information about server
+    pub server: ServerInfo,
+    /// Unique id of the last *elected* proposer we dealt with. Not needed
+    /// for correctness, exists for monitoring purposes.
+    #[serde(with = "hex")]
+    pub proposer_uuid: PgUuid,
+    /// Since which LSN this timeline generally starts. Safekeeper might have
+    /// joined later.
+    pub timeline_start_lsn: Lsn,
+    /// Since which LSN safekeeper has (had) WAL for this timeline.
+    /// All WAL segments next to one containing local_start_lsn are
+    /// filled with data from the beginning.
+    pub local_start_lsn: Lsn,
+    /// Part of WAL acknowledged by quorum *and available locally*. Always points
+    /// to record boundary.
+    pub commit_lsn: Lsn,
+    /// LSN that points to the end of the last backed up segment. Useful to
+    /// persist to avoid finding out offloading progress on boot.
+    pub backup_lsn: Lsn,
+    /// Minimal LSN which may be needed for recovery of some safekeeper (end_lsn
+    /// of last record streamed to everyone). Persisting it helps skipping
+    /// recovery in walproposer, generally we compute it from peers. In
+    /// walproposer proto called 'truncate_lsn'. Updates are currently drived
+    /// only by walproposer.
+    pub peer_horizon_lsn: Lsn,
+    /// LSN of the oldest known checkpoint made by pageserver and successfully
+    /// pushed to s3. We don't remove WAL beyond it. Persisted only for
+    /// informational purposes, we receive it from pageserver (or broker).
+    pub remote_consistent_lsn: Lsn,
+    /// Peers and their state as we remember it. Knowing peers themselves is
+    /// fundamental; but state is saved here only for informational purposes and
+    /// obviously can be stale. (Currently not saved at all, but let's provision
+    /// place to have less file version upgrades).
+    pub peers: PersistedPeers,
+    /// Holds names of partial segments uploaded to remote storage. Used to
+    /// clean up old objects without leaving garbage in remote storage.
+    pub partial_backup: wal_backup_partial::State,
+    /// Eviction state of the timeline. If it's Offloaded, we should download
+    /// WAL files from remote storage to serve the timeline.
+    pub eviction_state: EvictionState,
+}
+
 pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersistentState> {
     // migrate to storing full term history
     if version == 1 {
@@ -248,6 +337,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
         return Ok(TimelinePersistentState {
             tenant_id: oldstate.server.tenant_id,
             timeline_id: oldstate.server.timeline_id,
+            mconf: Configuration::empty(),
             acceptor_state: ac,
             server: ServerInfo {
                 pg_version: oldstate.server.pg_version,
@@ -261,9 +351,9 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
             backup_lsn: Lsn(0),
             peer_horizon_lsn: oldstate.truncate_lsn,
             remote_consistent_lsn: Lsn(0),
-            peers: PersistedPeers(vec![]),
             partial_backup: wal_backup_partial::State::default(),
             eviction_state: EvictionState::Present,
+            creation_ts: std::time::SystemTime::UNIX_EPOCH,
         });
     // migrate to hexing some ids
     } else if version == 2 {
@@ -277,6 +367,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
         return Ok(TimelinePersistentState {
             tenant_id: oldstate.server.tenant_id,
             timeline_id: oldstate.server.timeline_id,
+            mconf: Configuration::empty(),
             acceptor_state: oldstate.acceptor_state,
             server,
             proposer_uuid: oldstate.proposer_uuid,
@@ -286,9 +377,9 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
             backup_lsn: Lsn(0),
             peer_horizon_lsn: oldstate.truncate_lsn,
             remote_consistent_lsn: Lsn(0),
-            peers: PersistedPeers(vec![]),
             partial_backup: wal_backup_partial::State::default(),
             eviction_state: EvictionState::Present,
+            creation_ts: std::time::SystemTime::UNIX_EPOCH,
         });
     // migrate to moving tenant_id/timeline_id to the top and adding some lsns
     } else if version == 3 {
@@ -302,6 +393,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
         return Ok(TimelinePersistentState {
             tenant_id: oldstate.server.tenant_id,
             timeline_id: oldstate.server.timeline_id,
+            mconf: Configuration::empty(),
             acceptor_state: oldstate.acceptor_state,
             server,
             proposer_uuid: oldstate.proposer_uuid,
@@ -311,9 +403,9 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
             backup_lsn: Lsn(0),
             peer_horizon_lsn: oldstate.truncate_lsn,
             remote_consistent_lsn: Lsn(0),
-            peers: PersistedPeers(vec![]),
             partial_backup: wal_backup_partial::State::default(),
             eviction_state: EvictionState::Present,
+            creation_ts: std::time::SystemTime::UNIX_EPOCH,
         });
     // migrate to having timeline_start_lsn
     } else if version == 4 {
@@ -327,6 +419,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
         return Ok(TimelinePersistentState {
             tenant_id: oldstate.tenant_id,
             timeline_id: oldstate.timeline_id,
+            mconf: Configuration::empty(),
             acceptor_state: oldstate.acceptor_state,
             server,
             proposer_uuid: oldstate.proposer_uuid,
@@ -336,9 +429,9 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
             backup_lsn: Lsn::INVALID,
             peer_horizon_lsn: oldstate.peer_horizon_lsn,
             remote_consistent_lsn: Lsn(0),
-            peers: PersistedPeers(vec![]),
             partial_backup: wal_backup_partial::State::default(),
             eviction_state: EvictionState::Present,
+            creation_ts: std::time::SystemTime::UNIX_EPOCH,
         });
     } else if version == 5 {
         info!("reading safekeeper control file version {}", version);
@@ -372,6 +465,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
         return Ok(TimelinePersistentState {
             tenant_id: oldstate.tenant_id,
             timeline_id: oldstate.timeline_id,
+            mconf: Configuration::empty(),
             acceptor_state: oldstate.acceptor_state,
             server: oldstate.server,
             proposer_uuid: oldstate.proposer_uuid,
@@ -381,9 +475,9 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
             backup_lsn: oldstate.backup_lsn,
             peer_horizon_lsn: oldstate.peer_horizon_lsn,
             remote_consistent_lsn: oldstate.remote_consistent_lsn,
-            peers: oldstate.peers,
             partial_backup: wal_backup_partial::State::default(),
             eviction_state: EvictionState::Present,
+            creation_ts: std::time::SystemTime::UNIX_EPOCH,
         });
     } else if version == 8 {
         let oldstate = SafeKeeperStateV8::des(&buf[..buf.len()])?;
@@ -391,6 +485,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
         return Ok(TimelinePersistentState {
             tenant_id: oldstate.tenant_id,
             timeline_id: oldstate.timeline_id,
+            mconf: Configuration::empty(),
             acceptor_state: oldstate.acceptor_state,
             server: oldstate.server,
             proposer_uuid: oldstate.proposer_uuid,
@@ -400,9 +495,28 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
             backup_lsn: oldstate.backup_lsn,
             peer_horizon_lsn: oldstate.peer_horizon_lsn,
             remote_consistent_lsn: oldstate.remote_consistent_lsn,
-            peers: oldstate.peers,
             partial_backup: oldstate.partial_backup,
             eviction_state: EvictionState::Present,
+            creation_ts: std::time::SystemTime::UNIX_EPOCH,
+        });
+    } else if version == 9 {
+        let oldstate = TimelinePersistentStateV9::des(&buf[..buf.len()])?;
+        return Ok(TimelinePersistentState {
+            tenant_id: oldstate.tenant_id,
+            timeline_id: oldstate.timeline_id,
+            mconf: Configuration::empty(),
+            acceptor_state: oldstate.acceptor_state,
+            server: oldstate.server,
+            proposer_uuid: oldstate.proposer_uuid,
+            timeline_start_lsn: oldstate.timeline_start_lsn,
+            local_start_lsn: oldstate.local_start_lsn,
+            commit_lsn: oldstate.commit_lsn,
+            backup_lsn: oldstate.backup_lsn,
+            peer_horizon_lsn: oldstate.peer_horizon_lsn,
+            remote_consistent_lsn: oldstate.remote_consistent_lsn,
+            partial_backup: oldstate.partial_backup,
+            eviction_state: oldstate.eviction_state,
+            creation_ts: std::time::SystemTime::UNIX_EPOCH,
         });
     }
 
@@ -412,9 +526,11 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<TimelinePersiste
     bail!("unsupported safekeeper control file version {}", version)
 }
 
-pub fn downgrade_v9_to_v8(state: &TimelinePersistentState) -> SafeKeeperStateV8 {
-    assert!(state.eviction_state == EvictionState::Present);
-    SafeKeeperStateV8 {
+// Used as a temp hack to make forward compatibility test work. Should be
+// removed after PR adding v10 is merged.
+pub fn downgrade_v10_to_v9(state: &TimelinePersistentState) -> TimelinePersistentStateV9 {
+    assert!(state.mconf.generation == INVALID_GENERATION);
+    TimelinePersistentStateV9 {
         tenant_id: state.tenant_id,
         timeline_id: state.timeline_id,
         acceptor_state: state.acceptor_state.clone(),
@@ -426,8 +542,9 @@ pub fn downgrade_v9_to_v8(state: &TimelinePersistentState) -> SafeKeeperStateV8
         backup_lsn: state.backup_lsn,
         peer_horizon_lsn: state.peer_horizon_lsn,
         remote_consistent_lsn: state.remote_consistent_lsn,
-        peers: state.peers.clone(),
+        peers: PersistedPeers(vec![]),
         partial_backup: state.partial_backup.clone(),
+        eviction_state: state.eviction_state,
     }
 }
 
@@ -437,7 +554,7 @@ mod tests {
 
     use utils::{id::NodeId, Hex};
 
-    use crate::safekeeper::PersistedPeerInfo;
+    use crate::control_file_upgrade::PersistedPeerInfo;
 
     use super::*;
 

safekeeper/src/copy_timeline.rs

@@ -1,6 +1,7 @@
 use anyhow::{bail, Result};
 use camino::Utf8PathBuf;
 use postgres_ffi::{MAX_SEND_SIZE, WAL_SEGMENT_SIZE};
+use safekeeper_api::membership::Configuration;
 use std::sync::Arc;
 use tokio::{
     fs::OpenOptions,
@@ -147,10 +148,10 @@ pub async fn handle_request(
 
     let mut new_state = TimelinePersistentState::new(
         &request.destination_ttid,
+        Configuration::empty(),
         state.server.clone(),
-        vec![],
-        request.until_lsn,
         start_lsn,
+        request.until_lsn,
     )?;
     new_state.timeline_start_lsn = start_lsn;
     new_state.peer_horizon_lsn = request.until_lsn;

safekeeper/src/http/routes.rs

@@ -1,4 +1,5 @@
 use hyper::{Body, Request, Response, StatusCode};
+use safekeeper_api::models;
 use safekeeper_api::models::AcceptorStateStatus;
 use safekeeper_api::models::SafekeeperStatus;
 use safekeeper_api::models::TermSwitchApiEntry;
@@ -111,14 +112,15 @@ async fn timeline_create_handler(mut request: Request<Body>) -> Result<Response<
         system_id: request_data.system_id.unwrap_or(0),
         wal_seg_size: request_data.wal_seg_size.unwrap_or(WAL_SEGMENT_SIZE as u32),
     };
-    let local_start_lsn = request_data.local_start_lsn.unwrap_or_else(|| {
-        request_data
-            .commit_lsn
-            .segment_lsn(server_info.wal_seg_size as usize)
-    });
     let global_timelines = get_global_timelines(&request);
     global_timelines
-        .create(ttid, server_info, request_data.commit_lsn, local_start_lsn)
+        .create(
+            ttid,
+            request_data.mconf,
+            server_info,
+            request_data.start_lsn,
+            request_data.commit_lsn.unwrap_or(request_data.start_lsn),
+        )
         .await
         .map_err(ApiError::InternalServerError)?;
 
@@ -182,6 +184,7 @@ async fn timeline_status_handler(request: Request<Body>) -> Result<Response<Body
     let status = TimelineStatus {
         tenant_id: ttid.tenant_id,
         timeline_id: ttid.timeline_id,
+        mconf: state.mconf,
         acceptor_state: acc_state,
         pg_info: state.server,
         flush_lsn,
@@ -192,7 +195,7 @@ async fn timeline_status_handler(request: Request<Body>) -> Result<Response<Body
         peer_horizon_lsn: inmem.peer_horizon_lsn,
         remote_consistent_lsn: inmem.remote_consistent_lsn,
         peers: tli.get_peers(conf).await,
-        walsenders: tli.get_walsenders().get_all(),
+        walsenders: tli.get_walsenders().get_all_public(),
         walreceivers: tli.get_walreceivers().get_all(),
     };
     json_response(StatusCode::OK, status)
@@ -267,6 +270,28 @@ async fn timeline_snapshot_handler(request: Request<Body>) -> Result<Response<Bo
     Ok(response)
 }
 
+/// Consider switching timeline membership configuration to the provided one.
+async fn timeline_membership_handler(
+    mut request: Request<Body>,
+) -> Result<Response<Body>, ApiError> {
+    let ttid = TenantTimelineId::new(
+        parse_request_param(&request, "tenant_id")?,
+        parse_request_param(&request, "timeline_id")?,
+    );
+    check_permission(&request, Some(ttid.tenant_id))?;
+
+    let global_timelines = get_global_timelines(&request);
+    let tli = global_timelines.get(ttid).map_err(ApiError::from)?;
+
+    let data: models::TimelineMembershipSwitchRequest = json_request(&mut request).await?;
+    let response = tli
+        .membership_switch(data.mconf)
+        .await
+        .map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, response)
+}
+
 async fn timeline_copy_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
     check_permission(&request, None)?;
 
@@ -618,6 +643,10 @@ pub fn make_router(
             "/v1/tenant/:tenant_id/timeline/:timeline_id/snapshot/:destination_id",
             |r| request_span(r, timeline_snapshot_handler),
         )
+        .post(
+            "/v1/tenant/:tenant_id/timeline/:timeline_id/membership",
+            |r| request_span(r, timeline_membership_handler),
+        )
         .post(
             "/v1/tenant/:tenant_id/timeline/:source_timeline_id/copy",
             |r| request_span(r, timeline_copy_handler),

safekeeper/src/json_ctrl.rs

@@ -8,6 +8,7 @@
 
 use anyhow::Context;
 use postgres_backend::QueryError;
+use safekeeper_api::membership::Configuration;
 use safekeeper_api::{ServerInfo, Term};
 use serde::{Deserialize, Serialize};
 use tokio::io::{AsyncRead, AsyncWrite};
@@ -105,6 +106,7 @@ async fn prepare_safekeeper(
         .global_timelines
         .create(
             spg.ttid,
+            Configuration::empty(),
             ServerInfo {
                 pg_version,
                 wal_seg_size: WAL_SEGMENT_SIZE as u32,

safekeeper/src/lib.rs

@@ -108,6 +108,8 @@ pub struct SafeKeeperConf {
     pub control_file_save_interval: Duration,
     pub partial_backup_concurrency: usize,
     pub eviction_min_resident: Duration,
+    pub wal_reader_fanout: bool,
+    pub max_delta_for_fanout: Option<u64>,
 }
 
 impl SafeKeeperConf {
@@ -150,6 +152,8 @@ impl SafeKeeperConf {
             control_file_save_interval: Duration::from_secs(1),
             partial_backup_concurrency: 1,
             eviction_min_resident: Duration::ZERO,
+            wal_reader_fanout: false,
+            max_delta_for_fanout: None,
         }
     }
 }

safekeeper/src/metrics.rs

@@ -12,9 +12,9 @@ use metrics::{
     pow2_buckets,
     proto::MetricFamily,
     register_histogram, register_histogram_vec, register_int_counter, register_int_counter_pair,
-    register_int_counter_pair_vec, register_int_counter_vec, register_int_gauge, Gauge, GaugeVec,
-    Histogram, HistogramVec, IntCounter, IntCounterPair, IntCounterPairVec, IntCounterVec,
-    IntGauge, IntGaugeVec, DISK_FSYNC_SECONDS_BUCKETS,
+    register_int_counter_pair_vec, register_int_counter_vec, register_int_gauge,
+    register_int_gauge_vec, Gauge, GaugeVec, Histogram, HistogramVec, IntCounter, IntCounterPair,
+    IntCounterPairVec, IntCounterVec, IntGauge, IntGaugeVec, DISK_FSYNC_SECONDS_BUCKETS,
 };
 use once_cell::sync::Lazy;
 use postgres_ffi::XLogSegNo;
@@ -211,6 +211,14 @@ pub static WAL_RECEIVERS: Lazy<IntGauge> = Lazy::new(|| {
     )
     .expect("Failed to register safekeeper_wal_receivers")
 });
+pub static WAL_READERS: Lazy<IntGaugeVec> = Lazy::new(|| {
+    register_int_gauge_vec!(
+        "safekeeper_wal_readers",
+        "Number of active WAL readers (may serve pageservers or other safekeepers)",
+        &["kind", "target"]
+    )
+    .expect("Failed to register safekeeper_wal_receivers")
+});
 pub static WAL_RECEIVER_QUEUE_DEPTH: Lazy<Histogram> = Lazy::new(|| {
     // Use powers of two buckets, but add a bucket at 0 and the max queue size to track empty and
     // full queues respectively.
@@ -443,6 +451,7 @@ pub struct FullTimelineInfo {
     pub timeline_is_active: bool,
     pub num_computes: u32,
     pub last_removed_segno: XLogSegNo,
+    pub interpreted_wal_reader_tasks: usize,
 
     pub epoch_start_lsn: Lsn,
     pub mem_state: TimelineMemState,
@@ -472,6 +481,7 @@ pub struct TimelineCollector {
     disk_usage: GenericGaugeVec<AtomicU64>,
     acceptor_term: GenericGaugeVec<AtomicU64>,
     written_wal_bytes: GenericGaugeVec<AtomicU64>,
+    interpreted_wal_reader_tasks: GenericGaugeVec<AtomicU64>,
     written_wal_seconds: GaugeVec,
     flushed_wal_seconds: GaugeVec,
     collect_timeline_metrics: Gauge,
@@ -670,6 +680,16 @@ impl TimelineCollector {
         .unwrap();
         descs.extend(active_timelines_count.desc().into_iter().cloned());
 
+        let interpreted_wal_reader_tasks = GenericGaugeVec::new(
+            Opts::new(
+                "safekeeper_interpreted_wal_reader_tasks",
+                "Number of active interpreted wal reader tasks, grouped by timeline",
+            ),
+            &["tenant_id", "timeline_id"],
+        )
+        .unwrap();
+        descs.extend(interpreted_wal_reader_tasks.desc().into_iter().cloned());
+
         TimelineCollector {
             global_timelines,
             descs,
@@ -693,6 +713,7 @@ impl TimelineCollector {
             collect_timeline_metrics,
             timelines_count,
             active_timelines_count,
+            interpreted_wal_reader_tasks,
         }
     }
 }
@@ -721,6 +742,7 @@ impl Collector for TimelineCollector {
         self.disk_usage.reset();
         self.acceptor_term.reset();
         self.written_wal_bytes.reset();
+        self.interpreted_wal_reader_tasks.reset();
         self.written_wal_seconds.reset();
         self.flushed_wal_seconds.reset();
 
@@ -782,6 +804,9 @@ impl Collector for TimelineCollector {
             self.written_wal_bytes
                 .with_label_values(labels)
                 .set(tli.wal_storage.write_wal_bytes);
+            self.interpreted_wal_reader_tasks
+                .with_label_values(labels)
+                .set(tli.interpreted_wal_reader_tasks as u64);
             self.written_wal_seconds
                 .with_label_values(labels)
                 .set(tli.wal_storage.write_wal_seconds);
@@ -834,6 +859,7 @@ impl Collector for TimelineCollector {
         mfs.extend(self.disk_usage.collect());
         mfs.extend(self.acceptor_term.collect());
         mfs.extend(self.written_wal_bytes.collect());
+        mfs.extend(self.interpreted_wal_reader_tasks.collect());
         mfs.extend(self.written_wal_seconds.collect());
         mfs.extend(self.flushed_wal_seconds.collect());
 

safekeeper/src/receive_wal.rs

@@ -21,6 +21,7 @@ use postgres_backend::PostgresBackend;
 use postgres_backend::PostgresBackendReader;
 use postgres_backend::QueryError;
 use pq_proto::BeMessage;
+use safekeeper_api::membership::Configuration;
 use safekeeper_api::models::{ConnectionId, WalReceiverState, WalReceiverStatus};
 use safekeeper_api::ServerInfo;
 use std::future;
@@ -337,7 +338,13 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> NetworkReader<'_, IO> {
                 };
                 let tli = self
                     .global_timelines
-                    .create(self.ttid, server_info, Lsn::INVALID, Lsn::INVALID)
+                    .create(
+                        self.ttid,
+                        Configuration::empty(),
+                        server_info,
+                        Lsn::INVALID,
+                        Lsn::INVALID,
+                    )
                     .await
                     .context("create timeline")?;
                 tli.wal_residence_guard().await?

safekeeper/src/safekeeper.rs

@@ -7,7 +7,6 @@ use bytes::{Buf, BufMut, Bytes, BytesMut};
 use postgres_ffi::{TimeLineID, MAX_SEND_SIZE};
 use safekeeper_api::models::HotStandbyFeedback;
 use safekeeper_api::Term;
-use safekeeper_api::INVALID_TERM;
 use serde::{Deserialize, Serialize};
 use std::cmp::max;
 use std::cmp::min;
@@ -193,36 +192,6 @@ impl AcceptorState {
     }
 }
 
-#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
-pub struct PersistedPeerInfo {
-    /// LSN up to which safekeeper offloaded WAL to s3.
-    pub backup_lsn: Lsn,
-    /// Term of the last entry.
-    pub term: Term,
-    /// LSN of the last record.
-    pub flush_lsn: Lsn,
-    /// Up to which LSN safekeeper regards its WAL as committed.
-    pub commit_lsn: Lsn,
-}
-
-impl PersistedPeerInfo {
-    pub fn new() -> Self {
-        Self {
-            backup_lsn: Lsn::INVALID,
-            term: INVALID_TERM,
-            flush_lsn: Lsn(0),
-            commit_lsn: Lsn(0),
-        }
-    }
-}
-
-// make clippy happy
-impl Default for PersistedPeerInfo {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
 // protocol messages
 
 /// Initial Proposer -> Acceptor message
@@ -1010,7 +979,7 @@ where
 
     /// Update commit_lsn from peer safekeeper data.
     pub async fn record_safekeeper_info(&mut self, sk_info: &SafekeeperTimelineInfo) -> Result<()> {
-        if (Lsn(sk_info.commit_lsn) != Lsn::INVALID) && (sk_info.last_log_term != INVALID_TERM) {
+        if Lsn(sk_info.commit_lsn) != Lsn::INVALID {
             // Note: the check is too restrictive, generally we can update local
             // commit_lsn if our history matches (is part of) history of advanced
             // commit_lsn provider.
@@ -1025,12 +994,20 @@ where
 #[cfg(test)]
 mod tests {
     use futures::future::BoxFuture;
+
     use postgres_ffi::{XLogSegNo, WAL_SEGMENT_SIZE};
-    use safekeeper_api::ServerInfo;
+    use safekeeper_api::{
+        membership::{Configuration, MemberSet, SafekeeperId},
+        ServerInfo,
+    };
 
     use super::*;
-    use crate::state::{EvictionState, PersistedPeers, TimelinePersistentState};
-    use std::{ops::Deref, str::FromStr, time::Instant};
+    use crate::state::{EvictionState, TimelinePersistentState};
+    use std::{
+        ops::Deref,
+        str::FromStr,
+        time::{Instant, UNIX_EPOCH},
+    };
 
     // fake storage for tests
     struct InMemoryState {
@@ -1313,12 +1290,21 @@ mod tests {
 
     #[test]
     fn test_sk_state_bincode_serde_roundtrip() {
-        use utils::Hex;
         let tenant_id = TenantId::from_str("cf0480929707ee75372337efaa5ecf96").unwrap();
         let timeline_id = TimelineId::from_str("112ded66422aa5e953e5440fa5427ac4").unwrap();
         let state = TimelinePersistentState {
             tenant_id,
             timeline_id,
+            mconf: Configuration {
+                generation: 42,
+                members: MemberSet::new(vec![SafekeeperId {
+                    id: NodeId(1),
+                    host: "hehe.org".to_owned(),
+                    pg_port: 5432,
+                }])
+                .expect("duplicate member"),
+                new_members: None,
+            },
             acceptor_state: AcceptorState {
                 term: 42,
                 term_history: TermHistory(vec![TermLsn {
@@ -1342,70 +1328,13 @@ mod tests {
             backup_lsn: Lsn(1234567300),
             peer_horizon_lsn: Lsn(9999999),
             remote_consistent_lsn: Lsn(1234560000),
-            peers: PersistedPeers(vec![(
-                NodeId(1),
-                PersistedPeerInfo {
-                    backup_lsn: Lsn(1234567000),
-                    term: 42,
-                    flush_lsn: Lsn(1234567800 - 8),
-                    commit_lsn: Lsn(1234567600),
-                },
-            )]),
             partial_backup: crate::wal_backup_partial::State::default(),
             eviction_state: EvictionState::Present,
+            creation_ts: UNIX_EPOCH,
         };
 
         let ser = state.ser().unwrap();
 
-        #[rustfmt::skip]
-        let expected = [
-            // tenant_id as length prefixed hex
-            0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            0x63, 0x66, 0x30, 0x34, 0x38, 0x30, 0x39, 0x32, 0x39, 0x37, 0x30, 0x37, 0x65, 0x65, 0x37, 0x35, 0x33, 0x37, 0x32, 0x33, 0x33, 0x37, 0x65, 0x66, 0x61, 0x61, 0x35, 0x65, 0x63, 0x66, 0x39, 0x36,
-            // timeline_id as length prefixed hex
-            0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            0x31, 0x31, 0x32, 0x64, 0x65, 0x64, 0x36, 0x36, 0x34, 0x32, 0x32, 0x61, 0x61, 0x35, 0x65, 0x39, 0x35, 0x33, 0x65, 0x35, 0x34, 0x34, 0x30, 0x66, 0x61, 0x35, 0x34, 0x32, 0x37, 0x61, 0x63, 0x34,
-            // term
-            0x2a, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            // length prefix
-            0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            // unsure why this order is swapped
-            0x29, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            // pg_version
-            0x0e, 0x00, 0x00, 0x00,
-            // systemid
-            0x21, 0x43, 0x65, 0x87, 0x78, 0x56, 0x34, 0x12,
-            // wal_seg_size
-            0x78, 0x56, 0x34, 0x12,
-            // pguuid as length prefixed hex
-            0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            0x63, 0x34, 0x37, 0x61, 0x34, 0x32, 0x61, 0x35, 0x30, 0x66, 0x34, 0x34, 0x65, 0x35, 0x35, 0x33, 0x65, 0x39, 0x61, 0x35, 0x32, 0x61, 0x34, 0x32, 0x36, 0x36, 0x65, 0x64, 0x32, 0x64, 0x31, 0x31,
-
-            // timeline_start_lsn
-            0x00, 0x56, 0x34, 0x12, 0x00, 0x00, 0x00, 0x00,
-            0x12, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            0x78, 0x02, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00,
-            0x84, 0x00, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00,
-            0x7f, 0x96, 0x98, 0x00, 0x00, 0x00, 0x00, 0x00,
-            0x00, 0xe4, 0x95, 0x49, 0x00, 0x00, 0x00, 0x00,
-            // length prefix for persistentpeers
-            0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            // nodeid
-            0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            // backuplsn
-            0x58, 0xff, 0x95, 0x49, 0x00, 0x00, 0x00, 0x00,
-            0x2a, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            0x70, 0x02, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00,
-            0xb0, 0x01, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00,
-            // partial_backup
-            0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-            // eviction_state
-            0x00, 0x00, 0x00, 0x00,
-        ];
-
-        assert_eq!(Hex(&ser), Hex(&expected));
-
         let deser = TimelinePersistentState::des(&ser).unwrap();
 
         assert_eq!(deser, state);

safekeeper/src/send_interpreted_wal.rs

@@ -1,96 +1,330 @@
+use std::collections::HashMap;
+use std::fmt::Display;
+use std::sync::Arc;
 use std::time::Duration;
 
-use anyhow::Context;
+use anyhow::{anyhow, Context};
+use futures::future::Either;
 use futures::StreamExt;
 use pageserver_api::shard::ShardIdentity;
 use postgres_backend::{CopyStreamHandlerEnd, PostgresBackend};
-use postgres_ffi::MAX_SEND_SIZE;
+use postgres_ffi::waldecoder::WalDecodeError;
 use postgres_ffi::{get_current_timestamp, waldecoder::WalStreamDecoder};
 use pq_proto::{BeMessage, InterpretedWalRecordsBody, WalSndKeepAlive};
 use tokio::io::{AsyncRead, AsyncWrite};
+use tokio::sync::mpsc::error::SendError;
+use tokio::task::JoinHandle;
 use tokio::time::MissedTickBehavior;
+use tracing::{info_span, Instrument};
 use utils::lsn::Lsn;
 use utils::postgres_client::Compression;
 use utils::postgres_client::InterpretedFormat;
 use wal_decoder::models::{InterpretedWalRecord, InterpretedWalRecords};
 use wal_decoder::wire_format::ToWireFormat;
 
-use crate::send_wal::EndWatchView;
-use crate::wal_reader_stream::{WalBytes, WalReaderStreamBuilder};
+use crate::metrics::WAL_READERS;
+use crate::send_wal::{EndWatchView, WalSenderGuard};
+use crate::timeline::WalResidentTimeline;
+use crate::wal_reader_stream::{StreamingWalReader, WalBytes};
 
-/// Shard-aware interpreted record sender.
-/// This is used for sending WAL to the pageserver. Said WAL
-/// is pre-interpreted and filtered for the shard.
-pub(crate) struct InterpretedWalSender<'a, IO> {
-    pub(crate) format: InterpretedFormat,
-    pub(crate) compression: Option<Compression>,
-    pub(crate) pgb: &'a mut PostgresBackend<IO>,
-    pub(crate) wal_stream_builder: WalReaderStreamBuilder,
-    pub(crate) end_watch_view: EndWatchView,
-    pub(crate) shard: ShardIdentity,
-    pub(crate) pg_version: u32,
-    pub(crate) appname: Option<String>,
+/// Identifier used to differentiate between senders of the same
+/// shard.
+///
+/// In the steady state there's only one, but two pageservers may
+/// temporarily have the same shard attached and attempt to ingest
+/// WAL for it. See also [`ShardSenderId`].
+#[derive(Hash, Eq, PartialEq, Copy, Clone)]
+struct SenderId(u8);
+
+impl SenderId {
+    fn first() -> Self {
+        SenderId(0)
+    }
+
+    fn next(&self) -> Self {
+        SenderId(self.0.checked_add(1).expect("few senders"))
+    }
 }
 
-struct Batch {
+#[derive(Hash, Eq, PartialEq)]
+struct ShardSenderId {
+    shard: ShardIdentity,
+    sender_id: SenderId,
+}
+
+impl Display for ShardSenderId {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}{}", self.sender_id.0, self.shard.shard_slug())
+    }
+}
+
+impl ShardSenderId {
+    fn new(shard: ShardIdentity, sender_id: SenderId) -> Self {
+        ShardSenderId { shard, sender_id }
+    }
+
+    fn shard(&self) -> ShardIdentity {
+        self.shard
+    }
+}
+
+/// Shard-aware fan-out interpreted record reader.
+/// Reads WAL from disk, decodes it, intepretets it, and sends
+/// it to any [`InterpretedWalSender`] connected to it.
+/// Each [`InterpretedWalSender`] corresponds to one shard
+/// and gets interpreted records concerning that shard only.
+pub(crate) struct InterpretedWalReader {
+    wal_stream: StreamingWalReader,
+    shard_senders: HashMap<ShardIdentity, smallvec::SmallVec<[ShardSenderState; 1]>>,
+    shard_notification_rx: Option<tokio::sync::mpsc::UnboundedReceiver<AttachShardNotification>>,
+    state: Arc<std::sync::RwLock<InterpretedWalReaderState>>,
+    pg_version: u32,
+}
+
+/// A handle for [`InterpretedWalReader`] which allows for interacting with it
+/// when it runs as a separate tokio task.
+#[derive(Debug)]
+pub(crate) struct InterpretedWalReaderHandle {
+    join_handle: JoinHandle<Result<(), InterpretedWalReaderError>>,
+    state: Arc<std::sync::RwLock<InterpretedWalReaderState>>,
+    shard_notification_tx: tokio::sync::mpsc::UnboundedSender<AttachShardNotification>,
+}
+
+struct ShardSenderState {
+    sender_id: SenderId,
+    tx: tokio::sync::mpsc::Sender<Batch>,
+    next_record_lsn: Lsn,
+}
+
+/// State of [`InterpretedWalReader`] visible outside of the task running it.
+#[derive(Debug)]
+pub(crate) enum InterpretedWalReaderState {
+    Running { current_position: Lsn },
+    Done,
+}
+
+pub(crate) struct Batch {
     wal_end_lsn: Lsn,
     available_wal_end_lsn: Lsn,
     records: InterpretedWalRecords,
 }
 
-impl<IO: AsyncRead + AsyncWrite + Unpin> InterpretedWalSender<'_, IO> {
-    /// Send interpreted WAL to a receiver.
-    /// Stops when an error occurs or the receiver is caught up and there's no active compute.
-    ///
-    /// Err(CopyStreamHandlerEnd) is always returned; Result is used only for ?
-    /// convenience.
-    pub(crate) async fn run(self) -> Result<(), CopyStreamHandlerEnd> {
-        let mut wal_position = self.wal_stream_builder.start_pos();
-        let mut wal_decoder =
-            WalStreamDecoder::new(self.wal_stream_builder.start_pos(), self.pg_version);
+#[derive(thiserror::Error, Debug)]
+pub enum InterpretedWalReaderError {
+    /// Handler initiates the end of streaming.
+    #[error("decode error: {0}")]
+    Decode(#[from] WalDecodeError),
+    #[error("read or interpret error: {0}")]
+    ReadOrInterpret(#[from] anyhow::Error),
+    #[error("wal stream closed")]
+    WalStreamClosed,
+}
 
-        let stream = self.wal_stream_builder.build(MAX_SEND_SIZE).await?;
-        let mut stream = std::pin::pin!(stream);
+impl InterpretedWalReaderState {
+    fn current_position(&self) -> Option<Lsn> {
+        match self {
+            InterpretedWalReaderState::Running {
+                current_position, ..
+            } => Some(*current_position),
+            InterpretedWalReaderState::Done => None,
+        }
+    }
+}
 
-        let mut keepalive_ticker = tokio::time::interval(Duration::from_secs(1));
-        keepalive_ticker.set_missed_tick_behavior(MissedTickBehavior::Skip);
-        keepalive_ticker.reset();
+pub(crate) struct AttachShardNotification {
+    shard_id: ShardIdentity,
+    sender: tokio::sync::mpsc::Sender<Batch>,
+    start_pos: Lsn,
+}
 
-        let (tx, mut rx) = tokio::sync::mpsc::channel::<Batch>(2);
+impl InterpretedWalReader {
+    /// Spawn the reader in a separate tokio task and return a handle
+    pub(crate) fn spawn(
+        wal_stream: StreamingWalReader,
+        start_pos: Lsn,
+        tx: tokio::sync::mpsc::Sender<Batch>,
+        shard: ShardIdentity,
+        pg_version: u32,
+        appname: &Option<String>,
+    ) -> InterpretedWalReaderHandle {
+        let state = Arc::new(std::sync::RwLock::new(InterpretedWalReaderState::Running {
+            current_position: start_pos,
+        }));
+
+        let (shard_notification_tx, shard_notification_rx) = tokio::sync::mpsc::unbounded_channel();
+
+        let reader = InterpretedWalReader {
+            wal_stream,
+            shard_senders: HashMap::from([(
+                shard,
+                smallvec::smallvec![ShardSenderState {
+                    sender_id: SenderId::first(),
+                    tx,
+                    next_record_lsn: start_pos,
+                }],
+            )]),
+            shard_notification_rx: Some(shard_notification_rx),
+            state: state.clone(),
+            pg_version,
+        };
+
+        let metric = WAL_READERS
+            .get_metric_with_label_values(&["task", appname.as_deref().unwrap_or("safekeeper")])
+            .unwrap();
+
+        let join_handle = tokio::task::spawn(
+            async move {
+                metric.inc();
+                scopeguard::defer! {
+                    metric.dec();
+                }
+
+                let res = reader.run_impl(start_pos).await;
+                if let Err(ref err) = res {
+                    tracing::error!("Task finished with error: {err}");
+                }
+                res
+            }
+            .instrument(info_span!("interpreted wal reader")),
+        );
+
+        InterpretedWalReaderHandle {
+            join_handle,
+            state,
+            shard_notification_tx,
+        }
+    }
+
+    /// Construct the reader without spawning anything
+    /// Callers should drive the future returned by [`Self::run`].
+    pub(crate) fn new(
+        wal_stream: StreamingWalReader,
+        start_pos: Lsn,
+        tx: tokio::sync::mpsc::Sender<Batch>,
+        shard: ShardIdentity,
+        pg_version: u32,
+    ) -> InterpretedWalReader {
+        let state = Arc::new(std::sync::RwLock::new(InterpretedWalReaderState::Running {
+            current_position: start_pos,
+        }));
+
+        InterpretedWalReader {
+            wal_stream,
+            shard_senders: HashMap::from([(
+                shard,
+                smallvec::smallvec![ShardSenderState {
+                    sender_id: SenderId::first(),
+                    tx,
+                    next_record_lsn: start_pos,
+                }],
+            )]),
+            shard_notification_rx: None,
+            state: state.clone(),
+            pg_version,
+        }
+    }
+
+    /// Entry point for future (polling) based wal reader.
+    pub(crate) async fn run(
+        self,
+        start_pos: Lsn,
+        appname: &Option<String>,
+    ) -> Result<(), CopyStreamHandlerEnd> {
+        let metric = WAL_READERS
+            .get_metric_with_label_values(&["future", appname.as_deref().unwrap_or("safekeeper")])
+            .unwrap();
+
+        metric.inc();
+        scopeguard::defer! {
+            metric.dec();
+        }
+
+        let res = self.run_impl(start_pos).await;
+        if let Err(err) = res {
+            tracing::error!("Interpreted wal reader encountered error: {err}");
+        } else {
+            tracing::info!("Interpreted wal reader exiting");
+        }
+
+        Err(CopyStreamHandlerEnd::Other(anyhow!(
+            "interpreted wal reader finished"
+        )))
+    }
+
+    /// Send interpreted WAL to one or more [`InterpretedWalSender`]s
+    /// Stops when an error is encountered or when the [`InterpretedWalReaderHandle`]
+    /// goes out of scope.
+    async fn run_impl(mut self, start_pos: Lsn) -> Result<(), InterpretedWalReaderError> {
+        let defer_state = self.state.clone();
+        scopeguard::defer! {
+            *defer_state.write().unwrap() = InterpretedWalReaderState::Done;
+        }
+
+        let mut wal_decoder = WalStreamDecoder::new(start_pos, self.pg_version);
 
         loop {
             tokio::select! {
-                // Get some WAL from the stream and then: decode, interpret and push it down the
-                // pipeline.
-                wal = stream.next(), if tx.capacity() > 0 => {
-                    let WalBytes { wal, wal_start_lsn: _, wal_end_lsn, available_wal_end_lsn } = match wal {
-                        Some(some) => some?,
-                        None => { break; }
+                // Main branch for reading WAL and forwarding it
+                wal_or_reset = self.wal_stream.next() => {
+                    let wal = wal_or_reset.map(|wor| wor.get_wal().expect("reset handled in select branch below"));
+                    let WalBytes {
+                        wal,
+                        wal_start_lsn: _,
+                        wal_end_lsn,
+                        available_wal_end_lsn,
+                    } = match wal {
+                        Some(some) => some.map_err(InterpretedWalReaderError::ReadOrInterpret)?,
+                        None => {
+                            // [`StreamingWalReader::next`] is an endless stream of WAL.
+                            // It shouldn't ever finish unless it panicked or became internally
+                            // inconsistent.
+                            return Result::Err(InterpretedWalReaderError::WalStreamClosed);
+                        }
                     };
 
-                    wal_position = wal_end_lsn;
                     wal_decoder.feed_bytes(&wal);
 
-                    let mut records = Vec::new();
+                    // Deserialize and interpret WAL records from this batch of WAL.
+                    // Interpreted records for each shard are collected separately.
+                    let shard_ids = self.shard_senders.keys().copied().collect::<Vec<_>>();
+                    let mut records_by_sender: HashMap<ShardSenderId, Vec<InterpretedWalRecord>> = HashMap::new();
                     let mut max_next_record_lsn = None;
-                    while let Some((next_record_lsn, recdata)) = wal_decoder
-                        .poll_decode()
-                        .with_context(|| "Failed to decode WAL")?
+                    while let Some((next_record_lsn, recdata)) = wal_decoder.poll_decode()?
                     {
                         assert!(next_record_lsn.is_aligned());
                         max_next_record_lsn = Some(next_record_lsn);
 
-                        // Deserialize and interpret WAL record
                         let interpreted = InterpretedWalRecord::from_bytes_filtered(
                             recdata,
-                            &self.shard,
+                            &shard_ids,
                             next_record_lsn,
                             self.pg_version,
                         )
                         .with_context(|| "Failed to interpret WAL")?;
 
-                        if !interpreted.is_empty() {
-                            records.push(interpreted);
+                        for (shard, record) in interpreted {
+                            if record.is_empty() {
+                                continue;
+                            }
+
+                            let mut states_iter = self.shard_senders
+                                .get(&shard)
+                                .expect("keys collected above")
+                                .iter()
+                                .filter(|state| record.next_record_lsn > state.next_record_lsn)
+                                .peekable();
+                            while let Some(state) = states_iter.next() {
+                                let shard_sender_id = ShardSenderId::new(shard, state.sender_id);
+
+                                // The most commont case is one sender per shard. Peek and break to avoid the
+                                // clone in that situation.
+                                if states_iter.peek().is_none() {
+                                    records_by_sender.entry(shard_sender_id).or_default().push(record);
+                                    break;
+                                } else {
+                                    records_by_sender.entry(shard_sender_id).or_default().push(record.clone());
+                                }
+                            }
                         }
                     }
 
@@ -99,20 +333,170 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> InterpretedWalSender<'_, IO> {
                         None => { continue; }
                     };
 
-                    let batch = InterpretedWalRecords {
-                        records,
-                        next_record_lsn: Some(max_next_record_lsn),
-                    };
+                    // Update the current position such that new receivers can decide
+                    // whether to attach to us or spawn a new WAL reader.
+                    match &mut *self.state.write().unwrap() {
+                        InterpretedWalReaderState::Running { current_position, .. } => {
+                            *current_position = max_next_record_lsn;
+                        },
+                        InterpretedWalReaderState::Done => {
+                            unreachable!()
+                        }
+                    }
 
-                    tx.send(Batch {wal_end_lsn, available_wal_end_lsn, records: batch}).await.unwrap();
+                    // Send interpreted records downstream. Anything that has already been seen
+                    // by a shard is filtered out.
+                    let mut shard_senders_to_remove = Vec::new();
+                    for (shard, states) in &mut self.shard_senders {
+                        for state in states {
+                            if max_next_record_lsn <= state.next_record_lsn {
+                                continue;
+                            }
+
+                            let shard_sender_id = ShardSenderId::new(*shard, state.sender_id);
+                            let records = records_by_sender.remove(&shard_sender_id).unwrap_or_default();
+
+                            let batch = InterpretedWalRecords {
+                                records,
+                                next_record_lsn: Some(max_next_record_lsn),
+                            };
+
+                            let res = state.tx.send(Batch {
+                                wal_end_lsn,
+                                available_wal_end_lsn,
+                                records: batch,
+                            }).await;
+
+                            if res.is_err() {
+                                shard_senders_to_remove.push(shard_sender_id);
+                            } else {
+                                state.next_record_lsn = max_next_record_lsn;
+                            }
+                        }
+                    }
+
+                    // Clean up any shard senders that have dropped out.
+                    // This is inefficient, but such events are rare (connection to PS termination)
+                    // and the number of subscriptions on the same shards very small (only one
+                    // for the steady state).
+                    for to_remove in shard_senders_to_remove {
+                        let shard_senders = self.shard_senders.get_mut(&to_remove.shard()).expect("saw it above");
+                        if let Some(idx) = shard_senders.iter().position(|s| s.sender_id == to_remove.sender_id) {
+                            shard_senders.remove(idx);
+                            tracing::info!("Removed shard sender {}", to_remove);
+                        }
+
+                        if shard_senders.is_empty() {
+                            self.shard_senders.remove(&to_remove.shard());
+                        }
+                    }
                 },
-                // For a previously interpreted batch, serialize it and push it down the wire.
-                batch = rx.recv() => {
+                // Listen for new shards that want to attach to this reader.
+                // If the reader is not running as a task, then this is not supported
+                // (see the pending branch below).
+                notification = match self.shard_notification_rx.as_mut() {
+                        Some(rx) => Either::Left(rx.recv()),
+                        None => Either::Right(std::future::pending())
+                    } => {
+                    if let Some(n) = notification {
+                        let AttachShardNotification { shard_id, sender, start_pos } = n;
+
+                        // Update internal and external state, then reset the WAL stream
+                        // if required.
+                        let senders = self.shard_senders.entry(shard_id).or_default();
+                        let new_sender_id = match senders.last() {
+                            Some(sender) => sender.sender_id.next(),
+                            None => SenderId::first()
+                        };
+
+                        senders.push(ShardSenderState { sender_id: new_sender_id, tx: sender, next_record_lsn: start_pos});
+                        let current_pos = self.state.read().unwrap().current_position().unwrap();
+                        if start_pos < current_pos {
+                            self.wal_stream.reset(start_pos).await;
+                            wal_decoder = WalStreamDecoder::new(start_pos, self.pg_version);
+                        }
+
+                        tracing::info!(
+                            "Added shard sender {} with start_pos={} current_pos={}",
+                            ShardSenderId::new(shard_id, new_sender_id), start_pos, current_pos
+                        );
+                    }
+                }
+            }
+        }
+    }
+}
+
+impl InterpretedWalReaderHandle {
+    /// Fan-out the reader by attaching a new shard to it
+    pub(crate) fn fanout(
+        &self,
+        shard_id: ShardIdentity,
+        sender: tokio::sync::mpsc::Sender<Batch>,
+        start_pos: Lsn,
+    ) -> Result<(), SendError<AttachShardNotification>> {
+        self.shard_notification_tx.send(AttachShardNotification {
+            shard_id,
+            sender,
+            start_pos,
+        })
+    }
+
+    /// Get the current WAL position of the reader
+    pub(crate) fn current_position(&self) -> Option<Lsn> {
+        self.state.read().unwrap().current_position()
+    }
+
+    pub(crate) fn abort(&self) {
+        self.join_handle.abort()
+    }
+}
+
+impl Drop for InterpretedWalReaderHandle {
+    fn drop(&mut self) {
+        tracing::info!("Aborting interpreted wal reader");
+        self.abort()
+    }
+}
+
+pub(crate) struct InterpretedWalSender<'a, IO> {
+    pub(crate) format: InterpretedFormat,
+    pub(crate) compression: Option<Compression>,
+    pub(crate) appname: Option<String>,
+
+    pub(crate) tli: WalResidentTimeline,
+    pub(crate) start_lsn: Lsn,
+
+    pub(crate) pgb: &'a mut PostgresBackend<IO>,
+    pub(crate) end_watch_view: EndWatchView,
+    pub(crate) wal_sender_guard: Arc<WalSenderGuard>,
+    pub(crate) rx: tokio::sync::mpsc::Receiver<Batch>,
+}
+
+impl<IO: AsyncRead + AsyncWrite + Unpin> InterpretedWalSender<'_, IO> {
+    /// Send interpreted WAL records over the network.
+    /// Also manages keep-alives if nothing was sent for a while.
+    pub(crate) async fn run(mut self) -> Result<(), CopyStreamHandlerEnd> {
+        let mut keepalive_ticker = tokio::time::interval(Duration::from_secs(1));
+        keepalive_ticker.set_missed_tick_behavior(MissedTickBehavior::Skip);
+        keepalive_ticker.reset();
+
+        let mut wal_position = self.start_lsn;
+
+        loop {
+            tokio::select! {
+                batch = self.rx.recv() => {
                     let batch = match batch {
                         Some(b) => b,
-                        None => { break; }
+                        None => {
+                            return Result::Err(
+                                CopyStreamHandlerEnd::Other(anyhow!("Interpreted WAL reader exited early"))
+                            );
+                        }
                     };
 
+                    wal_position = batch.wal_end_lsn;
+
                     let buf = batch
                         .records
                         .to_wire(self.format, self.compression)
@@ -132,7 +516,21 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> InterpretedWalSender<'_, IO> {
                         })).await?;
                 }
                 // Send a periodic keep alive when the connection has been idle for a while.
+                // Since we've been idle, also check if we can stop streaming.
                 _ = keepalive_ticker.tick() => {
+                    if let Some(remote_consistent_lsn) = self.wal_sender_guard
+                        .walsenders()
+                        .get_ws_remote_consistent_lsn(self.wal_sender_guard.id())
+                    {
+                        if self.tli.should_walsender_stop(remote_consistent_lsn).await {
+                            // Stop streaming if the receivers are caught up and
+                            // there's no active compute. This causes the loop in
+                            // [`crate::send_interpreted_wal::InterpretedWalSender::run`]
+                            // to exit and terminate the WAL stream.
+                            break;
+                        }
+                    }
+
                     self.pgb
                         .write_message(&BeMessage::KeepAlive(WalSndKeepAlive {
                             wal_end: self.end_watch_view.get().0,
@@ -140,14 +538,259 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> InterpretedWalSender<'_, IO> {
                             request_reply: true,
                         }))
                         .await?;
-                }
+                },
             }
         }
 
-        // The loop above ends when the receiver is caught up and there's no more WAL to send.
         Err(CopyStreamHandlerEnd::ServerInitiated(format!(
             "ending streaming to {:?} at {}, receiver is caughtup and there is no computes",
             self.appname, wal_position,
         )))
     }
 }
+#[cfg(test)]
+mod tests {
+    use std::{collections::HashMap, str::FromStr, time::Duration};
+
+    use pageserver_api::shard::{ShardIdentity, ShardStripeSize};
+    use postgres_ffi::MAX_SEND_SIZE;
+    use tokio::sync::mpsc::error::TryRecvError;
+    use utils::{
+        id::{NodeId, TenantTimelineId},
+        lsn::Lsn,
+        shard::{ShardCount, ShardNumber},
+    };
+
+    use crate::{
+        send_interpreted_wal::{Batch, InterpretedWalReader},
+        test_utils::Env,
+        wal_reader_stream::StreamingWalReader,
+    };
+
+    #[tokio::test]
+    async fn test_interpreted_wal_reader_fanout() {
+        let _ = env_logger::builder().is_test(true).try_init();
+
+        const SIZE: usize = 8 * 1024;
+        const MSG_COUNT: usize = 200;
+        const PG_VERSION: u32 = 17;
+        const SHARD_COUNT: u8 = 2;
+
+        let start_lsn = Lsn::from_str("0/149FD18").unwrap();
+        let env = Env::new(true).unwrap();
+        let tli = env
+            .make_timeline(NodeId(1), TenantTimelineId::generate(), start_lsn)
+            .await
+            .unwrap();
+
+        let resident_tli = tli.wal_residence_guard().await.unwrap();
+        let end_watch = Env::write_wal(tli, start_lsn, SIZE, MSG_COUNT)
+            .await
+            .unwrap();
+        let end_pos = end_watch.get();
+
+        tracing::info!("Doing first round of reads ...");
+
+        let streaming_wal_reader = StreamingWalReader::new(
+            resident_tli,
+            None,
+            start_lsn,
+            end_pos,
+            end_watch,
+            MAX_SEND_SIZE,
+        );
+
+        let shard_0 = ShardIdentity::new(
+            ShardNumber(0),
+            ShardCount(SHARD_COUNT),
+            ShardStripeSize::default(),
+        )
+        .unwrap();
+
+        let shard_1 = ShardIdentity::new(
+            ShardNumber(1),
+            ShardCount(SHARD_COUNT),
+            ShardStripeSize::default(),
+        )
+        .unwrap();
+
+        let mut shards = HashMap::new();
+
+        for shard_number in 0..SHARD_COUNT {
+            let shard_id = ShardIdentity::new(
+                ShardNumber(shard_number),
+                ShardCount(SHARD_COUNT),
+                ShardStripeSize::default(),
+            )
+            .unwrap();
+            let (tx, rx) = tokio::sync::mpsc::channel::<Batch>(MSG_COUNT * 2);
+            shards.insert(shard_id, (Some(tx), Some(rx)));
+        }
+
+        let shard_0_tx = shards.get_mut(&shard_0).unwrap().0.take().unwrap();
+        let mut shard_0_rx = shards.get_mut(&shard_0).unwrap().1.take().unwrap();
+
+        let handle = InterpretedWalReader::spawn(
+            streaming_wal_reader,
+            start_lsn,
+            shard_0_tx,
+            shard_0,
+            PG_VERSION,
+            &Some("pageserver".to_string()),
+        );
+
+        tracing::info!("Reading all WAL with only shard 0 attached ...");
+
+        let mut shard_0_interpreted_records = Vec::new();
+        while let Some(batch) = shard_0_rx.recv().await {
+            shard_0_interpreted_records.push(batch.records);
+            if batch.wal_end_lsn == batch.available_wal_end_lsn {
+                break;
+            }
+        }
+
+        let shard_1_tx = shards.get_mut(&shard_1).unwrap().0.take().unwrap();
+        let mut shard_1_rx = shards.get_mut(&shard_1).unwrap().1.take().unwrap();
+
+        tracing::info!("Attaching shard 1 to the reader at start of WAL");
+        handle.fanout(shard_1, shard_1_tx, start_lsn).unwrap();
+
+        tracing::info!("Reading all WAL with shard 0 and shard 1 attached ...");
+
+        let mut shard_1_interpreted_records = Vec::new();
+        while let Some(batch) = shard_1_rx.recv().await {
+            shard_1_interpreted_records.push(batch.records);
+            if batch.wal_end_lsn == batch.available_wal_end_lsn {
+                break;
+            }
+        }
+
+        // This test uses logical messages. Those only go to shard 0. Check that the
+        // filtering worked and shard 1 did not get any.
+        assert!(shard_1_interpreted_records
+            .iter()
+            .all(|recs| recs.records.is_empty()));
+
+        // Shard 0 should not receive anything more since the reader is
+        // going through wal that it has already processed.
+        let res = shard_0_rx.try_recv();
+        if let Ok(ref ok) = res {
+            tracing::error!(
+                "Shard 0 received batch: wal_end_lsn={} available_wal_end_lsn={}",
+                ok.wal_end_lsn,
+                ok.available_wal_end_lsn
+            );
+        }
+        assert!(matches!(res, Err(TryRecvError::Empty)));
+
+        // Check that the next records lsns received by the two shards match up.
+        let shard_0_next_lsns = shard_0_interpreted_records
+            .iter()
+            .map(|recs| recs.next_record_lsn)
+            .collect::<Vec<_>>();
+        let shard_1_next_lsns = shard_1_interpreted_records
+            .iter()
+            .map(|recs| recs.next_record_lsn)
+            .collect::<Vec<_>>();
+        assert_eq!(shard_0_next_lsns, shard_1_next_lsns);
+
+        handle.abort();
+        let mut done = false;
+        for _ in 0..5 {
+            if handle.current_position().is_none() {
+                done = true;
+                break;
+            }
+            tokio::time::sleep(Duration::from_millis(1)).await;
+        }
+
+        assert!(done);
+    }
+
+    #[tokio::test]
+    async fn test_interpreted_wal_reader_same_shard_fanout() {
+        let _ = env_logger::builder().is_test(true).try_init();
+
+        const SIZE: usize = 8 * 1024;
+        const MSG_COUNT: usize = 200;
+        const PG_VERSION: u32 = 17;
+        const SHARD_COUNT: u8 = 2;
+        const ATTACHED_SHARDS: u8 = 4;
+
+        let start_lsn = Lsn::from_str("0/149FD18").unwrap();
+        let env = Env::new(true).unwrap();
+        let tli = env
+            .make_timeline(NodeId(1), TenantTimelineId::generate(), start_lsn)
+            .await
+            .unwrap();
+
+        let resident_tli = tli.wal_residence_guard().await.unwrap();
+        let end_watch = Env::write_wal(tli, start_lsn, SIZE, MSG_COUNT)
+            .await
+            .unwrap();
+        let end_pos = end_watch.get();
+
+        let streaming_wal_reader = StreamingWalReader::new(
+            resident_tli,
+            None,
+            start_lsn,
+            end_pos,
+            end_watch,
+            MAX_SEND_SIZE,
+        );
+
+        let shard_0 = ShardIdentity::new(
+            ShardNumber(0),
+            ShardCount(SHARD_COUNT),
+            ShardStripeSize::default(),
+        )
+        .unwrap();
+
+        let (tx, rx) = tokio::sync::mpsc::channel::<Batch>(MSG_COUNT * 2);
+        let mut batch_receivers = vec![rx];
+
+        let handle = InterpretedWalReader::spawn(
+            streaming_wal_reader,
+            start_lsn,
+            tx,
+            shard_0,
+            PG_VERSION,
+            &Some("pageserver".to_string()),
+        );
+
+        for _ in 0..(ATTACHED_SHARDS - 1) {
+            let (tx, rx) = tokio::sync::mpsc::channel::<Batch>(MSG_COUNT * 2);
+            handle.fanout(shard_0, tx, start_lsn).unwrap();
+            batch_receivers.push(rx);
+        }
+
+        loop {
+            let batch = batch_receivers.first_mut().unwrap().recv().await.unwrap();
+            for rx in batch_receivers.iter_mut().skip(1) {
+                let other_batch = rx.recv().await.unwrap();
+
+                assert_eq!(batch.wal_end_lsn, other_batch.wal_end_lsn);
+                assert_eq!(
+                    batch.available_wal_end_lsn,
+                    other_batch.available_wal_end_lsn
+                );
+            }
+
+            if batch.wal_end_lsn == batch.available_wal_end_lsn {
+                break;
+            }
+        }
+
+        handle.abort();
+        let mut done = false;
+        for _ in 0..5 {
+            if handle.current_position().is_none() {
+                done = true;
+                break;
+            }
+            tokio::time::sleep(Duration::from_millis(1)).await;
+        }
+
+        assert!(done);
+    }
+}

safekeeper/src/send_wal.rs

@@ -2,16 +2,18 @@
 //! with the "START_REPLICATION" message, and registry of walsenders.
 
 use crate::handler::SafekeeperPostgresHandler;
-use crate::metrics::RECEIVED_PS_FEEDBACKS;
+use crate::metrics::{RECEIVED_PS_FEEDBACKS, WAL_READERS};
 use crate::receive_wal::WalReceivers;
 use crate::safekeeper::TermLsn;
-use crate::send_interpreted_wal::InterpretedWalSender;
+use crate::send_interpreted_wal::{
+    Batch, InterpretedWalReader, InterpretedWalReaderHandle, InterpretedWalSender,
+};
 use crate::timeline::WalResidentTimeline;
-use crate::wal_reader_stream::WalReaderStreamBuilder;
+use crate::wal_reader_stream::StreamingWalReader;
 use crate::wal_storage::WalReader;
 use anyhow::{bail, Context as AnyhowContext};
 use bytes::Bytes;
-use futures::future::Either;
+use futures::FutureExt;
 use parking_lot::Mutex;
 use postgres_backend::PostgresBackend;
 use postgres_backend::{CopyStreamHandlerEnd, PostgresBackendReader, QueryError};
@@ -19,16 +21,16 @@ use postgres_ffi::get_current_timestamp;
 use postgres_ffi::{TimestampTz, MAX_SEND_SIZE};
 use pq_proto::{BeMessage, WalSndKeepAlive, XLogDataBody};
 use safekeeper_api::models::{
-    ConnectionId, HotStandbyFeedback, ReplicationFeedback, StandbyFeedback, StandbyReply,
-    WalSenderState, INVALID_FULL_TRANSACTION_ID,
+    HotStandbyFeedback, ReplicationFeedback, StandbyFeedback, StandbyReply,
+    INVALID_FULL_TRANSACTION_ID,
 };
 use safekeeper_api::Term;
 use tokio::io::{AsyncRead, AsyncWrite};
 use utils::failpoint_support;
-use utils::id::TenantTimelineId;
 use utils::pageserver_feedback::PageserverFeedback;
 use utils::postgres_client::PostgresClientProtocol;
 
+use itertools::Itertools;
 use std::cmp::{max, min};
 use std::net::SocketAddr;
 use std::sync::Arc;
@@ -50,6 +52,12 @@ pub struct WalSenders {
     walreceivers: Arc<WalReceivers>,
 }
 
+pub struct WalSendersTimelineMetricValues {
+    pub ps_feedback_counter: u64,
+    pub last_ps_feedback: PageserverFeedback,
+    pub interpreted_wal_reader_tasks: usize,
+}
+
 impl WalSenders {
     pub fn new(walreceivers: Arc<WalReceivers>) -> Arc<WalSenders> {
         Arc::new(WalSenders {
@@ -60,21 +68,8 @@ impl WalSenders {
 
     /// Register new walsender. Returned guard provides access to the slot and
     /// automatically deregisters in Drop.
-    fn register(
-        self: &Arc<WalSenders>,
-        ttid: TenantTimelineId,
-        addr: SocketAddr,
-        conn_id: ConnectionId,
-        appname: Option<String>,
-    ) -> WalSenderGuard {
+    fn register(self: &Arc<WalSenders>, walsender_state: WalSenderState) -> WalSenderGuard {
         let slots = &mut self.mutex.lock().slots;
-        let walsender_state = WalSenderState {
-            ttid,
-            addr,
-            conn_id,
-            appname,
-            feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()),
-        };
         // find empty slot or create new one
         let pos = if let Some(pos) = slots.iter().position(|s| s.is_none()) {
             slots[pos] = Some(walsender_state);
@@ -90,9 +85,79 @@ impl WalSenders {
         }
     }
 
+    fn create_or_update_interpreted_reader<
+        FUp: FnOnce(&Arc<InterpretedWalReaderHandle>) -> anyhow::Result<()>,
+        FNew: FnOnce() -> InterpretedWalReaderHandle,
+    >(
+        self: &Arc<WalSenders>,
+        id: WalSenderId,
+        start_pos: Lsn,
+        max_delta_for_fanout: Option<u64>,
+        update: FUp,
+        create: FNew,
+    ) -> anyhow::Result<()> {
+        let state = &mut self.mutex.lock();
+
+        let mut selected_interpreted_reader = None;
+        for slot in state.slots.iter().flatten() {
+            if let WalSenderState::Interpreted(slot_state) = slot {
+                if let Some(ref interpreted_reader) = slot_state.interpreted_wal_reader {
+                    let select = match (interpreted_reader.current_position(), max_delta_for_fanout)
+                    {
+                        (Some(pos), Some(max_delta)) => {
+                            let delta = pos.0.abs_diff(start_pos.0);
+                            delta <= max_delta
+                        }
+                        // Reader is not active
+                        (None, _) => false,
+                        // Gating fanout by max delta is disabled.
+                        // Attach to any active reader.
+                        (_, None) => true,
+                    };
+
+                    if select {
+                        selected_interpreted_reader = Some(interpreted_reader.clone());
+                        break;
+                    }
+                }
+            }
+        }
+
+        let slot = state.get_slot_mut(id);
+        let slot_state = match slot {
+            WalSenderState::Interpreted(s) => s,
+            WalSenderState::Vanilla(_) => unreachable!(),
+        };
+
+        let selected_or_new = match selected_interpreted_reader {
+            Some(selected) => {
+                update(&selected)?;
+                selected
+            }
+            None => Arc::new(create()),
+        };
+
+        slot_state.interpreted_wal_reader = Some(selected_or_new);
+
+        Ok(())
+    }
+
     /// Get state of all walsenders.
-    pub fn get_all(self: &Arc<WalSenders>) -> Vec<WalSenderState> {
-        self.mutex.lock().slots.iter().flatten().cloned().collect()
+    pub fn get_all_public(self: &Arc<WalSenders>) -> Vec<safekeeper_api::models::WalSenderState> {
+        self.mutex
+            .lock()
+            .slots
+            .iter()
+            .flatten()
+            .map(|state| match state {
+                WalSenderState::Vanilla(s) => {
+                    safekeeper_api::models::WalSenderState::Vanilla(s.clone())
+                }
+                WalSenderState::Interpreted(s) => {
+                    safekeeper_api::models::WalSenderState::Interpreted(s.public_state.clone())
+                }
+            })
+            .collect()
     }
 
     /// Get LSN of the most lagging pageserver receiver. Return None if there are no
@@ -103,7 +168,7 @@ impl WalSenders {
             .slots
             .iter()
             .flatten()
-            .filter_map(|s| match s.feedback {
+            .filter_map(|s| match s.get_feedback() {
                 ReplicationFeedback::Pageserver(feedback) => Some(feedback.last_received_lsn),
                 ReplicationFeedback::Standby(_) => None,
             })
@@ -111,9 +176,25 @@ impl WalSenders {
     }
 
     /// Returns total counter of pageserver feedbacks received and last feedback.
-    pub fn get_ps_feedback_stats(self: &Arc<WalSenders>) -> (u64, PageserverFeedback) {
+    pub fn info_for_metrics(self: &Arc<WalSenders>) -> WalSendersTimelineMetricValues {
         let shared = self.mutex.lock();
-        (shared.ps_feedback_counter, shared.last_ps_feedback)
+
+        let interpreted_wal_reader_tasks = shared
+            .slots
+            .iter()
+            .filter_map(|ss| match ss {
+                Some(WalSenderState::Interpreted(int)) => int.interpreted_wal_reader.as_ref(),
+                Some(WalSenderState::Vanilla(_)) => None,
+                None => None,
+            })
+            .unique_by(|reader| Arc::as_ptr(reader))
+            .count();
+
+        WalSendersTimelineMetricValues {
+            ps_feedback_counter: shared.ps_feedback_counter,
+            last_ps_feedback: shared.last_ps_feedback,
+            interpreted_wal_reader_tasks,
+        }
     }
 
     /// Get aggregated hot standby feedback (we send it to compute).
@@ -124,7 +205,7 @@ impl WalSenders {
     /// Record new pageserver feedback, update aggregated values.
     fn record_ps_feedback(self: &Arc<WalSenders>, id: WalSenderId, feedback: &PageserverFeedback) {
         let mut shared = self.mutex.lock();
-        shared.get_slot_mut(id).feedback = ReplicationFeedback::Pageserver(*feedback);
+        *shared.get_slot_mut(id).get_mut_feedback() = ReplicationFeedback::Pageserver(*feedback);
         shared.last_ps_feedback = *feedback;
         shared.ps_feedback_counter += 1;
         drop(shared);
@@ -143,10 +224,10 @@ impl WalSenders {
             "Record standby reply: ts={} apply_lsn={}",
             reply.reply_ts, reply.apply_lsn
         );
-        match &mut slot.feedback {
+        match &mut slot.get_mut_feedback() {
             ReplicationFeedback::Standby(sf) => sf.reply = *reply,
             ReplicationFeedback::Pageserver(_) => {
-                slot.feedback = ReplicationFeedback::Standby(StandbyFeedback {
+                *slot.get_mut_feedback() = ReplicationFeedback::Standby(StandbyFeedback {
                     reply: *reply,
                     hs_feedback: HotStandbyFeedback::empty(),
                 })
@@ -158,10 +239,10 @@ impl WalSenders {
     fn record_hs_feedback(self: &Arc<WalSenders>, id: WalSenderId, feedback: &HotStandbyFeedback) {
         let mut shared = self.mutex.lock();
         let slot = shared.get_slot_mut(id);
-        match &mut slot.feedback {
+        match &mut slot.get_mut_feedback() {
             ReplicationFeedback::Standby(sf) => sf.hs_feedback = *feedback,
             ReplicationFeedback::Pageserver(_) => {
-                slot.feedback = ReplicationFeedback::Standby(StandbyFeedback {
+                *slot.get_mut_feedback() = ReplicationFeedback::Standby(StandbyFeedback {
                     reply: StandbyReply::empty(),
                     hs_feedback: *feedback,
                 })
@@ -175,7 +256,7 @@ impl WalSenders {
     pub fn get_ws_remote_consistent_lsn(self: &Arc<WalSenders>, id: WalSenderId) -> Option<Lsn> {
         let shared = self.mutex.lock();
         let slot = shared.get_slot(id);
-        match slot.feedback {
+        match slot.get_feedback() {
             ReplicationFeedback::Pageserver(feedback) => Some(feedback.remote_consistent_lsn),
             _ => None,
         }
@@ -199,6 +280,47 @@ struct WalSendersShared {
     slots: Vec<Option<WalSenderState>>,
 }
 
+/// Safekeeper internal definitions of wal sender state
+///
+/// As opposed to [`safekeeper_api::models::WalSenderState`] these struct may
+/// include state that we don not wish to expose to the public api.
+#[derive(Debug, Clone)]
+pub(crate) enum WalSenderState {
+    Vanilla(VanillaWalSenderInternalState),
+    Interpreted(InterpretedWalSenderInternalState),
+}
+
+type VanillaWalSenderInternalState = safekeeper_api::models::VanillaWalSenderState;
+
+#[derive(Debug, Clone)]
+pub(crate) struct InterpretedWalSenderInternalState {
+    public_state: safekeeper_api::models::InterpretedWalSenderState,
+    interpreted_wal_reader: Option<Arc<InterpretedWalReaderHandle>>,
+}
+
+impl WalSenderState {
+    fn get_addr(&self) -> &SocketAddr {
+        match self {
+            WalSenderState::Vanilla(state) => &state.addr,
+            WalSenderState::Interpreted(state) => &state.public_state.addr,
+        }
+    }
+
+    fn get_feedback(&self) -> &ReplicationFeedback {
+        match self {
+            WalSenderState::Vanilla(state) => &state.feedback,
+            WalSenderState::Interpreted(state) => &state.public_state.feedback,
+        }
+    }
+
+    fn get_mut_feedback(&mut self) -> &mut ReplicationFeedback {
+        match self {
+            WalSenderState::Vanilla(state) => &mut state.feedback,
+            WalSenderState::Interpreted(state) => &mut state.public_state.feedback,
+        }
+    }
+}
+
 impl WalSendersShared {
     fn new() -> Self {
         WalSendersShared {
@@ -225,7 +347,7 @@ impl WalSendersShared {
         let mut agg = HotStandbyFeedback::empty();
         let mut reply_agg = StandbyReply::empty();
         for ws_state in self.slots.iter().flatten() {
-            if let ReplicationFeedback::Standby(standby_feedback) = ws_state.feedback {
+            if let ReplicationFeedback::Standby(standby_feedback) = ws_state.get_feedback() {
                 let hs_feedback = standby_feedback.hs_feedback;
                 // doing Option math like op1.iter().chain(op2.iter()).min()
                 // would be nicer, but we serialize/deserialize this struct
@@ -317,7 +439,7 @@ impl SafekeeperPostgresHandler {
     /// Wrapper around handle_start_replication_guts handling result. Error is
     /// handled here while we're still in walsender ttid span; with API
     /// extension, this can probably be moved into postgres_backend.
-    pub async fn handle_start_replication<IO: AsyncRead + AsyncWrite + Unpin>(
+    pub async fn handle_start_replication<IO: AsyncRead + AsyncWrite + Unpin + Send>(
         &mut self,
         pgb: &mut PostgresBackend<IO>,
         start_pos: Lsn,
@@ -342,7 +464,7 @@ impl SafekeeperPostgresHandler {
         Ok(())
     }
 
-    pub async fn handle_start_replication_guts<IO: AsyncRead + AsyncWrite + Unpin>(
+    pub async fn handle_start_replication_guts<IO: AsyncRead + AsyncWrite + Unpin + Send>(
         &mut self,
         pgb: &mut PostgresBackend<IO>,
         start_pos: Lsn,
@@ -352,12 +474,30 @@ impl SafekeeperPostgresHandler {
         let appname = self.appname.clone();
 
         // Use a guard object to remove our entry from the timeline when we are done.
-        let ws_guard = Arc::new(tli.get_walsenders().register(
-            self.ttid,
-            *pgb.get_peer_addr(),
-            self.conn_id,
-            self.appname.clone(),
-        ));
+        let ws_guard = match self.protocol() {
+            PostgresClientProtocol::Vanilla => Arc::new(tli.get_walsenders().register(
+                WalSenderState::Vanilla(VanillaWalSenderInternalState {
+                    ttid: self.ttid,
+                    addr: *pgb.get_peer_addr(),
+                    conn_id: self.conn_id,
+                    appname: self.appname.clone(),
+                    feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()),
+                }),
+            )),
+            PostgresClientProtocol::Interpreted { .. } => Arc::new(tli.get_walsenders().register(
+                WalSenderState::Interpreted(InterpretedWalSenderInternalState {
+                    public_state: safekeeper_api::models::InterpretedWalSenderState {
+                        ttid: self.ttid,
+                        shard: self.shard.unwrap(),
+                        addr: *pgb.get_peer_addr(),
+                        conn_id: self.conn_id,
+                        appname: self.appname.clone(),
+                        feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()),
+                    },
+                    interpreted_wal_reader: None,
+                }),
+            )),
+        };
 
         // Walsender can operate in one of two modes which we select by
         // application_name: give only committed WAL (used by pageserver) or all
@@ -403,7 +543,7 @@ impl SafekeeperPostgresHandler {
                     pgb,
                     // should succeed since we're already holding another guard
                     tli: tli.wal_residence_guard().await?,
-                    appname,
+                    appname: appname.clone(),
                     start_pos,
                     end_pos,
                     term,
@@ -413,7 +553,7 @@ impl SafekeeperPostgresHandler {
                     send_buf: vec![0u8; MAX_SEND_SIZE],
                 };
 
-                Either::Left(sender.run())
+                FutureExt::boxed(sender.run())
             }
             PostgresClientProtocol::Interpreted {
                 format,
@@ -421,27 +561,96 @@ impl SafekeeperPostgresHandler {
             } => {
                 let pg_version = tli.tli.get_state().await.1.server.pg_version / 10000;
                 let end_watch_view = end_watch.view();
-                let wal_stream_builder = WalReaderStreamBuilder {
-                    tli: tli.wal_residence_guard().await?,
-                    start_pos,
-                    end_pos,
-                    term,
-                    end_watch,
-                    wal_sender_guard: ws_guard.clone(),
-                };
+                let wal_residence_guard = tli.wal_residence_guard().await?;
+                let (tx, rx) = tokio::sync::mpsc::channel::<Batch>(2);
+                let shard = self.shard.unwrap();
 
-                let sender = InterpretedWalSender {
-                    format,
-                    compression,
-                    pgb,
-                    wal_stream_builder,
-                    end_watch_view,
-                    shard: self.shard.unwrap(),
-                    pg_version,
-                    appname,
-                };
+                if self.conf.wal_reader_fanout && !shard.is_unsharded() {
+                    let ws_id = ws_guard.id();
+                    ws_guard.walsenders().create_or_update_interpreted_reader(
+                        ws_id,
+                        start_pos,
+                        self.conf.max_delta_for_fanout,
+                        {
+                            let tx = tx.clone();
+                            |reader| {
+                                tracing::info!(
+                                    "Fanning out interpreted wal reader at {}",
+                                    start_pos
+                                );
+                                reader
+                                    .fanout(shard, tx, start_pos)
+                                    .with_context(|| "Failed to fan out reader")
+                            }
+                        },
+                        || {
+                            tracing::info!("Spawning interpreted wal reader at {}", start_pos);
 
-                Either::Right(sender.run())
+                            let wal_stream = StreamingWalReader::new(
+                                wal_residence_guard,
+                                term,
+                                start_pos,
+                                end_pos,
+                                end_watch,
+                                MAX_SEND_SIZE,
+                            );
+
+                            InterpretedWalReader::spawn(
+                                wal_stream, start_pos, tx, shard, pg_version, &appname,
+                            )
+                        },
+                    )?;
+
+                    let sender = InterpretedWalSender {
+                        format,
+                        compression,
+                        appname,
+                        tli: tli.wal_residence_guard().await?,
+                        start_lsn: start_pos,
+                        pgb,
+                        end_watch_view,
+                        wal_sender_guard: ws_guard.clone(),
+                        rx,
+                    };
+
+                    FutureExt::boxed(sender.run())
+                } else {
+                    let wal_reader = StreamingWalReader::new(
+                        wal_residence_guard,
+                        term,
+                        start_pos,
+                        end_pos,
+                        end_watch,
+                        MAX_SEND_SIZE,
+                    );
+
+                    let reader =
+                        InterpretedWalReader::new(wal_reader, start_pos, tx, shard, pg_version);
+
+                    let sender = InterpretedWalSender {
+                        format,
+                        compression,
+                        appname: appname.clone(),
+                        tli: tli.wal_residence_guard().await?,
+                        start_lsn: start_pos,
+                        pgb,
+                        end_watch_view,
+                        wal_sender_guard: ws_guard.clone(),
+                        rx,
+                    };
+
+                    FutureExt::boxed(async move {
+                        // Sender returns an Err on all code paths.
+                        // If the sender finishes first, we will drop the reader future.
+                        // If the reader finishes first, the sender will finish too since
+                        // the wal sender has dropped.
+                        let res = tokio::try_join!(sender.run(), reader.run(start_pos, &appname));
+                        match res.map(|_| ()) {
+                            Ok(_) => unreachable!("sender finishes with Err by convention"),
+                            err_res => err_res,
+                        }
+                    })
+                }
             }
         };
 
@@ -470,7 +679,8 @@ impl SafekeeperPostgresHandler {
             .clone();
         info!(
             "finished streaming to {}, feedback={:?}",
-            ws_state.addr, ws_state.feedback,
+            ws_state.get_addr(),
+            ws_state.get_feedback(),
         );
 
         // Join pg backend back.
@@ -578,6 +788,18 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
     /// Err(CopyStreamHandlerEnd) is always returned; Result is used only for ?
     /// convenience.
     async fn run(mut self) -> Result<(), CopyStreamHandlerEnd> {
+        let metric = WAL_READERS
+            .get_metric_with_label_values(&[
+                "future",
+                self.appname.as_deref().unwrap_or("safekeeper"),
+            ])
+            .unwrap();
+
+        metric.inc();
+        scopeguard::defer! {
+            metric.dec();
+        }
+
         loop {
             // Wait for the next portion if it is not there yet, or just
             // update our end of WAL available for sending value, we
@@ -813,7 +1035,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> ReplyReader<IO> {
 #[cfg(test)]
 mod tests {
     use safekeeper_api::models::FullTransactionId;
-    use utils::id::{TenantId, TimelineId};
+    use utils::id::{TenantId, TenantTimelineId, TimelineId};
 
     use super::*;
 
@@ -830,13 +1052,13 @@ mod tests {
 
     // add to wss specified feedback setting other fields to dummy values
     fn push_feedback(wss: &mut WalSendersShared, feedback: ReplicationFeedback) {
-        let walsender_state = WalSenderState {
+        let walsender_state = WalSenderState::Vanilla(VanillaWalSenderInternalState {
             ttid: mock_ttid(),
             addr: mock_addr(),
             conn_id: 1,
             appname: None,
             feedback,
-        };
+        });
         wss.slots.push(Some(walsender_state))
     }
 

safekeeper/src/state.rs

@@ -1,20 +1,25 @@
 //! Defines per timeline data stored persistently (SafeKeeperPersistentState)
 //! and its wrapper with in memory layer (SafekeeperState).
 
-use std::{cmp::max, ops::Deref};
+use std::{cmp::max, ops::Deref, time::SystemTime};
 
 use anyhow::{bail, Result};
 use postgres_ffi::WAL_SEGMENT_SIZE;
-use safekeeper_api::{models::TimelineTermBumpResponse, ServerInfo, Term};
+use safekeeper_api::{
+    membership::Configuration,
+    models::{TimelineMembershipSwitchResponse, TimelineTermBumpResponse},
+    ServerInfo, Term, INITIAL_TERM,
+};
 use serde::{Deserialize, Serialize};
+use tracing::info;
 use utils::{
-    id::{NodeId, TenantId, TenantTimelineId, TimelineId},
+    id::{TenantId, TenantTimelineId, TimelineId},
     lsn::Lsn,
 };
 
 use crate::{
     control_file,
-    safekeeper::{AcceptorState, PersistedPeerInfo, PgUuid, TermHistory, UNKNOWN_SERVER_VERSION},
+    safekeeper::{AcceptorState, PgUuid, TermHistory, TermLsn, UNKNOWN_SERVER_VERSION},
     timeline::TimelineError,
     wal_backup_partial::{self},
 };
@@ -27,6 +32,8 @@ pub struct TimelinePersistentState {
     pub tenant_id: TenantId,
     #[serde(with = "hex")]
     pub timeline_id: TimelineId,
+    /// Membership configuration.
+    pub mconf: Configuration,
     /// persistent acceptor state
     pub acceptor_state: AcceptorState,
     /// information about server
@@ -58,22 +65,15 @@ pub struct TimelinePersistentState {
     /// pushed to s3. We don't remove WAL beyond it. Persisted only for
     /// informational purposes, we receive it from pageserver (or broker).
     pub remote_consistent_lsn: Lsn,
-    /// Peers and their state as we remember it. Knowing peers themselves is
-    /// fundamental; but state is saved here only for informational purposes and
-    /// obviously can be stale. (Currently not saved at all, but let's provision
-    /// place to have less file version upgrades).
-    pub peers: PersistedPeers,
     /// Holds names of partial segments uploaded to remote storage. Used to
     /// clean up old objects without leaving garbage in remote storage.
     pub partial_backup: wal_backup_partial::State,
     /// Eviction state of the timeline. If it's Offloaded, we should download
     /// WAL files from remote storage to serve the timeline.
     pub eviction_state: EvictionState,
+    pub creation_ts: SystemTime,
 }
 
-#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
-pub struct PersistedPeers(pub Vec<(NodeId, PersistedPeerInfo)>);
-
 /// State of the local WAL files. Used to track current timeline state,
 /// that can be either WAL files are present on disk or last partial segment
 /// is offloaded to remote storage.
@@ -87,12 +87,14 @@ pub enum EvictionState {
 }
 
 impl TimelinePersistentState {
+    /// commit_lsn is the same as start_lsn in the normal creaiton; see
+    /// `TimelineCreateRequest` comments.`
     pub fn new(
         ttid: &TenantTimelineId,
+        mconf: Configuration,
         server_info: ServerInfo,
-        peers: Vec<NodeId>,
+        start_lsn: Lsn,
         commit_lsn: Lsn,
-        local_start_lsn: Lsn,
     ) -> anyhow::Result<TimelinePersistentState> {
         if server_info.wal_seg_size == 0 {
             bail!(TimelineError::UninitializedWalSegSize(*ttid));
@@ -102,49 +104,59 @@ impl TimelinePersistentState {
             bail!(TimelineError::UninitialinzedPgVersion(*ttid));
         }
 
-        if commit_lsn < local_start_lsn {
+        if commit_lsn < start_lsn {
             bail!(
-                "commit_lsn {} is smaller than local_start_lsn {}",
+                "commit_lsn {} is smaller than start_lsn {}",
                 commit_lsn,
-                local_start_lsn
+                start_lsn
             );
         }
 
+        // If we are given with init LSN, initialize term history with it. It
+        // ensures that walproposer always must be able to find a common point
+        // in histories; if it can't something is corrupted. Not having LSN here
+        // is so far left for legacy case where timeline is created by compute
+        // and LSN during creation is not known yet.
+        let term_history = if commit_lsn != Lsn::INVALID {
+            TermHistory(vec![TermLsn {
+                term: INITIAL_TERM,
+                lsn: start_lsn,
+            }])
+        } else {
+            TermHistory::empty()
+        };
+
         Ok(TimelinePersistentState {
             tenant_id: ttid.tenant_id,
             timeline_id: ttid.timeline_id,
+            mconf,
             acceptor_state: AcceptorState {
-                term: 0,
-                term_history: TermHistory::empty(),
+                term: INITIAL_TERM,
+                term_history,
             },
             server: server_info,
             proposer_uuid: [0; 16],
-            timeline_start_lsn: Lsn(0),
-            local_start_lsn,
+            timeline_start_lsn: start_lsn,
+            local_start_lsn: start_lsn,
             commit_lsn,
-            backup_lsn: local_start_lsn,
-            peer_horizon_lsn: local_start_lsn,
+            backup_lsn: start_lsn,
+            peer_horizon_lsn: start_lsn,
             remote_consistent_lsn: Lsn(0),
-            peers: PersistedPeers(
-                peers
-                    .iter()
-                    .map(|p| (*p, PersistedPeerInfo::new()))
-                    .collect(),
-            ),
             partial_backup: wal_backup_partial::State::default(),
             eviction_state: EvictionState::Present,
+            creation_ts: SystemTime::now(),
         })
     }
 
     pub fn empty() -> Self {
         TimelinePersistentState::new(
             &TenantTimelineId::empty(),
+            Configuration::empty(),
             ServerInfo {
                 pg_version: 170000, /* Postgres server version (major * 10000) */
                 system_id: 0,       /* Postgres system identifier */
                 wal_seg_size: WAL_SEGMENT_SIZE as u32,
             },
-            vec![],
             Lsn::INVALID,
             Lsn::INVALID,
         )
@@ -249,6 +261,31 @@ where
             current_term: after,
         })
     }
+
+    /// Switch into membership configuration `to` if it is higher than the
+    /// current one.
+    pub async fn membership_switch(
+        &mut self,
+        to: Configuration,
+    ) -> Result<TimelineMembershipSwitchResponse> {
+        let before = self.mconf.clone();
+        // Is switch allowed?
+        if to.generation <= self.mconf.generation {
+            info!(
+                "ignoring request to switch membership conf to lower {}, current conf {}",
+                to, self.mconf
+            );
+        } else {
+            let mut state = self.start_change();
+            state.mconf = to.clone();
+            self.finish_change(&state).await?;
+            info!("switched membership conf to {} from {}", to, before);
+        }
+        Ok(TimelineMembershipSwitchResponse {
+            previous_conf: before,
+            current_conf: self.mconf.clone(),
+        })
+    }
 }
 
 impl<CTRL> Deref for TimelineState<CTRL>

safekeeper/src/test_utils.rs

@@ -1,13 +1,19 @@
 use std::sync::Arc;
 
 use crate::rate_limit::RateLimiter;
-use crate::safekeeper::{ProposerAcceptorMessage, ProposerElected, SafeKeeper, TermHistory};
+use crate::receive_wal::WalAcceptor;
+use crate::safekeeper::{
+    AcceptorProposerMessage, AppendRequest, AppendRequestHeader, ProposerAcceptorMessage,
+    ProposerElected, SafeKeeper, TermHistory,
+};
+use crate::send_wal::EndWatch;
 use crate::state::{TimelinePersistentState, TimelineState};
 use crate::timeline::{get_timeline_dir, SharedState, StateSK, Timeline};
 use crate::timelines_set::TimelinesSet;
 use crate::wal_backup::remote_timeline_path;
-use crate::{control_file, wal_storage, SafeKeeperConf};
+use crate::{control_file, receive_wal, wal_storage, SafeKeeperConf};
 use camino_tempfile::Utf8TempDir;
+use postgres_ffi::v17::wal_generator::{LogicalMessageGenerator, WalGenerator};
 use tokio::fs::create_dir_all;
 use utils::id::{NodeId, TenantTimelineId};
 use utils::lsn::Lsn;
@@ -107,4 +113,59 @@ impl Env {
         );
         Ok(timeline)
     }
+
+    // This will be dead code when building a non-benchmark target with the
+    // benchmarking feature enabled.
+    #[allow(dead_code)]
+    pub(crate) async fn write_wal(
+        tli: Arc<Timeline>,
+        start_lsn: Lsn,
+        msg_size: usize,
+        msg_count: usize,
+    ) -> anyhow::Result<EndWatch> {
+        let (msg_tx, msg_rx) = tokio::sync::mpsc::channel(receive_wal::MSG_QUEUE_SIZE);
+        let (reply_tx, mut reply_rx) = tokio::sync::mpsc::channel(receive_wal::REPLY_QUEUE_SIZE);
+
+        let end_watch = EndWatch::Commit(tli.get_commit_lsn_watch_rx());
+
+        WalAcceptor::spawn(tli.wal_residence_guard().await?, msg_rx, reply_tx, Some(0));
+
+        let prefix = c"p";
+        let prefixlen = prefix.to_bytes_with_nul().len();
+        assert!(msg_size >= prefixlen);
+        let message = vec![0; msg_size - prefixlen];
+
+        let walgen =
+            &mut WalGenerator::new(LogicalMessageGenerator::new(prefix, &message), start_lsn);
+        for _ in 0..msg_count {
+            let (lsn, record) = walgen.next().unwrap();
+
+            let req = AppendRequest {
+                h: AppendRequestHeader {
+                    term: 1,
+                    term_start_lsn: start_lsn,
+                    begin_lsn: lsn,
+                    end_lsn: lsn + record.len() as u64,
+                    commit_lsn: lsn,
+                    truncate_lsn: Lsn(0),
+                    proposer_uuid: [0; 16],
+                },
+                wal_data: record,
+            };
+
+            let end_lsn = req.h.end_lsn;
+
+            let msg = ProposerAcceptorMessage::AppendRequest(req);
+            msg_tx.send(msg).await?;
+            while let Some(reply) = reply_rx.recv().await {
+                if let AcceptorProposerMessage::AppendResponse(resp) = reply {
+                    if resp.flush_lsn >= end_lsn {
+                        break;
+                    }
+                }
+            }
+        }
+
+        Ok(end_watch)
+    }
 }

safekeeper/src/timeline.rs

@@ -4,7 +4,10 @@
 use anyhow::{anyhow, bail, Result};
 use camino::{Utf8Path, Utf8PathBuf};
 use remote_storage::RemotePath;
-use safekeeper_api::models::{PeerInfo, TimelineTermBumpResponse};
+use safekeeper_api::membership::Configuration;
+use safekeeper_api::models::{
+    PeerInfo, TimelineMembershipSwitchResponse, TimelineTermBumpResponse,
+};
 use safekeeper_api::Term;
 use tokio::fs::{self};
 use tokio_util::sync::CancellationToken;
@@ -32,7 +35,7 @@ use crate::control_file;
 use crate::rate_limit::RateLimiter;
 use crate::receive_wal::WalReceivers;
 use crate::safekeeper::{AcceptorProposerMessage, ProposerAcceptorMessage, SafeKeeper, TermLsn};
-use crate::send_wal::WalSenders;
+use crate::send_wal::{WalSenders, WalSendersTimelineMetricValues};
 use crate::state::{EvictionState, TimelineMemState, TimelinePersistentState, TimelineState};
 use crate::timeline_guard::ResidenceGuard;
 use crate::timeline_manager::{AtomicStatus, ManagerCtl};
@@ -188,6 +191,13 @@ impl StateSK {
         self.state_mut().term_bump(to).await
     }
 
+    pub async fn membership_switch(
+        &mut self,
+        to: Configuration,
+    ) -> Result<TimelineMembershipSwitchResponse> {
+        self.state_mut().membership_switch(to).await
+    }
+
     /// Close open WAL files to release FDs.
     fn close_wal_store(&mut self) {
         if let StateSK::Loaded(sk) = self {
@@ -702,16 +712,22 @@ impl Timeline {
             return None;
         }
 
-        let (ps_feedback_count, last_ps_feedback) = self.walsenders.get_ps_feedback_stats();
+        let WalSendersTimelineMetricValues {
+            ps_feedback_counter,
+            last_ps_feedback,
+            interpreted_wal_reader_tasks,
+        } = self.walsenders.info_for_metrics();
+
         let state = self.read_shared_state().await;
         Some(FullTimelineInfo {
             ttid: self.ttid,
-            ps_feedback_count,
+            ps_feedback_count: ps_feedback_counter,
             last_ps_feedback,
             wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
             timeline_is_active: self.broker_active.load(Ordering::Relaxed),
             num_computes: self.walreceivers.get_num() as u32,
             last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed),
+            interpreted_wal_reader_tasks,
             epoch_start_lsn: state.sk.term_start_lsn(),
             mem_state: state.sk.state().inmem.clone(),
             persisted_state: TimelinePersistentState::clone(state.sk.state()),
@@ -730,7 +746,7 @@ impl Timeline {
         debug_dump::Memory {
             is_cancelled: self.is_cancelled(),
             peers_info_len: state.peers_info.0.len(),
-            walsenders: self.walsenders.get_all(),
+            walsenders: self.walsenders.get_all_public(),
             wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
             active: self.broker_active.load(Ordering::Relaxed),
             num_computes: self.walreceivers.get_num() as u32,
@@ -768,6 +784,14 @@ impl Timeline {
         state.sk.term_bump(to).await
     }
 
+    pub async fn membership_switch(
+        self: &Arc<Self>,
+        to: Configuration,
+    ) -> Result<TimelineMembershipSwitchResponse> {
+        let mut state = self.write_shared_state().await;
+        state.sk.membership_switch(to).await
+    }
+
     /// Guts of [`Self::wal_residence_guard`] and [`Self::try_wal_residence_guard`]
     async fn do_wal_residence_guard(
         self: &Arc<Self>,

safekeeper/src/timelines_global_map.rs

@@ -12,6 +12,7 @@ use crate::{control_file, wal_storage, SafeKeeperConf};
 use anyhow::{bail, Context, Result};
 use camino::Utf8PathBuf;
 use camino_tempfile::Utf8TempDir;
+use safekeeper_api::membership::Configuration;
 use safekeeper_api::ServerInfo;
 use serde::Serialize;
 use std::collections::HashMap;
@@ -214,9 +215,10 @@ impl GlobalTimelines {
     pub(crate) async fn create(
         &self,
         ttid: TenantTimelineId,
+        mconf: Configuration,
         server_info: ServerInfo,
+        start_lsn: Lsn,
         commit_lsn: Lsn,
-        local_start_lsn: Lsn,
     ) -> Result<Arc<Timeline>> {
         let (conf, _, _) = {
             let state = self.state.lock().unwrap();
@@ -239,8 +241,7 @@ impl GlobalTimelines {
 
         // TODO: currently we create only cfile. It would be reasonable to
         // immediately initialize first WAL segment as well.
-        let state =
-            TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn)?;
+        let state = TimelinePersistentState::new(&ttid, mconf, server_info, start_lsn, commit_lsn)?;
         control_file::FileStorage::create_new(&tmp_dir_path, state, conf.no_sync).await?;
         let timeline = self.load_temp_timeline(ttid, &tmp_dir_path, true).await?;
         Ok(timeline)

safekeeper/src/wal_reader_stream.rs

@@ -1,34 +1,16 @@
-use std::sync::Arc;
-
-use async_stream::try_stream;
-use bytes::Bytes;
-use futures::Stream;
-use postgres_backend::CopyStreamHandlerEnd;
-use safekeeper_api::Term;
-use std::time::Duration;
-use tokio::time::timeout;
-use utils::lsn::Lsn;
-
-use crate::{
-    send_wal::{EndWatch, WalSenderGuard},
-    timeline::WalResidentTimeline,
+use std::{
+    pin::Pin,
+    task::{Context, Poll},
 };
 
-pub(crate) struct WalReaderStreamBuilder {
-    pub(crate) tli: WalResidentTimeline,
-    pub(crate) start_pos: Lsn,
-    pub(crate) end_pos: Lsn,
-    pub(crate) term: Option<Term>,
-    pub(crate) end_watch: EndWatch,
-    pub(crate) wal_sender_guard: Arc<WalSenderGuard>,
-}
+use bytes::Bytes;
+use futures::{stream::BoxStream, Stream, StreamExt};
+use utils::lsn::Lsn;
 
-impl WalReaderStreamBuilder {
-    pub(crate) fn start_pos(&self) -> Lsn {
-        self.start_pos
-    }
-}
+use crate::{send_wal::EndWatch, timeline::WalResidentTimeline, wal_storage::WalReader};
+use safekeeper_api::Term;
 
+#[derive(PartialEq, Eq, Debug)]
 pub(crate) struct WalBytes {
     /// Raw PG WAL
     pub(crate) wal: Bytes,
@@ -44,106 +26,270 @@ pub(crate) struct WalBytes {
     pub(crate) available_wal_end_lsn: Lsn,
 }
 
-impl WalReaderStreamBuilder {
-    /// Builds a stream of Postgres WAL starting from [`Self::start_pos`].
-    /// The stream terminates when the receiver (pageserver) is fully caught up
-    /// and there's no active computes.
-    pub(crate) async fn build(
-        self,
-        buffer_size: usize,
-    ) -> anyhow::Result<impl Stream<Item = Result<WalBytes, CopyStreamHandlerEnd>>> {
-        // TODO(vlad): The code below duplicates functionality from [`crate::send_wal`].
-        // We can make the raw WAL sender use this stream too and remove the duplication.
-        let Self {
-            tli,
-            mut start_pos,
-            mut end_pos,
-            term,
-            mut end_watch,
-            wal_sender_guard,
-        } = self;
-        let mut wal_reader = tli.get_walreader(start_pos).await?;
-        let mut buffer = vec![0; buffer_size];
+struct PositionedWalReader {
+    start: Lsn,
+    end: Lsn,
+    reader: Option<WalReader>,
+}
 
-        const POLL_STATE_TIMEOUT: Duration = Duration::from_secs(1);
+/// A streaming WAL reader wrapper which can be reset while running
+pub(crate) struct StreamingWalReader {
+    stream: BoxStream<'static, WalOrReset>,
+    start_changed_tx: tokio::sync::watch::Sender<Lsn>,
+}
 
-        Ok(try_stream! {
-            loop {
-                let have_something_to_send = end_pos > start_pos;
+pub(crate) enum WalOrReset {
+    Wal(anyhow::Result<WalBytes>),
+    Reset(Lsn),
+}
 
-                if !have_something_to_send {
-                    // wait for lsn
-                    let res = timeout(POLL_STATE_TIMEOUT, end_watch.wait_for_lsn(start_pos, term)).await;
-                    match res {
-                        Ok(ok) => {
-                            end_pos = ok?;
-                        },
-                        Err(_) => {
-                            if let EndWatch::Commit(_) = end_watch {
-                                if let Some(remote_consistent_lsn) = wal_sender_guard
-                                    .walsenders()
-                                    .get_ws_remote_consistent_lsn(wal_sender_guard.id())
-                                {
-                                    if tli.should_walsender_stop(remote_consistent_lsn).await {
-                                        // Stop streaming if the receivers are caught up and
-                                        // there's no active compute. This causes the loop in
-                                        // [`crate::send_interpreted_wal::InterpretedWalSender::run`]
-                                        // to exit and terminate the WAL stream.
-                                        return;
-                                    }
-                                }
-                            }
-
-                            continue;
-                        }
-                    }
-                }
-
-
-                assert!(
-                    end_pos > start_pos,
-                    "nothing to send after waiting for WAL"
-                );
-
-                // try to send as much as available, capped by the buffer size
-                let mut chunk_end_pos = start_pos + buffer_size as u64;
-                // if we went behind available WAL, back off
-                if chunk_end_pos >= end_pos {
-                    chunk_end_pos = end_pos;
-                } else {
-                    // If sending not up to end pos, round down to page boundary to
-                    // avoid breaking WAL record not at page boundary, as protocol
-                    // demands. See walsender.c (XLogSendPhysical).
-                    chunk_end_pos = chunk_end_pos
-                        .checked_sub(chunk_end_pos.block_offset())
-                        .unwrap();
-                }
-                let send_size = (chunk_end_pos.0 - start_pos.0) as usize;
-                let buffer = &mut buffer[..send_size];
-                let send_size: usize;
-                {
-                    // If uncommitted part is being pulled, check that the term is
-                    // still the expected one.
-                    let _term_guard = if let Some(t) = term {
-                        Some(tli.acquire_term(t).await?)
-                    } else {
-                        None
-                    };
-                    // Read WAL into buffer. send_size can be additionally capped to
-                    // segment boundary here.
-                    send_size = wal_reader.read(buffer).await?
-                };
-                let wal = Bytes::copy_from_slice(&buffer[..send_size]);
-
-                yield WalBytes {
-                    wal,
-                    wal_start_lsn: start_pos,
-                    wal_end_lsn: start_pos + send_size as u64,
-                    available_wal_end_lsn: end_pos
-                };
-
-                start_pos += send_size as u64;
-            }
-        })
+impl WalOrReset {
+    pub(crate) fn get_wal(self) -> Option<anyhow::Result<WalBytes>> {
+        match self {
+            WalOrReset::Wal(wal) => Some(wal),
+            WalOrReset::Reset(_) => None,
+        }
+    }
+}
+
+impl StreamingWalReader {
+    pub(crate) fn new(
+        tli: WalResidentTimeline,
+        term: Option<Term>,
+        start: Lsn,
+        end: Lsn,
+        end_watch: EndWatch,
+        buffer_size: usize,
+    ) -> Self {
+        let (start_changed_tx, start_changed_rx) = tokio::sync::watch::channel(start);
+
+        let state = WalReaderStreamState {
+            tli,
+            wal_reader: PositionedWalReader {
+                start,
+                end,
+                reader: None,
+            },
+            term,
+            end_watch,
+            buffer: vec![0; buffer_size],
+            buffer_size,
+        };
+
+        // When a change notification is received while polling the internal
+        // reader, stop polling the read future and service the change.
+        let stream = futures::stream::unfold(
+            (state, start_changed_rx),
+            |(mut state, mut rx)| async move {
+                let wal_or_reset = tokio::select! {
+                    read_res = state.read() => { WalOrReset::Wal(read_res) },
+                    changed_res = rx.changed() => {
+                        if changed_res.is_err() {
+                            return None;
+                        }
+
+                        let new_start_pos = rx.borrow_and_update();
+                        WalOrReset::Reset(*new_start_pos)
+                    }
+                };
+
+                if let WalOrReset::Reset(lsn) = wal_or_reset {
+                    state.wal_reader.start = lsn;
+                    state.wal_reader.reader = None;
+                }
+
+                Some((wal_or_reset, (state, rx)))
+            },
+        )
+        .boxed();
+
+        Self {
+            stream,
+            start_changed_tx,
+        }
+    }
+
+    /// Reset the stream to a given position.
+    pub(crate) async fn reset(&mut self, start: Lsn) {
+        self.start_changed_tx.send(start).unwrap();
+        while let Some(wal_or_reset) = self.stream.next().await {
+            match wal_or_reset {
+                WalOrReset::Reset(at) => {
+                    // Stream confirmed the reset.
+                    // There may only one ongoing reset at any given time,
+                    // hence the assertion.
+                    assert_eq!(at, start);
+                    break;
+                }
+                WalOrReset::Wal(_) => {
+                    // Ignore wal generated before reset was handled
+                }
+            }
+        }
+    }
+}
+
+impl Stream for StreamingWalReader {
+    type Item = WalOrReset;
+
+    fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        Pin::new(&mut self.stream).poll_next(cx)
+    }
+}
+
+struct WalReaderStreamState {
+    tli: WalResidentTimeline,
+    wal_reader: PositionedWalReader,
+    term: Option<Term>,
+    end_watch: EndWatch,
+    buffer: Vec<u8>,
+    buffer_size: usize,
+}
+
+impl WalReaderStreamState {
+    async fn read(&mut self) -> anyhow::Result<WalBytes> {
+        // Create reader if needed
+        if self.wal_reader.reader.is_none() {
+            self.wal_reader.reader = Some(self.tli.get_walreader(self.wal_reader.start).await?);
+        }
+
+        let have_something_to_send = self.wal_reader.end > self.wal_reader.start;
+        if !have_something_to_send {
+            tracing::debug!(
+                "Waiting for wal: start={}, end={}",
+                self.wal_reader.end,
+                self.wal_reader.start
+            );
+            self.wal_reader.end = self
+                .end_watch
+                .wait_for_lsn(self.wal_reader.start, self.term)
+                .await?;
+            tracing::debug!(
+                "Done waiting for wal: start={}, end={}",
+                self.wal_reader.end,
+                self.wal_reader.start
+            );
+        }
+
+        assert!(
+            self.wal_reader.end > self.wal_reader.start,
+            "nothing to send after waiting for WAL"
+        );
+
+        // Calculate chunk size
+        let mut chunk_end_pos = self.wal_reader.start + self.buffer_size as u64;
+        if chunk_end_pos >= self.wal_reader.end {
+            chunk_end_pos = self.wal_reader.end;
+        } else {
+            chunk_end_pos = chunk_end_pos
+                .checked_sub(chunk_end_pos.block_offset())
+                .unwrap();
+        }
+
+        let send_size = (chunk_end_pos.0 - self.wal_reader.start.0) as usize;
+        let buffer = &mut self.buffer[..send_size];
+
+        // Read WAL
+        let send_size = {
+            let _term_guard = if let Some(t) = self.term {
+                Some(self.tli.acquire_term(t).await?)
+            } else {
+                None
+            };
+            self.wal_reader
+                .reader
+                .as_mut()
+                .unwrap()
+                .read(buffer)
+                .await?
+        };
+
+        let wal = Bytes::copy_from_slice(&buffer[..send_size]);
+        let result = WalBytes {
+            wal,
+            wal_start_lsn: self.wal_reader.start,
+            wal_end_lsn: self.wal_reader.start + send_size as u64,
+            available_wal_end_lsn: self.wal_reader.end,
+        };
+
+        self.wal_reader.start += send_size as u64;
+
+        Ok(result)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::str::FromStr;
+
+    use futures::StreamExt;
+    use postgres_ffi::MAX_SEND_SIZE;
+    use utils::{
+        id::{NodeId, TenantTimelineId},
+        lsn::Lsn,
+    };
+
+    use crate::{test_utils::Env, wal_reader_stream::StreamingWalReader};
+
+    #[tokio::test]
+    async fn test_streaming_wal_reader_reset() {
+        let _ = env_logger::builder().is_test(true).try_init();
+
+        const SIZE: usize = 8 * 1024;
+        const MSG_COUNT: usize = 200;
+
+        let start_lsn = Lsn::from_str("0/149FD18").unwrap();
+        let env = Env::new(true).unwrap();
+        let tli = env
+            .make_timeline(NodeId(1), TenantTimelineId::generate(), start_lsn)
+            .await
+            .unwrap();
+
+        let resident_tli = tli.wal_residence_guard().await.unwrap();
+        let end_watch = Env::write_wal(tli, start_lsn, SIZE, MSG_COUNT)
+            .await
+            .unwrap();
+        let end_pos = end_watch.get();
+
+        tracing::info!("Doing first round of reads ...");
+
+        let mut streaming_wal_reader = StreamingWalReader::new(
+            resident_tli,
+            None,
+            start_lsn,
+            end_pos,
+            end_watch,
+            MAX_SEND_SIZE,
+        );
+
+        let mut before_reset = Vec::new();
+        while let Some(wor) = streaming_wal_reader.next().await {
+            let wal = wor.get_wal().unwrap().unwrap();
+            let stop = wal.available_wal_end_lsn == wal.wal_end_lsn;
+            before_reset.push(wal);
+
+            if stop {
+                break;
+            }
+        }
+
+        tracing::info!("Resetting the WAL stream ...");
+
+        streaming_wal_reader.reset(start_lsn).await;
+
+        tracing::info!("Doing second round of reads ...");
+
+        let mut after_reset = Vec::new();
+        while let Some(wor) = streaming_wal_reader.next().await {
+            let wal = wor.get_wal().unwrap().unwrap();
+            let stop = wal.available_wal_end_lsn == wal.wal_end_lsn;
+            after_reset.push(wal);
+
+            if stop {
+                break;
+            }
+        }
+
+        assert_eq!(before_reset, after_reset);
     }
 }

safekeeper/tests/walproposer_sim/safekeeper.rs

@@ -21,7 +21,7 @@ use safekeeper::{
     wal_storage::Storage,
     SafeKeeperConf,
 };
-use safekeeper_api::ServerInfo;
+use safekeeper_api::{membership::Configuration, ServerInfo};
 use tracing::{debug, info_span, warn};
 use utils::{
     id::{NodeId, TenantId, TenantTimelineId, TimelineId},
@@ -96,8 +96,13 @@ impl GlobalMap {
         let commit_lsn = Lsn::INVALID;
         let local_start_lsn = Lsn::INVALID;
 
-        let state =
-            TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn)?;
+        let state = TimelinePersistentState::new(
+            &ttid,
+            Configuration::empty(),
+            server_info,
+            commit_lsn,
+            local_start_lsn,
+        )?;
 
         let disk_timeline = self.disk.put_state(&ttid, state);
         let control_store = DiskStateStorage::new(disk_timeline.clone());
@@ -173,6 +178,8 @@ pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
         control_file_save_interval: Duration::from_secs(1),
         partial_backup_concurrency: 1,
         eviction_min_resident: Duration::ZERO,
+        wal_reader_fanout: false,
+        max_delta_for_fanout: None,
     };
 
     let mut global = GlobalMap::new(disk, conf.clone())?;

storage_controller/src/http.rs

@@ -15,7 +15,7 @@ use metrics::{BuildInfo, NeonMetrics};
 use pageserver_api::controller_api::{
     MetadataHealthListOutdatedRequest, MetadataHealthListOutdatedResponse,
     MetadataHealthListUnhealthyResponse, MetadataHealthUpdateRequest, MetadataHealthUpdateResponse,
-    ShardsPreferredAzsRequest, TenantCreateRequest,
+    SafekeeperSchedulingPolicyRequest, ShardsPreferredAzsRequest, TenantCreateRequest,
 };
 use pageserver_api::models::{
     TenantConfigPatchRequest, TenantConfigRequest, TenantLocationConfigRequest,
@@ -653,6 +653,10 @@ async fn handle_tenant_list(
 ) -> Result<Response<Body>, ApiError> {
     check_permissions(&req, Scope::Admin)?;
 
+    let limit: Option<usize> = parse_query_param(&req, "limit")?;
+    let start_after: Option<TenantId> = parse_query_param(&req, "start_after")?;
+    tracing::info!("start_after: {:?}", start_after);
+
     match maybe_forward(req).await {
         ForwardOutcome::Forwarded(res) => {
             return res;
@@ -660,7 +664,7 @@ async fn handle_tenant_list(
         ForwardOutcome::NotForwarded(_req) => {}
     };
 
-    json_response(StatusCode::OK, service.tenant_list())
+    json_response(StatusCode::OK, service.tenant_list(limit, start_after))
 }
 
 async fn handle_node_register(req: Request<Body>) -> Result<Response<Body>, ApiError> {
@@ -1301,6 +1305,35 @@ async fn handle_upsert_safekeeper(mut req: Request<Body>) -> Result<Response<Bod
         .unwrap())
 }
 
+/// Sets the scheduling policy of the specified safekeeper
+async fn handle_safekeeper_scheduling_policy(
+    mut req: Request<Body>,
+) -> Result<Response<Body>, ApiError> {
+    check_permissions(&req, Scope::Admin)?;
+
+    let body = json_request::<SafekeeperSchedulingPolicyRequest>(&mut req).await?;
+    let id = parse_request_param::<i64>(&req, "id")?;
+
+    let req = match maybe_forward(req).await {
+        ForwardOutcome::Forwarded(res) => {
+            return res;
+        }
+        ForwardOutcome::NotForwarded(req) => req,
+    };
+
+    let state = get_state(&req);
+
+    state
+        .service
+        .set_safekeeper_scheduling_policy(id, body.scheduling_policy)
+        .await?;
+
+    Ok(Response::builder()
+        .status(StatusCode::NO_CONTENT)
+        .body(Body::empty())
+        .unwrap())
+}
+
 /// Common wrapper for request handlers that call into Service and will operate on tenants: they must only
 /// be allowed to run if Service has finished its initial reconciliation.
 async fn tenant_service_handler<R, H>(
@@ -1869,7 +1902,18 @@ pub fn make_router(
         })
         .post("/control/v1/safekeeper/:id", |r| {
             // id is in the body
-            named_request_span(r, handle_upsert_safekeeper, RequestName("v1_safekeeper"))
+            named_request_span(
+                r,
+                handle_upsert_safekeeper,
+                RequestName("v1_safekeeper_post"),
+            )
+        })
+        .post("/control/v1/safekeeper/:id/scheduling_policy", |r| {
+            named_request_span(
+                r,
+                handle_safekeeper_scheduling_policy,
+                RequestName("v1_safekeeper_status"),
+            )
         })
         // Tenant Shard operations
         .put("/control/v1/tenant/:tenant_shard_id/migrate", |r| {

storage_controller/src/persistence.rs

@@ -1104,6 +1104,37 @@ impl Persistence {
         })
         .await
     }
+
+    pub(crate) async fn set_safekeeper_scheduling_policy(
+        &self,
+        id_: i64,
+        scheduling_policy_: SkSchedulingPolicy,
+    ) -> Result<(), DatabaseError> {
+        use crate::schema::safekeepers::dsl::*;
+
+        self.with_conn(move |conn| -> DatabaseResult<()> {
+            #[derive(Insertable, AsChangeset)]
+            #[diesel(table_name = crate::schema::safekeepers)]
+            struct UpdateSkSchedulingPolicy<'a> {
+                id: i64,
+                scheduling_policy: &'a str,
+            }
+            let scheduling_policy_ = String::from(scheduling_policy_);
+
+            let rows_affected = diesel::update(safekeepers.filter(id.eq(id_)))
+                .set(scheduling_policy.eq(scheduling_policy_))
+                .execute(conn)?;
+
+            if rows_affected != 1 {
+                return Err(DatabaseError::Logical(format!(
+                    "unexpected number of rows ({rows_affected})",
+                )));
+            }
+
+            Ok(())
+        })
+        .await
+    }
 }
 
 /// Parts of [`crate::tenant_shard::TenantShard`] that are stored durably

storage_controller/src/service.rs

@@ -47,7 +47,7 @@ use pageserver_api::{
         AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
         NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
         SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest,
-        ShardsPreferredAzsResponse, TenantCreateRequest, TenantCreateResponse,
+        ShardsPreferredAzsResponse, SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse,
         TenantCreateResponseShard, TenantDescribeResponse, TenantDescribeResponseShard,
         TenantLocateResponse, TenantPolicyRequest, TenantShardMigrateRequest,
         TenantShardMigrateResponse,
@@ -4158,17 +4158,42 @@ impl Service {
         .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
     }
 
-    pub(crate) fn tenant_list(&self) -> Vec<TenantDescribeResponse> {
+    /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
+    /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
+    /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
+    /// in our external API.
+    pub(crate) fn tenant_list(
+        &self,
+        limit: Option<usize>,
+        start_after: Option<TenantId>,
+    ) -> Vec<TenantDescribeResponse> {
         let locked = self.inner.read().unwrap();
 
+        // Apply start_from parameter
+        let shard_range = match start_after {
+            None => locked.tenants.range(..),
+            Some(tenant_id) => locked.tenants.range(
+                TenantShardId {
+                    tenant_id,
+                    shard_number: ShardNumber(u8::MAX),
+                    shard_count: ShardCount(u8::MAX),
+                }..,
+            ),
+        };
+
         let mut result = Vec::new();
-        for (_tenant_id, tenant_shards) in
-            &locked.tenants.iter().group_by(|(id, _shard)| id.tenant_id)
-        {
+        for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
             result.push(
                 self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
                     .expect("Groups are always non-empty"),
             );
+
+            // Enforce `limit` parameter
+            if let Some(limit) = limit {
+                if result.len() >= limit {
+                    break;
+                }
+            }
         }
 
         result
@@ -7626,6 +7651,16 @@ impl Service {
         self.persistence.safekeeper_upsert(record).await
     }
 
+    pub(crate) async fn set_safekeeper_scheduling_policy(
+        &self,
+        id: i64,
+        scheduling_policy: SkSchedulingPolicy,
+    ) -> Result<(), DatabaseError> {
+        self.persistence
+            .set_safekeeper_scheduling_policy(id, scheduling_policy)
+            .await
+    }
+
     pub(crate) async fn update_shards_preferred_azs(
         &self,
         req: ShardsPreferredAzsRequest,

test_runner/fixtures/neon_fixtures.py

@@ -1884,7 +1884,10 @@ class NeonStorageController(MetricsGetter, LogUtils):
         )
         return response.json()
 
-    def tenant_list(self):
+    def tenant_shard_dump(self):
+        """
+        Debug listing API: dumps the internal map of tenant shards
+        """
         response = self.request(
             "GET",
             f"{self.api}/debug/v1/tenant",
@@ -1892,6 +1895,18 @@ class NeonStorageController(MetricsGetter, LogUtils):
         )
         return response.json()
 
+    def tenant_list(self, **kwargs):
+        """
+        Control API tenant listing: a vector of the same content returned by tenant_describe
+        """
+        response = self.request(
+            "GET",
+            f"{self.api}/control/v1/tenant",
+            headers=self.headers(TokenScope.ADMIN),
+            params=kwargs,
+        )
+        return response.json()
+
     def node_configure(self, node_id, body: dict[str, Any]):
         log.info(f"node_configure({node_id}, {body})")
         body["node_id"] = node_id
@@ -2238,7 +2253,7 @@ class NeonStorageController(MetricsGetter, LogUtils):
         """
         Get the intent and observed placements of all tenants known to the storage controller.
         """
-        tenants = self.tenant_list()
+        tenants = self.tenant_shard_dump()
 
         tenant_placement: defaultdict[str, dict[str, Any]] = defaultdict(
             lambda: {
@@ -2321,6 +2336,14 @@ class NeonStorageController(MetricsGetter, LogUtils):
             json=body,
         )
 
+    def safekeeper_scheduling_policy(self, id: int, scheduling_policy: str):
+        self.request(
+            "POST",
+            f"{self.api}/control/v1/safekeeper/{id}/scheduling_policy",
+            headers=self.headers(TokenScope.ADMIN),
+            json={"id": id, "scheduling_policy": scheduling_policy},
+        )
+
     def get_safekeeper(self, id: int) -> dict[str, Any] | None:
         try:
             response = self.request(
@@ -4120,7 +4143,7 @@ class Endpoint(PgProtocol, LogUtils):
 
     # Checkpoints running endpoint and returns pg_wal size in MB.
     def get_pg_wal_size(self):
-        log.info(f'checkpointing at LSN {self.safe_psql("select pg_current_wal_lsn()")[0][0]}')
+        log.info(f"checkpointing at LSN {self.safe_psql('select pg_current_wal_lsn()')[0][0]}")
         self.safe_psql("checkpoint")
         assert self.pgdata_dir is not None  # please mypy
         return get_dir_size(self.pgdata_dir / "pg_wal") / 1024 / 1024
@@ -4960,7 +4983,7 @@ def logical_replication_sync(
         if res:
             log.info(f"subscriber_lsn={res}")
             subscriber_lsn = Lsn(res)
-            log.info(f"Subscriber LSN={subscriber_lsn}, publisher LSN={ publisher_lsn}")
+            log.info(f"Subscriber LSN={subscriber_lsn}, publisher LSN={publisher_lsn}")
             if subscriber_lsn >= publisher_lsn:
                 return subscriber_lsn
         time.sleep(0.5)

test_runner/fixtures/pageserver/http.py

@@ -15,7 +15,6 @@ from requests.adapters import HTTPAdapter
 from urllib3.util.retry import Retry
 
 from fixtures.common_types import (
-    Id,
     Lsn,
     TenantId,
     TenantShardId,
@@ -25,7 +24,7 @@ from fixtures.common_types import (
 from fixtures.log_helper import log
 from fixtures.metrics import Metrics, MetricsGetter, parse_metrics
 from fixtures.pg_version import PgVersion
-from fixtures.utils import Fn
+from fixtures.utils import EnhancedJSONEncoder, Fn
 
 
 class PageserverApiException(Exception):
@@ -83,14 +82,6 @@ class TimelineCreateRequest:
     mode: TimelineCreateRequestMode
 
     def to_json(self) -> str:
-        class EnhancedJSONEncoder(json.JSONEncoder):
-            def default(self, o):
-                if dataclasses.is_dataclass(o) and not isinstance(o, type):
-                    return dataclasses.asdict(o)
-                elif isinstance(o, Id):
-                    return o.id.hex()
-                return super().default(o)
-
         # mode is flattened
         this = dataclasses.asdict(self)
         mode = this.pop("mode")

test_runner/fixtures/safekeeper/http.py

@@ -10,7 +10,7 @@ import requests
 from fixtures.common_types import Lsn, TenantId, TenantTimelineId, TimelineId
 from fixtures.log_helper import log
 from fixtures.metrics import Metrics, MetricsGetter, parse_metrics
-from fixtures.utils import wait_until
+from fixtures.utils import EnhancedJSONEncoder, wait_until
 
 if TYPE_CHECKING:
     from typing import Any
@@ -25,6 +25,7 @@ class Walreceiver:
 
 @dataclass
 class SafekeeperTimelineStatus:
+    mconf: Configuration | None
     term: int
     last_log_term: int
     pg_version: int  # Not exactly a PgVersion, safekeeper returns version as int, for example 150002 for 15.2
@@ -69,6 +70,56 @@ class TermBumpResponse:
         )
 
 
+@dataclass
+class SafekeeperId:
+    id: int
+    host: str
+    pg_port: int
+
+
+@dataclass
+class Configuration:
+    generation: int
+    members: list[SafekeeperId]
+    new_members: list[SafekeeperId] | None
+
+    @classmethod
+    def from_json(cls, d: dict[str, Any]) -> Configuration:
+        generation = d["generation"]
+        members = d["members"]
+        new_members = d.get("new_members")
+        return Configuration(generation, members, new_members)
+
+    def to_json(self) -> str:
+        return json.dumps(self, cls=EnhancedJSONEncoder)
+
+
+@dataclass
+class TimelineCreateRequest:
+    tenant_id: TenantId
+    timeline_id: TimelineId
+    mconf: Configuration
+    # not exactly PgVersion, for example 150002 for 15.2
+    pg_version: int
+    start_lsn: Lsn
+    commit_lsn: Lsn | None
+
+    def to_json(self) -> str:
+        return json.dumps(self, cls=EnhancedJSONEncoder)
+
+
+@dataclass
+class TimelineMembershipSwitchResponse:
+    previous_conf: Configuration
+    current_conf: Configuration
+
+    @classmethod
+    def from_json(cls, d: dict[str, Any]) -> TimelineMembershipSwitchResponse:
+        previous_conf = Configuration.from_json(d["previous_conf"])
+        current_conf = Configuration.from_json(d["current_conf"])
+        return TimelineMembershipSwitchResponse(previous_conf, current_conf)
+
+
 class SafekeeperHttpClient(requests.Session, MetricsGetter):
     HTTPError = requests.HTTPError
 
@@ -131,20 +182,8 @@ class SafekeeperHttpClient(requests.Session, MetricsGetter):
         resj = res.json()
         return [TenantTimelineId.from_json(ttidj) for ttidj in resj]
 
-    def timeline_create(
-        self,
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-        pg_version: int,  # Not exactly a PgVersion, safekeeper returns version as int, for example 150002 for 15.2
-        commit_lsn: Lsn,
-    ):
-        body = {
-            "tenant_id": str(tenant_id),
-            "timeline_id": str(timeline_id),
-            "pg_version": pg_version,
-            "commit_lsn": str(commit_lsn),
-        }
-        res = self.post(f"http://localhost:{self.port}/v1/tenant/timeline", json=body)
+    def timeline_create(self, r: TimelineCreateRequest):
+        res = self.post(f"http://localhost:{self.port}/v1/tenant/timeline", data=r.to_json())
         res.raise_for_status()
 
     def timeline_status(
@@ -154,7 +193,10 @@ class SafekeeperHttpClient(requests.Session, MetricsGetter):
         res.raise_for_status()
         resj = res.json()
         walreceivers = [Walreceiver(wr["conn_id"], wr["status"]) for wr in resj["walreceivers"]]
+        # It is always normally not None, it is allowed only to make forward compat tests happy.
+        mconf = Configuration.from_json(resj["mconf"]) if "mconf" in resj else None
         return SafekeeperTimelineStatus(
+            mconf=mconf,
             term=resj["acceptor_state"]["term"],
             last_log_term=resj["acceptor_state"]["epoch"],
             pg_version=resj["pg_info"]["pg_version"],
@@ -180,6 +222,11 @@ class SafekeeperHttpClient(requests.Session, MetricsGetter):
     def get_commit_lsn(self, tenant_id: TenantId, timeline_id: TimelineId) -> Lsn:
         return self.timeline_status(tenant_id, timeline_id).commit_lsn
 
+    # Get timeline membership configuration.
+    def get_membership(self, tenant_id: TenantId, timeline_id: TimelineId) -> Configuration:
+        # make mypy happy
+        return self.timeline_status(tenant_id, timeline_id).mconf  # type: ignore
+
     # only_local doesn't remove segments in the remote storage.
     def timeline_delete(
         self, tenant_id: TenantId, timeline_id: TimelineId, only_local: bool = False
@@ -226,6 +273,16 @@ class SafekeeperHttpClient(requests.Session, MetricsGetter):
         assert isinstance(res_json, dict)
         return res_json
 
+    def membership_switch(
+        self, tenant_id: TenantId, timeline_id: TimelineId, to: Configuration
+    ) -> TimelineMembershipSwitchResponse:
+        res = self.post(
+            f"http://localhost:{self.port}/v1/tenant/{tenant_id}/timeline/{timeline_id}/membership",
+            data=to.to_json(),
+        )
+        res.raise_for_status()
+        return TimelineMembershipSwitchResponse.from_json(res.json())
+
     def copy_timeline(self, tenant_id: TenantId, timeline_id: TimelineId, body: dict[str, Any]):
         res = self.post(
             f"http://localhost:{self.port}/v1/tenant/{tenant_id}/timeline/{timeline_id}/copy",

test_runner/fixtures/utils.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import contextlib
+import dataclasses
 import json
 import os
 import re
@@ -21,6 +22,7 @@ import zstandard
 from psycopg2.extensions import cursor
 from typing_extensions import override
 
+from fixtures.common_types import Id, Lsn
 from fixtures.log_helper import log
 from fixtures.pageserver.common_types import (
     parse_delta_layer,
@@ -605,6 +607,22 @@ class PropagatingThread(threading.Thread):
         return self.ret
 
 
+class EnhancedJSONEncoder(json.JSONEncoder):
+    """
+    Default json.JSONEncoder works only on primitive builtins. Extend it to any
+    dataclass plus our custom types.
+    """
+
+    def default(self, o):
+        if dataclasses.is_dataclass(o) and not isinstance(o, type):
+            return dataclasses.asdict(o)
+        elif isinstance(o, Id):
+            return o.id.hex()
+        elif isinstance(o, Lsn):
+            return str(o)  # standard hex notation
+        return super().default(o)
+
+
 def human_bytes(amt: float) -> str:
     """
     Render a bytes amount into nice IEC bytes string.

test_runner/fixtures/workload.py

@@ -53,6 +53,22 @@ class Workload:
         self._endpoint: Endpoint | None = None
         self._endpoint_opts = endpoint_opts or {}
 
+    def branch(
+        self,
+        timeline_id: TimelineId,
+        branch_name: str | None = None,
+        endpoint_opts: dict[str, Any] | None = None,
+    ) -> Workload:
+        """
+        Checkpoint the current status of the workload in case of branching
+        """
+        branch_workload = Workload(
+            self.env, self.tenant_id, timeline_id, branch_name, endpoint_opts
+        )
+        branch_workload.expect_rows = self.expect_rows
+        branch_workload.churn_cursor = self.churn_cursor
+        return branch_workload
+
     def reconfigure(self) -> None:
         """
         Request the endpoint to reconfigure based on location reported by storage controller

test_runner/regress/test_compaction.py

@@ -112,7 +112,11 @@ page_cache_size=10
 
 
 @skip_in_debug_build("only run with release build")
-def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder):
+@pytest.mark.parametrize(
+    "with_branches",
+    ["with_branches", "no_branches"],
+)
+def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder, with_branches: str):
     SMOKE_CONF = {
         # Run both gc and gc-compaction.
         "gc_period": "5s",
@@ -143,12 +147,17 @@ def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder):
     log.info("Writing initial data ...")
     workload.write_rows(row_count, env.pageserver.id)
 
+    child_workloads: list[Workload] = []
+
     for i in range(1, churn_rounds + 1):
         if i % 10 == 0:
             log.info(f"Running churn round {i}/{churn_rounds} ...")
-
-        if (i - 1) % 10 == 0:
-            # Run gc-compaction every 10 rounds to ensure the test doesn't take too long time.
+        if i % 10 == 5 and with_branches == "with_branches":
+            branch_name = f"child-{i}"
+            branch_timeline_id = env.create_branch(branch_name)
+            child_workloads.append(workload.branch(branch_timeline_id, branch_name))
+        if (i - 1) % 10 == 0 or (i - 1) % 10 == 1:
+            # Run gc-compaction twice every 10 rounds to ensure the test doesn't take too long time.
             ps_http.timeline_compact(
                 tenant_id,
                 timeline_id,
@@ -179,6 +188,9 @@ def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder):
 
     log.info("Validating at workload end ...")
     workload.validate(env.pageserver.id)
+    for child_workload in child_workloads:
+        log.info(f"Validating at branch {child_workload.branch_name}")
+        child_workload.validate(env.pageserver.id)
 
     # Run a legacy compaction+gc to ensure gc-compaction can coexist with legacy compaction.
     ps_http.timeline_checkpoint(tenant_id, timeline_id, wait_until_uploaded=True)

test_runner/regress/test_page_service_batching_regressions.py

@@ -1,60 +0,0 @@
-# NB: there are benchmarks that double-serve as tests inside the `performance` directory.
-
-import subprocess
-from pathlib import Path
-
-import pytest
-from fixtures.log_helper import log
-from fixtures.neon_fixtures import NeonEnvBuilder
-
-
-@pytest.mark.timeout(30)  # test takes <20s if pageserver impl is correct
-@pytest.mark.parametrize("kind", ["pageserver-stop", "tenant-detach"])
-def test_slow_flush(neon_env_builder: NeonEnvBuilder, neon_binpath: Path, kind: str):
-    def patch_pageserver_toml(config):
-        config["page_service_pipelining"] = {
-            "mode": "pipelined",
-            "max_batch_size": 32,
-            "execution": "concurrent-futures",
-        }
-
-    neon_env_builder.pageserver_config_override = patch_pageserver_toml
-    env = neon_env_builder.init_start()
-
-    log.info("make flush appear slow")
-
-    log.info("sending requests until pageserver accepts no more")
-    # TODO: extract this into a helper, like subprocess_capture,
-    # so that we capture the stderr from the helper somewhere.
-    child = subprocess.Popen(
-        [
-            neon_binpath / "test_helper_slow_client_reads",
-            env.pageserver.connstr(),
-            str(env.initial_tenant),
-            str(env.initial_timeline),
-        ],
-        bufsize=0,  # unbuffered
-        stdin=subprocess.PIPE,
-        stdout=subprocess.PIPE,
-    )
-    assert child.stdout is not None
-    buf = child.stdout.read(1)
-    if len(buf) != 1:
-        raise Exception("unexpected EOF")
-    if buf != b"R":
-        raise Exception(f"unexpected data: {buf!r}")
-    log.info("helper reports pageserver accepts no more requests")
-    log.info(
-        "assuming pageserver connection handle is in a state where TCP has backpressured pageserver=>client response flush() into userspace"
-    )
-
-    if kind == "pageserver-stop":
-        log.info("try to shut down the pageserver cleanly")
-        env.pageserver.stop()
-    elif kind == "tenant-detach":
-        log.info("try to shut down the tenant")
-        env.pageserver.tenant_detach(env.initial_tenant)
-    else:
-        raise ValueError(f"unexpected kind: {kind}")
-
-    log.info("shutdown did not time out, test passed")

test_runner/regress/test_storage_controller.py

@@ -113,6 +113,19 @@ def test_storage_controller_smoke(neon_env_builder: NeonEnvBuilder, combination)
     for tid in tenant_ids:
         env.create_tenant(tid, shard_count=shards_per_tenant)
 
+    # Tenant listing API should work
+    listed_tenants = env.storage_controller.tenant_list()
+    log.info(f"listed_tenants: {listed_tenants}")
+    assert set(t["tenant_id"] for t in listed_tenants) == set(str(t) for t in tenant_ids)
+    paged = env.storage_controller.tenant_list(limit=2, start_after=listed_tenants[0]["tenant_id"])
+    assert len(paged) == 2
+    assert paged[0] == listed_tenants[1]
+    assert paged[1] == listed_tenants[2]
+    paged = env.storage_controller.tenant_list(
+        limit=1000, start_after="ffffffffffffffffffffffffffffffff"
+    )
+    assert paged == []
+
     # Validate high level metrics
     assert (
         env.storage_controller.get_metric_value("storage_controller_tenant_shards")
@@ -1506,7 +1519,7 @@ class PageserverFailpoint(Failure):
 
 
 def build_node_to_tenants_map(env: NeonEnv) -> dict[int, list[TenantId]]:
-    tenants = env.storage_controller.tenant_list()
+    tenants = env.storage_controller.tenant_shard_dump()
 
     node_to_tenants: dict[int, list[TenantId]] = {}
     for t in tenants:
@@ -2631,7 +2644,7 @@ def test_storage_controller_step_down(neon_env_builder: NeonEnvBuilder):
     # Validate that the storcon attempts to forward the request, but stops.
     # when it realises it is still the current leader.
     with pytest.raises(StorageControllerApiException, match="Leader is stepped down instance"):
-        env.storage_controller.tenant_list()
+        env.storage_controller.tenant_shard_dump()
 
     # Validate that we can step down multiple times and the observed state
     # doesn't change.
@@ -2781,7 +2794,7 @@ def test_storage_controller_leadership_transfer(
         # Check that the stepped down instance forwards requests
         # to the new leader while it's still running.
         storage_controller_proxy.route_to(f"http://127.0.0.1:{storage_controller_1_port}")
-        env.storage_controller.tenant_list()
+        env.storage_controller.tenant_shard_dump()
         env.storage_controller.node_configure(env.pageservers[0].id, {"scheduling": "Pause"})
         status = env.storage_controller.node_status(env.pageservers[0].id)
         assert status["scheduling"] == "Pause"
@@ -3195,6 +3208,17 @@ def test_safekeeper_deployment_time_update(neon_env_builder: NeonEnvBuilder):
 
     assert eq_safekeeper_records(body, inserted_now)
 
+    # some small tests for the scheduling policy querying and returning APIs
+    newest_info = target.get_safekeeper(inserted["id"])
+    assert newest_info
+    assert newest_info["scheduling_policy"] == "Disabled"
+    target.safekeeper_scheduling_policy(inserted["id"], "Decomissioned")
+    newest_info = target.get_safekeeper(inserted["id"])
+    assert newest_info
+    assert newest_info["scheduling_policy"] == "Decomissioned"
+    # Ensure idempotency
+    target.safekeeper_scheduling_policy(inserted["id"], "Decomissioned")
+
 
 def eq_safekeeper_records(a: dict[str, Any], b: dict[str, Any]) -> bool:
     compared = [dict(a), dict(b)]

test_runner/regress/test_wal_acceptor.py

@@ -48,7 +48,12 @@ from fixtures.remote_storage import (
     default_remote_storage,
     s3_storage,
 )
-from fixtures.safekeeper.http import SafekeeperHttpClient
+from fixtures.safekeeper.http import (
+    Configuration,
+    SafekeeperHttpClient,
+    SafekeeperId,
+    TimelineCreateRequest,
+)
 from fixtures.safekeeper.utils import wait_walreceivers_absent
 from fixtures.utils import (
     PropagatingThread,
@@ -658,7 +663,13 @@ def test_s3_wal_replay(neon_env_builder: NeonEnvBuilder):
     for sk in env.safekeepers:
         sk.start()
         cli = sk.http_client()
-        cli.timeline_create(tenant_id, timeline_id, pg_version, last_lsn)
+        mconf = Configuration(generation=0, members=[], new_members=None)
+        # set start_lsn to the beginning of the first segment to allow reading
+        # WAL from there (could you intidb LSN as well).
+        r = TimelineCreateRequest(
+            tenant_id, timeline_id, mconf, pg_version, Lsn("0/1000000"), commit_lsn=last_lsn
+        )
+        cli.timeline_create(r)
         f_partial_path = (
             Path(sk.data_dir) / str(tenant_id) / str(timeline_id) / f_partial_saved.name
         )
@@ -2237,6 +2248,63 @@ def test_pull_timeline_while_evicted(neon_env_builder: NeonEnvBuilder):
     wait_until(unevicted_on_dest, interval=0.1, timeout=1.0)
 
 
+# Basic test for http API membership related calls: create timeline and switch
+# configuration. Normally these are called by storage controller, but this
+# allows to test them separately.
+@run_only_on_default_postgres("tests only safekeeper API")
+def test_membership_api(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.num_safekeepers = 1
+    env = neon_env_builder.init_start()
+
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
+
+    sk = env.safekeepers[0]
+    http_cli = sk.http_client()
+
+    sk_id_1 = SafekeeperId(env.safekeepers[0].id, "localhost", sk.port.pg_tenant_only)
+    sk_id_2 = SafekeeperId(11, "localhost", 5434)  # just a mock
+
+    # Request to switch before timeline creation should fail.
+    init_conf = Configuration(generation=1, members=[sk_id_1], new_members=None)
+    with pytest.raises(requests.exceptions.HTTPError):
+        http_cli.membership_switch(tenant_id, timeline_id, init_conf)
+
+    # Create timeline.
+    create_r = TimelineCreateRequest(
+        tenant_id, timeline_id, init_conf, 150002, Lsn("0/1000000"), commit_lsn=None
+    )
+    log.info(f"sending {create_r.to_json()}")
+    http_cli.timeline_create(create_r)
+
+    # Switch into some conf.
+    joint_conf = Configuration(generation=4, members=[sk_id_1], new_members=[sk_id_2])
+    resp = http_cli.membership_switch(tenant_id, timeline_id, joint_conf)
+    log.info(f"joint switch resp: {resp}")
+    assert resp.previous_conf.generation == 1
+    assert resp.current_conf.generation == 4
+
+    # Restart sk, conf should be preserved.
+    sk.stop().start()
+    after_restart = http_cli.get_membership(tenant_id, timeline_id)
+    log.info(f"conf after restart: {after_restart}")
+    assert after_restart.generation == 4
+
+    # Switch into disjoint conf.
+    non_joint = Configuration(generation=5, members=[sk_id_2], new_members=None)
+    resp = http_cli.membership_switch(tenant_id, timeline_id, non_joint)
+    log.info(f"non joint switch resp: {resp}")
+    assert resp.previous_conf.generation == 4
+    assert resp.current_conf.generation == 5
+
+    # Switch request to lower conf should be ignored.
+    lower_conf = Configuration(generation=3, members=[], new_members=None)
+    resp = http_cli.membership_switch(tenant_id, timeline_id, lower_conf)
+    log.info(f"lower switch resp: {resp}")
+    assert resp.previous_conf.generation == 5
+    assert resp.current_conf.generation == 5
+
+
 # In this test we check for excessive START_REPLICATION and START_WAL_PUSH queries
 # when compute is active, but there are no writes to the timeline. In that case
 # pageserver should maintain a single connection to safekeeper and don't attempt

vendor/revisions.json

@@ -1,18 +1,18 @@
 {
   "v17": [
     "17.2",
-    "0f8da73ed08d4fc4ee58cccea008c75bfb20baa8"
+    "a8dd6e779dde907778006adb436b557ad652fb97"
   ],
   "v16": [
     "16.6",
-    "f63b141cfb0c813725a6b2574049565bff643018"
+    "d674efd776f59d78e8fa1535bd2f95c3e6984fca"
   ],
   "v15": [
     "15.10",
-    "d3141e17a7155e3d07c8deba4a10c748a29ba1e6"
+    "dd0b28d6fbad39e227f3b77296fcca879af8b3a9"
   ],
   "v14": [
     "14.15",
-    "210a0ba3afd8134ea910b203f274b165bd4f05d7"
+    "46082f20884f087a2d974b33ac65d63af26142bd"
   ]
 }