fix deps

This is a no-op for the neon deployment

* Introduce the concept image consistent lsn: of the largest LSN below
which all pages have been redone successfully
* Use the image consistent LSN for forced image layer creations
* Optionally expose the image consistent LSN via the timeline describe
HTTP endpoint
* Add a sharded timeline describe endpoint to storcon

---------

Co-authored-by: Chen Luo <chen.luo@databricks.com>

Cargo.lock

@@ -1009,6 +1009,12 @@ dependencies = [
  "generic-array",
 ]
 
+[[package]]
+name = "boxcar"
+version = "0.2.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "26c4925bc979b677330a8c7fe7a8c94af2dbb4a2d37b4a20a80d884400f46baa"
+
 [[package]]
 name = "bstr"
 version = "1.5.0"
@@ -1242,14 +1248,14 @@ checksum = "afb84c814227b90d6895e01398aee0d8033c00e7466aca416fb6a8e0eb19d8a7"
 
 [[package]]
 name = "clashmap"
-version = "1.0.0"
+version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "93bd59c81e2bd87a775ae2de75f070f7e2bfe97363a6ad652f46824564c23e4d"
+checksum = "e8a055b1f1bf558eae4959f6dd77cf2d7d50ae1483928e60ef21ca5a24fd4321"
 dependencies = [
  "crossbeam-utils",
  "hashbrown 0.15.2",
  "lock_api",
- "parking_lot_core 0.9.8",
+ "parking_lot_core 0.9.10",
  "polonius-the-crab",
  "replace_with",
 ]
@@ -1753,19 +1759,6 @@ dependencies = [
  "syn 2.0.100",
 ]
 
-[[package]]
-name = "dashmap"
-version = "5.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6943ae99c34386c84a470c499d3414f66502a41340aa895406e0d2e4a207b91d"
-dependencies = [
- "cfg-if",
- "hashbrown 0.14.5",
- "lock_api",
- "once_cell",
- "parking_lot_core 0.9.8",
-]
-
 [[package]]
 name = "dashmap"
 version = "6.1.0"
@@ -1777,7 +1770,7 @@ dependencies = [
  "hashbrown 0.14.5",
  "lock_api",
  "once_cell",
- "parking_lot_core 0.9.8",
+ "parking_lot_core 0.9.10",
 ]
 
 [[package]]
@@ -2343,6 +2336,12 @@ version = "1.0.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1"
 
+[[package]]
+name = "foldhash"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d9c4f5dac5e15c24eb999c26181a6ca40b39fe946cbe4c263c7209467bc83af2"
+
 [[package]]
 name = "form_urlencoded"
 version = "1.2.1"
@@ -2582,7 +2581,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "842dc78579ce01e6a1576ad896edc92fca002dd60c9c3746b7fc2bec6fb429d0"
 dependencies = [
  "cfg-if",
- "dashmap 6.1.0",
+ "dashmap",
  "futures-sink",
  "futures-timer",
  "futures-util",
@@ -3288,7 +3287,7 @@ dependencies = [
  "clap",
  "crossbeam-channel",
  "crossbeam-utils",
- "dashmap 6.1.0",
+ "dashmap",
  "env_logger",
  "indexmap 2.9.0",
  "itoa",
@@ -3545,16 +3544,6 @@ dependencies = [
  "libc",
 ]
 
-[[package]]
-name = "lasso"
-version = "0.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4644821e1c3d7a560fe13d842d13f587c07348a1a05d3a797152d41c90c56df2"
-dependencies = [
- "dashmap 5.5.0",
- "hashbrown 0.13.2",
-]
-
 [[package]]
 name = "lazy_static"
 version = "1.5.0"
@@ -3674,17 +3663,17 @@ checksum = "490cc448043f947bae3cbee9c203358d62dbee0db12107a74be5c30ccfd09771"
 
 [[package]]
 name = "measured"
-version = "0.0.22"
+version = "0.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3051f3a030d55d680cdef6ca50e80abd1182f8da29f2344a7c9cb575721138f0"
+checksum = "d22ae866c28b9c59afaeb488ad6e3bd148570cf5a2bacf6c4845def5b9a03470"
 dependencies = [
  "bytes",
  "crossbeam-utils",
  "hashbrown 0.14.5",
  "itoa",
- "lasso",
  "measured-derive",
  "memchr",
+ "paracord",
  "parking_lot 0.12.1",
  "rustc-hash 1.1.0",
  "ryu",
@@ -3692,9 +3681,9 @@ dependencies = [
 
 [[package]]
 name = "measured-derive"
-version = "0.0.22"
+version = "0.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9e6777fc80a575f9503d908c8b498782a6c3ee88a06cb416dc3941401e43b94"
+checksum = "16d734ed9dbca22d87d56b54d220f254ce921cb5cce97d4a960075af0131d076"
 dependencies = [
  "heck",
  "proc-macro2",
@@ -3704,15 +3693,26 @@ dependencies = [
 
 [[package]]
 name = "measured-process"
-version = "0.0.22"
+version = "0.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7c4b80445aeb08e832d87bf1830049a924cdc1d6b7ef40b6b9b365bff17bf8ec"
+checksum = "f71a318d2b9edcded1e5e0ccf3c9e1e1614217d7f07933631c771daa717743aa"
 dependencies = [
  "libc",
  "measured",
  "procfs",
 ]
 
+[[package]]
+name = "measured-tokio"
+version = "0.0.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5e79a936051c484268e2d71a5dd01219096c62c8ff09afad95f07c14a3b0c580"
+dependencies = [
+ "itoa",
+ "measured",
+ "tokio",
+]
+
 [[package]]
 name = "memchr"
 version = "2.6.4"
@@ -4294,6 +4294,7 @@ dependencies = [
  "humantime-serde",
  "pageserver_api",
  "pageserver_client",
+ "pageserver_client_grpc",
  "pageserver_page_api",
  "rand 0.8.5",
  "reqwest",
@@ -4323,6 +4324,7 @@ dependencies = [
  "pageserver_api",
  "postgres_ffi",
  "remote_storage",
+ "serde",
  "serde_json",
  "svg_fmt",
  "thiserror 1.0.69",
@@ -4499,6 +4501,7 @@ name = "pageserver_client_grpc"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "arc-swap",
  "bytes",
  "compute_api",
  "futures",
@@ -4506,6 +4509,7 @@ dependencies = [
  "pageserver_page_api",
  "tokio",
  "tokio-stream",
+ "tokio-util",
  "tonic 0.13.1",
  "tracing",
  "utils",
@@ -4567,6 +4571,21 @@ dependencies = [
  "seize",
 ]
 
+[[package]]
+name = "paracord"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5e41e924113f9a05eecd561c6e695648f243a03b3ad8d9b3eff689342b95023b"
+dependencies = [
+ "boxcar",
+ "clashmap",
+ "foldhash",
+ "hashbrown 0.15.2",
+ "serde",
+ "sync_wrapper 1.0.1",
+ "typed-arena",
+]
+
 [[package]]
 name = "parking"
 version = "2.1.1"
@@ -4591,7 +4610,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3742b2c103b9f06bc9fff0a37ff4912935851bee6d36f3c02bcc755bcfec228f"
 dependencies = [
  "lock_api",
- "parking_lot_core 0.9.8",
+ "parking_lot_core 0.9.10",
 ]
 
 [[package]]
@@ -4610,15 +4629,15 @@ dependencies = [
 
 [[package]]
 name = "parking_lot_core"
-version = "0.9.8"
+version = "0.9.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "93f00c865fe7cabf650081affecd3871070f26767e7b2070a3ffae14c654b447"
+checksum = "1e401f977ab385c9e4e3ab30627d6f26d00e2c73eef317493c4ec6d468726cf8"
 dependencies = [
  "cfg-if",
  "libc",
- "redox_syscall 0.3.5",
+ "redox_syscall 0.5.10",
  "smallvec",
- "windows-targets 0.48.0",
+ "windows-targets 0.52.6",
 ]
 
 [[package]]
@@ -5324,13 +5343,14 @@ dependencies = [
  "itoa",
  "jose-jwa",
  "jose-jwk",
- "lasso",
  "measured",
+ "measured-tokio",
  "metrics",
  "once_cell",
  "opentelemetry",
  "p256 0.13.2",
  "papaya",
+ "paracord",
  "parking_lot 0.12.1",
  "parquet",
  "parquet_derive",
@@ -5633,6 +5653,15 @@ dependencies = [
  "bitflags 1.3.2",
 ]
 
+[[package]]
+name = "redox_syscall"
+version = "0.5.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b8c0c260b63a8219631167be35e6a988e9554dbd323f8bd08439c8ed1302bd1"
+dependencies = [
+ "bitflags 2.8.0",
+]
+
 [[package]]
 name = "regex"
 version = "1.10.2"
@@ -6893,12 +6922,12 @@ dependencies = [
  "hyper 0.14.30",
  "itertools 0.10.5",
  "json-structural-diff",
- "lasso",
  "measured",
  "metrics",
  "once_cell",
  "pageserver_api",
  "pageserver_client",
+ "paracord",
  "postgres_connection",
  "posthog_client_lite",
  "rand 0.8.5",
@@ -6987,6 +7016,7 @@ dependencies = [
  "pageserver_api",
  "pageserver_client",
  "reqwest",
+ "safekeeper_api",
  "serde_json",
  "storage_controller_client",
  "tokio",
@@ -8012,6 +8042,12 @@ dependencies = [
  "static_assertions",
 ]
 
+[[package]]
+name = "typed-arena"
+version = "2.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6af6ae20167a9ece4bcb41af5b80f8a1f1df981f6391189ce00fd257af04126a"
+
 [[package]]
 name = "typed-json"
 version = "0.1.1"
@@ -8756,6 +8792,7 @@ dependencies = [
  "num-traits",
  "once_cell",
  "p256 0.13.2",
+ "paracord",
  "parquet",
  "prettyplease",
  "proc-macro2",

Cargo.toml

@@ -128,11 +128,11 @@ itertools = "0.10"
 itoa = "1.0.11"
 jemalloc_pprof = { version = "0.7", features = ["symbolize", "flamegraph"] }
 jsonwebtoken = "9"
-lasso = "0.7"
 libc = "0.2"
 md5 = "0.7.0"
-measured = { version = "0.0.22", features=["lasso"] }
-measured-process = { version = "0.0.22" }
+measured = { version = "0.0.23", features = ["paracord"] }
+measured-process = { version = "0.0.23" }
+measured-tokio = { version = "0.0.23" }
 memoffset = "0.9"
 nix = { version = "0.30.1", features = ["dir", "fs", "mman", "process", "socket", "signal", "poll"] }
 # Do not update to >= 7.0.0, at least. The update will have a significant impact
@@ -145,6 +145,7 @@ opentelemetry = "0.27"
 opentelemetry_sdk = "0.27"
 opentelemetry-otlp = { version = "0.27", default-features = false, features = ["http-proto", "trace", "http", "reqwest-client"] }
 opentelemetry-semantic-conventions = "0.27"
+paracord = { version = "0.1.0", features = ["serde"] }
 parking_lot = "0.12"
 parquet = { version = "53", default-features = false, features = ["zstd"] }
 parquet_derive = "53"
@@ -262,6 +263,7 @@ neon-shmem = { version = "0.1", path = "./libs/neon-shmem/" }
 pageserver = { path = "./pageserver" }
 pageserver_api = { version = "0.1", path = "./libs/pageserver_api/" }
 pageserver_client = { path = "./pageserver/client" }
+pageserver_client_grpc = { path = "./pageserver/client_grpc" }
 pageserver_compaction = { version = "0.1", path = "./pageserver/compaction/" }
 pageserver_page_api = { path = "./pageserver/page_api" }
 postgres_backend = { version = "0.1", path = "./libs/postgres_backend/" }

compute_tools/README.md

@@ -46,11 +46,14 @@ stateDiagram-v2
   Configuration --> Failed : Failed to configure the compute
   Configuration --> Running : Compute has been configured
   Empty --> Init : Compute spec is immediately available
-  Empty --> TerminationPending : Requested termination
+  Empty --> TerminationPendingFast : Requested termination
+  Empty --> TerminationPendingImmediate : Requested termination
   Init --> Failed : Failed to start Postgres
   Init --> Running : Started Postgres
-  Running --> TerminationPending : Requested termination
-  TerminationPending --> Terminated : Terminated compute
+  Running --> TerminationPendingFast : Requested termination
+  Running --> TerminationPendingImmediate : Requested termination
+  TerminationPendingFast --> Terminated compute with 30s delay for cplane to inspect status
+  TerminationPendingImmediate --> Terminated : Terminated compute immediately
   Failed --> [*] : Compute exited
   Terminated --> [*] : Compute exited
 ```

compute_tools/src/compute.rs

@@ -956,14 +956,20 @@ impl ComputeNode {
             None
         };
 
-        let mut delay_exit = false;
         let mut state = self.state.lock().unwrap();
         state.terminate_flush_lsn = lsn;
-        if let ComputeStatus::TerminationPending { mode } = state.status {
+
+        let delay_exit = state.status == ComputeStatus::TerminationPendingFast;
+        if state.status == ComputeStatus::TerminationPendingFast
+            || state.status == ComputeStatus::TerminationPendingImmediate
+        {
+            info!(
+                "Changing compute status from {} to {}",
+                state.status,
+                ComputeStatus::Terminated
+            );
             state.status = ComputeStatus::Terminated;
             self.state_changed.notify_all();
-            // we were asked to terminate gracefully, don't exit to avoid restart
-            delay_exit = mode == compute_api::responses::TerminateMode::Fast
         }
         drop(state);
 
@@ -1805,6 +1811,8 @@ impl ComputeNode {
             tls_config,
         )?;
 
+        self.pg_reload_conf()?;
+
         if !spec.skip_pg_catalog_updates {
             let max_concurrent_connections = spec.reconfigure_concurrency;
             // Temporarily reset max_cluster_size in config
@@ -1824,10 +1832,9 @@ impl ComputeNode {
 
                 Ok(())
             })?;
+            self.pg_reload_conf()?;
         }
 
-        self.pg_reload_conf()?;
-
         let unknown_op = "unknown".to_string();
         let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
         info!(
@@ -1900,7 +1907,8 @@ impl ComputeNode {
 
                             // exit loop
                             ComputeStatus::Failed
-                            | ComputeStatus::TerminationPending { .. }
+                            | ComputeStatus::TerminationPendingFast
+                            | ComputeStatus::TerminationPendingImmediate
                             | ComputeStatus::Terminated => break 'cert_update,
 
                             // wait

compute_tools/src/compute_prewarm.rs

@@ -70,7 +70,7 @@ impl ComputeNode {
             }
         };
         let row = match client
-            .query_one("select * from get_prewarm_info()", &[])
+            .query_one("select * from neon.get_prewarm_info()", &[])
             .await
         {
             Ok(row) => row,
@@ -146,7 +146,7 @@ impl ComputeNode {
         ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
             .context("connecting to postgres")?
-            .query_one("select prewarm_local_cache($1)", &[&uncompressed])
+            .query_one("select neon.prewarm_local_cache($1)", &[&uncompressed])
             .await
             .context("loading LFC state into postgres")
             .map(|_| ())
@@ -196,7 +196,7 @@ impl ComputeNode {
         ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
             .context("connecting to postgres")?
-            .query_one("select get_local_cache_state()", &[])
+            .query_one("select neon.get_local_cache_state()", &[])
             .await
             .context("querying LFC state")?
             .try_get::<usize, &[u8]>(0)

compute_tools/src/http/openapi_spec.yaml

@@ -371,9 +371,28 @@ paths:
       summary: Terminate Postgres and wait for it to exit
       description: ""
       operationId: terminate
+      parameters:
+        - name: mode
+          in: query
+          description: "Terminate mode: fast (wait 30s before returning) and immediate"
+          required: false
+          schema:
+            type: string
+            enum: ["fast", "immediate"]
+            default: fast
       responses:
         200:
           description: Result
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/TerminateResponse"
+        201:
+          description: Result if compute is already terminated
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/TerminateResponse"
         412:
           description: "wrong state"
           content:
@@ -530,11 +549,14 @@ components:
       type: string
       enum:
         - empty
-        - init
-        - failed
-        - running
         - configuration_pending
+        - init
+        - running
         - configuration
+        - failed
+        - termination_pending_fast
+        - termination_pending_immediate
+        - terminated
       example: running
 
     ExtensionInstallRequest:
@@ -660,6 +682,17 @@ components:
           description: Role name.
           example: "neon"
 
+    TerminateResponse:
+      type: object
+      required:
+        - lsn
+      properties:
+        lsn:
+          type: string
+          nullable: true
+          description: "last WAL flush LSN"
+          example: "0/028F10D8"
+
     SetRoleGrantsResponse:
       type: object
       required:

compute_tools/src/http/routes/terminate.rs

@@ -3,7 +3,7 @@ use crate::http::JsonResponse;
 use axum::extract::State;
 use axum::response::Response;
 use axum_extra::extract::OptionalQuery;
-use compute_api::responses::{ComputeStatus, TerminateResponse};
+use compute_api::responses::{ComputeStatus, TerminateMode, TerminateResponse};
 use http::StatusCode;
 use serde::Deserialize;
 use std::sync::Arc;
@@ -12,7 +12,7 @@ use tracing::info;
 
 #[derive(Deserialize, Default)]
 pub struct TerminateQuery {
-    mode: compute_api::responses::TerminateMode,
+    mode: TerminateMode,
 }
 
 /// Terminate the compute.
@@ -24,16 +24,16 @@ pub(in crate::http) async fn terminate(
     {
         let mut state = compute.state.lock().unwrap();
         if state.status == ComputeStatus::Terminated {
-            return JsonResponse::success(StatusCode::CREATED, state.terminate_flush_lsn);
+            let response = TerminateResponse {
+                lsn: state.terminate_flush_lsn,
+            };
+            return JsonResponse::success(StatusCode::CREATED, response);
         }
 
         if !matches!(state.status, ComputeStatus::Empty | ComputeStatus::Running) {
             return JsonResponse::invalid_status(state.status);
         }
-        state.set_status(
-            ComputeStatus::TerminationPending { mode },
-            &compute.state_changed,
-        );
+        state.set_status(mode.into(), &compute.state_changed);
     }
 
     forward_termination_signal(false);

compute_tools/src/metrics.rs

@@ -108,7 +108,7 @@ pub(crate) static LFC_PREWARMS: Lazy<IntCounter> = Lazy::new(|| {
 pub(crate) static LFC_PREWARM_ERRORS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "compute_ctl_lfc_prewarm_errors_total",
-        "Total number of LFC prewarms errors requested by compute_ctl or autoprewarm option",
+        "Total number of LFC prewarm errors",
     )
     .expect("failed to define a metric")
 });
@@ -124,7 +124,7 @@ pub(crate) static LFC_OFFLOADS: Lazy<IntCounter> = Lazy::new(|| {
 pub(crate) static LFC_OFFLOAD_ERRORS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "compute_ctl_lfc_offload_errors_total",
-        "Total number of LFC offload errors requested by compute_ctl or lfc_offload_period_seconds option",
+        "Total number of LFC offload errors",
     )
     .expect("failed to define a metric")
 });

compute_tools/src/migrations/0002-alter_roles.sql

@@ -1,3 +1,16 @@
+-- On December 8th, 2023, an engineering escalation (INC-110) was opened after
+-- it was found that BYPASSRLS was being applied to all roles.
+--
+-- PR that introduced the issue: https://github.com/neondatabase/neon/pull/5657
+-- Subsequent commit on main: https://github.com/neondatabase/neon/commit/ad99fa5f0393e2679e5323df653c508ffa0ac072
+--
+-- NOBYPASSRLS and INHERIT are the defaults for a Postgres role, but because it
+-- isn't easy to know if a Postgres cluster is affected by the issue, we need to
+-- keep the migration around for a long time, if not indefinitely, so any
+-- cluster can be fixed.
+--
+-- Branching is the gift that keeps on giving...
+
 DO $$
 DECLARE
     role_name text;

compute_tools/src/migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql

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

compute_tools/src/migrations/tests/0012-grant_pg_signal_backend_to_neon_superuser.sql

@@ -0,0 +1,23 @@
+DO $$
+DECLARE
+    signal_backend record;
+BEGIN
+    SELECT pg_has_role('neon_superuser', 'pg_signal_backend', 'member') AS member,
+            admin_option AS admin
+        INTO signal_backend
+        FROM pg_auth_members
+        WHERE roleid = 'pg_signal_backend'::regrole
+            AND member = 'neon_superuser'::regrole;
+
+    IF signal_backend IS NULL THEN
+        RAISE EXCEPTION 'no entry in pg_auth_members for neon_superuser and pg_signal_backend';
+    END IF;
+
+    IF signal_backend.member IS NULL OR NOT signal_backend.member THEN
+        RAISE EXCEPTION 'neon_superuser is not a member of pg_signal_backend';
+    END IF;
+
+    IF signal_backend.admin IS NULL OR NOT signal_backend.admin THEN
+        RAISE EXCEPTION 'neon_superuser cannot grant pg_signal_backend';
+    END IF;
+END $$;

compute_tools/src/monitor.rs

@@ -84,7 +84,8 @@ impl ComputeMonitor {
         if matches!(
             compute_status,
             ComputeStatus::Terminated
-                | ComputeStatus::TerminationPending { .. }
+                | ComputeStatus::TerminationPendingFast
+                | ComputeStatus::TerminationPendingImmediate
                 | ComputeStatus::Failed
         ) {
             info!(

compute_tools/src/spec.rs

@@ -197,6 +197,7 @@ pub async fn handle_migrations(client: &mut Client) -> Result<()> {
         include_str!(
             "./migrations/0011-grant_pg_show_replication_origin_status_to_neon_superuser.sql"
         ),
+        include_str!("./migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql"),
     ];
 
     MigrationRunner::new(client, &migrations)

control_plane/src/broker.rs

@@ -36,7 +36,7 @@ impl StorageBroker {
     pub async fn start(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
         let broker = &self.env.broker;
 
-        print!("Starting neon broker at {}", broker.client_url());
+        println!("Starting neon broker at {}", broker.client_url());
 
         let mut args = Vec::new();
 

control_plane/src/endpoint.rs

@@ -922,7 +922,8 @@ impl Endpoint {
                         ComputeStatus::Empty
                         | ComputeStatus::ConfigurationPending
                         | ComputeStatus::Configuration
-                        | ComputeStatus::TerminationPending { .. }
+                        | ComputeStatus::TerminationPendingFast
+                        | ComputeStatus::TerminationPendingImmediate
                         | ComputeStatus::Terminated => {
                             bail!("unexpected compute status: {:?}", state.status)
                         }

control_plane/src/pageserver.rs

@@ -303,7 +303,7 @@ impl PageServerNode {
     async fn start_node(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
         // TODO: using a thread here because start_process() is not async but we need to call check_status()
         let datadir = self.repo_path();
-        print!(
+        println!(
             "Starting pageserver node {} at '{}' in {:?}, retrying for {:?}",
             self.conf.id,
             self.pg_connection_config.raw_address(),
@@ -452,6 +452,12 @@ impl PageServerNode {
                 .map(|x| x.parse::<usize>())
                 .transpose()
                 .context("Failed to parse 'image_creation_threshold' as non zero integer")?,
+            // HADRON
+            image_layer_force_creation_period: settings
+                .remove("image_layer_force_creation_period")
+                .map(humantime::parse_duration)
+                .transpose()
+                .context("Failed to parse 'image_layer_force_creation_period' as duration")?,
             image_layer_creation_check_threshold: settings
                 .remove("image_layer_creation_check_threshold")
                 .map(|x| x.parse::<u8>())

control_plane/src/safekeeper.rs

@@ -127,7 +127,7 @@ impl SafekeeperNode {
         extra_opts: &[String],
         retry_timeout: &Duration,
     ) -> anyhow::Result<()> {
-        print!(
+        println!(
             "Starting safekeeper at '{}' in '{}', retrying for {:?}",
             self.pg_connection_config.raw_address(),
             self.datadir_path().display(),

control_plane/src/storage_controller.rs

@@ -660,7 +660,7 @@ impl StorageController {
             ));
         }
 
-        println!("Starting storage controller");
+        println!("Starting storage controller at {scheme}://{host}:{listen_port}");
 
         background_process::start_process(
             COMMAND,

control_plane/storcon_cli/Cargo.toml

@@ -14,6 +14,7 @@ humantime.workspace = true
 pageserver_api.workspace = true
 pageserver_client.workspace = true
 reqwest.workspace = true
+safekeeper_api.workspace=true
 serde_json = { workspace = true, features = ["raw_value"] }
 storage_controller_client.workspace = true
 tokio.workspace = true

control_plane/storcon_cli/src/main.rs

@@ -11,7 +11,7 @@ use pageserver_api::controller_api::{
     PlacementPolicy, SafekeeperDescribeResponse, SafekeeperSchedulingPolicyRequest,
     ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
     SkSchedulingPolicy, TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
-    TenantShardMigrateRequest, TenantShardMigrateResponse,
+    TenantShardMigrateRequest, TenantShardMigrateResponse, TimelineSafekeeperMigrateRequest,
 };
 use pageserver_api::models::{
     EvictionPolicy, EvictionPolicyLayerAccessThreshold, ShardParameters, TenantConfig,
@@ -21,6 +21,7 @@ use pageserver_api::models::{
 use pageserver_api::shard::{ShardStripeSize, TenantShardId};
 use pageserver_client::mgmt_api::{self};
 use reqwest::{Certificate, Method, StatusCode, Url};
+use safekeeper_api::models::TimelineLocateResponse;
 use storage_controller_client::control_api::Client;
 use utils::id::{NodeId, TenantId, TimelineId};
 
@@ -279,6 +280,23 @@ enum Command {
         #[arg(long)]
         concurrency: Option<usize>,
     },
+    /// Locate safekeepers for a timeline from the storcon DB.
+    TimelineLocate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+    },
+    /// Migrate a timeline to a new set of safekeepers
+    TimelineSafekeeperMigrate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+        /// Example: --new-sk-set 1,2,3
+        #[arg(long, required = true, value_delimiter = ',')]
+        new_sk_set: Vec<NodeId>,
+    },
 }
 
 #[derive(Parser)]
@@ -1324,7 +1342,7 @@ async fn main() -> anyhow::Result<()> {
             concurrency,
         } => {
             let mut path = format!(
-                "/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
+                "v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
             );
 
             if let Some(c) = concurrency {
@@ -1335,6 +1353,41 @@ async fn main() -> anyhow::Result<()> {
                 .dispatch::<(), ()>(Method::POST, path, None)
                 .await?;
         }
+        Command::TimelineLocate {
+            tenant_id,
+            timeline_id,
+        } => {
+            let path = format!("debug/v1/tenant/{tenant_id}/timeline/{timeline_id}/locate");
+
+            let resp = storcon_client
+                .dispatch::<(), TimelineLocateResponse>(Method::GET, path, None)
+                .await?;
+
+            let sk_set = resp.sk_set.iter().map(|id| id.0 as i64).collect::<Vec<_>>();
+            let new_sk_set = resp
+                .new_sk_set
+                .as_ref()
+                .map(|ids| ids.iter().map(|id| id.0 as i64).collect::<Vec<_>>());
+
+            println!("generation = {}", resp.generation);
+            println!("sk_set = {sk_set:?}");
+            println!("new_sk_set = {new_sk_set:?}");
+        }
+        Command::TimelineSafekeeperMigrate {
+            tenant_id,
+            timeline_id,
+            new_sk_set,
+        } => {
+            let path = format!("v1/tenant/{tenant_id}/timeline/{timeline_id}/safekeeper_migrate");
+
+            storcon_client
+                .dispatch::<_, ()>(
+                    Method::POST,
+                    path,
+                    Some(TimelineSafekeeperMigrateRequest { new_sk_set }),
+                )
+                .await?;
+        }
     }
 
     Ok(())

deny.toml

@@ -52,6 +52,7 @@ exceptions = [
     # Zlib license has some restrictions if we decide to change sth
     { allow = ["Zlib"], name = "const_format_proc_macros", version = "*" },
     { allow = ["Zlib"], name = "const_format", version = "*" },
+    { allow = ["Zlib"], name = "foldhash", version = "*" },
 ]
 
 [licenses.private]

libs/compute_api/src/responses.rs

@@ -121,6 +121,15 @@ pub enum TerminateMode {
     Immediate,
 }
 
+impl From<TerminateMode> for ComputeStatus {
+    fn from(mode: TerminateMode) -> Self {
+        match mode {
+            TerminateMode::Fast => ComputeStatus::TerminationPendingFast,
+            TerminateMode::Immediate => ComputeStatus::TerminationPendingImmediate,
+        }
+    }
+}
+
 #[derive(Serialize, Clone, Copy, Debug, Deserialize, PartialEq, Eq)]
 #[serde(rename_all = "snake_case")]
 pub enum ComputeStatus {
@@ -141,7 +150,9 @@ pub enum ComputeStatus {
     // control-plane to terminate it.
     Failed,
     // Termination requested
-    TerminationPending { mode: TerminateMode },
+    TerminationPendingFast,
+    // Termination requested, without waiting 30s before returning from /terminate
+    TerminationPendingImmediate,
     // Terminated Postgres
     Terminated,
 }
@@ -160,7 +171,10 @@ impl Display for ComputeStatus {
             ComputeStatus::Running => f.write_str("running"),
             ComputeStatus::Configuration => f.write_str("configuration"),
             ComputeStatus::Failed => f.write_str("failed"),
-            ComputeStatus::TerminationPending { .. } => f.write_str("termination-pending"),
+            ComputeStatus::TerminationPendingFast => f.write_str("termination-pending-fast"),
+            ComputeStatus::TerminationPendingImmediate => {
+                f.write_str("termination-pending-immediate")
+            }
             ComputeStatus::Terminated => f.write_str("terminated"),
         }
     }

libs/http-utils/src/endpoint.rs

@@ -20,6 +20,7 @@ use tokio_stream::wrappers::ReceiverStream;
 use tokio_util::io::ReaderStream;
 use tracing::{Instrument, debug, info, info_span, warn};
 use utils::auth::{AuthError, Claims, SwappableJwtAuth};
+use utils::metrics_collector::{METRICS_COLLECTOR, METRICS_STALE_MILLIS};
 
 use crate::error::{ApiError, api_error_handler, route_error_handler};
 use crate::request::{get_query_param, parse_query_param};
@@ -250,9 +251,28 @@ impl std::io::Write for ChannelWriter {
     }
 }
 
-pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+pub async fn prometheus_metrics_handler(
+    req: Request<Body>,
+    force_metric_collection_on_scrape: bool,
+) -> Result<Response<Body>, ApiError> {
     SERVE_METRICS_COUNT.inc();
 
+    // HADRON
+    let requested_use_latest = parse_query_param(&req, "use_latest")?;
+
+    let use_latest = match requested_use_latest {
+        None => force_metric_collection_on_scrape,
+        Some(true) => true,
+        Some(false) => {
+            if force_metric_collection_on_scrape {
+                // We don't cache in this case
+                true
+            } else {
+                false
+            }
+        }
+    };
+
     let started_at = std::time::Instant::now();
 
     let (tx, rx) = mpsc::channel(1);
@@ -277,12 +297,18 @@ pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<
 
         let _span = span.entered();
 
-        let metrics = metrics::gather();
+        // HADRON
+        let collected = if use_latest {
+            // Skip caching the results if we always force metric collection on scrape.
+            METRICS_COLLECTOR.run_once(!force_metric_collection_on_scrape)
+        } else {
+            METRICS_COLLECTOR.last_collected()
+        };
 
         let gathered_at = std::time::Instant::now();
 
         let res = encoder
-            .encode(&metrics, &mut writer)
+            .encode(&collected.metrics, &mut writer)
             .and_then(|_| writer.flush().map_err(|e| e.into()));
 
         // this instant is not when we finally got the full response sent, sending is done by hyper
@@ -295,6 +321,10 @@ pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<
         let encoded_in = encoded_at - gathered_at - writer.wait_time();
         let total = encoded_at - started_at;
 
+        // HADRON
+        let staleness_ms = (encoded_at - collected.collected_at).as_millis();
+        METRICS_STALE_MILLIS.set(staleness_ms as i64);
+
         match res {
             Ok(()) => {
                 tracing::info!(
@@ -303,6 +333,7 @@ pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<
                     spawning_ms = spawned_in.as_millis(),
                     collection_ms = collected_in.as_millis(),
                     encoding_ms = encoded_in.as_millis(),
+                    stalenss_ms = staleness_ms,
                     "responded /metrics"
                 );
             }

libs/pageserver_api/src/config.rs

@@ -272,6 +272,9 @@ pub struct ConfigToml {
     pub timeline_import_config: TimelineImportConfig,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub basebackup_cache_config: Option<BasebackupCacheConfig>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_layer_generation_large_timeline_threshold: Option<u64>,
+    pub force_metric_collection_on_scrape: bool,
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -561,6 +564,11 @@ pub struct TenantConfigToml {
     pub gc_period: Duration,
     // Delta layer churn threshold to create L1 image layers.
     pub image_creation_threshold: usize,
+    // HADRON
+    // When the timeout is reached, PageServer will (1) force compact any remaining L0 deltas and
+    // (2) create image layers if there are any L1 deltas.
+    #[serde(with = "humantime_serde")]
+    pub image_layer_force_creation_period: Option<Duration>,
     // Determines how much history is retained, to allow
     // branching and read replicas at an older point in time.
     // The unit is time.
@@ -823,6 +831,8 @@ impl Default for ConfigToml {
             },
             basebackup_cache_config: None,
             posthog_config: None,
+            image_layer_generation_large_timeline_threshold: Some(2 * 1024 * 1024 * 1024),
+            force_metric_collection_on_scrape: true,
         }
     }
 }
@@ -916,6 +926,7 @@ impl Default for TenantConfigToml {
             gc_period: humantime::parse_duration(DEFAULT_GC_PERIOD)
                 .expect("cannot parse default gc period"),
             image_creation_threshold: DEFAULT_IMAGE_CREATION_THRESHOLD,
+            image_layer_force_creation_period: None,
             pitr_interval: humantime::parse_duration(DEFAULT_PITR_INTERVAL)
                 .expect("cannot parse default PITR interval"),
             walreceiver_connect_timeout: humantime::parse_duration(

libs/pageserver_api/src/controller_api.rs

@@ -10,7 +10,7 @@ use serde::{Deserialize, Serialize};
 use utils::id::{NodeId, TenantId, TimelineId};
 use utils::lsn::Lsn;
 
-use crate::models::{PageserverUtilization, ShardParameters, TenantConfig};
+use crate::models::{PageserverUtilization, ShardParameters, TenantConfig, TimelineInfo};
 use crate::shard::{ShardStripeSize, TenantShardId};
 
 #[derive(Serialize, Deserialize, Debug)]
@@ -126,6 +126,13 @@ pub struct TenantDescribeResponse {
     pub config: TenantConfig,
 }
 
+#[derive(Serialize, Deserialize, Debug)]
+pub struct TenantTimelineDescribeResponse {
+    pub shards: Vec<TimelineInfo>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_consistent_lsn: Option<Lsn>,
+}
+
 #[derive(Serialize, Deserialize, Debug)]
 pub struct NodeShardResponse {
     pub node_id: NodeId,

libs/pageserver_api/src/models.rs

@@ -597,6 +597,9 @@ pub struct TenantConfigPatch {
     pub gc_period: FieldPatch<String>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub image_creation_threshold: FieldPatch<usize>,
+    // HADRON
+    #[serde(skip_serializing_if = "FieldPatch::is_noop")]
+    pub image_layer_force_creation_period: FieldPatch<String>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub pitr_interval: FieldPatch<String>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
@@ -700,6 +703,11 @@ pub struct TenantConfig {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub image_creation_threshold: Option<usize>,
 
+    // HADRON
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[serde(with = "humantime_serde")]
+    pub image_layer_force_creation_period: Option<Duration>,
+
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(with = "humantime_serde")]
     pub pitr_interval: Option<Duration>,
@@ -798,6 +806,7 @@ impl TenantConfig {
             mut gc_horizon,
             mut gc_period,
             mut image_creation_threshold,
+            mut image_layer_force_creation_period,
             mut pitr_interval,
             mut walreceiver_connect_timeout,
             mut lagging_wal_timeout,
@@ -861,6 +870,11 @@ impl TenantConfig {
         patch
             .image_creation_threshold
             .apply(&mut image_creation_threshold);
+        // HADRON
+        patch
+            .image_layer_force_creation_period
+            .map(|v| humantime::parse_duration(&v))?
+            .apply(&mut image_layer_force_creation_period);
         patch
             .pitr_interval
             .map(|v| humantime::parse_duration(&v))?
@@ -942,6 +956,7 @@ impl TenantConfig {
             gc_horizon,
             gc_period,
             image_creation_threshold,
+            image_layer_force_creation_period,
             pitr_interval,
             walreceiver_connect_timeout,
             lagging_wal_timeout,
@@ -1016,6 +1031,9 @@ impl TenantConfig {
             image_creation_threshold: self
                 .image_creation_threshold
                 .unwrap_or(global_conf.image_creation_threshold),
+            image_layer_force_creation_period: self
+                .image_layer_force_creation_period
+                .or(global_conf.image_layer_force_creation_period),
             pitr_interval: self.pitr_interval.unwrap_or(global_conf.pitr_interval),
             walreceiver_connect_timeout: self
                 .walreceiver_connect_timeout
@@ -1604,6 +1622,9 @@ pub struct TimelineInfo {
 
     /// Whether the timeline is invisible in synthetic size calculations.
     pub is_invisible: Option<bool>,
+    // HADRON: the largest LSN below which all page updates have been included in the image layers.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_consistent_lsn: Option<Lsn>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]

libs/safekeeper_api/src/models.rs

@@ -11,7 +11,7 @@ use utils::id::{NodeId, TenantId, TenantTimelineId, TimelineId};
 use utils::lsn::Lsn;
 use utils::pageserver_feedback::PageserverFeedback;
 
-use crate::membership::Configuration;
+use crate::membership::{Configuration, SafekeeperGeneration};
 use crate::{ServerInfo, Term};
 
 #[derive(Debug, Serialize, Deserialize)]
@@ -311,3 +311,12 @@ pub struct PullTimelineResponse {
     pub safekeeper_host: Option<String>,
     // TODO: add more fields?
 }
+
+/// Response to a timeline locate request.
+/// Storcon-only API.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct TimelineLocateResponse {
+    pub generation: SafekeeperGeneration,
+    pub sk_set: Vec<NodeId>,
+    pub new_sk_set: Option<Vec<NodeId>>,
+}

libs/utils/src/lib.rs

@@ -99,6 +99,8 @@ pub mod elapsed_accum;
 #[cfg(target_os = "linux")]
 pub mod linux_socket_ioctl;
 
+pub mod metrics_collector;
+
 // Re-export used in macro. Avoids adding git-version as dep in target crates.
 #[doc(hidden)]
 pub use git_version;

libs/utils/src/metrics_collector.rs

@@ -0,0 +1,75 @@
+use std::{
+    sync::{Arc, RwLock},
+    time::{Duration, Instant},
+};
+
+use metrics::{IntGauge, proto::MetricFamily, register_int_gauge};
+use once_cell::sync::Lazy;
+
+pub static METRICS_STALE_MILLIS: Lazy<IntGauge> = Lazy::new(|| {
+    register_int_gauge!(
+        "metrics_metrics_stale_milliseconds",
+        "The current metrics stale time in milliseconds"
+    )
+    .expect("failed to define a metric")
+});
+
+#[derive(Debug)]
+pub struct CollectedMetrics {
+    pub metrics: Vec<MetricFamily>,
+    pub collected_at: Instant,
+}
+
+impl CollectedMetrics {
+    fn new(metrics: Vec<MetricFamily>) -> Self {
+        Self {
+            metrics,
+            collected_at: Instant::now(),
+        }
+    }
+}
+
+#[derive(Debug)]
+pub struct MetricsCollector {
+    last_collected: RwLock<Arc<CollectedMetrics>>,
+}
+
+impl MetricsCollector {
+    pub fn new() -> Self {
+        Self {
+            last_collected: RwLock::new(Arc::new(CollectedMetrics::new(vec![]))),
+        }
+    }
+
+    #[tracing::instrument(name = "metrics_collector", skip_all)]
+    pub fn run_once(&self, cache_metrics: bool) -> Arc<CollectedMetrics> {
+        let started = Instant::now();
+        let metrics = metrics::gather();
+        let collected = Arc::new(CollectedMetrics::new(metrics));
+        if cache_metrics {
+            let mut guard = self.last_collected.write().unwrap();
+            *guard = collected.clone();
+        }
+        tracing::info!(
+            "Collected {} metric families in {} ms",
+            collected.metrics.len(),
+            started.elapsed().as_millis()
+        );
+        collected
+    }
+
+    pub fn last_collected(&self) -> Arc<CollectedMetrics> {
+        self.last_collected.read().unwrap().clone()
+    }
+}
+
+impl Default for MetricsCollector {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+// Interval for metrics collection. Currently hard-coded to be the same as the metrics scape interval from the obs agent
+pub static METRICS_COLLECTION_INTERVAL: Duration = Duration::from_secs(30);
+
+pub static METRICS_COLLECTOR: Lazy<MetricsCollector> = Lazy::new(MetricsCollector::default);

libs/walproposer/src/api_bindings.rs

@@ -428,6 +428,12 @@ pub fn empty_shmem() -> crate::bindings::WalproposerShmemState {
         shard_number: 0,
     };
 
+    let empty_wal_rate_limiter = crate::bindings::WalRateLimiter {
+        should_limit: crate::bindings::pg_atomic_uint32 { value: 0 },
+        sent_bytes: 0,
+        last_recorded_time_us: 0,
+    };
+
     crate::bindings::WalproposerShmemState {
         propEpochStartLsn: crate::bindings::pg_atomic_uint64 { value: 0 },
         donor_name: [0; 64],
@@ -441,6 +447,7 @@ pub fn empty_shmem() -> crate::bindings::WalproposerShmemState {
         num_shards: 0,
         replica_promote: false,
         min_ps_feedback: empty_feedback,
+        wal_rate_limiter: empty_wal_rate_limiter,
     }
 }
 

pageserver/client/src/mgmt_api.rs

@@ -1,4 +1,4 @@
-use std::collections::HashMap;
+use std::collections::{BTreeMap, HashMap};
 use std::error::Error as _;
 use std::time::Duration;
 
@@ -251,6 +251,70 @@ impl Client {
         Ok(())
     }
 
+    pub async fn tenant_timeline_compact(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        force_image_layer_creation: bool,
+        must_force_image_layer_creation: bool,
+        scheduled: bool,
+        wait_until_done: bool,
+    ) -> Result<()> {
+        let mut path = reqwest::Url::parse(&format!(
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/compact",
+            self.mgmt_api_endpoint
+        ))
+        .expect("Cannot build URL");
+
+        if force_image_layer_creation {
+            path.query_pairs_mut()
+                .append_pair("force_image_layer_creation", "true");
+        }
+
+        if must_force_image_layer_creation {
+            path.query_pairs_mut()
+                .append_pair("must_force_image_layer_creation", "true");
+        }
+
+        if scheduled {
+            path.query_pairs_mut().append_pair("scheduled", "true");
+        }
+        if wait_until_done {
+            path.query_pairs_mut()
+                .append_pair("wait_until_scheduled_compaction_done", "true");
+            path.query_pairs_mut()
+                .append_pair("wait_until_uploaded", "true");
+        }
+        self.request(Method::PUT, path, ()).await?;
+        Ok(())
+    }
+
+    /* BEGIN_HADRON */
+    pub async fn tenant_timeline_describe(
+        &self,
+        tenant_shard_id: &TenantShardId,
+        timeline_id: &TimelineId,
+    ) -> Result<TimelineInfo> {
+        let mut path = reqwest::Url::parse(&format!(
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}",
+            self.mgmt_api_endpoint
+        ))
+        .expect("Cannot build URL");
+        path.query_pairs_mut()
+            .append_pair("include-image-consistent-lsn", "true");
+
+        let response: reqwest::Response = self.request(Method::GET, path, ()).await?;
+        let body = response.json().await.map_err(Error::ReceiveBody)?;
+        Ok(body)
+    }
+
+    pub async fn list_tenant_visible_size(&self) -> Result<BTreeMap<TenantShardId, u64>> {
+        let uri = format!("{}/v1/list_tenant_visible_size", self.mgmt_api_endpoint);
+        let resp = self.get(&uri).await?;
+        resp.json().await.map_err(Error::ReceiveBody)
+    }
+    /* END_HADRON */
+
     pub async fn tenant_scan_remote_storage(
         &self,
         tenant_id: TenantId,

pageserver/client_grpc/Cargo.toml

@@ -4,8 +4,12 @@ version = "0.1.0"
 edition.workspace = true
 license.workspace = true
 
+[features]
+testing = ["pageserver_api/testing"]
+
 [dependencies]
 anyhow.workspace = true
+arc-swap.workspace = true
 bytes.workspace = true
 compute_api.workspace = true
 futures.workspace = true
@@ -13,6 +17,7 @@ pageserver_api.workspace = true
 pageserver_page_api.workspace = true
 tokio.workspace = true
 tokio-stream.workspace = true
+tokio-util.workspace = true
 tonic.workspace = true
 tracing.workspace = true
 utils.workspace = true

pageserver/client_grpc/src/client.rs

@@ -3,8 +3,10 @@ use std::num::NonZero;
 use std::sync::Arc;
 
 use anyhow::anyhow;
+use arc_swap::ArcSwap;
 use futures::stream::FuturesUnordered;
 use futures::{FutureExt as _, StreamExt as _};
+use tonic::codec::CompressionEncoding;
 use tracing::instrument;
 
 use crate::pool::{ChannelPool, ClientGuard, ClientPool, StreamGuard, StreamPool};
@@ -55,28 +57,85 @@ const MAX_BULK_STREAM_QUEUE_DEPTH: NonZero<usize> = NonZero::new(4).unwrap();
 /// TODO: this client does not support base backups or LSN leases, as these are only used by
 /// compute_ctl. Consider adding this, but LSN leases need concurrent requests on all shards.
 pub struct PageserverClient {
-    // TODO: support swapping out the shard map, e.g. via an ArcSwap.
-    shards: Shards,
+    /// The tenant ID.
+    tenant_id: TenantId,
+    /// The timeline ID.
+    timeline_id: TimelineId,
+    /// The JWT auth token for this tenant, if any.
+    auth_token: Option<String>,
+    /// The compression to use, if any.
+    compression: Option<CompressionEncoding>,
+    /// The shards for this tenant.
+    shards: ArcSwap<Shards>,
+    /// The retry configuration.
     retry: Retry,
 }
 
 impl PageserverClient {
     /// Creates a new Pageserver client for a given tenant and timeline. Uses the Pageservers given
-    /// in the shard map, which must be complete and must use gRPC URLs.
+    /// in the shard spec, which must be complete and must use gRPC URLs.
     pub fn new(
         tenant_id: TenantId,
         timeline_id: TimelineId,
-        shard_map: HashMap<ShardIndex, String>,
-        stripe_size: ShardStripeSize,
+        shard_spec: ShardSpec,
         auth_token: Option<String>,
+        compression: Option<CompressionEncoding>,
     ) -> anyhow::Result<Self> {
-        let shards = Shards::new(tenant_id, timeline_id, shard_map, stripe_size, auth_token)?;
+        let shards = Shards::new(
+            tenant_id,
+            timeline_id,
+            shard_spec,
+            auth_token.clone(),
+            compression,
+        )?;
         Ok(Self {
-            shards,
+            tenant_id,
+            timeline_id,
+            auth_token,
+            compression,
+            shards: ArcSwap::new(Arc::new(shards)),
             retry: Retry,
         })
     }
 
+    /// Updates the shards from the given shard spec. In-flight requests will complete using the
+    /// existing shards, but may retry with the new shards if they fail.
+    ///
+    /// TODO: verify that in-flight requests are allowed to complete, and that the old pools are
+    /// properly spun down and dropped afterwards.
+    pub fn update_shards(&self, shard_spec: ShardSpec) -> anyhow::Result<()> {
+        // Validate the shard spec. We should really use `ArcSwap::rcu` for this, to avoid races
+        // with concurrent updates, but that involves creating a new `Shards` on every attempt,
+        // which spins up a bunch of Tokio tasks and such. These should already be checked elsewhere
+        // in the stack, and if they're violated then we already have problems elsewhere, so a
+        // best-effort but possibly-racy check is okay here.
+        let old = self.shards.load_full();
+        if shard_spec.count < old.count {
+            return Err(anyhow!(
+                "can't reduce shard count from {} to {}",
+                old.count,
+                shard_spec.count
+            ));
+        }
+        if !old.count.is_unsharded() && shard_spec.stripe_size != old.stripe_size {
+            return Err(anyhow!(
+                "can't change stripe size from {} to {}",
+                old.stripe_size,
+                shard_spec.stripe_size
+            ));
+        }
+
+        let shards = Shards::new(
+            self.tenant_id,
+            self.timeline_id,
+            shard_spec,
+            self.auth_token.clone(),
+            self.compression,
+        )?;
+        self.shards.store(Arc::new(shards));
+        Ok(())
+    }
+
     /// Returns whether a relation exists.
     #[instrument(skip_all, fields(rel=%req.rel, lsn=%req.read_lsn))]
     pub async fn check_rel_exists(
@@ -84,9 +143,9 @@ impl PageserverClient {
         req: page_api::CheckRelExistsRequest,
     ) -> tonic::Result<page_api::CheckRelExistsResponse> {
         self.retry
-            .with(async || {
+            .with(async |_| {
                 // Relation metadata is only available on shard 0.
-                let mut client = self.shards.get_zero().client().await?;
+                let mut client = self.shards.load_full().get_zero().client().await?;
                 client.check_rel_exists(req).await
             })
             .await
@@ -99,16 +158,17 @@ impl PageserverClient {
         req: page_api::GetDbSizeRequest,
     ) -> tonic::Result<page_api::GetDbSizeResponse> {
         self.retry
-            .with(async || {
+            .with(async |_| {
                 // Relation metadata is only available on shard 0.
-                let mut client = self.shards.get_zero().client().await?;
+                let mut client = self.shards.load_full().get_zero().client().await?;
                 client.get_db_size(req).await
             })
             .await
     }
 
-    /// Fetches pages. The `request_id` must be unique across all in-flight requests. Automatically
-    /// splits requests that straddle shard boundaries, and assembles the responses.
+    /// Fetches pages. The `request_id` must be unique across all in-flight requests, and the
+    /// `attempt` must be 0 (incremented on retry). Automatically splits requests that straddle
+    /// shard boundaries, and assembles the responses.
     ///
     /// Unlike `page_api::Client`, this automatically converts `status_code` into `tonic::Status`
     /// errors. All responses will have `GetPageStatusCode::Ok`.
@@ -128,72 +188,96 @@ impl PageserverClient {
         if req.block_numbers.is_empty() {
             return Err(tonic::Status::invalid_argument("no block number"));
         }
+        // The request attempt must be 0. The client will increment it internally.
+        if req.request_id.attempt != 0 {
+            return Err(tonic::Status::invalid_argument("request attempt must be 0"));
+        }
 
+        // The shards may change while we're fetching pages. We execute the request using a stable
+        // view of the shards (especially important for requests that span shards), but retry the
+        // top-level (pre-split) request to pick up shard changes. This can lead to unnecessary
+        // retries and re-splits in some cases where requests span shards, but these are expected to
+        // be rare.
+        //
+        // TODO: the gRPC server and client doesn't yet properly support shard splits. Revisit this
+        // once we figure out how to handle these.
+        self.retry
+            .with(async |attempt| {
+                let mut req = req.clone();
+                req.request_id.attempt = attempt as u32;
+                Self::get_page_with_shards(req, &self.shards.load_full()).await
+            })
+            .await
+    }
+
+    /// Fetches pages using the given shards. This uses a stable view of the shards, regardless of
+    /// concurrent shard updates. Does not retry internally, but is retried by `get_page()`.
+    async fn get_page_with_shards(
+        req: page_api::GetPageRequest,
+        shards: &Shards,
+    ) -> tonic::Result<page_api::GetPageResponse> {
         // Fast path: request is for a single shard.
         if let Some(shard_id) =
-            GetPageSplitter::is_single_shard(&req, self.shards.count, self.shards.stripe_size)
+            GetPageSplitter::for_single_shard(&req, shards.count, shards.stripe_size)
         {
-            return self.get_page_for_shard(shard_id, req).await;
+            return Self::get_page_with_shard(req, shards.get(shard_id)?).await;
         }
 
         // Request spans multiple shards. Split it, dispatch concurrent per-shard requests, and
         // reassemble the responses.
-        //
-        // TODO: when we support shard map updates, we need to detect when it changes and re-split
-        // the request on errors.
-        let mut splitter = GetPageSplitter::split(req, self.shards.count, self.shards.stripe_size);
+        let mut splitter = GetPageSplitter::split(req, shards.count, shards.stripe_size);
 
-        let mut shard_requests: FuturesUnordered<_> = splitter
-            .drain_requests()
-            .map(|(shard_id, shard_req)| {
-                // NB: each request will retry internally.
-                self.get_page_for_shard(shard_id, shard_req)
-                    .map(move |result| result.map(|resp| (shard_id, resp)))
-            })
-            .collect();
+        let mut shard_requests = FuturesUnordered::new();
+        for (shard_id, shard_req) in splitter.drain_requests() {
+            let future = Self::get_page_with_shard(shard_req, shards.get(shard_id)?)
+                .map(move |result| result.map(|resp| (shard_id, resp)));
+            shard_requests.push(future);
+        }
 
         while let Some((shard_id, shard_response)) = shard_requests.next().await.transpose()? {
             splitter.add_response(shard_id, shard_response)?;
         }
 
-        splitter.assemble_response()
+        splitter.get_response()
     }
 
-    /// Fetches pages that belong to the given shard.
-    #[instrument(skip_all, fields(shard = %shard_id))]
-    async fn get_page_for_shard(
-        &self,
-        shard_id: ShardIndex,
+    /// Fetches pages on the given shard. Does not retry internally.
+    async fn get_page_with_shard(
         req: page_api::GetPageRequest,
+        shard: &Shard,
     ) -> tonic::Result<page_api::GetPageResponse> {
-        let resp = self
-            .retry
-            .with(async || {
-                let stream = self
-                    .shards
-                    .get(shard_id)?
-                    .stream(req.request_class.is_bulk())
-                    .await;
-                let resp = stream.send(req.clone()).await?;
+        let stream = shard.stream(req.request_class.is_bulk()).await;
+        let resp = stream.send(req.clone()).await?;
 
-                // Convert per-request errors into a tonic::Status.
-                if resp.status_code != page_api::GetPageStatusCode::Ok {
-                    return Err(tonic::Status::new(
-                        resp.status_code.into(),
-                        resp.reason.unwrap_or_else(|| String::from("unknown error")),
-                    ));
-                }
+        // Convert per-request errors into a tonic::Status.
+        if resp.status_code != page_api::GetPageStatusCode::Ok {
+            return Err(tonic::Status::new(
+                resp.status_code.into(),
+                resp.reason.unwrap_or_else(|| String::from("unknown error")),
+            ));
+        }
 
-                Ok(resp)
-            })
-            .await?;
-
-        // Make sure we got the right number of pages.
-        // NB: check outside of the retry loop, since we don't want to retry this.
-        let (expected, actual) = (req.block_numbers.len(), resp.page_images.len());
-        if expected != actual {
+        // Check that we received the expected pages.
+        if req.rel != resp.rel {
             return Err(tonic::Status::internal(format!(
-                "expected {expected} pages for shard {shard_id}, got {actual}",
+                "shard {} returned wrong relation, expected {} got {}",
+                shard.id, req.rel, resp.rel
+            )));
+        }
+        if !req
+            .block_numbers
+            .iter()
+            .copied()
+            .eq(resp.pages.iter().map(|p| p.block_number))
+        {
+            return Err(tonic::Status::internal(format!(
+                "shard {} returned wrong pages, expected {:?} got {:?}",
+                shard.id,
+                req.block_numbers,
+                resp.pages
+                    .iter()
+                    .map(|page| page.block_number)
+                    .collect::<Vec<_>>()
             )));
         }
 
@@ -207,9 +291,9 @@ impl PageserverClient {
         req: page_api::GetRelSizeRequest,
     ) -> tonic::Result<page_api::GetRelSizeResponse> {
         self.retry
-            .with(async || {
+            .with(async |_| {
                 // Relation metadata is only available on shard 0.
-                let mut client = self.shards.get_zero().client().await?;
+                let mut client = self.shards.load_full().get_zero().client().await?;
                 client.get_rel_size(req).await
             })
             .await
@@ -222,50 +306,53 @@ impl PageserverClient {
         req: page_api::GetSlruSegmentRequest,
     ) -> tonic::Result<page_api::GetSlruSegmentResponse> {
         self.retry
-            .with(async || {
+            .with(async |_| {
                 // SLRU segments are only available on shard 0.
-                let mut client = self.shards.get_zero().client().await?;
+                let mut client = self.shards.load_full().get_zero().client().await?;
                 client.get_slru_segment(req).await
             })
             .await
     }
 }
 
-/// Tracks the tenant's shards.
-struct Shards {
+/// Shard specification for a PageserverClient.
+pub struct ShardSpec {
+    /// Maps shard indices to gRPC URLs.
+    ///
+    /// INVARIANT: every shard 0..count is present, and shard 0 is always present.
+    /// INVARIANT: every URL is valid and uses grpc:// scheme.
+    urls: HashMap<ShardIndex, String>,
     /// The shard count.
     ///
     /// NB: this is 0 for unsharded tenants, following `ShardIndex::unsharded()` convention.
     count: ShardCount,
-    /// The stripe size. Only used for sharded tenants.
+    /// The stripe size for these shards.
     stripe_size: ShardStripeSize,
-    /// Shards by shard index.
-    ///
-    /// NB: unsharded tenants use count 0, like `ShardIndex::unsharded()`.
-    ///
-    /// INVARIANT: every shard 0..count is present.
-    /// INVARIANT: shard 0 is always present.
-    map: HashMap<ShardIndex, Shard>,
 }
 
-impl Shards {
-    /// Creates a new set of shards based on a shard map.
-    fn new(
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-        shard_map: HashMap<ShardIndex, String>,
-        stripe_size: ShardStripeSize,
-        auth_token: Option<String>,
+impl ShardSpec {
+    /// Creates a new shard spec with the given URLs and stripe size. All shards must be given.
+    /// The stripe size may be omitted for unsharded tenants.
+    pub fn new(
+        urls: HashMap<ShardIndex, String>,
+        stripe_size: Option<ShardStripeSize>,
     ) -> anyhow::Result<Self> {
-        let count = match shard_map.len() {
+        // Compute the shard count.
+        let count = match urls.len() {
             0 => return Err(anyhow!("no shards provided")),
             1 => ShardCount::new(0), // NB: unsharded tenants use 0, like `ShardIndex::unsharded()`
             n if n > u8::MAX as usize => return Err(anyhow!("too many shards: {n}")),
             n => ShardCount::new(n as u8),
         };
 
-        let mut map = HashMap::new();
-        for (shard_id, url) in shard_map {
+        // Determine the stripe size. It doesn't matter for unsharded tenants.
+        if stripe_size.is_none() && !count.is_unsharded() {
+            return Err(anyhow!("stripe size must be given for sharded tenants"));
+        }
+        let stripe_size = stripe_size.unwrap_or_default();
+
+        // Validate the shard spec.
+        for (shard_id, url) in &urls {
             // The shard index must match the computed shard count, even for unsharded tenants.
             if shard_id.shard_count != count {
                 return Err(anyhow!("invalid shard index {shard_id}, expected {count}"));
@@ -276,21 +363,72 @@ impl Shards {
             }
             // The above conditions guarantee that we have all shards 0..count: len() matches count,
             // shard number < count, and numbers are unique (via hashmap).
-            let shard = Shard::new(url, tenant_id, timeline_id, shard_id, auth_token.clone())?;
-            map.insert(shard_id, shard);
+
+            // Validate the URL.
+            if PageserverProtocol::from_connstring(url)? != PageserverProtocol::Grpc {
+                return Err(anyhow!("invalid shard URL {url}: must use gRPC"));
+            }
         }
 
         Ok(Self {
+            urls,
             count,
             stripe_size,
-            map,
+        })
+    }
+}
+
+/// Tracks the tenant's shards.
+struct Shards {
+    /// Shards by shard index.
+    ///
+    /// INVARIANT: every shard 0..count is present.
+    /// INVARIANT: shard 0 is always present.
+    by_index: HashMap<ShardIndex, Shard>,
+    /// The shard count.
+    ///
+    /// NB: this is 0 for unsharded tenants, following `ShardIndex::unsharded()` convention.
+    count: ShardCount,
+    /// The stripe size. Only used for sharded tenants.
+    stripe_size: ShardStripeSize,
+}
+
+impl Shards {
+    /// Creates a new set of shards based on a shard spec.
+    fn new(
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        shard_spec: ShardSpec,
+        auth_token: Option<String>,
+        compression: Option<CompressionEncoding>,
+    ) -> anyhow::Result<Self> {
+        // NB: the shard spec has already been validated when constructed.
+        let mut shards = HashMap::with_capacity(shard_spec.urls.len());
+        for (shard_id, url) in shard_spec.urls {
+            shards.insert(
+                shard_id,
+                Shard::new(
+                    url,
+                    tenant_id,
+                    timeline_id,
+                    shard_id,
+                    auth_token.clone(),
+                    compression,
+                )?,
+            );
+        }
+
+        Ok(Self {
+            by_index: shards,
+            count: shard_spec.count,
+            stripe_size: shard_spec.stripe_size,
         })
     }
 
     /// Looks up the given shard.
     #[allow(clippy::result_large_err)] // TODO: check perf impact
     fn get(&self, shard_id: ShardIndex) -> tonic::Result<&Shard> {
-        self.map
+        self.by_index
             .get(&shard_id)
             .ok_or_else(|| tonic::Status::not_found(format!("unknown shard {shard_id}")))
     }
@@ -312,6 +450,8 @@ impl Shards {
 ///   * Bulk client pool: unbounded.
 ///     * Bulk stream pool: MAX_BULK_STREAMS and MAX_BULK_STREAM_QUEUE_DEPTH.
 struct Shard {
+    /// The shard ID.
+    id: ShardIndex,
     /// Unary gRPC client pool.
     client_pool: Arc<ClientPool>,
     /// GetPage stream pool.
@@ -328,12 +468,8 @@ impl Shard {
         timeline_id: TimelineId,
         shard_id: ShardIndex,
         auth_token: Option<String>,
+        compression: Option<CompressionEncoding>,
     ) -> anyhow::Result<Self> {
-        // Sanity-check that the URL uses gRPC.
-        if PageserverProtocol::from_connstring(&url)? != PageserverProtocol::Grpc {
-            return Err(anyhow!("invalid shard URL {url}: must use gRPC"));
-        }
-
         // Common channel pool for unary and stream requests. Bounded by client/stream pools.
         let channel_pool = ChannelPool::new(url.clone(), MAX_CLIENTS_PER_CHANNEL)?;
 
@@ -344,6 +480,7 @@ impl Shard {
             timeline_id,
             shard_id,
             auth_token.clone(),
+            compression,
             Some(MAX_UNARY_CLIENTS),
         );
 
@@ -356,6 +493,7 @@ impl Shard {
                 timeline_id,
                 shard_id,
                 auth_token.clone(),
+                compression,
                 None, // unbounded, limited by stream pool
             ),
             Some(MAX_STREAMS),
@@ -371,6 +509,7 @@ impl Shard {
                 timeline_id,
                 shard_id,
                 auth_token,
+                compression,
                 None, // unbounded, limited by stream pool
             ),
             Some(MAX_BULK_STREAMS),
@@ -378,6 +517,7 @@ impl Shard {
         );
 
         Ok(Self {
+            id: shard_id,
             client_pool,
             stream_pool,
             bulk_stream_pool,

pageserver/client_grpc/src/lib.rs

@@ -3,4 +3,4 @@ mod pool;
 mod retry;
 mod split;
 
-pub use client::PageserverClient;
+pub use client::{PageserverClient, ShardSpec};

pageserver/client_grpc/src/pool.rs

@@ -34,10 +34,13 @@ use std::num::NonZero;
 use std::ops::{Deref, DerefMut};
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex, Weak};
+use std::time::{Duration, Instant};
 
 use futures::StreamExt as _;
 use tokio::sync::mpsc::{Receiver, Sender};
 use tokio::sync::{OwnedSemaphorePermit, Semaphore, mpsc, oneshot};
+use tokio_util::sync::CancellationToken;
+use tonic::codec::CompressionEncoding;
 use tonic::transport::{Channel, Endpoint};
 use tracing::{error, warn};
 
@@ -45,6 +48,25 @@ use pageserver_page_api as page_api;
 use utils::id::{TenantId, TimelineId};
 use utils::shard::ShardIndex;
 
+/// Reap channels/clients/streams that have been idle for this long.
+///
+/// TODO: this is per-pool. For nested pools, it can take up to 3x as long for a TCP connection to
+/// be reaped. First, we must wait for an idle stream to be reaped, which marks its client as idle.
+/// Then, we must wait for the idle client to be reaped, which marks its channel as idle. Then, we
+/// must wait for the idle channel to be reaped. Is that a problem? Maybe not, we just have to
+/// account for it when setting the reap threshold. Alternatively, we can immediately reap empty
+/// channels, and/or stream pool clients.
+const REAP_IDLE_THRESHOLD: Duration = match cfg!(any(test, feature = "testing")) {
+    false => Duration::from_secs(180),
+    true => Duration::from_secs(1), // exercise reaping in tests
+};
+
+/// Reap idle resources with this interval.
+const REAP_IDLE_INTERVAL: Duration = match cfg!(any(test, feature = "testing")) {
+    false => Duration::from_secs(10),
+    true => Duration::from_secs(1), // exercise reaping in tests
+};
+
 /// A gRPC channel pool, for a single Pageserver. A channel is shared by many clients (via HTTP/2
 /// stream multiplexing), up to `clients_per_channel` -- a new channel will be spun up beyond this.
 /// The pool does not limit the number of channels, and instead relies on `ClientPool` or
@@ -52,7 +74,6 @@ use utils::shard::ShardIndex;
 ///
 /// The pool is always wrapped in an outer `Arc`, to allow long-lived guards across tasks/threads.
 ///
-/// TODO: reap idle channels.
 /// TODO: consider prewarming a set of channels, to avoid initial connection latency.
 /// TODO: consider adding a circuit breaker for errors and fail fast.
 pub struct ChannelPool {
@@ -62,6 +83,8 @@ pub struct ChannelPool {
     max_clients_per_channel: NonZero<usize>,
     /// Open channels.
     channels: Mutex<BTreeMap<ChannelID, ChannelEntry>>,
+    /// Reaps idle channels.
+    idle_reaper: Reaper,
     /// Channel ID generator.
     next_channel_id: AtomicUsize,
 }
@@ -73,6 +96,9 @@ struct ChannelEntry {
     channel: Channel,
     /// Number of clients using this channel.
     clients: usize,
+    /// The channel has been idle (no clients) since this time. None if channel is in use.
+    /// INVARIANT: Some if clients == 0, otherwise None.
+    idle_since: Option<Instant>,
 }
 
 impl ChannelPool {
@@ -82,12 +108,15 @@ impl ChannelPool {
         E: TryInto<Endpoint> + Send + Sync + 'static,
         <E as TryInto<Endpoint>>::Error: std::error::Error + Send + Sync,
     {
-        Ok(Arc::new(Self {
+        let pool = Arc::new(Self {
             endpoint: endpoint.try_into()?,
             max_clients_per_channel,
             channels: Mutex::default(),
+            idle_reaper: Reaper::new(REAP_IDLE_THRESHOLD, REAP_IDLE_INTERVAL),
             next_channel_id: AtomicUsize::default(),
-        }))
+        });
+        pool.idle_reaper.spawn(&pool);
+        Ok(pool)
     }
 
     /// Acquires a gRPC channel for a client. Multiple clients may acquire the same channel.
@@ -116,8 +145,14 @@ impl ChannelPool {
                 entry.clients <= self.max_clients_per_channel.get(),
                 "channel overflow"
             );
+            assert_eq!(
+                entry.idle_since.is_some(),
+                entry.clients == 0,
+                "incorrect channel idle state"
+            );
             if entry.clients < self.max_clients_per_channel.get() {
                 entry.clients += 1;
+                entry.idle_since = None;
                 return ChannelGuard {
                     pool: Arc::downgrade(self),
                     id,
@@ -134,6 +169,7 @@ impl ChannelPool {
         let entry = ChannelEntry {
             channel: channel.clone(),
             clients: 1, // account for the guard below
+            idle_since: None,
         };
         channels.insert(id, entry);
 
@@ -145,6 +181,20 @@ impl ChannelPool {
     }
 }
 
+impl Reapable for ChannelPool {
+    /// Reaps channels that have been idle since before the cutoff.
+    fn reap_idle(&self, cutoff: Instant) {
+        self.channels.lock().unwrap().retain(|_, entry| {
+            let Some(idle_since) = entry.idle_since else {
+                assert_ne!(entry.clients, 0, "empty channel not marked idle");
+                return true;
+            };
+            assert_eq!(entry.clients, 0, "idle channel has clients");
+            idle_since >= cutoff
+        })
+    }
+}
+
 /// Tracks a channel acquired from the pool. The owned inner channel can be obtained with `take()`,
 /// since the gRPC client requires an owned `Channel`.
 pub struct ChannelGuard {
@@ -167,10 +217,15 @@ impl Drop for ChannelGuard {
         let Some(pool) = self.pool.upgrade() else {
             return; // pool was dropped
         };
+
         let mut channels = pool.channels.lock().unwrap();
         let entry = channels.get_mut(&self.id).expect("unknown channel");
+        assert!(entry.idle_since.is_none(), "active channel marked idle");
         assert!(entry.clients > 0, "channel underflow");
         entry.clients -= 1;
+        if entry.clients == 0 {
+            entry.idle_since = Some(Instant::now()); // mark channel as idle
+        }
     }
 }
 
@@ -179,8 +234,6 @@ impl Drop for ChannelGuard {
 /// number of concurrent clients to `max_clients` via semaphore.
 ///
 /// The pool is always wrapped in an outer `Arc`, to allow long-lived guards across tasks/threads.
-///
-/// TODO: reap idle clients.
 pub struct ClientPool {
     /// Tenant ID.
     tenant_id: TenantId,
@@ -190,6 +243,8 @@ pub struct ClientPool {
     shard_id: ShardIndex,
     /// Authentication token, if any.
     auth_token: Option<String>,
+    /// Compression to use.
+    compression: Option<CompressionEncoding>,
     /// Channel pool to acquire channels from.
     channel_pool: Arc<ChannelPool>,
     /// Limits the max number of concurrent clients for this pool. None if the pool is unbounded.
@@ -201,6 +256,8 @@ pub struct ClientPool {
     /// lower-ordered channels. This allows us to free up and reap higher-numbered channels as idle
     /// clients are reaped.
     idle: Mutex<BTreeMap<ClientID, ClientEntry>>,
+    /// Reaps idle clients.
+    idle_reaper: Reaper,
     /// Unique client ID generator.
     next_client_id: AtomicUsize,
 }
@@ -212,6 +269,9 @@ struct ClientEntry {
     client: page_api::Client,
     /// The channel guard for the channel used by the client.
     channel_guard: ChannelGuard,
+    /// The client has been idle since this time. All clients in `ClientPool::idle` are idle by
+    /// definition, so this is the time when it was added back to the pool.
+    idle_since: Instant,
 }
 
 impl ClientPool {
@@ -224,18 +284,23 @@ impl ClientPool {
         timeline_id: TimelineId,
         shard_id: ShardIndex,
         auth_token: Option<String>,
+        compression: Option<CompressionEncoding>,
         max_clients: Option<NonZero<usize>>,
     ) -> Arc<Self> {
-        Arc::new(Self {
+        let pool = Arc::new(Self {
             tenant_id,
             timeline_id,
             shard_id,
             auth_token,
+            compression,
             channel_pool,
             idle: Mutex::default(),
+            idle_reaper: Reaper::new(REAP_IDLE_THRESHOLD, REAP_IDLE_INTERVAL),
             limiter: max_clients.map(|max| Arc::new(Semaphore::new(max.get()))),
             next_client_id: AtomicUsize::default(),
-        })
+        });
+        pool.idle_reaper.spawn(&pool);
+        pool
     }
 
     /// Gets a client from the pool, or creates a new one if necessary. Connections are established
@@ -271,7 +336,7 @@ impl ClientPool {
             self.timeline_id,
             self.shard_id,
             self.auth_token.clone(),
-            None,
+            self.compression,
         )?;
 
         Ok(ClientGuard {
@@ -287,6 +352,16 @@ impl ClientPool {
     }
 }
 
+impl Reapable for ClientPool {
+    /// Reaps clients that have been idle since before the cutoff.
+    fn reap_idle(&self, cutoff: Instant) {
+        self.idle
+            .lock()
+            .unwrap()
+            .retain(|_, entry| entry.idle_since >= cutoff)
+    }
+}
+
 /// A client acquired from the pool. The inner client can be accessed via Deref. The client is
 /// returned to the pool when dropped.
 pub struct ClientGuard {
@@ -317,9 +392,11 @@ impl Drop for ClientGuard {
         let Some(pool) = self.pool.upgrade() else {
             return; // pool was dropped
         };
+
         let entry = ClientEntry {
             client: self.client.take().expect("dropped once"),
             channel_guard: self.channel_guard.take().expect("dropped once"),
+            idle_since: Instant::now(),
         };
         pool.idle.lock().unwrap().insert(self.id, entry);
 
@@ -334,7 +411,6 @@ impl Drop for ClientGuard {
 /// a single request and await the response. Internally, requests are multiplexed across streams and
 /// channels. This allows proper queue depth enforcement and response routing.
 ///
-/// TODO: reap idle streams.
 /// TODO: consider making this generic over request and response types; not currently needed.
 pub struct StreamPool {
     /// The client pool to acquire clients from. Must be unbounded.
@@ -344,7 +420,7 @@ pub struct StreamPool {
     /// Incoming requests will be sent over an existing stream with available capacity. If all
     /// streams are full, a new one is spun up and added to the pool (up to `max_streams`). Each
     /// stream has an associated Tokio task that processes requests and responses.
-    streams: Arc<Mutex<HashMap<StreamID, StreamEntry>>>,
+    streams: Mutex<HashMap<StreamID, StreamEntry>>,
     /// The max number of concurrent streams, or None if unbounded.
     max_streams: Option<NonZero<usize>>,
     /// The max number of concurrent requests per stream.
@@ -352,6 +428,8 @@ pub struct StreamPool {
     /// Limits the max number of concurrent requests, given by `max_streams * max_queue_depth`.
     /// None if the pool is unbounded.
     limiter: Option<Arc<Semaphore>>,
+    /// Reaps idle streams.
+    idle_reaper: Reaper,
     /// Stream ID generator.
     next_stream_id: AtomicUsize,
 }
@@ -364,9 +442,11 @@ type ResponseSender = oneshot::Sender<tonic::Result<page_api::GetPageResponse>>;
 struct StreamEntry {
     /// Sends caller requests to the stream task. The stream task exits when this is dropped.
     sender: RequestSender,
-    /// Number of in-flight requests on this stream. This is an atomic to allow decrementing it on
-    /// completion without acquiring the `StreamPool::streams` lock.
-    queue_depth: Arc<AtomicUsize>,
+    /// Number of in-flight requests on this stream.
+    queue_depth: usize,
+    /// The time when this stream went idle (queue_depth == 0).
+    /// INVARIANT: Some if queue_depth == 0, otherwise None.
+    idle_since: Option<Instant>,
 }
 
 impl StreamPool {
@@ -383,16 +463,19 @@ impl StreamPool {
         max_queue_depth: NonZero<usize>,
     ) -> Arc<Self> {
         assert!(client_pool.limiter.is_none(), "bounded client pool");
-        Arc::new(Self {
+        let pool = Arc::new(Self {
             client_pool,
-            streams: Arc::default(),
+            streams: Mutex::default(),
             limiter: max_streams.map(|max_streams| {
                 Arc::new(Semaphore::new(max_streams.get() * max_queue_depth.get()))
             }),
             max_streams,
             max_queue_depth,
+            idle_reaper: Reaper::new(REAP_IDLE_THRESHOLD, REAP_IDLE_INTERVAL),
             next_stream_id: AtomicUsize::default(),
-        })
+        });
+        pool.idle_reaper.spawn(&pool);
+        pool
     }
 
     /// Acquires an available stream from the pool, or spins up a new stream async if all streams
@@ -412,8 +495,8 @@ impl StreamPool {
     /// * Allow concurrent clients to join onto streams while they're spun up.
     /// * Allow spinning up multiple streams concurrently, but don't overshoot limits.
     ///
-    /// For now, we just do something simple and functional, but very inefficient (linear scan).
-    pub async fn get(&self) -> StreamGuard {
+    /// For now, we just do something simple but inefficient (linear scan under mutex).
+    pub async fn get(self: &Arc<Self>) -> StreamGuard {
         // Acquire a permit if the pool is bounded.
         let mut permit = None;
         if let Some(limiter) = self.limiter.clone() {
@@ -422,23 +505,23 @@ impl StreamPool {
         let mut streams = self.streams.lock().unwrap();
 
         // Look for a pooled stream with available capacity.
-        for entry in streams.values() {
+        for (&id, entry) in streams.iter_mut() {
             assert!(
-                entry.queue_depth.load(Ordering::Relaxed) <= self.max_queue_depth.get(),
+                entry.queue_depth <= self.max_queue_depth.get(),
                 "stream queue overflow"
             );
-            if entry
-                .queue_depth
-                .fetch_update(Ordering::SeqCst, Ordering::SeqCst, |queue_depth| {
-                    // Increment the queue depth via compare-and-swap.
-                    // TODO: review ordering.
-                    (queue_depth < self.max_queue_depth.get()).then_some(queue_depth + 1)
-                })
-                .is_ok()
-            {
+            assert_eq!(
+                entry.idle_since.is_some(),
+                entry.queue_depth == 0,
+                "incorrect stream idle state"
+            );
+            if entry.queue_depth < self.max_queue_depth.get() {
+                entry.queue_depth += 1;
+                entry.idle_since = None;
                 return StreamGuard {
+                    pool: Arc::downgrade(self),
+                    id,
                     sender: entry.sender.clone(),
-                    queue_depth: entry.queue_depth.clone(),
                     permit,
                 };
             }
@@ -448,11 +531,11 @@ impl StreamPool {
         // return the guard, while spinning up the stream task async. This allows other callers to
         // join onto this stream and also create additional streams concurrently if this fills up.
         let id = self.next_stream_id.fetch_add(1, Ordering::Relaxed);
-        let queue_depth = Arc::new(AtomicUsize::new(1)); // reserve quota for this caller
         let (req_tx, req_rx) = mpsc::channel(self.max_queue_depth.get());
         let entry = StreamEntry {
             sender: req_tx.clone(),
-            queue_depth: queue_depth.clone(),
+            queue_depth: 1, // reserve quota for this caller
+            idle_since: None,
         };
         streams.insert(id, entry);
 
@@ -461,20 +544,23 @@ impl StreamPool {
         };
 
         let client_pool = self.client_pool.clone();
-        let streams = self.streams.clone();
+        let pool = Arc::downgrade(self);
 
         tokio::spawn(async move {
             if let Err(err) = Self::run_stream(client_pool, req_rx).await {
                 error!("stream failed: {err}");
             }
-            // Remove stream from pool on exit.
-            let entry = streams.lock().unwrap().remove(&id);
-            assert!(entry.is_some(), "unknown stream ID: {id}");
+            // Remove stream from pool on exit. Weak reference to avoid holding the pool alive.
+            if let Some(pool) = pool.upgrade() {
+                let entry = pool.streams.lock().unwrap().remove(&id);
+                assert!(entry.is_some(), "unknown stream ID: {id}");
+            }
         });
 
         StreamGuard {
+            pool: Arc::downgrade(self),
+            id,
             sender: req_tx,
-            queue_depth,
             permit,
         }
     }
@@ -505,6 +591,10 @@ impl StreamPool {
 
         // Track caller response channels by request ID. If the task returns early, these response
         // channels will be dropped and the waiting callers will receive an error.
+        //
+        // NB: this will leak entries if the server doesn't respond to a request (by request ID).
+        // It shouldn't happen, and if it does it will often hold onto queue depth quota anyway and
+        // block further use. But we could consider reaping closed channels after some time.
         let mut callers = HashMap::new();
 
         // Process requests and responses.
@@ -552,11 +642,26 @@ impl StreamPool {
     }
 }
 
+impl Reapable for StreamPool {
+    /// Reaps streams that have been idle since before the cutoff.
+    fn reap_idle(&self, cutoff: Instant) {
+        self.streams.lock().unwrap().retain(|_, entry| {
+            let Some(idle_since) = entry.idle_since else {
+                assert_ne!(entry.queue_depth, 0, "empty stream not marked idle");
+                return true;
+            };
+            assert_eq!(entry.queue_depth, 0, "idle stream has requests");
+            idle_since >= cutoff
+        });
+    }
+}
+
 /// A pooled stream reference. Can be used to send a single request, to properly enforce queue
 /// depth. Queue depth is already reserved and will be returned on drop.
 pub struct StreamGuard {
+    pool: Weak<StreamPool>,
+    id: StreamID,
     sender: RequestSender,
-    queue_depth: Arc<AtomicUsize>,
     permit: Option<OwnedSemaphorePermit>, // None if pool is unbounded
 }
 
@@ -588,11 +693,87 @@ impl StreamGuard {
 
 impl Drop for StreamGuard {
     fn drop(&mut self) {
+        let Some(pool) = self.pool.upgrade() else {
+            return; // pool was dropped
+        };
+
         // Release the queue depth reservation on drop. This can prematurely decrement it if dropped
         // before the response is received, but that's okay.
-        let prev_queue_depth = self.queue_depth.fetch_sub(1, Ordering::SeqCst);
-        assert!(prev_queue_depth > 0, "stream queue underflow");
+        //
+        // TODO: actually, it's probably not okay. Queue depth release should be moved into the
+        // stream task, such that it continues to account for the queue depth slot until the server
+        // responds. Otherwise, if a slow request times out and keeps blocking the stream, the
+        // server will keep waiting on it and we can pile on subsequent requests (including the
+        // timeout retry) in the same stream and get blocked. But we may also want to avoid blocking
+        // requests on e.g. LSN waits and layer downloads, instead returning early to free up the
+        // stream. Or just scale out streams with a queue depth of 1 to sidestep all head-of-line
+        // blocking. TBD.
+        let mut streams = pool.streams.lock().unwrap();
+        let entry = streams.get_mut(&self.id).expect("unknown stream");
+        assert!(entry.idle_since.is_none(), "active stream marked idle");
+        assert!(entry.queue_depth > 0, "stream queue underflow");
+        entry.queue_depth -= 1;
+        if entry.queue_depth == 0 {
+            entry.idle_since = Some(Instant::now()); // mark stream as idle
+        }
 
         _ = self.permit; // returned on drop, referenced for visibility
     }
 }
+
+/// Periodically reaps idle resources from a pool.
+struct Reaper {
+    /// The task check interval.
+    interval: Duration,
+    /// The threshold for reaping idle resources.
+    threshold: Duration,
+    /// Cancels the reaper task. Cancelled when the reaper is dropped.
+    cancel: CancellationToken,
+}
+
+impl Reaper {
+    /// Creates a new reaper.
+    pub fn new(threshold: Duration, interval: Duration) -> Self {
+        Self {
+            cancel: CancellationToken::new(),
+            threshold,
+            interval,
+        }
+    }
+
+    /// Spawns a task to periodically reap idle resources from the given task pool. The task is
+    /// cancelled when the reaper is dropped.
+    pub fn spawn(&self, pool: &Arc<impl Reapable>) {
+        // NB: hold a weak pool reference, otherwise the task will prevent dropping the pool.
+        let pool = Arc::downgrade(pool);
+        let cancel = self.cancel.clone();
+        let (interval, threshold) = (self.interval, self.threshold);
+
+        tokio::spawn(async move {
+            loop {
+                tokio::select! {
+                    _ = tokio::time::sleep(interval) => {
+                        let Some(pool) = pool.upgrade() else {
+                            return; // pool was dropped
+                        };
+                        pool.reap_idle(Instant::now() - threshold);
+                    }
+
+                    _ = cancel.cancelled() => return,
+                }
+            }
+        });
+    }
+}
+
+impl Drop for Reaper {
+    fn drop(&mut self) {
+        self.cancel.cancel(); // cancel reaper task
+    }
+}
+
+/// A reapable resource pool.
+trait Reapable: Send + Sync + 'static {
+    /// Reaps resources that have been idle since before the given cutoff.
+    fn reap_idle(&self, cutoff: Instant);
+}

pageserver/client_grpc/src/retry.rs

@@ -23,14 +23,14 @@ impl Retry {
     /// If true, log successful requests. For debugging.
     const LOG_SUCCESS: bool = false;
 
-    /// Runs the given async closure with timeouts and retries (exponential backoff). Logs errors,
-    /// using the current tracing span for context.
+    /// Runs the given async closure with timeouts and retries (exponential backoff), passing the
+    /// attempt number starting at 0. Logs errors, using the current tracing span for context.
     ///
     /// Only certain gRPC status codes are retried, see [`Self::should_retry`]. For default
     /// timeouts, see [`Self::REQUEST_TIMEOUT`] and [`Self::TOTAL_TIMEOUT`].
     pub async fn with<T, F, O>(&self, mut f: F) -> tonic::Result<T>
     where
-        F: FnMut() -> O,
+        F: FnMut(usize) -> O, // takes attempt number, starting at 0
         O: Future<Output = tonic::Result<T>>,
     {
         let started = Instant::now();
@@ -47,7 +47,7 @@ impl Retry {
                 }
 
                 let request_started = Instant::now();
-                tokio::time::timeout(Self::REQUEST_TIMEOUT, f())
+                tokio::time::timeout(Self::REQUEST_TIMEOUT, f(retries))
                     .await
                     .map_err(|_| {
                         tonic::Status::deadline_exceeded(format!(
@@ -131,7 +131,6 @@ impl Retry {
             tonic::Code::Aborted => true,
             tonic::Code::Cancelled => true,
             tonic::Code::DeadlineExceeded => true, // maybe transient slowness
-            tonic::Code::Internal => true,         // maybe transient failure?
             tonic::Code::ResourceExhausted => true,
             tonic::Code::Unavailable => true,
 
@@ -139,6 +138,10 @@ impl Retry {
             tonic::Code::AlreadyExists => false,
             tonic::Code::DataLoss => false,
             tonic::Code::FailedPrecondition => false,
+            // NB: don't retry Internal. It is intended for serious errors such as invariant
+            // violations, and is also used for client-side invariant checks that would otherwise
+            // result in retry loops.
+            tonic::Code::Internal => false,
             tonic::Code::InvalidArgument => false,
             tonic::Code::NotFound => false,
             tonic::Code::OutOfRange => false,

pageserver/client_grpc/src/split.rs

@@ -5,27 +5,24 @@ use bytes::Bytes;
 use pageserver_api::key::rel_block_to_key;
 use pageserver_api::shard::{ShardStripeSize, key_to_shard_number};
 use pageserver_page_api as page_api;
-use utils::shard::{ShardCount, ShardIndex};
+use utils::shard::{ShardCount, ShardIndex, ShardNumber};
 
 /// Splits GetPageRequests that straddle shard boundaries and assembles the responses.
 /// TODO: add tests for this.
 pub struct GetPageSplitter {
-    /// The original request ID. Used for all shard requests.
-    request_id: page_api::RequestID,
     /// Split requests by shard index.
     requests: HashMap<ShardIndex, page_api::GetPageRequest>,
-    /// Maps the offset in `GetPageRequest::block_numbers` to the owning shard. Used to assemble
-    /// the response pages in the same order as the original request.
+    /// The response being assembled. Preallocated with empty pages, to be filled in.
+    response: page_api::GetPageResponse,
+    /// Maps the offset in `request.block_numbers` and `response.pages` to the owning shard. Used
+    /// to assemble the response pages in the same order as the original request.
     block_shards: Vec<ShardIndex>,
-    /// Page responses by shard index. Will be assembled into a single response.
-    responses: HashMap<ShardIndex, Vec<Bytes>>,
 }
 
 impl GetPageSplitter {
     /// Checks if the given request only touches a single shard, and returns the shard ID. This is
     /// the common case, so we check first in order to avoid unnecessary allocations and overhead.
-    /// The caller must ensure that the request has at least one block number, or this will panic.
-    pub fn is_single_shard(
+    pub fn for_single_shard(
         req: &page_api::GetPageRequest,
         count: ShardCount,
         stripe_size: ShardStripeSize,
@@ -35,8 +32,12 @@ impl GetPageSplitter {
             return Some(ShardIndex::unsharded());
         }
 
-        // Find the base shard index for the first page, and compare with the rest.
-        let key = rel_block_to_key(req.rel, *req.block_numbers.first().expect("no pages"));
+        // Find the first page's shard, for comparison. If there are no pages, just return the first
+        // shard (caller likely checked already, otherwise the server will reject it).
+        let Some(&first_page) = req.block_numbers.first() else {
+            return Some(ShardIndex::new(ShardNumber(0), count));
+        };
+        let key = rel_block_to_key(req.rel, first_page);
         let shard_number = key_to_shard_number(count, stripe_size, &key);
 
         req.block_numbers
@@ -57,19 +58,19 @@ impl GetPageSplitter {
     ) -> Self {
         // The caller should make sure we don't split requests unnecessarily.
         debug_assert!(
-            Self::is_single_shard(&req, count, stripe_size).is_none(),
+            Self::for_single_shard(&req, count, stripe_size).is_none(),
             "unnecessary request split"
         );
 
         // Split the requests by shard index.
         let mut requests = HashMap::with_capacity(2); // common case
         let mut block_shards = Vec::with_capacity(req.block_numbers.len());
-        for blkno in req.block_numbers {
+        for &blkno in &req.block_numbers {
             let key = rel_block_to_key(req.rel, blkno);
             let shard_number = key_to_shard_number(count, stripe_size, &key);
             let shard_id = ShardIndex::new(shard_number, count);
 
-            let shard_req = requests
+            requests
                 .entry(shard_id)
                 .or_insert_with(|| page_api::GetPageRequest {
                     request_id: req.request_id,
@@ -77,27 +78,47 @@ impl GetPageSplitter {
                     rel: req.rel,
                     read_lsn: req.read_lsn,
                     block_numbers: Vec::new(),
-                });
-            shard_req.block_numbers.push(blkno);
+                })
+                .block_numbers
+                .push(blkno);
             block_shards.push(shard_id);
         }
 
-        Self {
+        // Construct a response to be populated by shard responses. Preallocate empty page slots
+        // with the expected block numbers.
+        let response = page_api::GetPageResponse {
             request_id: req.request_id,
-            responses: HashMap::with_capacity(requests.len()),
+            status_code: page_api::GetPageStatusCode::Ok,
+            reason: None,
+            rel: req.rel,
+            pages: req
+                .block_numbers
+                .into_iter()
+                .map(|block_number| {
+                    page_api::Page {
+                        block_number,
+                        image: Bytes::new(), // empty page slot to be filled in
+                    }
+                })
+                .collect(),
+        };
+
+        Self {
             requests,
+            response,
             block_shards,
         }
     }
 
-    /// Drains the per-shard requests, moving them out of the hashmap to avoid extra allocations.
+    /// Drains the per-shard requests, moving them out of the splitter to avoid extra allocations.
     pub fn drain_requests(
         &mut self,
     ) -> impl Iterator<Item = (ShardIndex, page_api::GetPageRequest)> {
         self.requests.drain()
     }
 
-    /// Adds a response from the given shard.
+    /// Adds a response from the given shard. The response must match the request ID and have an OK
+    /// status code. A response must not already exist for the given shard ID.
     #[allow(clippy::result_large_err)]
     pub fn add_response(
         &mut self,
@@ -105,68 +126,84 @@ impl GetPageSplitter {
         response: page_api::GetPageResponse,
     ) -> tonic::Result<()> {
         // The caller should already have converted status codes into tonic::Status.
-        assert_eq!(response.status_code, page_api::GetPageStatusCode::Ok);
-
-        // Make sure the response matches the request ID.
-        if response.request_id != self.request_id {
+        if response.status_code != page_api::GetPageStatusCode::Ok {
             return Err(tonic::Status::internal(format!(
-                "response ID {} does not match request ID {}",
-                response.request_id, self.request_id
+                "unexpected non-OK response for shard {shard_id}: {} {}",
+                response.status_code,
+                response.reason.unwrap_or_default()
             )));
         }
 
-        // Add the response data to the map.
-        let old = self.responses.insert(shard_id, response.page_images);
-
-        if old.is_some() {
+        if response.request_id != self.response.request_id {
             return Err(tonic::Status::internal(format!(
-                "duplicate response for shard {shard_id}",
+                "response ID mismatch for shard {shard_id}: expected {}, got {}",
+                self.response.request_id, response.request_id
+            )));
+        }
+
+        // Place the shard response pages into the assembled response, in request order.
+        let mut pages = response.pages.into_iter();
+
+        for (i, &s) in self.block_shards.iter().enumerate() {
+            if shard_id != s {
+                continue;
+            }
+
+            let Some(slot) = self.response.pages.get_mut(i) else {
+                return Err(tonic::Status::internal(format!(
+                    "no block_shards slot {i} for shard {shard_id}"
+                )));
+            };
+            let Some(page) = pages.next() else {
+                return Err(tonic::Status::internal(format!(
+                    "missing page {} in shard {shard_id} response",
+                    slot.block_number
+                )));
+            };
+            if page.block_number != slot.block_number {
+                return Err(tonic::Status::internal(format!(
+                    "shard {shard_id} returned wrong page at index {i}, expected {} got {}",
+                    slot.block_number, page.block_number
+                )));
+            }
+            if !slot.image.is_empty() {
+                return Err(tonic::Status::internal(format!(
+                    "shard {shard_id} returned duplicate page {} at index {i}",
+                    slot.block_number
+                )));
+            }
+
+            *slot = page;
+        }
+
+        // Make sure we've consumed all pages from the shard response.
+        if let Some(extra_page) = pages.next() {
+            return Err(tonic::Status::internal(format!(
+                "shard {shard_id} returned extra page: {}",
+                extra_page.block_number
             )));
         }
 
         Ok(())
     }
 
-    /// Assembles the shard responses into a single response. Responses must be present for all
-    /// relevant shards, and the total number of pages must match the original request.
+    /// Fetches the final, assembled response.
     #[allow(clippy::result_large_err)]
-    pub fn assemble_response(self) -> tonic::Result<page_api::GetPageResponse> {
-        let mut response = page_api::GetPageResponse {
-            request_id: self.request_id,
-            status_code: page_api::GetPageStatusCode::Ok,
-            reason: None,
-            page_images: Vec::with_capacity(self.block_shards.len()),
-        };
-
-        // Set up per-shard page iterators we can pull from.
-        let mut shard_responses = HashMap::with_capacity(self.responses.len());
-        for (shard_id, responses) in self.responses {
-            shard_responses.insert(shard_id, responses.into_iter());
-        }
-
-        // Reassemble the responses in the same order as the original request.
-        for shard_id in &self.block_shards {
-            let page = shard_responses
-                .get_mut(shard_id)
-                .ok_or_else(|| {
-                    tonic::Status::internal(format!("missing response for shard {shard_id}"))
-                })?
-                .next()
-                .ok_or_else(|| {
-                    tonic::Status::internal(format!("missing page from shard {shard_id}"))
-                })?;
-            response.page_images.push(page);
-        }
-
-        // Make sure there are no additional pages.
-        for (shard_id, mut pages) in shard_responses {
-            if pages.next().is_some() {
+    pub fn get_response(self) -> tonic::Result<page_api::GetPageResponse> {
+        // Check that the response is complete.
+        for (i, page) in self.response.pages.iter().enumerate() {
+            if page.image.is_empty() {
                 return Err(tonic::Status::internal(format!(
-                    "extra pages returned from shard {shard_id}"
+                    "missing page {} for shard {}",
+                    page.block_number,
+                    self.block_shards
+                        .get(i)
+                        .map(|s| s.to_string())
+                        .unwrap_or_else(|| "?".to_string())
                 )));
             }
         }
 
-        Ok(response)
+        Ok(self.response)
     }
 }

pageserver/ctl/Cargo.toml

@@ -17,6 +17,7 @@ pageserver = { path = ".." }
 pageserver_api.workspace = true
 remote_storage = { path = "../../libs/remote_storage" }
 postgres_ffi.workspace = true
+serde.workspace = true
 thiserror.workspace = true
 tokio.workspace = true
 tokio-util.workspace = true

pageserver/ctl/src/download_remote_object.rs

@@ -0,0 +1,85 @@
+use camino::Utf8PathBuf;
+use clap::Parser;
+use tokio_util::sync::CancellationToken;
+
+/// Download a specific object from remote storage to a local file.
+///
+/// The remote storage configuration is supplied via the `REMOTE_STORAGE_CONFIG` environment
+/// variable, in the same TOML format that the pageserver itself understands. This allows the
+/// command to work with any cloud supported by the `remote_storage` crate (currently AWS S3,
+/// Azure Blob Storage and local files), as long as the credentials are available via the
+/// standard environment variables expected by the underlying SDKs.
+///
+/// Examples for setting the environment variable:
+///
+/// ```bash
+/// # AWS S3 (region can also be provided via AWS_REGION)
+/// export REMOTE_STORAGE_CONFIG='remote_storage = { bucket_name = "my-bucket", bucket_region = "us-east-2" }'
+///
+/// # Azure Blob Storage (account key picked up from AZURE_STORAGE_ACCOUNT_KEY)
+/// export REMOTE_STORAGE_CONFIG='remote_storage = { container = "my-container", account = "my-account" }'
+/// ```
+#[derive(Parser)]
+pub(crate) struct DownloadRemoteObjectCmd {
+    /// Key / path of the object to download (relative to the remote storage prefix).
+    ///
+    /// Examples:
+    ///   "wal/3aa8f.../00000001000000000000000A"
+    ///   "pageserver/v1/tenants/<tenant_id>/timelines/<timeline_id>/layer_12345"
+    pub remote_path: String,
+
+    /// Path of the local file to create. Existing file will be overwritten.
+    ///
+    /// Examples:
+    ///   "./segment"
+    ///   "/tmp/layer_12345.parquet"
+    pub output_file: Utf8PathBuf,
+}
+
+pub(crate) async fn main(cmd: &DownloadRemoteObjectCmd) -> anyhow::Result<()> {
+    use remote_storage::{DownloadOpts, GenericRemoteStorage, RemotePath, RemoteStorageConfig};
+
+    // Fetch remote storage configuration from the environment
+    let config_str = std::env::var("REMOTE_STORAGE_CONFIG").map_err(|_| {
+        anyhow::anyhow!(
+            "'REMOTE_STORAGE_CONFIG' environment variable must be set to a valid remote storage TOML config"
+        )
+    })?;
+
+    let config = RemoteStorageConfig::from_toml_str(&config_str)?;
+
+    // Initialise remote storage client
+    let storage = GenericRemoteStorage::from_config(&config).await?;
+
+    // RemotePath must be relative – leading slashes confuse the parser.
+    let remote_path_str = cmd.remote_path.trim_start_matches('/');
+    let remote_path = RemotePath::from_string(remote_path_str)?;
+
+    let cancel = CancellationToken::new();
+
+    println!(
+        "Downloading '{remote_path}' from remote storage bucket {:?} ...",
+        config.storage.bucket_name()
+    );
+
+    // Start the actual download
+    let download = storage
+        .download(&remote_path, &DownloadOpts::default(), &cancel)
+        .await?;
+
+    // Stream to file
+    let mut reader = tokio_util::io::StreamReader::new(download.download_stream);
+    let tmp_path = cmd.output_file.with_extension("tmp");
+    let mut file = tokio::fs::File::create(&tmp_path).await?;
+    tokio::io::copy(&mut reader, &mut file).await?;
+    file.sync_all().await?;
+    // Atomically move into place
+    tokio::fs::rename(&tmp_path, &cmd.output_file).await?;
+
+    println!(
+        "Downloaded to '{}'. Last modified: {:?}, etag: {}",
+        cmd.output_file, download.last_modified, download.etag
+    );
+
+    Ok(())
+}

pageserver/ctl/src/index_part.rs

@@ -1,14 +1,16 @@
 use std::str::FromStr;
 
-use anyhow::Context;
+use anyhow::{Context, Ok};
 use camino::Utf8PathBuf;
 use pageserver::tenant::{
     IndexPart,
     layer_map::{LayerMap, SearchResult},
-    remote_timeline_client::remote_layer_path,
-    storage_layer::{PersistentLayerDesc, ReadableLayerWeak},
+    remote_timeline_client::{index::LayerFileMetadata, remote_layer_path},
+    storage_layer::{LayerName, LayerVisibilityHint, PersistentLayerDesc, ReadableLayerWeak},
 };
 use pageserver_api::key::Key;
+use serde::Serialize;
+use std::collections::BTreeMap;
 use utils::{
     id::{TenantId, TimelineId},
     lsn::Lsn,
@@ -33,6 +35,31 @@ pub(crate) enum IndexPartCmd {
         #[arg(long)]
         lsn: String,
     },
+    /// List all visible delta and image layers at the latest LSN.
+    ListVisibleLayers {
+        #[arg(long)]
+        path: Utf8PathBuf,
+    },
+}
+
+fn create_layer_map_from_index_part(
+    index_part: &IndexPart,
+    tenant_shard_id: TenantShardId,
+    timeline_id: TimelineId,
+) -> LayerMap {
+    let mut layer_map = LayerMap::default();
+    {
+        let mut updates = layer_map.batch_update();
+        for (key, value) in index_part.layer_metadata.iter() {
+            updates.insert_historic(PersistentLayerDesc::from_filename(
+                tenant_shard_id,
+                timeline_id,
+                key.clone(),
+                value.file_size,
+            ));
+        }
+    }
+    layer_map
 }
 
 async fn search_layers(
@@ -49,18 +76,7 @@ async fn search_layers(
         let bytes = tokio::fs::read(path).await?;
         IndexPart::from_json_bytes(&bytes).unwrap()
     };
-    let mut layer_map = LayerMap::default();
-    {
-        let mut updates = layer_map.batch_update();
-        for (key, value) in index_json.layer_metadata.iter() {
-            updates.insert_historic(PersistentLayerDesc::from_filename(
-                tenant_shard_id,
-                timeline_id,
-                key.clone(),
-                value.file_size,
-            ));
-        }
-    }
+    let layer_map = create_layer_map_from_index_part(&index_json, tenant_shard_id, timeline_id);
     let key = Key::from_hex(key)?;
 
     let lsn = Lsn::from_str(lsn).unwrap();
@@ -98,6 +114,69 @@ async fn search_layers(
     Ok(())
 }
 
+#[derive(Debug, Clone, Serialize)]
+struct VisibleLayers {
+    pub total_images: u64,
+    pub total_image_bytes: u64,
+    pub total_deltas: u64,
+    pub total_delta_bytes: u64,
+    pub layer_metadata: BTreeMap<LayerName, LayerFileMetadata>,
+}
+
+impl VisibleLayers {
+    pub fn new() -> Self {
+        Self {
+            layer_metadata: BTreeMap::new(),
+            total_images: 0,
+            total_image_bytes: 0,
+            total_deltas: 0,
+            total_delta_bytes: 0,
+        }
+    }
+
+    pub fn add_layer(&mut self, name: LayerName, layer: LayerFileMetadata) {
+        match name {
+            LayerName::Image(_) => {
+                self.total_images += 1;
+                self.total_image_bytes += layer.file_size;
+            }
+            LayerName::Delta(_) => {
+                self.total_deltas += 1;
+                self.total_delta_bytes += layer.file_size;
+            }
+        }
+        self.layer_metadata.insert(name, layer);
+    }
+}
+
+async fn list_visible_layers(path: &Utf8PathBuf) -> anyhow::Result<()> {
+    let tenant_id = TenantId::generate();
+    let tenant_shard_id = TenantShardId::unsharded(tenant_id);
+    let timeline_id = TimelineId::generate();
+
+    let bytes = tokio::fs::read(path).await.context("read file")?;
+    let index_part = IndexPart::from_json_bytes(&bytes).context("deserialize")?;
+    let layer_map = create_layer_map_from_index_part(&index_part, tenant_shard_id, timeline_id);
+    let mut visible_layers = VisibleLayers::new();
+    let (layers, _key_space) = layer_map.get_visibility(Vec::new());
+    for (layer, visibility) in layers {
+        if visibility == LayerVisibilityHint::Visible {
+            visible_layers.add_layer(
+                layer.layer_name(),
+                index_part
+                    .layer_metadata
+                    .get(&layer.layer_name())
+                    .unwrap()
+                    .clone(),
+            );
+        }
+    }
+    let output = serde_json::to_string_pretty(&visible_layers).context("serialize output")?;
+    println!("{output}");
+
+    Ok(())
+}
+
 pub(crate) async fn main(cmd: &IndexPartCmd) -> anyhow::Result<()> {
     match cmd {
         IndexPartCmd::Dump { path } => {
@@ -114,5 +193,6 @@ pub(crate) async fn main(cmd: &IndexPartCmd) -> anyhow::Result<()> {
             key,
             lsn,
         } => search_layers(tenant_id, timeline_id, path, key, lsn).await,
+        IndexPartCmd::ListVisibleLayers { path } => list_visible_layers(path).await,
     }
 }

pageserver/ctl/src/main.rs

@@ -4,6 +4,7 @@
 //!
 //! Separate, `metadata` subcommand allows to print and update pageserver's metadata file.
 
+mod download_remote_object;
 mod draw_timeline_dir;
 mod index_part;
 mod key;
@@ -16,6 +17,7 @@ use std::time::{Duration, SystemTime};
 
 use camino::{Utf8Path, Utf8PathBuf};
 use clap::{Parser, Subcommand};
+use download_remote_object::DownloadRemoteObjectCmd;
 use index_part::IndexPartCmd;
 use layers::LayerCmd;
 use page_trace::PageTraceCmd;
@@ -63,6 +65,7 @@ enum Commands {
     /// Debug print a hex key found from logs
     Key(key::DescribeKeyCommand),
     PageTrace(PageTraceCmd),
+    DownloadRemoteObject(DownloadRemoteObjectCmd),
 }
 
 /// Read and update pageserver metadata file
@@ -185,6 +188,9 @@ async fn main() -> anyhow::Result<()> {
         }
         Commands::Key(dkc) => dkc.execute(),
         Commands::PageTrace(cmd) => page_trace::main(&cmd)?,
+        Commands::DownloadRemoteObject(cmd) => {
+            download_remote_object::main(&cmd).await?;
+        }
     };
     Ok(())
 }

pageserver/page_api/proto/page_service.proto

@@ -153,7 +153,7 @@ message GetDbSizeResponse {
 message GetPageRequest {
   // A request ID. Will be included in the response. Should be unique for
   // in-flight requests on the stream.
-  uint64 request_id = 1;
+  RequestID request_id = 1;
   // The request class.
   GetPageClass request_class = 2;
   // The LSN to read at.
@@ -177,6 +177,14 @@ message GetPageRequest {
   repeated uint32 block_number = 5;
 }
 
+// A Request ID. Should be unique for in-flight requests on a stream. Included in the response.
+message RequestID {
+  // The base request ID.
+  uint64 id = 1;
+  // The request attempt. Starts at 0, incremented on each retry.
+  uint32 attempt = 2;
+}
+
 // A GetPageRequest class. Primarily intended for observability, but may also be
 // used for prioritization in the future.
 enum GetPageClass {
@@ -199,13 +207,26 @@ enum GetPageClass {
 // the entire batch is ready, so no one can make use of the individual pages.
 message GetPageResponse {
   // The original request's ID.
-  uint64 request_id = 1;
-  // The response status code.
+  RequestID request_id = 1;
+  // The response status code. If not OK, the rel and page fields will be empty.
   GetPageStatusCode status_code = 2;
   // A string describing the status, if any.
   string reason = 3;
-  // The 8KB page images, in the same order as the request. Empty if status_code != OK.
-  repeated bytes page_image = 4;
+  // The relation that the pages belong to.
+  RelTag rel = 4;
+  // The page(s), in the same order as the request.
+  repeated Page page = 5;
+}
+
+// A page.
+//
+// TODO: it would be slightly more efficient (but less convenient) to have separate arrays of block
+// numbers and images, but given the 8KB page size it's probably negligible. Benchmark it anyway.
+message Page {
+  // The page number.
+  uint32 block_number = 1;
+  // The materialized page image, as an 8KB byte vector.
+  bytes image = 2;
 }
 
 // A GetPageResponse status code.

pageserver/page_api/src/client.rs

@@ -1,4 +1,5 @@
 use anyhow::Context as _;
+use futures::future::ready;
 use futures::{Stream, StreamExt as _, TryStreamExt as _};
 use tokio::io::AsyncRead;
 use tokio_util::io::StreamReader;
@@ -110,7 +111,7 @@ impl Client {
     ) -> tonic::Result<impl Stream<Item = tonic::Result<GetPageResponse>> + Send + 'static> {
         let reqs = reqs.map(proto::GetPageRequest::from);
         let resps = self.inner.get_pages(reqs).await?.into_inner();
-        Ok(resps.map_ok(GetPageResponse::from))
+        Ok(resps.and_then(|resp| ready(GetPageResponse::try_from(resp).map_err(|err| err.into()))))
     }
 
     /// Returns the size of a relation, as # of blocks.

pageserver/page_api/src/model.rs

@@ -356,7 +356,10 @@ impl TryFrom<proto::GetPageRequest> for GetPageRequest {
             return Err(ProtocolError::Missing("block_number"));
         }
         Ok(Self {
-            request_id: pb.request_id,
+            request_id: pb
+                .request_id
+                .ok_or(ProtocolError::Missing("request_id"))?
+                .into(),
             request_class: pb.request_class.into(),
             read_lsn: pb
                 .read_lsn
@@ -371,7 +374,7 @@ impl TryFrom<proto::GetPageRequest> for GetPageRequest {
 impl From<GetPageRequest> for proto::GetPageRequest {
     fn from(request: GetPageRequest) -> Self {
         Self {
-            request_id: request.request_id,
+            request_id: Some(request.request_id.into()),
             request_class: request.request_class.into(),
             read_lsn: Some(request.read_lsn.into()),
             rel: Some(request.rel.into()),
@@ -380,8 +383,51 @@ impl From<GetPageRequest> for proto::GetPageRequest {
     }
 }
 
-/// A GetPage request ID.
-pub type RequestID = u64;
+/// A GetPage request ID and retry attempt. Should be unique for in-flight requests on a stream.
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Hash, PartialOrd, Ord)]
+pub struct RequestID {
+    /// The base request ID.
+    pub id: u64,
+    // The request attempt. Starts at 0, incremented on each retry.
+    pub attempt: u32,
+}
+
+impl RequestID {
+    /// Creates a new RequestID with the given ID and an initial attempt of 0.
+    pub fn new(id: u64) -> Self {
+        Self { id, attempt: 0 }
+    }
+}
+
+impl Display for RequestID {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}.{}", self.id, self.attempt)
+    }
+}
+
+impl From<proto::RequestId> for RequestID {
+    fn from(pb: proto::RequestId) -> Self {
+        Self {
+            id: pb.id,
+            attempt: pb.attempt,
+        }
+    }
+}
+
+impl From<u64> for RequestID {
+    fn from(id: u64) -> Self {
+        Self::new(id)
+    }
+}
+
+impl From<RequestID> for proto::RequestId {
+    fn from(request_id: RequestID) -> Self {
+        Self {
+            id: request_id.id,
+            attempt: request_id.attempt,
+        }
+    }
+}
 
 /// A GetPage request class.
 #[derive(Clone, Copy, Debug, strum_macros::Display)]
@@ -456,32 +502,41 @@ impl From<GetPageClass> for i32 {
 pub struct GetPageResponse {
     /// The original request's ID.
     pub request_id: RequestID,
-    /// The response status code.
+    /// The response status code. If not OK, the `rel` and `pages` fields will be empty.
     pub status_code: GetPageStatusCode,
     /// A string describing the status, if any.
     pub reason: Option<String>,
-    /// The 8KB page images, in the same order as the request. Empty if status != OK.
-    pub page_images: Vec<Bytes>,
+    /// The relation that the pages belong to.
+    pub rel: RelTag,
+    // The page(s), in the same order as the request.
+    pub pages: Vec<Page>,
 }
 
-impl From<proto::GetPageResponse> for GetPageResponse {
-    fn from(pb: proto::GetPageResponse) -> Self {
-        Self {
-            request_id: pb.request_id,
+impl TryFrom<proto::GetPageResponse> for GetPageResponse {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::GetPageResponse) -> Result<Self, ProtocolError> {
+        Ok(Self {
+            request_id: pb
+                .request_id
+                .ok_or(ProtocolError::Missing("request_id"))?
+                .into(),
             status_code: pb.status_code.into(),
             reason: Some(pb.reason).filter(|r| !r.is_empty()),
-            page_images: pb.page_image,
-        }
+            rel: pb.rel.ok_or(ProtocolError::Missing("rel"))?.try_into()?,
+            pages: pb.page.into_iter().map(Page::from).collect(),
+        })
     }
 }
 
 impl From<GetPageResponse> for proto::GetPageResponse {
     fn from(response: GetPageResponse) -> Self {
         Self {
-            request_id: response.request_id,
+            request_id: Some(response.request_id.into()),
             status_code: response.status_code.into(),
             reason: response.reason.unwrap_or_default(),
-            page_image: response.page_images,
+            rel: Some(response.rel.into()),
+            page: response.pages.into_iter().map(proto::Page::from).collect(),
         }
     }
 }
@@ -514,11 +569,39 @@ impl GetPageResponse {
             request_id,
             status_code,
             reason: Some(status.message().to_string()),
-            page_images: Vec::new(),
+            rel: RelTag::default(),
+            pages: Vec::new(),
         })
     }
 }
 
+// A page.
+#[derive(Clone, Debug)]
+pub struct Page {
+    /// The page number.
+    pub block_number: u32,
+    /// The materialized page image, as an 8KB byte vector.
+    pub image: Bytes,
+}
+
+impl From<proto::Page> for Page {
+    fn from(pb: proto::Page) -> Self {
+        Self {
+            block_number: pb.block_number,
+            image: pb.image,
+        }
+    }
+}
+
+impl From<Page> for proto::Page {
+    fn from(page: Page) -> Self {
+        Self {
+            block_number: page.block_number,
+            image: page.image,
+        }
+    }
+}
+
 /// A GetPage response status code.
 ///
 /// These are effectively equivalent to gRPC statuses. However, we use a bidirectional stream

pageserver/pagebench/Cargo.toml

@@ -27,8 +27,9 @@ tokio-util.workspace = true
 tonic.workspace = true
 url.workspace = true
 
-pageserver_client.workspace = true
 pageserver_api.workspace = true
+pageserver_client.workspace = true
+pageserver_client_grpc.workspace = true
 pageserver_page_api.workspace = true
 utils = { path = "../../libs/utils/" }
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -10,12 +10,14 @@ use anyhow::Context;
 use async_trait::async_trait;
 use bytes::Bytes;
 use camino::Utf8PathBuf;
+use futures::stream::FuturesUnordered;
 use futures::{Stream, StreamExt as _};
 use pageserver_api::key::Key;
 use pageserver_api::keyspace::KeySpaceAccum;
 use pageserver_api::pagestream_api::{PagestreamGetPageRequest, PagestreamRequest};
 use pageserver_api::reltag::RelTag;
 use pageserver_api::shard::TenantShardId;
+use pageserver_client_grpc::{self as client_grpc, ShardSpec};
 use pageserver_page_api as page_api;
 use rand::prelude::*;
 use tokio::task::JoinSet;
@@ -37,6 +39,10 @@ pub(crate) struct Args {
     /// Pageserver connection string. Supports postgresql:// and grpc:// protocols.
     #[clap(long, default_value = "postgres://postgres@localhost:64000")]
     page_service_connstring: String,
+    /// Use the rich gRPC Pageserver client `client_grpc::PageserverClient`, rather than the basic
+    /// no-frills `page_api::Client`. Only valid with grpc:// connstrings.
+    #[clap(long)]
+    rich_client: bool,
     #[clap(long)]
     pageserver_jwt: Option<String>,
     #[clap(long, default_value = "1")]
@@ -332,6 +338,7 @@ async fn main_impl(
             let client: Box<dyn Client> = match scheme.as_str() {
                 "postgresql" | "postgres" => {
                     assert!(!args.compression, "libpq does not support compression");
+                    assert!(!args.rich_client, "rich client requires grpc://");
                     Box::new(
                         LibpqClient::new(&args.page_service_connstring, worker_id.timeline)
                             .await
@@ -339,6 +346,16 @@ async fn main_impl(
                     )
                 }
 
+                "grpc" if args.rich_client => Box::new(
+                    RichGrpcClient::new(
+                        &args.page_service_connstring,
+                        worker_id.timeline,
+                        args.compression,
+                    )
+                    .await
+                    .unwrap(),
+                ),
+
                 "grpc" => Box::new(
                     GrpcClient::new(
                         &args.page_service_connstring,
@@ -657,7 +674,7 @@ impl Client for GrpcClient {
         blks: Vec<u32>,
     ) -> anyhow::Result<()> {
         let req = page_api::GetPageRequest {
-            request_id: req_id,
+            request_id: req_id.into(),
             request_class: page_api::GetPageClass::Normal,
             read_lsn: page_api::ReadLsn {
                 request_lsn: req_lsn,
@@ -677,6 +694,79 @@ impl Client for GrpcClient {
             "unexpected status code: {}",
             resp.status_code,
         );
-        Ok((resp.request_id, resp.page_images))
+        Ok((
+            resp.request_id.id,
+            resp.pages.into_iter().map(|p| p.image).collect(),
+        ))
+    }
+}
+
+/// A rich gRPC Pageserver client.
+struct RichGrpcClient {
+    inner: Arc<client_grpc::PageserverClient>,
+    requests: FuturesUnordered<
+        Pin<Box<dyn Future<Output = anyhow::Result<page_api::GetPageResponse>> + Send>>,
+    >,
+}
+
+impl RichGrpcClient {
+    async fn new(
+        connstring: &str,
+        ttid: TenantTimelineId,
+        compression: bool,
+    ) -> anyhow::Result<Self> {
+        let inner = Arc::new(client_grpc::PageserverClient::new(
+            ttid.tenant_id,
+            ttid.timeline_id,
+            ShardSpec::new(
+                [(ShardIndex::unsharded(), connstring.to_string())].into(),
+                None,
+            )?,
+            None,
+            compression.then_some(tonic::codec::CompressionEncoding::Zstd),
+        )?);
+        Ok(Self {
+            inner,
+            requests: FuturesUnordered::new(),
+        })
+    }
+}
+
+#[async_trait]
+impl Client for RichGrpcClient {
+    async fn send_get_page(
+        &mut self,
+        req_id: u64,
+        req_lsn: Lsn,
+        mod_lsn: Lsn,
+        rel: RelTag,
+        blks: Vec<u32>,
+    ) -> anyhow::Result<()> {
+        let req = page_api::GetPageRequest {
+            request_id: req_id.into(),
+            request_class: page_api::GetPageClass::Normal,
+            read_lsn: page_api::ReadLsn {
+                request_lsn: req_lsn,
+                not_modified_since_lsn: Some(mod_lsn),
+            },
+            rel,
+            block_numbers: blks,
+        };
+        let inner = self.inner.clone();
+        self.requests.push(Box::pin(async move {
+            inner
+                .get_page(req)
+                .await
+                .map_err(|err| anyhow::anyhow!("{err}"))
+        }));
+        Ok(())
+    }
+
+    async fn recv_get_page(&mut self) -> anyhow::Result<(u64, Vec<Bytes>)> {
+        let resp = self.requests.next().await.unwrap()?;
+        Ok((
+            resp.request_id.id,
+            resp.pages.into_iter().map(|p| p.image).collect(),
+        ))
     }
 }

pageserver/src/bin/pageserver.rs

@@ -29,8 +29,8 @@ use pageserver::task_mgr::{
 };
 use pageserver::tenant::{TenantSharedResources, mgr, secondary};
 use pageserver::{
-    CancellableTask, ConsumptionMetricsTasks, HttpEndpointListener, HttpsEndpointListener, http,
-    page_cache, page_service, task_mgr, virtual_file,
+    CancellableTask, ConsumptionMetricsTasks, HttpEndpointListener, HttpsEndpointListener,
+    MetricsCollectionTask, http, page_cache, page_service, task_mgr, virtual_file,
 };
 use postgres_backend::AuthType;
 use remote_storage::GenericRemoteStorage;
@@ -41,6 +41,7 @@ use tracing_utils::OtelGuard;
 use utils::auth::{JwtAuth, SwappableJwtAuth};
 use utils::crashsafe::syncfs;
 use utils::logging::TracingErrorLayerEnablement;
+use utils::metrics_collector::{METRICS_COLLECTION_INTERVAL, METRICS_COLLECTOR};
 use utils::sentry_init::init_sentry;
 use utils::{failpoint_support, logging, project_build_tag, project_git_version, tcp_listener};
 
@@ -763,6 +764,41 @@ fn start_pageserver(
         (http_task, https_task)
     };
 
+    /* BEGIN_HADRON */
+    let metrics_collection_task = {
+        let cancel = shutdown_pageserver.child_token();
+        let task = crate::BACKGROUND_RUNTIME.spawn({
+            let cancel = cancel.clone();
+            let background_jobs_barrier = background_jobs_barrier.clone();
+            async move {
+                if conf.force_metric_collection_on_scrape {
+                    return;
+                }
+
+                // first wait until background jobs are cleared to launch.
+                tokio::select! {
+                    _ = cancel.cancelled() => { return; },
+                    _ = background_jobs_barrier.wait() => {}
+                };
+                let mut interval = tokio::time::interval(METRICS_COLLECTION_INTERVAL);
+                loop {
+                    tokio::select! {
+                        _ = cancel.cancelled() => {
+                            tracing::info!("cancelled metrics collection task, exiting...");
+                             break;
+                        },
+                        _ = interval.tick() => {}
+                    }
+                    tokio::task::spawn_blocking(|| {
+                        METRICS_COLLECTOR.run_once(true);
+                    });
+                }
+            }
+        });
+        MetricsCollectionTask(CancellableTask { task, cancel })
+    };
+    /* END_HADRON */
+
     let consumption_metrics_tasks = {
         let cancel = shutdown_pageserver.child_token();
         let task = crate::BACKGROUND_RUNTIME.spawn({
@@ -844,6 +880,7 @@ fn start_pageserver(
             https_endpoint_listener,
             page_service,
             page_service_grpc,
+            metrics_collection_task,
             consumption_metrics_tasks,
             disk_usage_eviction_task,
             &tenant_manager,

pageserver/src/config.rs

@@ -252,6 +252,14 @@ pub struct PageServerConf {
     pub timeline_import_config: pageserver_api::config::TimelineImportConfig,
 
     pub basebackup_cache_config: Option<pageserver_api::config::BasebackupCacheConfig>,
+
+    /// Defines what is a big tenant for the purpose of image layer generation.
+    /// See Timeline::should_check_if_image_layers_required
+    pub image_layer_generation_large_timeline_threshold: Option<u64>,
+
+    /// Controls whether to collect all metrics on each scrape or to return potentially stale
+    /// results.
+    pub force_metric_collection_on_scrape: bool,
 }
 
 /// Token for authentication to safekeepers
@@ -432,6 +440,8 @@ impl PageServerConf {
             posthog_config,
             timeline_import_config,
             basebackup_cache_config,
+            image_layer_generation_large_timeline_threshold,
+            force_metric_collection_on_scrape,
         } = config_toml;
 
         let mut conf = PageServerConf {
@@ -490,6 +500,8 @@ impl PageServerConf {
             dev_mode,
             timeline_import_config,
             basebackup_cache_config,
+            image_layer_generation_large_timeline_threshold,
+            force_metric_collection_on_scrape,
 
             // ------------------------------------------------------------
             // fields that require additional validation or custom handling

pageserver/src/http/routes.rs

@@ -2,7 +2,9 @@
 //! Management HTTP API
 //!
 use std::cmp::Reverse;
-use std::collections::{BinaryHeap, HashMap};
+use std::collections::BTreeMap;
+use std::collections::BinaryHeap;
+use std::collections::HashMap;
 use std::str::FromStr;
 use std::sync::Arc;
 use std::time::Duration;
@@ -79,8 +81,8 @@ use crate::tenant::storage_layer::{IoConcurrency, LayerAccessStatsReset, LayerNa
 use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 use crate::tenant::timeline::offload::{OffloadError, offload_timeline};
 use crate::tenant::timeline::{
-    CompactFlags, CompactOptions, CompactRequest, CompactionError, MarkInvisibleRequest, Timeline,
-    WaitLsnTimeout, WaitLsnWaiter, import_pgdata,
+    CompactFlags, CompactOptions, CompactRequest, MarkInvisibleRequest, Timeline, WaitLsnTimeout,
+    WaitLsnWaiter, import_pgdata,
 };
 use crate::tenant::{
     GetTimelineError, LogicalSizeCalculationCause, OffloadedTimeline, PageReconstructError,
@@ -395,6 +397,7 @@ async fn build_timeline_info(
     timeline: &Arc<Timeline>,
     include_non_incremental_logical_size: bool,
     force_await_initial_logical_size: bool,
+    include_image_consistent_lsn: bool,
     ctx: &RequestContext,
 ) -> anyhow::Result<TimelineInfo> {
     crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id();
@@ -419,6 +422,10 @@ async fn build_timeline_info(
                 .await?,
         );
     }
+    // HADRON
+    if include_image_consistent_lsn {
+        info.image_consistent_lsn = Some(timeline.compute_image_consistent_lsn().await?);
+    }
     Ok(info)
 }
 
@@ -508,6 +515,8 @@ async fn build_timeline_info_common(
         is_invisible: Some(is_invisible),
 
         walreceiver_status,
+        // HADRON
+        image_consistent_lsn: None,
     };
     Ok(info)
 }
@@ -710,6 +719,8 @@ async fn timeline_list_handler(
         parse_query_param(&request, "include-non-incremental-logical-size")?;
     let force_await_initial_logical_size: Option<bool> =
         parse_query_param(&request, "force-await-initial-logical-size")?;
+    let include_image_consistent_lsn: Option<bool> =
+        parse_query_param(&request, "include-image-consistent-lsn")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
     let state = get_state(&request);
@@ -730,6 +741,7 @@ async fn timeline_list_handler(
                 &timeline,
                 include_non_incremental_logical_size.unwrap_or(false),
                 force_await_initial_logical_size.unwrap_or(false),
+                include_image_consistent_lsn.unwrap_or(false),
                 &ctx,
             )
             .instrument(info_span!("build_timeline_info", timeline_id = %timeline.timeline_id))
@@ -758,6 +770,9 @@ async fn timeline_and_offloaded_list_handler(
         parse_query_param(&request, "include-non-incremental-logical-size")?;
     let force_await_initial_logical_size: Option<bool> =
         parse_query_param(&request, "force-await-initial-logical-size")?;
+    let include_image_consistent_lsn: Option<bool> =
+        parse_query_param(&request, "include-image-consistent-lsn")?;
+
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
     let state = get_state(&request);
@@ -778,6 +793,7 @@ async fn timeline_and_offloaded_list_handler(
                 &timeline,
                 include_non_incremental_logical_size.unwrap_or(false),
                 force_await_initial_logical_size.unwrap_or(false),
+                include_image_consistent_lsn.unwrap_or(false),
                 &ctx,
             )
             .instrument(info_span!("build_timeline_info", timeline_id = %timeline.timeline_id))
@@ -962,6 +978,9 @@ async fn timeline_detail_handler(
         parse_query_param(&request, "include-non-incremental-logical-size")?;
     let force_await_initial_logical_size: Option<bool> =
         parse_query_param(&request, "force-await-initial-logical-size")?;
+    // HADRON
+    let include_image_consistent_lsn: Option<bool> =
+        parse_query_param(&request, "include-image-consistent-lsn")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
     // Logical size calculation needs downloading.
@@ -982,6 +1001,7 @@ async fn timeline_detail_handler(
             &timeline,
             include_non_incremental_logical_size.unwrap_or(false),
             force_await_initial_logical_size.unwrap_or(false),
+            include_image_consistent_lsn.unwrap_or(false),
             ctx,
         )
         .await
@@ -2500,9 +2520,10 @@ async fn timeline_checkpoint_handler(
                 .compact(&cancel, flags, &ctx)
                 .await
                 .map_err(|e|
-                    match e {
-                        CompactionError::ShuttingDown => ApiError::ShuttingDown,
-                        CompactionError::Other(e) => ApiError::InternalServerError(e),
+                    if e.is_cancel() {
+                        ApiError::ShuttingDown
+                    } else {
+                        ApiError::InternalServerError(e.into_anyhow())
                     }
                 )?;
         }
@@ -3213,6 +3234,30 @@ async fn get_utilization(
         .map_err(ApiError::InternalServerError)
 }
 
+/// HADRON
+async fn list_tenant_visible_size_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    check_permission(&request, None)?;
+    let state = get_state(&request);
+
+    let mut map = BTreeMap::new();
+    for (tenant_shard_id, slot) in state.tenant_manager.list() {
+        match slot {
+            TenantSlot::Attached(tenant) => {
+                let visible_size = tenant.get_visible_size();
+                map.insert(tenant_shard_id, visible_size);
+            }
+            TenantSlot::Secondary(_) | TenantSlot::InProgress(_) => {
+                continue;
+            }
+        }
+    }
+
+    json_response(StatusCode::OK, map)
+}
+
 async fn list_aux_files(
     mut request: Request<Body>,
     _cancel: CancellationToken,
@@ -3616,6 +3661,7 @@ async fn activate_post_import_handler(
         let timeline_info = build_timeline_info(
             &timeline, false, // include_non_incremental_logical_size,
             false, // force_await_initial_logical_size
+            false, // include_image_consistent_lsn
             &ctx,
         )
         .await
@@ -3937,9 +3983,14 @@ pub fn make_router(
         .expect("construct launch timestamp header middleware"),
     );
 
+    let force_metric_collection_on_scrape = state.conf.force_metric_collection_on_scrape;
+
+    let prometheus_metrics_handler_wrapper =
+        move |req| prometheus_metrics_handler(req, force_metric_collection_on_scrape);
+
     Ok(router
         .data(state)
-        .get("/metrics", |r| request_span(r, prometheus_metrics_handler))
+        .get("/metrics", move |r| request_span(r, prometheus_metrics_handler_wrapper))
         .get("/profile/cpu", |r| request_span(r, profile_cpu_handler))
         .get("/profile/heap", |r| request_span(r, profile_heap_handler))
         .get("/v1/status", |r| api_handler(r, status_handler))
@@ -4145,6 +4196,7 @@ pub fn make_router(
         .put("/v1/io_engine", |r| api_handler(r, put_io_engine_handler))
         .put("/v1/io_mode", |r| api_handler(r, put_io_mode_handler))
         .get("/v1/utilization", |r| api_handler(r, get_utilization))
+        .get("/v1/list_tenant_visible_size", |r| api_handler(r, list_tenant_visible_size_handler))
         .post(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/ingest_aux_files",
             |r| testing_api_handler("ingest_aux_files", r, ingest_aux_files),

pageserver/src/lib.rs

@@ -73,6 +73,9 @@ pub struct HttpEndpointListener(pub CancellableTask);
 pub struct HttpsEndpointListener(pub CancellableTask);
 pub struct ConsumptionMetricsTasks(pub CancellableTask);
 pub struct DiskUsageEvictionTask(pub CancellableTask);
+// HADRON
+pub struct MetricsCollectionTask(pub CancellableTask);
+
 impl CancellableTask {
     pub async fn shutdown(self) {
         self.cancel.cancel();
@@ -87,6 +90,7 @@ pub async fn shutdown_pageserver(
     https_listener: Option<HttpsEndpointListener>,
     page_service: page_service::Listener,
     grpc_task: Option<CancellableTask>,
+    metrics_collection_task: MetricsCollectionTask,
     consumption_metrics_worker: ConsumptionMetricsTasks,
     disk_usage_eviction_task: Option<DiskUsageEvictionTask>,
     tenant_manager: &TenantManager,
@@ -211,6 +215,14 @@ pub async fn shutdown_pageserver(
     // Best effort to persist any outstanding deletions, to avoid leaking objects
     deletion_queue.shutdown(Duration::from_secs(5)).await;
 
+    // HADRON
+    timed(
+        metrics_collection_task.0.shutdown(),
+        "shutdown metrics collections metrics",
+        Duration::from_secs(1),
+    )
+    .await;
+
     timed(
         consumption_metrics_worker.0.shutdown(),
         "shutdown consumption metrics",

pageserver/src/metrics.rs

@@ -2847,6 +2847,24 @@ pub(crate) static MISROUTED_PAGESTREAM_REQUESTS: Lazy<IntCounter> = Lazy::new(||
     .expect("failed to define a metric")
 });
 
+// Global counter for PageStream request results by outcome. Outcomes are divided into 3 categories:
+// - success
+// - internal_error: errors that indicate bugs in the storage cluster (e.g. page reconstruction errors, misrouted requests, LSN timeout errors)
+// - other_error: transient error conditions that are expected in normal operation or indicate bugs with other parts of the system (e.g. error due to pageserver shutdown, malformed requests etc.)
+pub(crate) static PAGESTREAM_HANDLER_RESULTS_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "pageserver_pagestream_handler_results_total",
+        "Number of pageserver pagestream handler results by outcome (success, internal_error, other_error)",
+        &["outcome"]
+    )
+    .expect("failed to define a metric")
+});
+
+// Constants for pageserver_pagestream_handler_results_total's outcome labels
+pub(crate) const PAGESTREAM_HANDLER_OUTCOME_SUCCESS: &str = "success";
+pub(crate) const PAGESTREAM_HANDLER_OUTCOME_INTERNAL_ERROR: &str = "internal_error";
+pub(crate) const PAGESTREAM_HANDLER_OUTCOME_OTHER_ERROR: &str = "other_error";
+
 // Metrics collected on WAL redo operations
 //
 // We collect the time spent in actual WAL redo ('redo'), and time waiting

pageserver/src/page_service.rs

@@ -70,7 +70,7 @@ use crate::context::{
 };
 use crate::metrics::{
     self, COMPUTE_COMMANDS_COUNTERS, ComputeCommandKind, GetPageBatchBreakReason, LIVE_CONNECTIONS,
-    MISROUTED_PAGESTREAM_REQUESTS, SmgrOpTimer, TimelineMetrics,
+    MISROUTED_PAGESTREAM_REQUESTS, PAGESTREAM_HANDLER_RESULTS_TOTAL, SmgrOpTimer, TimelineMetrics,
 };
 use crate::pgdatadir_mapping::{LsnRange, Version};
 use crate::span::{
@@ -1441,20 +1441,57 @@ impl PageServerHandler {
             let (response_msg, ctx) = match handler_result {
                 Err(e) => match &e.err {
                     PageStreamError::Shutdown => {
+                        // BEGIN HADRON
+                        PAGESTREAM_HANDLER_RESULTS_TOTAL
+                            .with_label_values(&[metrics::PAGESTREAM_HANDLER_OUTCOME_OTHER_ERROR])
+                            .inc();
+                        // END HADRON
+
                         // If we fail to fulfil a request during shutdown, which may be _because_ of
                         // shutdown, then do not send the error to the client.  Instead just drop the
                         // connection.
                         span.in_scope(|| info!("dropping connection due to shutdown"));
                         return Err(QueryError::Shutdown);
                     }
-                    PageStreamError::Reconnect(reason) => {
-                        span.in_scope(|| info!("handler requested reconnect: {reason}"));
+                    PageStreamError::Reconnect(_reason) => {
+                        span.in_scope(|| {
+                            // BEGIN HADRON
+                            // We can get here because the compute node is pointing at the wrong PS. We
+                            // already have a metric to keep track of this so suppressing this log to
+                            // reduce log spam. The information in this log message is not going to be that
+                            // helpful given the volume of logs that can be generated.
+                            // info!("handler requested reconnect: {reason}")
+                            // END HADRON
+                        });
+                        // BEGIN HADRON
+                        PAGESTREAM_HANDLER_RESULTS_TOTAL
+                            .with_label_values(&[
+                                metrics::PAGESTREAM_HANDLER_OUTCOME_INTERNAL_ERROR,
+                            ])
+                            .inc();
+                        // END HADRON
                         return Err(QueryError::Reconnect);
                     }
                     PageStreamError::Read(_)
                     | PageStreamError::LsnTimeout(_)
                     | PageStreamError::NotFound(_)
                     | PageStreamError::BadRequest(_) => {
+                        // BEGIN HADRON
+                        if let PageStreamError::Read(_) | PageStreamError::LsnTimeout(_) = &e.err {
+                            PAGESTREAM_HANDLER_RESULTS_TOTAL
+                                .with_label_values(&[
+                                    metrics::PAGESTREAM_HANDLER_OUTCOME_INTERNAL_ERROR,
+                                ])
+                                .inc();
+                        } else {
+                            PAGESTREAM_HANDLER_RESULTS_TOTAL
+                                .with_label_values(&[
+                                    metrics::PAGESTREAM_HANDLER_OUTCOME_OTHER_ERROR,
+                                ])
+                                .inc();
+                        }
+                        // END HADRON
+
                         // print the all details to the log with {:#}, but for the client the
                         // error message is enough.  Do not log if shutting down, as the anyhow::Error
                         // here includes cancellation which is not an error.
@@ -1472,7 +1509,15 @@ impl PageServerHandler {
                         )
                     }
                 },
-                Ok((response_msg, _op_timer_already_observed, ctx)) => (response_msg, Some(ctx)),
+                Ok((response_msg, _op_timer_already_observed, ctx)) => {
+                    // BEGIN HADRON
+                    PAGESTREAM_HANDLER_RESULTS_TOTAL
+                        .with_label_values(&[metrics::PAGESTREAM_HANDLER_OUTCOME_SUCCESS])
+                        .inc();
+                    // END HADRON
+
+                    (response_msg, Some(ctx))
+                }
             };
 
             let ctx = ctx.map(|req_ctx| {
@@ -3293,9 +3338,12 @@ impl GrpcPageServiceHandler {
     }
 
     /// Generates a PagestreamRequest header from a ReadLsn and request ID.
-    fn make_hdr(read_lsn: page_api::ReadLsn, req_id: u64) -> PagestreamRequest {
+    fn make_hdr(
+        read_lsn: page_api::ReadLsn,
+        req_id: Option<page_api::RequestID>,
+    ) -> PagestreamRequest {
         PagestreamRequest {
-            reqid: req_id,
+            reqid: req_id.map(|r| r.id).unwrap_or_default(),
             request_lsn: read_lsn.request_lsn,
             not_modified_since: read_lsn
                 .not_modified_since_lsn
@@ -3405,7 +3453,7 @@ impl GrpcPageServiceHandler {
 
             batch.push(BatchedGetPageRequest {
                 req: PagestreamGetPageRequest {
-                    hdr: Self::make_hdr(req.read_lsn, req.request_id),
+                    hdr: Self::make_hdr(req.read_lsn, Some(req.request_id)),
                     rel: req.rel,
                     blkno,
                 },
@@ -3435,12 +3483,16 @@ impl GrpcPageServiceHandler {
             request_id: req.request_id,
             status_code: page_api::GetPageStatusCode::Ok,
             reason: None,
-            page_images: Vec::with_capacity(results.len()),
+            rel: req.rel,
+            pages: Vec::with_capacity(results.len()),
         };
 
         for result in results {
             match result {
-                Ok((PagestreamBeMessage::GetPage(r), _, _)) => resp.page_images.push(r.page),
+                Ok((PagestreamBeMessage::GetPage(r), _, _)) => resp.pages.push(page_api::Page {
+                    block_number: r.req.blkno,
+                    image: r.page,
+                }),
                 Ok((resp, _, _)) => {
                     return Err(tonic::Status::internal(format!(
                         "unexpected response: {resp:?}"
@@ -3483,7 +3535,7 @@ impl proto::PageService for GrpcPageServiceHandler {
         span_record!(rel=%req.rel, lsn=%req.read_lsn);
 
         let req = PagestreamExistsRequest {
-            hdr: Self::make_hdr(req.read_lsn, 0),
+            hdr: Self::make_hdr(req.read_lsn, None),
             rel: req.rel,
         };
 
@@ -3633,7 +3685,7 @@ impl proto::PageService for GrpcPageServiceHandler {
         span_record!(db_oid=%req.db_oid, lsn=%req.read_lsn);
 
         let req = PagestreamDbSizeRequest {
-            hdr: Self::make_hdr(req.read_lsn, 0),
+            hdr: Self::make_hdr(req.read_lsn, None),
             dbnode: req.db_oid,
         };
 
@@ -3683,7 +3735,7 @@ impl proto::PageService for GrpcPageServiceHandler {
                 .await?
                 .downgrade();
             while let Some(req) = reqs.message().await? {
-                let req_id = req.request_id;
+                let req_id = req.request_id.map(page_api::RequestID::from).unwrap_or_default();
                 let result = Self::get_page(&ctx, &timeline, req, io_concurrency.clone())
                     .instrument(span.clone()) // propagate request span
                     .await;
@@ -3722,7 +3774,7 @@ impl proto::PageService for GrpcPageServiceHandler {
         span_record!(rel=%req.rel, lsn=%req.read_lsn);
 
         let req = PagestreamNblocksRequest {
-            hdr: Self::make_hdr(req.read_lsn, 0),
+            hdr: Self::make_hdr(req.read_lsn, None),
             rel: req.rel,
         };
 
@@ -3755,7 +3807,7 @@ impl proto::PageService for GrpcPageServiceHandler {
         span_record!(kind=%req.kind, segno=%req.segno, lsn=%req.read_lsn);
 
         let req = PagestreamGetSlruSegmentRequest {
-            hdr: Self::make_hdr(req.read_lsn, 0),
+            hdr: Self::make_hdr(req.read_lsn, None),
             kind: req.kind as u8,
             segno: req.segno,
         };

pageserver/src/tenant.rs

@@ -3291,7 +3291,7 @@ impl TenantShard {
                         // Ignore this, we likely raced with unarchival.
                         OffloadError::NotArchived => Ok(()),
                         OffloadError::AlreadyInProgress => Ok(()),
-                        OffloadError::Cancelled => Err(CompactionError::ShuttingDown),
+                        OffloadError::Cancelled => Err(CompactionError::new_cancelled()),
                         // don't break the anyhow chain
                         OffloadError::Other(err) => Err(CompactionError::Other(err)),
                     })?;
@@ -3321,16 +3321,13 @@ impl TenantShard {
 
     /// Trips the compaction circuit breaker if appropriate.
     pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
-        match err {
-            err if err.is_cancel() => {}
-            CompactionError::ShuttingDown => (),
-            CompactionError::Other(err) => {
-                self.compaction_circuit_breaker
-                    .lock()
-                    .unwrap()
-                    .fail(&CIRCUIT_BREAKERS_BROKEN, err);
-            }
+        if err.is_cancel() {
+            return;
         }
+        self.compaction_circuit_breaker
+            .lock()
+            .unwrap()
+            .fail(&CIRCUIT_BREAKERS_BROKEN, err);
     }
 
     /// Cancel scheduled compaction tasks
@@ -4174,6 +4171,15 @@ impl TenantShard {
             .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
     }
 
+    // HADRON
+    pub fn get_image_creation_timeout(&self) -> Option<Duration> {
+        let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
+        tenant_conf.image_layer_force_creation_period.or(self
+            .conf
+            .default_tenant_conf
+            .image_layer_force_creation_period)
+    }
+
     pub fn get_pitr_interval(&self) -> Duration {
         let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
         tenant_conf
@@ -5713,6 +5719,16 @@ impl TenantShard {
             .unwrap_or(0)
     }
 
+    /// HADRON
+    /// Return the visible size of all timelines in this tenant.
+    pub(crate) fn get_visible_size(&self) -> u64 {
+        let timelines = self.timelines.lock().unwrap();
+        timelines
+            .values()
+            .map(|t| t.metrics.visible_physical_size_gauge.get())
+            .sum()
+    }
+
     /// Builds a new tenant manifest, and uploads it if it differs from the last-known tenant
     /// manifest in `Self::remote_tenant_manifest`.
     ///
@@ -12800,6 +12816,40 @@ mod tests {
                 },
             ]
         );
+
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_get_force_image_creation_lsn() -> anyhow::Result<()> {
+        let tenant_conf = pageserver_api::models::TenantConfig {
+            pitr_interval: Some(Duration::from_secs(7 * 3600)),
+            image_layer_force_creation_period: Some(Duration::from_secs(3600)),
+            ..Default::default()
+        };
+
+        let tenant_id = TenantId::generate();
+
+        let harness = TenantHarness::create_custom(
+            "test_get_force_image_creation_lsn",
+            tenant_conf,
+            tenant_id,
+            ShardIdentity::unsharded(),
+            Generation::new(1),
+        )
+        .await?;
+        let (tenant, ctx) = harness.load().await;
+        let timeline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
+            .await?;
+        timeline.gc_info.write().unwrap().cutoffs.time = Some(Lsn(100));
+        {
+            let writer = timeline.writer().await;
+            writer.finish_write(Lsn(5000));
+        }
+
+        let image_creation_lsn = timeline.get_force_image_creation_lsn().unwrap();
+        assert_eq!(image_creation_lsn, Lsn(4300));
         Ok(())
     }
 }

pageserver/src/tenant/layer_map.rs

@@ -46,10 +46,11 @@
 mod historic_layer_coverage;
 mod layer_coverage;
 
-use std::collections::{HashMap, VecDeque};
+use std::collections::{BTreeMap, HashMap, VecDeque};
 use std::iter::Peekable;
 use std::ops::Range;
 use std::sync::Arc;
+use std::time::Instant;
 
 use anyhow::Result;
 use historic_layer_coverage::BufferedHistoricLayerCoverage;
@@ -904,6 +905,103 @@ impl LayerMap {
         max_stacked_deltas
     }
 
+    /* BEGIN_HADRON */
+    /**
+     * Compute the image consistent LSN, the largest LSN below which all pages have been redone successfully.
+     * It works by first finding the latest image layers and store them into a map. Then for each delta layer,
+     * find all overlapping image layers in order to potentially increase the image LSN in case there are gaps
+     * (e.g., if an image is created at LSN 100 but the delta layer spans LSN [150, 200], then we can increase
+     * image LSN to 150 because there is no WAL record in between).
+     * Finally, the image consistent LSN is computed by taking the minimum of all image layers.
+     */
+    pub fn compute_image_consistent_lsn(&self, disk_consistent_lsn: Lsn) -> Lsn {
+        struct ImageLayerInfo {
+            // creation LSN of the image layer
+            image_lsn: Lsn,
+            // the current minimum LSN of newer delta layers with overlapping key ranges
+            min_delta_lsn: Lsn,
+        }
+        let started_at = Instant::now();
+
+        let min_l0_deltas_lsn = {
+            let l0_deltas = self.level0_deltas();
+            l0_deltas
+                .iter()
+                .map(|layer| layer.get_lsn_range().start)
+                .min()
+                .unwrap_or(disk_consistent_lsn)
+        };
+        let global_key_range = Key::MIN..Key::MAX;
+
+        // step 1: collect all most recent image layers into a map
+        // map: end key to image_layer_info
+        let mut image_map: BTreeMap<Key, ImageLayerInfo> = BTreeMap::new();
+        for (img_range, img) in self.image_coverage(&global_key_range, disk_consistent_lsn) {
+            let img_lsn = img.map(|layer| layer.get_lsn_range().end).unwrap_or(Lsn(0));
+            image_map.insert(
+                img_range.end,
+                ImageLayerInfo {
+                    image_lsn: img_lsn,
+                    min_delta_lsn: min_l0_deltas_lsn,
+                },
+            );
+        }
+
+        // step 2: go through all delta layers, and update the image layer info with overlapping
+        // key ranges
+        for layer in self.historic.iter() {
+            if !layer.is_delta {
+                continue;
+            }
+            let delta_key_range = layer.get_key_range();
+            let delta_lsn_range = layer.get_lsn_range();
+            for (img_end_key, img_info) in image_map.range_mut(delta_key_range.start..Key::MAX) {
+                debug_assert!(img_end_key >= &delta_key_range.start);
+                if delta_lsn_range.end > img_info.image_lsn {
+                    // the delta layer includes WAL records after the image
+                    // it's possibel that the delta layer's start LSN < image LSN, which will be simply ignored by step 3
+                    img_info.min_delta_lsn =
+                        std::cmp::min(img_info.min_delta_lsn, delta_lsn_range.start);
+                }
+                if img_end_key >= &delta_key_range.end {
+                    // we have fully processed all overlapping image layers
+                    break;
+                }
+            }
+        }
+
+        // step 3, go through all image layers and find the image consistent LSN
+        let mut img_consistent_lsn = min_l0_deltas_lsn.checked_sub(Lsn(1)).unwrap();
+        let mut prev_key = Key::MIN;
+        for (img_key, img_info) in image_map {
+            tracing::debug!(
+                "Image layer {:?}:{} has min delta lsn {}",
+                Range {
+                    start: prev_key,
+                    end: img_key,
+                },
+                img_info.image_lsn,
+                img_info.min_delta_lsn,
+            );
+            let image_lsn = std::cmp::max(
+                img_info.image_lsn,
+                img_info.min_delta_lsn.checked_sub(Lsn(1)).unwrap_or(Lsn(0)),
+            );
+            img_consistent_lsn = std::cmp::min(img_consistent_lsn, image_lsn);
+            prev_key = img_key;
+        }
+        tracing::info!(
+            "computed image_consistent_lsn {} for disk_consistent_lsn {} in {}ms. Processed {} layrs in total.",
+            img_consistent_lsn,
+            disk_consistent_lsn,
+            started_at.elapsed().as_millis(),
+            self.historic.len()
+        );
+        img_consistent_lsn
+    }
+
+    /* END_HADRON */
+
     /// Return all L0 delta layers
     pub fn level0_deltas(&self) -> &Vec<Arc<PersistentLayerDesc>> {
         &self.l0_delta_layers
@@ -1579,6 +1677,138 @@ mod tests {
             LayerVisibilityHint::Visible
         ));
     }
+
+    /* BEGIN_HADRON */
+    #[test]
+    fn test_compute_image_consistent_lsn() {
+        let mut layer_map = LayerMap::default();
+
+        let disk_consistent_lsn = Lsn(1000);
+        // case 1: empty layer map
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(
+            disk_consistent_lsn.checked_sub(Lsn(1)).unwrap(),
+            image_consistent_lsn
+        );
+
+        // case 2: only L0 delta layer
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(100),
+                Lsn(900)..Lsn(990),
+                true,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(100),
+                Lsn(850)..Lsn(899),
+                true,
+            ));
+        }
+
+        // should use min L0 delta LSN - 1 as image consistent LSN
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(849), image_consistent_lsn);
+
+        // case 3: 3 images, no L1 delta
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(40),
+                Lsn(100)..Lsn(100),
+                false,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(40)..Key::from_i128(70),
+                Lsn(200)..Lsn(200),
+                false,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(70)..Key::from_i128(100),
+                Lsn(150)..Lsn(150),
+                false,
+            ));
+        }
+        // should use min L0 delta LSN - 1 as image consistent LSN
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(849), image_consistent_lsn);
+
+        // case 4: 3 images with 1 L1 delta
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(50),
+                Lsn(300)..Lsn(350),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(299), image_consistent_lsn);
+
+        // case 5: 3 images with 1 more L1 delta with smaller LSN
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(50)..Key::from_i128(72),
+                Lsn(200)..Lsn(300),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(199), image_consistent_lsn);
+
+        // case 6: 3 images with more newer L1 deltas (no impact on final results)
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(30),
+                Lsn(400)..Lsn(500),
+                true,
+            ));
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(35)..Key::from_i128(100),
+                Lsn(450)..Lsn(600),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(199), image_consistent_lsn);
+
+        // case 7: 3 images with more older L1 deltas (no impact on final results)
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(40),
+                Lsn(0)..Lsn(50),
+                true,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(50)..Key::from_i128(100),
+                Lsn(10)..Lsn(60),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(199), image_consistent_lsn);
+
+        // case 8: 3 images with one more L1 delta with overlapping LSN range
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(50),
+                Lsn(50)..Lsn(250),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(100), image_consistent_lsn);
+    }
+
+    /* END_HADRON */
 }
 
 #[cfg(test)]

pageserver/src/tenant/storage_layer/layer_name.rs

@@ -225,7 +225,7 @@ impl fmt::Display for ImageLayerName {
 /// storage and object names in remote storage consist of the LayerName plus some extra qualifiers
 /// that uniquely identify the physical incarnation of a layer (see [crate::tenant::remote_timeline_client::remote_layer_path])
 /// and [`crate::tenant::storage_layer::layer::local_layer_path`])
-#[derive(Debug, PartialEq, Eq, Hash, Clone)]
+#[derive(Debug, PartialEq, Eq, Hash, Clone, Ord, PartialOrd)]
 pub enum LayerName {
     Image(ImageLayerName),
     Delta(DeltaLayerName),

pageserver/src/tenant/tasks.rs

@@ -17,17 +17,14 @@ use tracing::*;
 use utils::backoff::exponential_backoff_duration;
 use utils::completion::Barrier;
 use utils::pausable_failpoint;
-use utils::sync::gate::GateError;
 
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::metrics::{self, BackgroundLoopSemaphoreMetricsRecorder, TENANT_TASK_EVENTS};
 use crate::task_mgr::{self, BACKGROUND_RUNTIME, TOKIO_WORKER_THREADS, TaskKind};
-use crate::tenant::blob_io::WriteBlobError;
 use crate::tenant::throttle::Stats;
 use crate::tenant::timeline::CompactionError;
 use crate::tenant::timeline::compaction::CompactionOutcome;
 use crate::tenant::{TenantShard, TenantState};
-use crate::virtual_file::owned_buffers_io::write::FlushTaskError;
 
 /// Semaphore limiting concurrent background tasks (across all tenants).
 ///
@@ -310,45 +307,12 @@ pub(crate) fn log_compaction_error(
     task_cancelled: bool,
     degrade_to_warning: bool,
 ) {
-    use CompactionError::*;
+    let is_cancel = err.is_cancel();
 
-    use crate::tenant::PageReconstructError;
-    use crate::tenant::upload_queue::NotInitialized;
-
-    let level = match err {
-        e if e.is_cancel() => return,
-        ShuttingDown => return,
-        _ if task_cancelled => Level::INFO,
-        Other(err) => {
-            let root_cause = err.root_cause();
-
-            let upload_queue = root_cause
-                .downcast_ref::<NotInitialized>()
-                .is_some_and(|e| e.is_stopping());
-            let timeline = root_cause
-                .downcast_ref::<PageReconstructError>()
-                .is_some_and(|e| e.is_cancel());
-            let buffered_writer_flush_task_canelled = root_cause
-                .downcast_ref::<FlushTaskError>()
-                .is_some_and(|e| e.is_cancel());
-            let write_blob_cancelled = root_cause
-                .downcast_ref::<WriteBlobError>()
-                .is_some_and(|e| e.is_cancel());
-            let gate_closed = root_cause
-                .downcast_ref::<GateError>()
-                .is_some_and(|e| e.is_cancel());
-            let is_stopping = upload_queue
-                || timeline
-                || buffered_writer_flush_task_canelled
-                || write_blob_cancelled
-                || gate_closed;
-
-            if is_stopping {
-                Level::INFO
-            } else {
-                Level::ERROR
-            }
-        }
+    let level = if is_cancel || task_cancelled {
+        Level::INFO
+    } else {
+        Level::ERROR
     };
 
     if let Some((error_count, sleep_duration)) = retry_info {

pageserver/src/tenant/timeline.rs

@@ -1002,7 +1002,7 @@ impl From<WaitLsnError> for tonic::Status {
 impl From<CreateImageLayersError> for CompactionError {
     fn from(e: CreateImageLayersError) -> Self {
         match e {
-            CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
+            CreateImageLayersError::Cancelled => CompactionError::new_cancelled(),
             CreateImageLayersError::Other(e) => {
                 CompactionError::Other(e.context("create image layers"))
             }
@@ -2117,12 +2117,7 @@ impl Timeline {
         match &result {
             Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
             Err(e) if e.is_cancel() => {}
-            Err(CompactionError::ShuttingDown) => {
-                // Covered by the `Err(e) if e.is_cancel()` branch.
-            }
-            Err(CompactionError::Other(_)) => {
-                self.compaction_failed.store(true, AtomicOrdering::Relaxed)
-            }
+            Err(_) => self.compaction_failed.store(true, AtomicOrdering::Relaxed),
         };
 
         result
@@ -2851,6 +2846,18 @@ impl Timeline {
             .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
     }
 
+    // HADRON
+    fn get_image_layer_force_creation_period(&self) -> Option<Duration> {
+        let tenant_conf = self.tenant_conf.load();
+        tenant_conf
+            .tenant_conf
+            .image_layer_force_creation_period
+            .or(self
+                .conf
+                .default_tenant_conf
+                .image_layer_force_creation_period)
+    }
+
     fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
         let tenant_conf = &self.tenant_conf.load();
         tenant_conf
@@ -3120,7 +3127,6 @@ impl Timeline {
                 repartition_threshold: 0,
                 last_image_layer_creation_check_at: AtomicLsn::new(0),
                 last_image_layer_creation_check_instant: Mutex::new(None),
-
                 last_received_wal: Mutex::new(None),
                 rel_size_latest_cache: RwLock::new(HashMap::new()),
                 rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
@@ -5041,6 +5047,7 @@ impl Timeline {
                 .create_image_layers(
                     &partitions,
                     self.initdb_lsn,
+                    None,
                     ImageLayerCreationMode::Initial,
                     ctx,
                     LastImageLayerCreationStatus::Initial,
@@ -5312,14 +5319,19 @@ impl Timeline {
     }
 
     // Is it time to create a new image layer for the given partition? True if we want to generate.
-    async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
+    async fn time_for_new_image_layer(
+        &self,
+        partition: &KeySpace,
+        lsn: Lsn,
+        force_image_creation_lsn: Option<Lsn>,
+    ) -> bool {
         let threshold = self.get_image_creation_threshold();
 
         let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
         let Ok(layers) = guard.layer_map() else {
             return false;
         };
-
+        let mut min_image_lsn: Lsn = Lsn::MAX;
         let mut max_deltas = 0;
         for part_range in &partition.ranges {
             let image_coverage = layers.image_coverage(part_range, lsn);
@@ -5354,9 +5366,25 @@ impl Timeline {
                         return true;
                     }
                 }
+                min_image_lsn = min(min_image_lsn, img_lsn);
             }
         }
 
+        // HADRON
+        // for child timelines, we consider all pages up to ancestor_LSN are redone successfully by the parent timeline
+        min_image_lsn = min_image_lsn.max(self.get_ancestor_lsn());
+        if min_image_lsn < force_image_creation_lsn.unwrap_or(Lsn(0)) && max_deltas > 0 {
+            info!(
+                "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}, num deltas: {}",
+                partition.ranges[0].start,
+                partition.ranges[0].end,
+                min_image_lsn,
+                force_image_creation_lsn.unwrap(),
+                max_deltas
+            );
+            return true;
+        }
+
         debug!(
             max_deltas,
             "none of the partitioned ranges had >= {threshold} deltas"
@@ -5582,7 +5610,7 @@ impl Timeline {
     ///        suffer from the lack of image layers
     ///     2. For small tenants (that can mostly fit in RAM), we use a much longer interval
     fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
-        const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
+        let large_timeline_threshold = self.conf.image_layer_generation_large_timeline_threshold;
 
         let last_checks_at = self.last_image_layer_creation_check_at.load();
         let distance = lsn
@@ -5596,12 +5624,12 @@ impl Timeline {
         let mut time_based_decision = false;
         let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
         if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
-            let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
-            {
-                self.get_checkpoint_timeout()
-            } else {
-                Duration::from_secs(3600 * 48)
-            };
+            let check_required_after =
+                if Some(Into::<u64>::into(&logical_size)) >= large_timeline_threshold {
+                    self.get_checkpoint_timeout()
+                } else {
+                    Duration::from_secs(3600 * 48)
+                };
 
             time_based_decision = match *last_check_instant {
                 Some(last_check) => {
@@ -5629,10 +5657,12 @@ impl Timeline {
     /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
     ///
     /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
+    #[allow(clippy::too_many_arguments)]
     async fn create_image_layers(
         self: &Arc<Timeline>,
         partitioning: &KeyPartitioning,
         lsn: Lsn,
+        force_image_creation_lsn: Option<Lsn>,
         mode: ImageLayerCreationMode,
         ctx: &RequestContext,
         last_status: LastImageLayerCreationStatus,
@@ -5736,7 +5766,11 @@ impl Timeline {
             } else if let ImageLayerCreationMode::Try = mode {
                 // check_for_image_layers = false -> skip
                 // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
-                if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
+                if !check_for_image_layers
+                    || !self
+                        .time_for_new_image_layer(partition, lsn, force_image_creation_lsn)
+                        .await
+                {
                     start = img_range.end;
                     continue;
                 }
@@ -6057,26 +6091,88 @@ impl Drop for Timeline {
     }
 }
 
-/// Top-level failure to compact.
-#[derive(Debug, thiserror::Error)]
-pub(crate) enum CompactionError {
-    #[error("The timeline or pageserver is shutting down")]
-    ShuttingDown,
-    #[error(transparent)]
-    Other(anyhow::Error),
-}
+pub(crate) use compaction_error::CompactionError;
+/// In a private mod to enforce that [`CompactionError::is_cancel`] is used
+/// instead of `match`ing on [`CompactionError::ShuttingDown`].
+mod compaction_error {
+    use utils::sync::gate::GateError;
 
-impl CompactionError {
-    /// Errors that can be ignored, i.e., cancel and shutdown.
-    pub fn is_cancel(&self) -> bool {
-        matches!(self, Self::ShuttingDown)
+    use crate::{
+        pgdatadir_mapping::CollectKeySpaceError,
+        tenant::{PageReconstructError, blob_io::WriteBlobError, upload_queue::NotInitialized},
+        virtual_file::owned_buffers_io::write::FlushTaskError,
+    };
+
+    /// Top-level failure to compact. Use [`Self::is_cancel`].
+    #[derive(Debug, thiserror::Error)]
+    pub(crate) enum CompactionError {
+        /// Use [`Self::is_cancel`] instead of checking for this variant.
+        #[error("The timeline or pageserver is shutting down")]
+        #[allow(private_interfaces)]
+        ShuttingDown(ForbidMatching), // private ForbidMatching enforces use of [`Self::is_cancel`].
+        #[error(transparent)]
+        Other(anyhow::Error),
     }
 
-    pub fn from_collect_keyspace(err: CollectKeySpaceError) -> Self {
-        if err.is_cancel() {
-            Self::ShuttingDown
-        } else {
-            Self::Other(err.into_anyhow())
+    #[derive(Debug)]
+    struct ForbidMatching;
+
+    impl CompactionError {
+        pub fn new_cancelled() -> Self {
+            Self::ShuttingDown(ForbidMatching)
+        }
+        /// Errors that can be ignored, i.e., cancel and shutdown.
+        pub fn is_cancel(&self) -> bool {
+            let other = match self {
+                CompactionError::ShuttingDown(_) => return true,
+                CompactionError::Other(other) => other,
+            };
+
+            // The write path of compaction in particular often lacks differentiated
+            // handling errors stemming from cancellation from other errors.
+            // So, if requested, we also check the ::Other variant by downcasting.
+            // The list below has been found empirically from flaky tests and production logs.
+            // The process is simple: on ::Other(), compaction will print the enclosed
+            // anyhow::Error in debug mode, i.e., with backtrace. That backtrace contains the
+            // line where the write path / compaction code does undifferentiated error handling
+            // from a non-anyhow type to an anyhow type. Add the type to the list of downcasts
+            // below, following the same is_cancel() pattern.
+
+            let root_cause = other.root_cause();
+
+            let upload_queue = root_cause
+                .downcast_ref::<NotInitialized>()
+                .is_some_and(|e| e.is_stopping());
+            let timeline = root_cause
+                .downcast_ref::<PageReconstructError>()
+                .is_some_and(|e| e.is_cancel());
+            let buffered_writer_flush_task_canelled = root_cause
+                .downcast_ref::<FlushTaskError>()
+                .is_some_and(|e| e.is_cancel());
+            let write_blob_cancelled = root_cause
+                .downcast_ref::<WriteBlobError>()
+                .is_some_and(|e| e.is_cancel());
+            let gate_closed = root_cause
+                .downcast_ref::<GateError>()
+                .is_some_and(|e| e.is_cancel());
+            upload_queue
+                || timeline
+                || buffered_writer_flush_task_canelled
+                || write_blob_cancelled
+                || gate_closed
+        }
+        pub fn into_anyhow(self) -> anyhow::Error {
+            match self {
+                CompactionError::ShuttingDown(ForbidMatching) => anyhow::Error::new(self),
+                CompactionError::Other(e) => e,
+            }
+        }
+        pub fn from_collect_keyspace(err: CollectKeySpaceError) -> Self {
+            if err.is_cancel() {
+                Self::new_cancelled()
+            } else {
+                Self::Other(err.into_anyhow())
+            }
         }
     }
 }
@@ -6088,7 +6184,7 @@ impl From<super::upload_queue::NotInitialized> for CompactionError {
                 CompactionError::Other(anyhow::anyhow!(value))
             }
             super::upload_queue::NotInitialized::ShuttingDown
-            | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
+            | super::upload_queue::NotInitialized::Stopped => CompactionError::new_cancelled(),
         }
     }
 }
@@ -6098,7 +6194,7 @@ impl From<super::storage_layer::layer::DownloadError> for CompactionError {
         match e {
             super::storage_layer::layer::DownloadError::TimelineShutdown
             | super::storage_layer::layer::DownloadError::DownloadCancelled => {
-                CompactionError::ShuttingDown
+                CompactionError::new_cancelled()
             }
             super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
             | super::storage_layer::layer::DownloadError::DownloadRequired
@@ -6117,14 +6213,14 @@ impl From<super::storage_layer::layer::DownloadError> for CompactionError {
 
 impl From<layer_manager::Shutdown> for CompactionError {
     fn from(_: layer_manager::Shutdown) -> Self {
-        CompactionError::ShuttingDown
+        CompactionError::new_cancelled()
     }
 }
 
 impl From<super::storage_layer::errors::PutError> for CompactionError {
     fn from(e: super::storage_layer::errors::PutError) -> Self {
         if e.is_cancel() {
-            CompactionError::ShuttingDown
+            CompactionError::new_cancelled()
         } else {
             CompactionError::Other(e.into_anyhow())
         }
@@ -6223,7 +6319,7 @@ impl Timeline {
         let mut guard = tokio::select! {
             guard = self.layers.write(LayerManagerLockHolder::Compaction) => guard,
             _ = self.cancel.cancelled() => {
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
         };
 
@@ -7050,6 +7146,19 @@ impl Timeline {
             .unwrap()
             .clone()
     }
+
+    /* BEGIN_HADRON */
+    pub(crate) async fn compute_image_consistent_lsn(&self) -> anyhow::Result<Lsn> {
+        let guard = self
+            .layers
+            .read(LayerManagerLockHolder::ComputeImageConsistentLsn)
+            .await;
+        let layer_map = guard.layer_map()?;
+        let disk_consistent_lsn = self.get_disk_consistent_lsn();
+
+        Ok(layer_map.compute_image_consistent_lsn(disk_consistent_lsn))
+    }
+    /* END_HADRON */
 }
 
 impl Timeline {

pageserver/src/tenant/timeline/compaction.rs

@@ -4,6 +4,7 @@
 //!
 //! The old legacy algorithm is implemented directly in `timeline.rs`.
 
+use std::cmp::min;
 use std::collections::{BinaryHeap, HashMap, HashSet, VecDeque};
 use std::ops::{Deref, Range};
 use std::sync::Arc;
@@ -572,8 +573,8 @@ impl GcCompactionQueue {
         }
         match res {
             Ok(res) => Ok(res),
-            Err(CompactionError::ShuttingDown) => Err(CompactionError::ShuttingDown),
-            Err(CompactionError::Other(_)) => {
+            Err(e) if e.is_cancel() => Err(e),
+            Err(_) => {
                 // There are some cases where traditional gc might collect some layer
                 // files causing gc-compaction cannot read the full history of the key.
                 // This needs to be resolved in the long-term by improving the compaction
@@ -1260,13 +1261,16 @@ impl Timeline {
         // Is the timeline being deleted?
         if self.is_stopping() {
             trace!("Dropping out of compaction on timeline shutdown");
-            return Err(CompactionError::ShuttingDown);
+            return Err(CompactionError::new_cancelled());
         }
 
         let target_file_size = self.get_checkpoint_distance();
 
         // Define partitioning schema if needed
 
+        // HADRON
+        let force_image_creation_lsn = self.get_force_image_creation_lsn();
+
         // 1. L0 Compact
         let l0_outcome = {
             let timer = self.metrics.compact_time_histo.start_timer();
@@ -1274,6 +1278,7 @@ impl Timeline {
                 .compact_level0(
                     target_file_size,
                     options.flags.contains(CompactFlags::ForceL0Compaction),
+                    force_image_creation_lsn,
                     ctx,
                 )
                 .await?;
@@ -1376,6 +1381,7 @@ impl Timeline {
                     .create_image_layers(
                         &partitioning,
                         lsn,
+                        force_image_creation_lsn,
                         mode,
                         &image_ctx,
                         self.last_image_layer_creation_status
@@ -1472,6 +1478,41 @@ impl Timeline {
         Ok(CompactionOutcome::Done)
     }
 
+    /* BEGIN_HADRON */
+    // Get the force image creation LSN based on gc_cutoff_lsn.
+    // Note that this is an estimation and the workload rate may suddenly change. When that happens,
+    // the force image creation may be too early or too late, but eventually it should be able to catch up.
+    pub(crate) fn get_force_image_creation_lsn(self: &Arc<Self>) -> Option<Lsn> {
+        let image_creation_period = self.get_image_layer_force_creation_period()?;
+        let current_lsn = self.get_last_record_lsn();
+        let pitr_lsn = self.gc_info.read().unwrap().cutoffs.time?;
+        let pitr_interval = self.get_pitr_interval();
+        if pitr_lsn == Lsn::INVALID || pitr_interval.is_zero() {
+            tracing::warn!(
+                "pitr LSN/interval not found, skipping force image creation LSN calculation"
+            );
+            return None;
+        }
+
+        let delta_lsn = current_lsn.checked_sub(pitr_lsn).unwrap().0
+            * image_creation_period.as_secs()
+            / pitr_interval.as_secs();
+        let force_image_creation_lsn = current_lsn.checked_sub(delta_lsn).unwrap_or(Lsn(0));
+
+        tracing::info!(
+            "Tenant shard {} computed force_image_creation_lsn: {}. Current lsn: {}, image_layer_force_creation_period: {:?}, GC cutoff: {}, PITR interval: {:?}",
+            self.tenant_shard_id,
+            force_image_creation_lsn,
+            current_lsn,
+            image_creation_period,
+            pitr_lsn,
+            pitr_interval
+        );
+
+        Some(force_image_creation_lsn)
+    }
+    /* END_HADRON */
+
     /// Check for layers that are elegible to be rewritten:
     /// - Shard splitting: After a shard split, ancestor layers beyond pitr_interval, so that
     ///   we don't indefinitely retain keys in this shard that aren't needed.
@@ -1624,7 +1665,7 @@ impl Timeline {
 
         for (i, layer) in layers_to_rewrite.into_iter().enumerate() {
             if self.cancel.is_cancelled() {
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
 
             info!(layer=%layer, "rewriting layer after shard split: {}/{}", i, total);
@@ -1722,7 +1763,7 @@ impl Timeline {
                     Ok(()) => {},
                     Err(WaitCompletionError::NotInitialized(ni)) => return Err(CompactionError::from(ni)),
                     Err(WaitCompletionError::UploadQueueShutDownOrStopped) => {
-                        return Err(CompactionError::ShuttingDown);
+                        return Err(CompactionError::new_cancelled());
                     }
                 },
                 // Don't wait if there's L0 compaction to do. We don't need to update the outcome
@@ -1801,6 +1842,7 @@ impl Timeline {
         self: &Arc<Self>,
         target_file_size: u64,
         force_compaction_ignore_threshold: bool,
+        force_compaction_lsn: Option<Lsn>,
         ctx: &RequestContext,
     ) -> Result<CompactionOutcome, CompactionError> {
         let CompactLevel0Phase1Result {
@@ -1821,6 +1863,7 @@ impl Timeline {
                 stats,
                 target_file_size,
                 force_compaction_ignore_threshold,
+                force_compaction_lsn,
                 &ctx,
             )
             .instrument(phase1_span)
@@ -1843,6 +1886,7 @@ impl Timeline {
         mut stats: CompactLevel0Phase1StatsBuilder,
         target_file_size: u64,
         force_compaction_ignore_threshold: bool,
+        force_compaction_lsn: Option<Lsn>,
         ctx: &RequestContext,
     ) -> Result<CompactLevel0Phase1Result, CompactionError> {
         let begin = tokio::time::Instant::now();
@@ -1872,11 +1916,28 @@ impl Timeline {
                     return Ok(CompactLevel0Phase1Result::default());
                 }
             } else {
-                debug!(
-                    level0_deltas = level0_deltas.len(),
-                    threshold, "too few deltas to compact"
-                );
-                return Ok(CompactLevel0Phase1Result::default());
+                // HADRON
+                let min_lsn = level0_deltas
+                    .iter()
+                    .map(|a| a.get_lsn_range().start)
+                    .reduce(min);
+                if force_compaction_lsn.is_some()
+                    && min_lsn.is_some()
+                    && min_lsn.unwrap() < force_compaction_lsn.unwrap()
+                {
+                    info!(
+                        "forcing L0 compaction of {} L0 deltas. Min lsn: {}, force compaction lsn: {}",
+                        level0_deltas.len(),
+                        min_lsn.unwrap(),
+                        force_compaction_lsn.unwrap()
+                    );
+                } else {
+                    debug!(
+                        level0_deltas = level0_deltas.len(),
+                        threshold, "too few deltas to compact"
+                    );
+                    return Ok(CompactLevel0Phase1Result::default());
+                }
             }
         }
 
@@ -1985,7 +2046,7 @@ impl Timeline {
             let mut all_keys = Vec::new();
             for l in deltas_to_compact.iter() {
                 if self.cancel.is_cancelled() {
-                    return Err(CompactionError::ShuttingDown);
+                    return Err(CompactionError::new_cancelled());
                 }
                 let delta = l.get_as_delta(ctx).await.map_err(CompactionError::Other)?;
                 let keys = delta
@@ -2078,7 +2139,7 @@ impl Timeline {
         stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
 
         if self.cancel.is_cancelled() {
-            return Err(CompactionError::ShuttingDown);
+            return Err(CompactionError::new_cancelled());
         }
 
         stats.read_lock_drop_micros = stats.read_lock_held_compute_holes_micros.till_now();
@@ -2186,7 +2247,7 @@ impl Timeline {
                 // avoid hitting the cancellation token on every key. in benches, we end up
                 // shuffling an order of million keys per layer, this means we'll check it
                 // around tens of times per layer.
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
 
             let same_key = prev_key == Some(key);
@@ -2271,7 +2332,7 @@ impl Timeline {
                 if writer.is_none() {
                     if self.cancel.is_cancelled() {
                         // to be somewhat responsive to cancellation, check for each new layer
-                        return Err(CompactionError::ShuttingDown);
+                        return Err(CompactionError::new_cancelled());
                     }
                     // Create writer if not initiaized yet
                     writer = Some(
@@ -2527,7 +2588,7 @@ impl Timeline {
         // Is the timeline being deleted?
         if self.is_stopping() {
             trace!("Dropping out of compaction on timeline shutdown");
-            return Err(CompactionError::ShuttingDown);
+            return Err(CompactionError::new_cancelled());
         }
 
         let (dense_ks, _sparse_ks) = self
@@ -3189,7 +3250,7 @@ impl Timeline {
         let gc_lock = async {
             tokio::select! {
                 guard = self.gc_lock.lock() => Ok(guard),
-                _ = cancel.cancelled() => Err(CompactionError::ShuttingDown),
+                _ = cancel.cancelled() => Err(CompactionError::new_cancelled()),
             }
         };
 
@@ -3462,7 +3523,7 @@ impl Timeline {
             }
             total_layer_size += layer.layer_desc().file_size;
             if cancel.is_cancelled() {
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
             let should_yield = yield_for_l0
                 && self
@@ -3609,7 +3670,7 @@ impl Timeline {
             }
 
             if cancel.is_cancelled() {
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
 
             let should_yield = yield_for_l0

pageserver/src/tenant/timeline/layer_manager.rs

@@ -47,6 +47,7 @@ pub(crate) enum LayerManagerLockHolder {
     ImportPgData,
     DetachAncestor,
     Eviction,
+    ComputeImageConsistentLsn,
     #[cfg(test)]
     Testing,
 }

pgxn/neon/libpagestore.c

@@ -1410,7 +1410,7 @@ pg_init_libpagestore(void)
 							"sharding stripe size",
 							NULL,
 							&stripe_size,
-							32768, 1, INT_MAX,
+							2048, 1, INT_MAX,
 							PGC_SIGHUP,
 							GUC_UNIT_BLOCKS,
 							NULL, NULL, NULL);

pgxn/neon/neon_ddl_handler.c

@@ -953,7 +953,9 @@ neon_fmgr_hook(FmgrHookEventType event, FmgrInfo *flinfo, Datum *private)
 
 	/*
 	 * Fire Event Trigger if both function owner and current user are
-	 * superuser, or none of them are.
+	 * superuser. Allow executing Event Trigger function that belongs to a
+	 * superuser when connected as a non-superuser, even when the function is
+	 * SECURITY DEFINER.
 	 */
     else if (event == FHET_START
 		/* still enable it to pass pg_regress tests */
@@ -976,32 +978,7 @@ neon_fmgr_hook(FmgrHookEventType event, FmgrInfo *flinfo, Datum *private)
 		function_is_owned_by_super = superuser_arg(function_owner);
 
 		/*
-		 * 1. Refuse to run SECURITY DEFINER function that belongs to a
-		 * superuser when the current user is not a superuser itself.
-		 */
-		if (!role_is_super
-			&& function_is_owned_by_super
-			&& function_is_secdef)
-		{
-			char *func_name = get_func_name(flinfo->fn_oid);
-
-			ereport(WARNING,
-					(errmsg("Skipping Event Trigger"),
-					 errdetail("Event Trigger function \"%s\" is owned by \"%s\" "
-							   "and is SECURITY DEFINER",
-							   func_name,
-							   GetUserNameFromId(function_owner, false))));
-
-			/*
-			 * we can't skip execution directly inside the fmgr_hook so
-			 * instead we change the event trigger function to a noop
-			 * function.
-			 */
-			force_noop(flinfo);
-		}
-
-		/*
-		 * 2. Refuse to run functions that belongs to a non-superuser when the
+		 * Refuse to run functions that belongs to a non-superuser when the
 		 * current user is a superuser.
 		 *
 		 * We could run a SECURITY DEFINER user-function here and be safe with
@@ -1009,7 +986,7 @@ neon_fmgr_hook(FmgrHookEventType event, FmgrInfo *flinfo, Datum *private)
 		 * infrastructure maintenance operations, where we prefer to skip
 		 * running user-defined code.
 		 */
-		else if (role_is_super && !function_is_owned_by_super)
+		if (role_is_super && !function_is_owned_by_super)
 		{
 			char *func_name = get_func_name(flinfo->fn_oid);
 

pgxn/neon/walproposer.h

@@ -376,6 +376,18 @@ typedef struct PageserverFeedback
 	uint32		shard_number;
 } PageserverFeedback;
 
+/* BEGIN_HADRON */
+typedef struct WalRateLimiter
+{
+	/* If the value is 1, PG backends will hit backpressure. */
+	pg_atomic_uint32 should_limit;
+	/* The number of bytes sent in the current second. */
+	uint64		sent_bytes;
+	/* The last recorded time in microsecond. */
+	TimestampTz last_recorded_time_us;
+} WalRateLimiter;
+/* END_HADRON */
+
 typedef struct WalproposerShmemState
 {
 	pg_atomic_uint64 propEpochStartLsn;
@@ -395,6 +407,11 @@ typedef struct WalproposerShmemState
 
 	/* aggregated feedback with min LSNs across shards */
 	PageserverFeedback min_ps_feedback;
+
+	/* BEGIN_HADRON */
+	/* The WAL rate limiter */
+	WalRateLimiter wal_rate_limiter;
+	/* END_HADRON */
 } WalproposerShmemState;
 
 /*

pgxn/neon/walproposer_pg.c

@@ -66,6 +66,9 @@ int			wal_acceptor_reconnect_timeout = 1000;
 int			wal_acceptor_connection_timeout = 10000;
 int			safekeeper_proto_version = 3;
 char	   *safekeeper_conninfo_options = "";
+/* BEGIN_HADRON */
+int         databricks_max_wal_mb_per_second = -1;
+/* END_HADRON */
 
 /* Set to true in the walproposer bgw. */
 static bool am_walproposer;
@@ -252,6 +255,18 @@ nwp_register_gucs(void)
 							PGC_POSTMASTER,
 							0,
 							NULL, NULL, NULL);
+
+    /* BEGIN_HADRON */
+    DefineCustomIntVariable(
+                            "databricks.max_wal_mb_per_second",
+                            "The maximum WAL MB per second allowed. If breached, sending WAL hit the backpressure. Setting to -1 disables the limit.",
+                            NULL,
+                            &databricks_max_wal_mb_per_second,
+                            -1, -1, INT_MAX,
+                            PGC_SUSET,
+                            GUC_UNIT_MB,
+                            NULL, NULL, NULL);
+    /* END_HADRON */
 }
 
 
@@ -393,6 +408,7 @@ assign_neon_safekeepers(const char *newval, void *extra)
 static uint64
 backpressure_lag_impl(void)
 {
+	struct WalproposerShmemState* state = NULL;
 	if (max_replication_apply_lag > 0 || max_replication_flush_lag > 0 || max_replication_write_lag > 0)
 	{
 		XLogRecPtr	writePtr;
@@ -426,6 +442,18 @@ backpressure_lag_impl(void)
 			return (myFlushLsn - applyPtr - max_replication_apply_lag * MB);
 		}
 	}
+
+	/* BEGIN_HADRON */
+	if (databricks_max_wal_mb_per_second == -1) {
+		return 0;
+	}
+
+	state = GetWalpropShmemState();
+	if (state != NULL && pg_atomic_read_u32(&state->wal_rate_limiter.should_limit) == 1)
+	{
+		return 1;
+	}
+	/* END_HADRON */
 	return 0;
 }
 
@@ -472,6 +500,9 @@ WalproposerShmemInit(void)
 		pg_atomic_init_u64(&walprop_shared->mineLastElectedTerm, 0);
 		pg_atomic_init_u64(&walprop_shared->backpressureThrottlingTime, 0);
 		pg_atomic_init_u64(&walprop_shared->currentClusterSize, 0);
+		/* BEGIN_HADRON */
+		pg_atomic_init_u32(&walprop_shared->wal_rate_limiter.should_limit, 0);
+		/* END_HADRON */
 	}
 	LWLockRelease(AddinShmemInitLock);
 
@@ -487,6 +518,9 @@ WalproposerShmemInit_SyncSafekeeper(void)
 	pg_atomic_init_u64(&walprop_shared->propEpochStartLsn, 0);
 	pg_atomic_init_u64(&walprop_shared->mineLastElectedTerm, 0);
 	pg_atomic_init_u64(&walprop_shared->backpressureThrottlingTime, 0);
+	/* BEGIN_HADRON */
+	pg_atomic_init_u32(&walprop_shared->wal_rate_limiter.should_limit, 0);
+	/* END_HADRON */
 }
 
 #define BACK_PRESSURE_DELAY 10000L // 0.01 sec
@@ -521,7 +555,6 @@ backpressure_throttling_impl(void)
 	if (lag == 0)
 		return retry;
 
-
 	old_status = get_ps_display(&len);
 	new_status = (char *) palloc(len + 64 + 1);
 	memcpy(new_status, old_status, len);
@@ -1458,6 +1491,8 @@ XLogBroadcastWalProposer(WalProposer *wp)
 {
 	XLogRecPtr	startptr;
 	XLogRecPtr	endptr;
+	struct WalproposerShmemState *state = NULL;
+	TimestampTz now = 0;
 
 	/* Start from the last sent position */
 	startptr = sentPtr;
@@ -1502,13 +1537,36 @@ XLogBroadcastWalProposer(WalProposer *wp)
 	 * that arbitrary LSN is eventually reported as written, flushed and
 	 * applied, so that it can measure the elapsed time.
 	 */
-	LagTrackerWrite(endptr, GetCurrentTimestamp());
+	now = GetCurrentTimestamp();
+	LagTrackerWrite(endptr, now);
 
 	/* Do we have any work to do? */
 	Assert(startptr <= endptr);
 	if (endptr <= startptr)
 		return;
 
+	/* BEGIN_HADRON */
+	state = GetWalpropShmemState();
+	if (databricks_max_wal_mb_per_second != -1 && state != NULL)
+	{
+		uint64 max_wal_bytes = (uint64) databricks_max_wal_mb_per_second * 1024 * 1024;
+		struct WalRateLimiter *limiter = &state->wal_rate_limiter;
+
+		if (now - limiter->last_recorded_time_us > USECS_PER_SEC)
+		{
+			/* Reset the rate limiter */
+			limiter->last_recorded_time_us = now;
+			limiter->sent_bytes = 0;
+			pg_atomic_exchange_u32(&limiter->should_limit, 0);
+		}
+		limiter->sent_bytes += (endptr - startptr);
+		if (limiter->sent_bytes > max_wal_bytes)
+		{
+			pg_atomic_exchange_u32(&limiter->should_limit, 1);
+		}
+	}
+	/* END_HADRON */
+
 	WalProposerBroadcast(wp, startptr, endptr);
 	sentPtr = endptr;
 

proxy/Cargo.toml

@@ -48,12 +48,13 @@ indexmap = { workspace = true, features = ["serde"] }
 ipnet.workspace = true
 itertools.workspace = true
 itoa.workspace = true
-lasso = { workspace = true, features = ["multi-threaded"] }
-measured = { workspace = true, features = ["lasso"] }
+measured = { workspace = true, features = ["paracord"] }
+measured-tokio.workspace = true
 metrics.workspace = true
 once_cell.workspace = true
 opentelemetry = { workspace = true, features = ["trace"] }
 papaya = "0.2.0"
+paracord.workspace = true
 parking_lot.workspace = true
 parquet.workspace = true
 parquet_derive.workspace = true
@@ -127,4 +128,4 @@ rstest.workspace = true
 walkdir.workspace = true
 rand_distr = "0.4"
 tokio-postgres.workspace = true
-tracing-test = "0.2"
+tracing-test = "0.2"

proxy/src/auth/backend/jwt.rs

@@ -87,7 +87,12 @@ impl JwkCacheEntry {
         self.key_sets
             .values()
             // make sure our requested role has access to the key set
-            .filter(|key_set| key_set.role_names.iter().any(|role| **role == **role_name))
+            .filter(|key_set| {
+                key_set
+                    .role_names
+                    .iter()
+                    .any(|role| *role.as_str() == **role_name)
+            })
             // try and find the requested key-id in the key set
             .find_map(|key_set| {
                 key_set

proxy/src/auth/backend/local.rs

@@ -12,7 +12,7 @@ use crate::context::RequestContext;
 use crate::control_plane::NodeInfo;
 use crate::control_plane::messages::{ColdStartInfo, EndpointJwksResponse, MetricsAuxInfo};
 use crate::http;
-use crate::intern::{BranchIdTag, EndpointIdTag, InternId, ProjectIdTag};
+use crate::intern::{BranchIdInt, EndpointIdInt, ProjectIdInt};
 use crate::types::EndpointId;
 use crate::url::ApiUrl;
 
@@ -38,9 +38,9 @@ impl LocalBackend {
                 },
                 // TODO(conrad): make this better reflect compute info rather than endpoint info.
                 aux: MetricsAuxInfo {
-                    endpoint_id: EndpointIdTag::get_interner().get_or_intern("local"),
-                    project_id: ProjectIdTag::get_interner().get_or_intern("local"),
-                    branch_id: BranchIdTag::get_interner().get_or_intern("local"),
+                    endpoint_id: EndpointIdInt::new("local"),
+                    project_id: ProjectIdInt::new("local"),
+                    branch_id: BranchIdInt::new("local"),
                     compute_id: "local".into(),
                     cold_start_info: ColdStartInfo::WarmCached,
                 },

proxy/src/binary/local_proxy.rs

@@ -192,6 +192,7 @@ pub async fn run() -> anyhow::Result<()> {
             jemalloc,
             neon_metrics,
             proxy: crate::metrics::Metrics::get(),
+            tokio: measured_tokio::RuntimeCollector::current(),
         },
     ));
 

proxy/src/binary/proxy.rs

@@ -513,6 +513,7 @@ pub async fn run() -> anyhow::Result<()> {
             jemalloc,
             neon_metrics,
             proxy: crate::metrics::Metrics::get(),
+            tokio: measured_tokio::RuntimeCollector::current(),
         },
     ));
     maintenance_tasks.spawn(control_plane::mgmt::task_main(mgmt_listener));

proxy/src/cache/project_info.rs

@@ -194,7 +194,7 @@ impl ProjectInfoCacheImpl {
         &self,
         endpoint_id: &EndpointId,
     ) -> Option<Ref<'_, EndpointIdInt, EndpointInfo>> {
-        let endpoint_id = EndpointIdInt::get(endpoint_id)?;
+        let endpoint_id = EndpointIdInt::try_new_existing(endpoint_id)?;
         self.cache.get(&endpoint_id)
     }
 
@@ -204,7 +204,7 @@ impl ProjectInfoCacheImpl {
         role_name: &RoleName,
     ) -> Option<RoleAccessControl> {
         let valid_since = self.get_cache_times();
-        let role_name = RoleNameInt::get(role_name)?;
+        let role_name = RoleNameInt::try_new_existing(role_name)?;
         let endpoint_info = self.get_endpoint_cache(endpoint_id)?;
         endpoint_info.get_role_secret(role_name, valid_since)
     }
@@ -297,10 +297,10 @@ impl ProjectInfoCacheImpl {
     }
 
     pub fn maybe_invalidate_role_secret(&self, endpoint_id: &EndpointId, role_name: &RoleName) {
-        let Some(endpoint_id) = EndpointIdInt::get(endpoint_id) else {
+        let Some(endpoint_id) = EndpointIdInt::try_new_existing(endpoint_id) else {
             return;
         };
-        let Some(role_name) = RoleNameInt::get(role_name) else {
+        let Some(role_name) = RoleNameInt::try_new_existing(role_name) else {
             return;
         };
 

proxy/src/http/health_server.rs

@@ -71,6 +71,8 @@ pub struct AppMetrics {
     pub neon_metrics: NeonMetrics,
     #[metric(flatten)]
     pub proxy: &'static crate::metrics::Metrics,
+    #[metric(namespace = "tokio")]
+    pub tokio: measured_tokio::RuntimeCollector,
 }
 
 async fn prometheus_metrics_handler(

proxy/src/intern.rs

@@ -1,272 +1,59 @@
-use std::hash::BuildHasherDefault;
-use std::marker::PhantomData;
-use std::num::NonZeroUsize;
-use std::ops::Index;
-use std::sync::OnceLock;
-
-use lasso::{Capacity, MemoryLimits, Spur, ThreadedRodeo};
-use rustc_hash::FxHasher;
+use paracord::custom_key;
 
 use crate::types::{AccountId, BranchId, EndpointId, ProjectId, RoleName};
 
-pub trait InternId: Sized + 'static {
-    fn get_interner() -> &'static StringInterner<Self>;
-}
+custom_key!(pub struct RoleNameInt);
+custom_key!(pub struct EndpointIdInt);
+custom_key!(pub struct BranchIdInt);
+custom_key!(pub struct ProjectIdInt);
+custom_key!(pub struct AccountIdInt);
 
-pub struct StringInterner<Id> {
-    inner: ThreadedRodeo<Spur, BuildHasherDefault<FxHasher>>,
-    _id: PhantomData<Id>,
-}
-
-#[derive(PartialEq, Debug, Clone, Copy, Eq, Hash)]
-pub struct InternedString<Id> {
-    inner: Spur,
-    _id: PhantomData<Id>,
-}
-
-impl<Id: InternId> std::fmt::Display for InternedString<Id> {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        self.as_str().fmt(f)
-    }
-}
-
-impl<Id: InternId> InternedString<Id> {
-    pub(crate) fn as_str(&self) -> &'static str {
-        Id::get_interner().inner.resolve(&self.inner)
-    }
-    pub(crate) fn get(s: &str) -> Option<Self> {
-        Id::get_interner().get(s)
-    }
-}
-
-impl<Id: InternId> AsRef<str> for InternedString<Id> {
-    fn as_ref(&self) -> &str {
-        self.as_str()
-    }
-}
-
-impl<Id: InternId> std::ops::Deref for InternedString<Id> {
-    type Target = str;
-    fn deref(&self) -> &str {
-        self.as_str()
-    }
-}
-
-impl<'de, Id: InternId> serde::de::Deserialize<'de> for InternedString<Id> {
-    fn deserialize<D: serde::de::Deserializer<'de>>(d: D) -> Result<Self, D::Error> {
-        struct Visitor<Id>(PhantomData<Id>);
-        impl<Id: InternId> serde::de::Visitor<'_> for Visitor<Id> {
-            type Value = InternedString<Id>;
-
-            fn expecting(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-                formatter.write_str("a string")
-            }
-
-            fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
-            where
-                E: serde::de::Error,
-            {
-                Ok(Id::get_interner().get_or_intern(v))
-            }
-        }
-        d.deserialize_str(Visitor::<Id>(PhantomData))
-    }
-}
-
-impl<Id: InternId> serde::Serialize for InternedString<Id> {
-    fn serialize<S: serde::Serializer>(&self, s: S) -> Result<S::Ok, S::Error> {
-        self.as_str().serialize(s)
-    }
-}
-
-impl<Id: InternId> StringInterner<Id> {
-    pub(crate) fn new() -> Self {
-        StringInterner {
-            inner: ThreadedRodeo::with_capacity_memory_limits_and_hasher(
-                Capacity::new(2500, NonZeroUsize::new(1 << 16).expect("value is nonzero")),
-                // unbounded
-                MemoryLimits::for_memory_usage(usize::MAX),
-                BuildHasherDefault::<FxHasher>::default(),
-            ),
-            _id: PhantomData,
-        }
-    }
-
-    #[cfg(test)]
-    fn len(&self) -> usize {
-        self.inner.len()
-    }
-
-    #[cfg(test)]
-    fn current_memory_usage(&self) -> usize {
-        self.inner.current_memory_usage()
-    }
-
-    pub(crate) fn get_or_intern(&self, s: &str) -> InternedString<Id> {
-        InternedString {
-            inner: self.inner.get_or_intern(s),
-            _id: PhantomData,
-        }
-    }
-
-    pub(crate) fn get(&self, s: &str) -> Option<InternedString<Id>> {
-        Some(InternedString {
-            inner: self.inner.get(s)?,
-            _id: PhantomData,
-        })
-    }
-}
-
-impl<Id: InternId> Index<InternedString<Id>> for StringInterner<Id> {
-    type Output = str;
-
-    fn index(&self, index: InternedString<Id>) -> &Self::Output {
-        self.inner.resolve(&index.inner)
-    }
-}
-
-impl<Id: InternId> Default for StringInterner<Id> {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
-#[derive(Copy, Clone, Debug, PartialEq, Eq, Hash)]
-pub struct RoleNameTag;
-impl InternId for RoleNameTag {
-    fn get_interner() -> &'static StringInterner<Self> {
-        static ROLE_NAMES: OnceLock<StringInterner<RoleNameTag>> = OnceLock::new();
-        ROLE_NAMES.get_or_init(Default::default)
-    }
-}
-pub type RoleNameInt = InternedString<RoleNameTag>;
 impl From<&RoleName> for RoleNameInt {
     fn from(value: &RoleName) -> Self {
-        RoleNameTag::get_interner().get_or_intern(value)
+        RoleNameInt::new(value)
     }
 }
 
-#[derive(Copy, Clone, Debug, PartialEq, Eq, Hash)]
-pub struct EndpointIdTag;
-impl InternId for EndpointIdTag {
-    fn get_interner() -> &'static StringInterner<Self> {
-        static ROLE_NAMES: OnceLock<StringInterner<EndpointIdTag>> = OnceLock::new();
-        ROLE_NAMES.get_or_init(Default::default)
-    }
-}
-pub type EndpointIdInt = InternedString<EndpointIdTag>;
 impl From<&EndpointId> for EndpointIdInt {
     fn from(value: &EndpointId) -> Self {
-        EndpointIdTag::get_interner().get_or_intern(value)
+        EndpointIdInt::new(value)
     }
 }
 impl From<EndpointId> for EndpointIdInt {
     fn from(value: EndpointId) -> Self {
-        EndpointIdTag::get_interner().get_or_intern(&value)
+        EndpointIdInt::new(&value)
     }
 }
 
-#[derive(Copy, Clone, Debug, PartialEq, Eq, Hash)]
-pub struct BranchIdTag;
-impl InternId for BranchIdTag {
-    fn get_interner() -> &'static StringInterner<Self> {
-        static ROLE_NAMES: OnceLock<StringInterner<BranchIdTag>> = OnceLock::new();
-        ROLE_NAMES.get_or_init(Default::default)
-    }
-}
-pub type BranchIdInt = InternedString<BranchIdTag>;
 impl From<&BranchId> for BranchIdInt {
     fn from(value: &BranchId) -> Self {
-        BranchIdTag::get_interner().get_or_intern(value)
+        BranchIdInt::new(value)
     }
 }
 impl From<BranchId> for BranchIdInt {
     fn from(value: BranchId) -> Self {
-        BranchIdTag::get_interner().get_or_intern(&value)
+        BranchIdInt::new(&value)
     }
 }
 
-#[derive(Copy, Clone, Debug, PartialEq, Eq, Hash)]
-pub struct ProjectIdTag;
-impl InternId for ProjectIdTag {
-    fn get_interner() -> &'static StringInterner<Self> {
-        static ROLE_NAMES: OnceLock<StringInterner<ProjectIdTag>> = OnceLock::new();
-        ROLE_NAMES.get_or_init(Default::default)
-    }
-}
-pub type ProjectIdInt = InternedString<ProjectIdTag>;
 impl From<&ProjectId> for ProjectIdInt {
     fn from(value: &ProjectId) -> Self {
-        ProjectIdTag::get_interner().get_or_intern(value)
+        ProjectIdInt::new(value)
     }
 }
 impl From<ProjectId> for ProjectIdInt {
     fn from(value: ProjectId) -> Self {
-        ProjectIdTag::get_interner().get_or_intern(&value)
+        ProjectIdInt::new(&value)
     }
 }
 
-#[derive(Copy, Clone, Debug, PartialEq, Eq, Hash)]
-pub struct AccountIdTag;
-impl InternId for AccountIdTag {
-    fn get_interner() -> &'static StringInterner<Self> {
-        static ROLE_NAMES: OnceLock<StringInterner<AccountIdTag>> = OnceLock::new();
-        ROLE_NAMES.get_or_init(Default::default)
-    }
-}
-pub type AccountIdInt = InternedString<AccountIdTag>;
 impl From<&AccountId> for AccountIdInt {
     fn from(value: &AccountId) -> Self {
-        AccountIdTag::get_interner().get_or_intern(value)
+        AccountIdInt::new(value)
     }
 }
 impl From<AccountId> for AccountIdInt {
     fn from(value: AccountId) -> Self {
-        AccountIdTag::get_interner().get_or_intern(&value)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::sync::OnceLock;
-
-    use super::InternId;
-    use crate::intern::StringInterner;
-
-    struct MyId;
-    impl InternId for MyId {
-        fn get_interner() -> &'static StringInterner<Self> {
-            pub(crate) static ROLE_NAMES: OnceLock<StringInterner<MyId>> = OnceLock::new();
-            ROLE_NAMES.get_or_init(Default::default)
-        }
-    }
-
-    #[test]
-    fn push_many_strings() {
-        use rand::rngs::StdRng;
-        use rand::{Rng, SeedableRng};
-        use rand_distr::Zipf;
-
-        let endpoint_dist = Zipf::new(500000, 0.8).unwrap();
-        let endpoints = StdRng::seed_from_u64(272488357).sample_iter(endpoint_dist);
-
-        let interner = MyId::get_interner();
-
-        const N: usize = 100_000;
-        let mut verify = Vec::with_capacity(N);
-        for endpoint in endpoints.take(N) {
-            let endpoint = format!("ep-string-interning-{endpoint}");
-            let key = interner.get_or_intern(&endpoint);
-            verify.push((endpoint, key));
-        }
-
-        for (s, key) in verify {
-            assert_eq!(interner[key], s);
-        }
-
-        // 2031616/59861 = 34 bytes per string
-        assert_eq!(interner.len(), 59_861);
-        // will have other overhead for the internal hashmaps that are not accounted for.
-        assert_eq!(interner.current_memory_usage(), 2_031_616);
+        AccountIdInt::new(&value)
     }
 }

proxy/src/metrics.rs

@@ -1,6 +1,5 @@
 use std::sync::{Arc, OnceLock};
 
-use lasso::ThreadedRodeo;
 use measured::label::{
     FixedCardinalitySet, LabelGroupSet, LabelName, LabelSet, LabelValue, StaticLabelSet,
 };
@@ -11,6 +10,7 @@ use measured::{
     MetricGroup,
 };
 use metrics::{CounterPairAssoc, CounterPairVec, HyperLogLogVec};
+use paracord::ParaCord;
 use tokio::time::{self, Instant};
 
 use crate::control_plane::messages::ColdStartInfo;
@@ -222,7 +222,7 @@ pub enum CacheOutcome {
 #[derive(LabelGroup)]
 #[label(set = ConsoleRequestSet)]
 pub struct ConsoleRequest<'a> {
-    #[label(dynamic_with = ThreadedRodeo, default)]
+    #[label(dynamic_with = ParaCord, default)]
     pub request: &'a str,
 }
 
@@ -345,7 +345,7 @@ pub struct ConnectionFailuresBreakdownGroup {
 #[derive(LabelGroup, Copy, Clone)]
 #[label(set = RedisErrorsSet)]
 pub struct RedisErrors<'a> {
-    #[label(dynamic_with = ThreadedRodeo, default)]
+    #[label(dynamic_with = ParaCord, default)]
     pub channel: &'a str,
 }
 

proxy/src/types.rs

@@ -1,4 +1,4 @@
-use crate::intern::{EndpointIdInt, EndpointIdTag, InternId};
+use crate::intern::EndpointIdInt;
 
 macro_rules! smol_str_wrapper {
     ($name:ident) => {
@@ -85,7 +85,7 @@ impl EndpointId {
 
     #[must_use]
     pub fn normalize_intern(&self) -> EndpointIdInt {
-        EndpointIdTag::get_interner().get_or_intern(self.normalize_str())
+        EndpointIdInt::new(self.normalize_str())
     }
 }
 

safekeeper/src/bin/safekeeper.rs

@@ -37,6 +37,7 @@ use tracing::*;
 use utils::auth::{JwtAuth, Scope, SwappableJwtAuth};
 use utils::id::NodeId;
 use utils::logging::{self, LogFormat, SecretString};
+use utils::metrics_collector::{METRICS_COLLECTION_INTERVAL, METRICS_COLLECTOR};
 use utils::sentry_init::init_sentry;
 use utils::{pid_file, project_build_tag, project_git_version, tcp_listener};
 
@@ -243,6 +244,11 @@ struct Args {
     #[arg(long)]
     enable_tls_wal_service_api: bool,
 
+    /// Controls whether to collect all metrics on each scrape or to return potentially stale
+    /// results.
+    #[arg(long, default_value_t = true)]
+    force_metric_collection_on_scrape: bool,
+
     /// Run in development mode (disables security checks)
     #[arg(long, help = "Run in development mode (disables security checks)")]
     dev: bool,
@@ -428,6 +434,7 @@ async fn main() -> anyhow::Result<()> {
         ssl_ca_certs,
         use_https_safekeeper_api: args.use_https_safekeeper_api,
         enable_tls_wal_service_api: args.enable_tls_wal_service_api,
+        force_metric_collection_on_scrape: args.force_metric_collection_on_scrape,
     });
 
     // initialize sentry if SENTRY_DSN is provided
@@ -640,6 +647,26 @@ async fn start_safekeeper(conf: Arc<SafeKeeperConf>) -> Result<()> {
         .map(|res| ("broker main".to_owned(), res));
     tasks_handles.push(Box::pin(broker_task_handle));
 
+    /* BEGIN_HADRON */
+    if conf.force_metric_collection_on_scrape {
+        let metrics_handle = current_thread_rt
+            .as_ref()
+            .unwrap_or_else(|| BACKGROUND_RUNTIME.handle())
+            .spawn(async move {
+                let mut interval: tokio::time::Interval =
+                    tokio::time::interval(METRICS_COLLECTION_INTERVAL);
+                loop {
+                    interval.tick().await;
+                    tokio::task::spawn_blocking(|| {
+                        METRICS_COLLECTOR.run_once(true);
+                    });
+                }
+            })
+            .map(|res| ("broker main".to_owned(), res));
+        tasks_handles.push(Box::pin(metrics_handle));
+    }
+    /* END_HADRON */
+
     set_build_info_metric(GIT_VERSION, BUILD_TAG);
 
     // TODO: update tokio-stream, convert to real async Stream with

safekeeper/src/http/routes.rs

@@ -699,6 +699,11 @@ pub fn make_router(
         }))
     }
 
+    let force_metric_collection_on_scrape = conf.force_metric_collection_on_scrape;
+
+    let prometheus_metrics_handler_wrapper =
+        move |req| prometheus_metrics_handler(req, force_metric_collection_on_scrape);
+
     // NB: on any changes do not forget to update the OpenAPI spec
     // located nearby (/safekeeper/src/http/openapi_spec.yaml).
     let auth = conf.http_auth.clone();
@@ -706,7 +711,9 @@ pub fn make_router(
         .data(conf)
         .data(global_timelines)
         .data(auth)
-        .get("/metrics", |r| request_span(r, prometheus_metrics_handler))
+        .get("/metrics", move |r| {
+            request_span(r, prometheus_metrics_handler_wrapper)
+        })
         .get("/profile/cpu", |r| request_span(r, profile_cpu_handler))
         .get("/profile/heap", |r| request_span(r, profile_heap_handler))
         .get("/v1/status", |r| request_span(r, status_handler))

safekeeper/src/lib.rs

@@ -134,6 +134,7 @@ pub struct SafeKeeperConf {
     pub ssl_ca_certs: Vec<Pem>,
     pub use_https_safekeeper_api: bool,
     pub enable_tls_wal_service_api: bool,
+    pub force_metric_collection_on_scrape: bool,
 }
 
 impl SafeKeeperConf {
@@ -183,6 +184,7 @@ impl SafeKeeperConf {
             ssl_ca_certs: Vec::new(),
             use_https_safekeeper_api: false,
             enable_tls_wal_service_api: false,
+            force_metric_collection_on_scrape: true,
         }
     }
 }

safekeeper/src/metrics.rs

@@ -59,6 +59,15 @@ pub static FLUSH_WAL_SECONDS: Lazy<Histogram> = Lazy::new(|| {
     .expect("Failed to register safekeeper_flush_wal_seconds histogram")
 });
 /* BEGIN_HADRON */
+// Counter of all ProposerAcceptorMessage requests received
+pub static PROPOSER_ACCEPTOR_MESSAGES_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "safekeeper_proposer_acceptor_messages_total",
+        "Total number of ProposerAcceptorMessage requests received by the Safekeeper.",
+        &["outcome"]
+    )
+    .expect("Failed to register safekeeper_proposer_acceptor_messages_total counter")
+});
 pub static WAL_DISK_IO_ERRORS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "safekeeper_wal_disk_io_errors",

safekeeper/src/safekeeper.rs

@@ -24,7 +24,7 @@ use utils::id::{NodeId, TenantId, TimelineId};
 use utils::lsn::Lsn;
 use utils::pageserver_feedback::PageserverFeedback;
 
-use crate::metrics::MISC_OPERATION_SECONDS;
+use crate::metrics::{MISC_OPERATION_SECONDS, PROPOSER_ACCEPTOR_MESSAGES_TOTAL};
 use crate::state::TimelineState;
 use crate::{control_file, wal_storage};
 
@@ -938,7 +938,7 @@ where
         &mut self,
         msg: &ProposerAcceptorMessage,
     ) -> Result<Option<AcceptorProposerMessage>> {
-        match msg {
+        let res = match msg {
             ProposerAcceptorMessage::Greeting(msg) => self.handle_greeting(msg).await,
             ProposerAcceptorMessage::VoteRequest(msg) => self.handle_vote_request(msg).await,
             ProposerAcceptorMessage::Elected(msg) => self.handle_elected(msg).await,
@@ -949,7 +949,20 @@ where
                 self.handle_append_request(msg, false).await
             }
             ProposerAcceptorMessage::FlushWAL => self.handle_flush().await,
-        }
+        };
+
+        // BEGIN HADRON
+        match &res {
+            Ok(_) => PROPOSER_ACCEPTOR_MESSAGES_TOTAL
+                .with_label_values(&["success"])
+                .inc(),
+            Err(_) => PROPOSER_ACCEPTOR_MESSAGES_TOTAL
+                .with_label_values(&["error"])
+                .inc(),
+        };
+
+        res
+        // END HADRON
     }
 
     /// Handle initial message from proposer: check its sanity and send my

safekeeper/src/wal_backup.rs

@@ -166,7 +166,7 @@ fn hadron_determine_offloader(mgr: &Manager, state: &StateSnapshot) -> (Option<N
 
     let backup_lag = state.commit_lsn.checked_sub(state.backup_lsn);
     if backup_lag.is_none() {
-        info!("Backup lag is None. Skipping re-election.");
+        debug!("Backup lag is None. Skipping re-election.");
         return (offloader, election_dbg_str);
     }
 

safekeeper/tests/walproposer_sim/safekeeper.rs

@@ -190,6 +190,7 @@ pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
         ssl_ca_certs: Vec::new(),
         use_https_safekeeper_api: false,
         enable_tls_wal_service_api: false,
+        force_metric_collection_on_scrape: true,
     };
 
     let mut global = GlobalMap::new(disk, conf.clone())?;

storage_controller/Cargo.toml

@@ -31,10 +31,10 @@ humantime.workspace = true
 humantime-serde.workspace = true
 itertools.workspace = true
 json-structural-diff.workspace = true
-lasso.workspace = true
 once_cell.workspace = true
 pageserver_api.workspace = true
 pageserver_client.workspace = true
+paracord.workspace = true
 postgres_connection.workspace = true
 posthog_client_lite.workspace = true
 rand.workspace = true
@@ -72,4 +72,4 @@ http-utils = { path = "../libs/http-utils/" }
 utils = { path = "../libs/utils/" }
 metrics = { path = "../libs/metrics/" }
 control_plane = { path = "../control_plane" }
-workspace_hack = { version = "0.1", path = "../workspace_hack" }
+workspace_hack = { version = "0.1", path = "../workspace_hack" }

storage_controller/src/http.rs

@@ -850,6 +850,31 @@ async fn handle_tenant_describe(
     json_response(StatusCode::OK, service.tenant_describe(tenant_id)?)
 }
 
+/* BEGIN_HADRON */
+async fn handle_tenant_timeline_describe(
+    service: Arc<Service>,
+    req: Request<Body>,
+) -> Result<Response<Body>, ApiError> {
+    check_permissions(&req, Scope::Scrubber)?;
+
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    let timeline_id: TimelineId = parse_request_param(&req, "timeline_id")?;
+    match maybe_forward(req).await {
+        ForwardOutcome::Forwarded(res) => {
+            return res;
+        }
+        ForwardOutcome::NotForwarded(_req) => {}
+    };
+
+    json_response(
+        StatusCode::OK,
+        service
+            .tenant_timeline_describe(tenant_id, timeline_id)
+            .await?,
+    )
+}
+/* END_HADRON */
+
 async fn handle_tenant_list(
     service: Arc<Service>,
     req: Request<Body>,
@@ -2480,6 +2505,13 @@ pub fn make_router(
             )
         })
         // Timeline operations
+        .get("/control/v1/tenant/:tenant_id/timeline/:timeline_id", |r| {
+            tenant_service_handler(
+                r,
+                handle_tenant_timeline_describe,
+                RequestName("v1_tenant_timeline_describe"),
+            )
+        })
         .delete("/v1/tenant/:tenant_id/timeline/:timeline_id", |r| {
             tenant_service_handler(
                 r,

storage_controller/src/metrics.rs

@@ -183,9 +183,9 @@ impl Default for StorageControllerMetrics {
 #[derive(measured::LabelGroup, Clone)]
 #[label(set = NodeLabelGroupSet)]
 pub(crate) struct NodeLabelGroup<'a> {
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) az: &'a str,
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) node_id: &'a str,
 }
 
@@ -198,7 +198,7 @@ pub(crate) struct ReconcileCompleteLabelGroup {
 #[derive(measured::LabelGroup)]
 #[label(set = HttpRequestStatusLabelGroupSet)]
 pub(crate) struct HttpRequestStatusLabelGroup<'a> {
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) path: &'a str,
     pub(crate) method: Method,
     pub(crate) status: StatusCode,
@@ -207,7 +207,7 @@ pub(crate) struct HttpRequestStatusLabelGroup<'a> {
 #[derive(measured::LabelGroup)]
 #[label(set = HttpRequestLatencyLabelGroupSet)]
 pub(crate) struct HttpRequestLatencyLabelGroup<'a> {
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) path: &'a str,
     pub(crate) method: Method,
 }
@@ -215,9 +215,9 @@ pub(crate) struct HttpRequestLatencyLabelGroup<'a> {
 #[derive(measured::LabelGroup, Clone)]
 #[label(set = PageserverRequestLabelGroupSet)]
 pub(crate) struct PageserverRequestLabelGroup<'a> {
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) pageserver_id: &'a str,
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) path: &'a str,
     pub(crate) method: Method,
 }
@@ -225,9 +225,9 @@ pub(crate) struct PageserverRequestLabelGroup<'a> {
 #[derive(measured::LabelGroup, Clone)]
 #[label(set = SafekeeperRequestLabelGroupSet)]
 pub(crate) struct SafekeeperRequestLabelGroup<'a> {
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) safekeeper_id: &'a str,
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) path: &'a str,
     pub(crate) method: Method,
 }
@@ -254,11 +254,11 @@ pub(crate) struct LeadershipStatusGroup {
 #[derive(measured::LabelGroup, Clone)]
 #[label(set = ReconcileLongRunningLabelGroupSet)]
 pub(crate) struct ReconcileLongRunningLabelGroup<'a> {
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) tenant_id: &'a str,
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) shard_number: &'a str,
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) sequence: &'a str,
 }
 
@@ -282,11 +282,11 @@ pub(crate) enum Method {
 #[derive(measured::LabelGroup, Clone)]
 #[label(set = SafekeeperReconcilerLabelGroupSet)]
 pub(crate) struct SafekeeperReconcilerLabelGroup<'a> {
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) sk_az: &'a str,
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) sk_node_id: &'a str,
-    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    #[label(dynamic_with = paracord::ParaCord, default)]
     pub(crate) sk_hostname: &'a str,
 }
 

storage_controller/src/pageserver_client.rs

@@ -86,6 +86,23 @@ impl PageserverClient {
         )
     }
 
+    /* BEGIN_HADRON */
+    pub(crate) async fn tenant_timeline_describe(
+        &self,
+        tenant_shard_id: &TenantShardId,
+        timeline_id: &TimelineId,
+    ) -> Result<TimelineInfo> {
+        measured_request!(
+            "tenant_timeline_describe",
+            crate::metrics::Method::Get,
+            &self.node_id_label,
+            self.inner
+                .tenant_timeline_describe(tenant_shard_id, timeline_id,)
+                .await
+        )
+    }
+    /* END_HADRON */
+
     pub(crate) async fn tenant_scan_remote_storage(
         &self,
         tenant_id: TenantId,

storage_controller/src/service.rs

@@ -32,7 +32,7 @@ use pageserver_api::controller_api::{
     ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
     SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse, TenantCreateResponseShard,
     TenantDescribeResponse, TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest,
-    TenantShardMigrateRequest, TenantShardMigrateResponse,
+    TenantShardMigrateRequest, TenantShardMigrateResponse, TenantTimelineDescribeResponse,
 };
 use pageserver_api::models::{
     self, DetachBehavior, LocationConfig, LocationConfigListResponse, LocationConfigMode, LsnLease,
@@ -1984,11 +1984,14 @@ impl Service {
         });
 
         // Check that there is enough safekeepers configured that we can create new timelines
-        let test_sk_res = this.safekeepers_for_new_timeline().await;
+        let test_sk_res_str = match this.safekeepers_for_new_timeline().await {
+            Ok(v) => format!("Ok({v:?})"),
+            Err(v) => format!("Err({v:})"),
+        };
         tracing::info!(
             timeline_safekeeper_count = config.timeline_safekeeper_count,
             timelines_onto_safekeepers = config.timelines_onto_safekeepers,
-            "viability test result (test timeline creation on safekeepers): {test_sk_res:?}",
+            "viability test result (test timeline creation on safekeepers): {test_sk_res_str}",
         );
 
         Ok(this)
@@ -4758,6 +4761,7 @@ impl Service {
         )
         .await;
 
+        let mut retry_if_not_attached = false;
         let targets = {
             let locked = self.inner.read().unwrap();
             let mut targets = Vec::new();
@@ -4774,6 +4778,24 @@ impl Service {
                         .expect("Pageservers may not be deleted while referenced");
 
                     targets.push((*tenant_shard_id, node.clone()));
+
+                    if let Some(location) = shard.observed.locations.get(node_id) {
+                        if let Some(ref conf) = location.conf {
+                            if conf.mode != LocationConfigMode::AttachedSingle
+                                && conf.mode != LocationConfigMode::AttachedMulti
+                            {
+                                // If the shard is attached as secondary, we need to retry if 404.
+                                retry_if_not_attached = true;
+                            }
+                            // If the shard is attached as primary, we should succeed.
+                        } else {
+                            // Location conf is not available yet, retry if 404.
+                            retry_if_not_attached = true;
+                        }
+                    } else {
+                        // The shard is not attached to the intended pageserver yet, retry if 404.
+                        retry_if_not_attached = true;
+                    }
                 }
             }
             targets
@@ -4804,6 +4826,18 @@ impl Service {
                         valid_until = Some(lease.valid_until);
                     }
                 }
+                Err(mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _))
+                    if retry_if_not_attached =>
+                {
+                    // This is expected if the attach is not finished yet. Return 503 so that the client can retry.
+                    return Err(ApiError::ResourceUnavailable(
+                        format!(
+                            "Timeline is not attached to the pageserver {} yet, please retry",
+                            node.get_id()
+                        )
+                        .into(),
+                    ));
+                }
                 Err(e) => {
                     return Err(passthrough_api_error(&node, e));
                 }
@@ -5452,6 +5486,92 @@ impl Service {
         .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
     }
 
+    /* BEGIN_HADRON */
+    pub(crate) async fn tenant_timeline_describe(
+        &self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> Result<TenantTimelineDescribeResponse, ApiError> {
+        self.tenant_remote_mutation(tenant_id, |locations| async move {
+            if locations.0.is_empty() {
+                return Err(ApiError::NotFound(
+                    anyhow::anyhow!("Tenant not found").into(),
+                ));
+            };
+
+            let locations: Vec<(TenantShardId, Node)> = locations
+                .0
+                .iter()
+                .map(|t| (*t.0, t.1.latest.node.clone()))
+                .collect();
+            let mut futs = FuturesUnordered::new();
+
+            for (shard_id, node) in locations {
+                futs.push({
+                    async move {
+                        let result = node
+                            .with_client_retries(
+                                |client| async move {
+                                    client
+                                        .tenant_timeline_describe(&shard_id, &timeline_id)
+                                        .await
+                                },
+                                &self.http_client,
+                                &self.config.pageserver_jwt_token,
+                                3,
+                                3,
+                                Duration::from_secs(30),
+                                &self.cancel,
+                            )
+                            .await;
+                        (result, shard_id, node.get_id())
+                    }
+                });
+            }
+
+            let mut results: Vec<TimelineInfo> = Vec::new();
+            while let Some((result, tenant_shard_id, node_id)) = futs.next().await {
+                match result {
+                    Some(Ok(timeline_info)) => results.push(timeline_info),
+                    Some(Err(e)) => {
+                        tracing::warn!(
+                            "Failed to describe tenant {} timeline {} for pageserver {}: {e}",
+                            tenant_shard_id,
+                            timeline_id,
+                            node_id,
+                        );
+                        return Err(ApiError::ResourceUnavailable(format!("{e}").into()));
+                    }
+                    None => return Err(ApiError::Cancelled),
+                }
+            }
+            let mut image_consistent_lsn: Option<Lsn> = Some(Lsn::MAX);
+            for timeline_info in &results {
+                if let Some(tline_image_consistent_lsn) = timeline_info.image_consistent_lsn {
+                    image_consistent_lsn = Some(std::cmp::min(
+                        image_consistent_lsn.unwrap(),
+                        tline_image_consistent_lsn,
+                    ));
+                } else {
+                    tracing::warn!(
+                        "Timeline {} on shard {} does not have image consistent lsn",
+                        timeline_info.timeline_id,
+                        timeline_info.tenant_id
+                    );
+                    image_consistent_lsn = None;
+                    break;
+                }
+            }
+
+            Ok(TenantTimelineDescribeResponse {
+                shards: results,
+                image_consistent_lsn,
+            })
+        })
+        .await?
+    }
+    /* END_HADRON */
+
     /// 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

storage_controller/src/service/safekeeper_service.rs

@@ -25,7 +25,8 @@ use pageserver_api::models::{SafekeeperInfo, SafekeepersInfo, TimelineInfo};
 use safekeeper_api::PgVersionId;
 use safekeeper_api::membership::{self, MemberSet, SafekeeperGeneration};
 use safekeeper_api::models::{
-    PullTimelineRequest, TimelineMembershipSwitchRequest, TimelineMembershipSwitchResponse,
+    PullTimelineRequest, TimelineLocateResponse, TimelineMembershipSwitchRequest,
+    TimelineMembershipSwitchResponse,
 };
 use safekeeper_api::{INITIAL_TERM, Term};
 use safekeeper_client::mgmt_api;
@@ -37,13 +38,6 @@ use utils::lsn::Lsn;
 
 use super::Service;
 
-#[derive(serde::Serialize, serde::Deserialize, Clone)]
-pub struct TimelineLocateResponse {
-    pub generation: SafekeeperGeneration,
-    pub sk_set: Vec<NodeId>,
-    pub new_sk_set: Option<Vec<NodeId>>,
-}
-
 impl Service {
     fn make_member_set(safekeepers: &[Safekeeper]) -> Result<MemberSet, ApiError> {
         let members = safekeepers

test_runner/fixtures/neon_fixtures.py

@@ -1795,6 +1795,33 @@ def neon_env_builder(
         record_property("preserve_database_files", builder.preserve_database_files)
 
 
+@pytest.fixture(scope="function")
+def neon_env_builder_local(
+    neon_env_builder: NeonEnvBuilder,
+    test_output_dir: Path,
+    pg_distrib_dir: Path,
+) -> NeonEnvBuilder:
+    """
+    Fixture to create a Neon environment for test with its own pg_install copy.
+
+    This allows the test to edit the list of available extensions in the
+    local instance of Postgres used for the test, and install extensions via
+    downloading them when a remote extension is tested, for instance, or
+    copying files around for local extension testing.
+    """
+    test_local_pginstall = test_output_dir / "pg_install"
+    log.info(f"copy {pg_distrib_dir} to {test_local_pginstall}")
+
+    # We can't copy only the version that we are currently testing because other
+    # binaries like the storage controller need specific Postgres versions.
+    shutil.copytree(pg_distrib_dir, test_local_pginstall)
+
+    neon_env_builder.pg_distrib_dir = test_local_pginstall
+    log.info(f"local neon_env_builder.pg_distrib_dir: {neon_env_builder.pg_distrib_dir}")
+
+    return neon_env_builder
+
+
 @dataclass
 class PageserverPort:
     pg: int
@@ -2315,6 +2342,20 @@ class NeonStorageController(MetricsGetter, LogUtils):
         response.raise_for_status()
         return response.json()
 
+    # HADRON
+    def tenant_timeline_describe(
+        self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ):
+        response = self.request(
+            "GET",
+            f"{self.api}/control/v1/tenant/{tenant_id}/timeline/{timeline_id}",
+            headers=self.headers(TokenScope.ADMIN),
+        )
+        response.raise_for_status()
+        return response.json()
+
     def nodes(self):
         """
         :return: list of {"id": ""}

test_runner/fixtures/pageserver/http.py

@@ -333,6 +333,13 @@ class PageserverHttpClient(requests.Session, MetricsGetter):
         res = self.post(f"http://localhost:{self.port}/v1/reload_auth_validation_keys")
         self.verbose_error(res)
 
+    def list_tenant_visible_size(self) -> dict[TenantShardId, int]:
+        res = self.get(f"http://localhost:{self.port}/v1/list_tenant_visible_size")
+        self.verbose_error(res)
+        res_json = res.json()
+        assert isinstance(res_json, dict)
+        return res_json
+
     def tenant_list(self) -> list[dict[Any, Any]]:
         res = self.get(f"http://localhost:{self.port}/v1/tenant")
         self.verbose_error(res)
@@ -1002,7 +1009,7 @@ class PageserverHttpClient(requests.Session, MetricsGetter):
 
     def get_metrics_str(self) -> str:
         """You probably want to use get_metrics() instead."""
-        res = self.get(f"http://localhost:{self.port}/metrics")
+        res = self.get(f"http://localhost:{self.port}/metrics?use_latest=true")
         self.verbose_error(res)
         return res.text
 

test_runner/fixtures/port_distributor.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import re
 import socket
 from contextlib import closing
+from itertools import cycle
 
 from fixtures.log_helper import log
 
@@ -34,15 +35,23 @@ def can_bind(host: str, port: int) -> bool:
 
 class PortDistributor:
     def __init__(self, base_port: int, port_number: int):
-        self.iterator = iter(range(base_port, base_port + port_number))
+        self.base_port = base_port
+        self.port_number = port_number
+        self.cycle = cycle(range(base_port, base_port + port_number))
         self.port_map: dict[int, int] = {}
 
     def get_port(self) -> int:
-        for port in self.iterator:
+        checked = 0
+        for port in self.cycle:
             if can_bind("localhost", port):
                 return port
+            elif checked < self.port_number:
+                checked += 1
+            else:
+                break
+
         raise RuntimeError(
-            "port range configured for test is exhausted, consider enlarging the range"
+            f"port range ({self.base_port}..{self.base_port + self.port_number}) configured for test is exhausted, consider enlarging the range"
         )
 
     def replace_with_new_port(self, value: int | str) -> int | str:

test_runner/fixtures/safekeeper/http.py

@@ -143,7 +143,7 @@ class SafekeeperHttpClient(requests.Session, MetricsGetter):
 
     def get_metrics_str(self) -> str:
         """You probably want to use get_metrics() instead."""
-        request_result = self.get(f"http://localhost:{self.port}/metrics")
+        request_result = self.get(f"http://localhost:{self.port}/metrics?use_latest=true")
         request_result.raise_for_status()
         return request_result.text
 

test_runner/regress/data/test_event_trigger_extension/test_event_trigger_extension--1.0.sql

@@ -0,0 +1,32 @@
+\echo Use "CREATE EXTENSION test_event_trigger_extension" to load this file. \quit
+
+CREATE SCHEMA event_trigger;
+
+create sequence if not exists event_trigger.seq_schema_version as int cycle;
+
+create or replace function event_trigger.increment_schema_version()
+    returns event_trigger
+    security definer
+    language plpgsql
+as $$
+begin
+    perform pg_catalog.nextval('event_trigger.seq_schema_version');
+end;
+$$;
+
+create or replace function event_trigger.get_schema_version()
+    returns int
+    security definer
+    language sql
+as $$
+    select last_value from event_trigger.seq_schema_version;
+$$;
+
+-- On DDL event, increment the schema version number
+create event trigger event_trigger_watch_ddl
+    on ddl_command_end
+    execute procedure event_trigger.increment_schema_version();
+
+create event trigger event_trigger_watch_drop
+    on sql_drop
+    execute procedure event_trigger.increment_schema_version();

test_runner/regress/data/test_event_trigger_extension/test_event_trigger_extension.control

@@ -0,0 +1,8 @@
+default_version = '1.0'
+comment = 'Test extension with Event Trigger'
+
+# make sure the extension objects are owned by the bootstrap user
+# to check that the SECURITY DEFINER event trigger function is still
+# called during non-superuser DDL events.
+superuser = true
+trusted = true

test_runner/regress/test_attach_tenant_config.py

@@ -165,6 +165,7 @@ def test_fully_custom_config(positive_env: NeonEnv):
         "gc_horizon": 23 * (1024 * 1024),
         "gc_period": "2h 13m",
         "image_creation_threshold": 7,
+        "image_layer_force_creation_period": "1m",
         "pitr_interval": "1m",
         "lagging_wal_timeout": "23m",
         "lazy_slru_download": True,

test_runner/regress/test_compaction.py

@@ -944,3 +944,143 @@ def test_image_layer_compression(neon_env_builder: NeonEnvBuilder, enabled: bool
                 f"SELECT count(*) FROM foo WHERE id={v} and val=repeat('abcde{v:0>3}', 500)"
             )
             assert res[0][0] == 1
+
+
+# BEGIN_HADRON
+def get_layer_map(env, tenant_shard_id, timeline_id, ps_id):
+    client = env.pageservers[ps_id].http_client()
+    layer_map = client.layer_map_info(tenant_shard_id, timeline_id)
+    image_layer_count = 0
+    delta_layer_count = 0
+    for layer in layer_map.historic_layers:
+        if layer.kind == "Image":
+            image_layer_count += 1
+        elif layer.kind == "Delta":
+            delta_layer_count += 1
+    return image_layer_count, delta_layer_count
+
+
+def test_image_layer_force_creation_period(neon_env_builder: NeonEnvBuilder):
+    """
+    Tests that page server can force creating new images if image_layer_force_creation_period is enabled
+    """
+    # use large knobs to disable L0 compaction/image creation except for the force image creation
+    tenant_conf = {
+        "compaction_threshold": "100",
+        "image_creation_threshold": "100",
+        "image_layer_creation_check_threshold": "1",
+        "checkpoint_distance": 10 * 1024,
+        "checkpoint_timeout": "1s",
+        "image_layer_force_creation_period": "1s",
+        "pitr_interval": "10s",
+        "gc_period": "1s",
+        "compaction_period": "1s",
+        "lsn_lease_length": "1s",
+    }
+
+    # consider every tenant large to run the image layer generation check more eagerly
+    neon_env_builder.pageserver_config_override = (
+        "image_layer_generation_large_timeline_threshold=0"
+    )
+
+    neon_env_builder.num_pageservers = 1
+    neon_env_builder.num_safekeepers = 1
+    env = neon_env_builder.init_start(initial_tenant_conf=tenant_conf)
+
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
+
+    endpoint = env.endpoints.create_start("main")
+    endpoint.safe_psql("CREATE TABLE foo (id INTEGER PRIMARY KEY, val text)")
+    # Generate some rows.
+    for v in range(10):
+        endpoint.safe_psql(f"INSERT INTO foo (id, val) VALUES ({v}, repeat('abcde{v:0>3}', 500))")
+
+    # Sleep a bit such that the inserts are considered when calculating the forced image layer creation LSN.
+    time.sleep(2)
+
+    def check_force_image_creation():
+        ps_http = env.pageserver.http_client()
+        ps_http.timeline_compact(tenant_id, timeline_id)
+        image, delta = get_layer_map(env, tenant_id, timeline_id, 0)
+        log.info(f"images: {image}, deltas: {delta}")
+        assert image > 0
+
+        env.pageserver.assert_log_contains("forcing L0 compaction of")
+        env.pageserver.assert_log_contains("forcing image creation for partitioned range")
+
+    wait_until(check_force_image_creation)
+
+    endpoint.stop_and_destroy()
+
+    env.pageserver.allowed_errors.append(
+        ".*created delta file of size.*larger than double of target.*"
+    )
+
+
+def test_image_consistent_lsn(neon_env_builder: NeonEnvBuilder):
+    """
+    Test the /v1/tenant/<tenant_id>/timeline/<timeline_id> endpoint and the computation of image_consistent_lsn
+    """
+    # use large knobs to disable L0 compaction/image creation except for the force image creation
+    tenant_conf = {
+        "compaction_threshold": "100",
+        "image_creation_threshold": "100",
+        "image_layer_creation_check_threshold": "1",
+        "checkpoint_distance": 10 * 1024,
+        "checkpoint_timeout": "1s",
+        "image_layer_force_creation_period": "1s",
+        "pitr_interval": "10s",
+        "gc_period": "1s",
+        "compaction_period": "1s",
+        "lsn_lease_length": "1s",
+    }
+
+    neon_env_builder.num_pageservers = 2
+    neon_env_builder.num_safekeepers = 1
+    env = neon_env_builder.init_start(
+        initial_tenant_conf=tenant_conf,
+        initial_tenant_shard_count=4,
+        initial_tenant_shard_stripe_size=1,
+    )
+
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
+
+    endpoint = env.endpoints.create_start("main")
+    endpoint.safe_psql("CREATE TABLE foo (id INTEGER, val text)")
+    for v in range(10):
+        endpoint.safe_psql(
+            f"INSERT INTO foo (id, val) VALUES ({v}, repeat('abcde{v:0>3}', 500))", log_query=False
+        )
+
+    response = env.storage_controller.tenant_timeline_describe(tenant_id, timeline_id)
+    shards = response["shards"]
+    for shard in shards:
+        assert shard["image_consistent_lsn"] is not None
+    image_consistent_lsn = response["image_consistent_lsn"]
+    assert image_consistent_lsn is not None
+
+    # do more writes and wait for image_consistent_lsn to advance
+    for v in range(100):
+        endpoint.safe_psql(
+            f"INSERT INTO foo (id, val) VALUES ({v}, repeat('abcde{v:0>3}', 500))", log_query=False
+        )
+
+    def check_image_consistent_lsn_advanced():
+        response = env.storage_controller.tenant_timeline_describe(tenant_id, timeline_id)
+        new_image_consistent_lsn = response["image_consistent_lsn"]
+        shards = response["shards"]
+        for shard in shards:
+            print(f"shard {shard['tenant_id']} image_consistent_lsn{shard['image_consistent_lsn']}")
+        assert new_image_consistent_lsn != image_consistent_lsn
+
+    wait_until(check_image_consistent_lsn_advanced)
+
+    endpoint.stop_and_destroy()
+
+    for ps in env.pageservers:
+        ps.allowed_errors.append(".*created delta file of size.*larger than double of target.*")
+
+
+# END_HADRON

test_runner/regress/test_download_extensions.py

@@ -2,7 +2,6 @@ from __future__ import annotations
 
 import os
 import platform
-import shutil
 import tarfile
 from enum import StrEnum
 from pathlib import Path
@@ -31,27 +30,6 @@ if TYPE_CHECKING:
     from werkzeug.wrappers.request import Request
 
 
-# use neon_env_builder_local fixture to override the default neon_env_builder fixture
-# and use a test-specific pg_install instead of shared one
-@pytest.fixture(scope="function")
-def neon_env_builder_local(
-    neon_env_builder: NeonEnvBuilder,
-    test_output_dir: Path,
-    pg_distrib_dir: Path,
-) -> NeonEnvBuilder:
-    test_local_pginstall = test_output_dir / "pg_install"
-    log.info(f"copy {pg_distrib_dir} to {test_local_pginstall}")
-
-    # We can't copy only the version that we are currently testing because other
-    # binaries like the storage controller need specific Postgres versions.
-    shutil.copytree(pg_distrib_dir, test_local_pginstall)
-
-    neon_env_builder.pg_distrib_dir = test_local_pginstall
-    log.info(f"local neon_env_builder.pg_distrib_dir: {neon_env_builder.pg_distrib_dir}")
-
-    return neon_env_builder
-
-
 @final
 class RemoteExtension(StrEnum):
     SQL_ONLY = "test_extension_sql_only"

test_runner/regress/test_event_trigger_extension.py

@@ -0,0 +1,102 @@
+from __future__ import annotations
+
+import shutil
+from pathlib import Path
+from typing import TYPE_CHECKING, cast
+
+import pytest
+from fixtures.log_helper import log
+from fixtures.paths import BASE_DIR
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+    from fixtures.neon_fixtures import (
+        NeonEnvBuilder,
+    )
+    from fixtures.pg_version import PgVersion
+
+
+# use neon_env_builder_local fixture to override the default neon_env_builder fixture
+# and use a test-specific pg_install instead of shared one
+@pytest.fixture(scope="function")
+def neon_env_builder_event_trigger_extension(
+    neon_env_builder_local: NeonEnvBuilder,
+    test_output_dir: Path,
+    pg_version: PgVersion,
+) -> NeonEnvBuilder:
+    test_local_pginstall = test_output_dir / "pg_install"
+
+    # Now copy the SQL only extension test_event_trigger_extension in the local
+    # pginstall extension directory on-disk
+    test_event_trigger_extension_dir = (
+        BASE_DIR / "test_runner" / "regress" / "data" / "test_event_trigger_extension"
+    )
+
+    test_local_extension_dir = (
+        test_local_pginstall / f"v{pg_version}" / "share" / "postgresql" / "extension"
+    )
+
+    log.info(f"copy {test_event_trigger_extension_dir} to {test_local_extension_dir}")
+
+    for f in [
+        test_event_trigger_extension_dir / "test_event_trigger_extension.control",
+        test_event_trigger_extension_dir / "test_event_trigger_extension--1.0.sql",
+    ]:
+        shutil.copy(f, test_local_extension_dir)
+
+    return neon_env_builder_local
+
+
+def test_event_trigger_extension(neon_env_builder_event_trigger_extension: NeonEnvBuilder):
+    """
+    Test installing an extension that contains an Event Trigger.
+
+    The Event Trigger function is owned by the extension owner, which at
+    CREATE EXTENSION is going to be the Postgres bootstrap user, per the
+    extension control file where both superuser = true and trusted = true.
+
+    Also this function is SECURTY DEFINER, to allow for making changes to
+    the extension SQL objects, in our case a sequence.
+
+    This test makes sure that the event trigger function is fired correctly
+    by non-privileged user DDL actions such as CREATE TABLE.
+    """
+    env = neon_env_builder_event_trigger_extension.init_start()
+    env.create_branch("test_event_trigger_extension")
+
+    endpoint = env.endpoints.create_start("test_event_trigger_extension")
+    extension = "test_event_trigger_extension"
+    database = "test_event_trigger_extension"
+
+    endpoint.safe_psql(f"CREATE DATABASE {database}")
+    endpoint.safe_psql(f"CREATE EXTENSION {extension}", dbname=database)
+
+    # check that the extension is owned by the bootstrap superuser (cloud_admin)
+    pg_bootstrap_superuser_name = "cloud_admin"
+    with endpoint.connect(dbname=database) as pg_conn:
+        with pg_conn.cursor() as cur:
+            cur.execute(
+                f"select rolname from pg_roles r join pg_extension e on r.oid = e.extowner where extname = '{extension}'"
+            )
+            owner = cast("tuple[str]", cur.fetchone())[0]
+            assert owner == pg_bootstrap_superuser_name, (
+                f"extension {extension} is not owned by bootstrap user '{pg_bootstrap_superuser_name}'"
+            )
+
+    # test that the SQL-only Event Trigger (SECURITY DEFINER function) runs
+    # correctly now that the extension has been installed
+    #
+    # create table to trigger the event trigger, twice, check sequence count
+    with endpoint.connect(dbname=database) as pg_conn:
+        log.info("creating SQL objects (tables)")
+        with pg_conn.cursor() as cur:
+            cur.execute("CREATE TABLE foo1(id int primary key)")
+            cur.execute("CREATE TABLE foo2(id int)")
+
+            cur.execute("SELECT event_trigger.get_schema_version()")
+            res = cast("tuple[int]", cur.fetchone())
+            ver = res[0]
+
+            log.info(f"schema version is now {ver}")
+            assert ver == 2, "schema version is not 2"

test_runner/regress/test_lfc_prewarm.py

@@ -1,6 +1,7 @@
 import random
 import threading
 from enum import StrEnum
+from time import sleep
 from typing import Any
 
 import pytest
@@ -24,18 +25,7 @@ OFFLOAD_LABEL = "compute_ctl_lfc_offloads_total"
 OFFLOAD_ERR_LABEL = "compute_ctl_lfc_offload_errors_total"
 METHOD_VALUES = [e for e in PrewarmMethod]
 METHOD_IDS = [e.value for e in PrewarmMethod]
-
-
-def check_pinned_entries(cur: Cursor):
-    """
-    Wait till none of LFC buffers are pinned
-    """
-
-    def none_pinned():
-        cur.execute("select lfc_value from neon_lfc_stats where lfc_key='file_cache_chunks_pinned'")
-        assert cur.fetchall()[0][0] == 0
-
-    wait_until(none_pinned)
+AUTOOFFLOAD_INTERVAL_SECS = 2
 
 
 def prom_parse(client: EndpointHttpClient) -> dict[str, float]:
@@ -49,9 +39,18 @@ def prom_parse(client: EndpointHttpClient) -> dict[str, float]:
 
 
 def offload_lfc(method: PrewarmMethod, client: EndpointHttpClient, cur: Cursor) -> Any:
+    if method == PrewarmMethod.POSTGRES:
+        cur.execute("select neon.get_local_cache_state()")
+        return cur.fetchall()[0][0]
+
     if method == PrewarmMethod.AUTOPREWARM:
+        # With autoprewarm, we need to be sure LFC was offloaded after all writes
+        # finish, so we sleep. Otherwise we'll have less prewarmed pages than we want
+        sleep(AUTOOFFLOAD_INTERVAL_SECS)
         client.offload_lfc_wait()
-    elif method == PrewarmMethod.COMPUTE_CTL:
+        return
+
+    if method == PrewarmMethod.COMPUTE_CTL:
         status = client.prewarm_lfc_status()
         assert status["status"] == "not_prewarmed"
         assert "error" not in status
@@ -60,11 +59,9 @@ def offload_lfc(method: PrewarmMethod, client: EndpointHttpClient, cur: Cursor)
         parsed = prom_parse(client)
         desired = {OFFLOAD_LABEL: 1, PREWARM_LABEL: 0, OFFLOAD_ERR_LABEL: 0, PREWARM_ERR_LABEL: 0}
         assert parsed == desired, f"{parsed=} != {desired=}"
-    elif method == PrewarmMethod.POSTGRES:
-        cur.execute("select get_local_cache_state()")
-        return cur.fetchall()[0][0]
-    else:
-        raise AssertionError(f"{method} not in PrewarmMethod")
+        return
+
+    raise AssertionError(f"{method} not in PrewarmMethod")
 
 
 def prewarm_endpoint(
@@ -75,7 +72,7 @@ def prewarm_endpoint(
     elif method == PrewarmMethod.COMPUTE_CTL:
         client.prewarm_lfc()
     elif method == PrewarmMethod.POSTGRES:
-        cur.execute("select prewarm_local_cache(%s)", (lfc_state,))
+        cur.execute("select neon.prewarm_local_cache(%s)", (lfc_state,))
 
 
 def check_prewarmed(
@@ -106,21 +103,20 @@ def test_lfc_prewarm(neon_simple_env: NeonEnv, method: PrewarmMethod):
         "neon.file_cache_size_limit=1GB",
         "neon.file_cache_prewarm_limit=1000",
     ]
-    offload_secs = 2
 
     if method == PrewarmMethod.AUTOPREWARM:
         endpoint = env.endpoints.create_start(
             branch_name="main",
             config_lines=cfg,
             autoprewarm=True,
-            offload_lfc_interval_seconds=offload_secs,
+            offload_lfc_interval_seconds=AUTOOFFLOAD_INTERVAL_SECS,
         )
     else:
         endpoint = env.endpoints.create_start(branch_name="main", config_lines=cfg)
 
     pg_conn = endpoint.connect()
     pg_cur = pg_conn.cursor()
-    pg_cur.execute("create extension neon")
+    pg_cur.execute("create schema neon; create extension neon with schema neon")
     pg_cur.execute("create database lfc")
 
     lfc_conn = endpoint.connect(dbname="lfc")
@@ -135,7 +131,7 @@ def test_lfc_prewarm(neon_simple_env: NeonEnv, method: PrewarmMethod):
 
     endpoint.stop()
     if method == PrewarmMethod.AUTOPREWARM:
-        endpoint.start(autoprewarm=True, offload_lfc_interval_seconds=offload_secs)
+        endpoint.start(autoprewarm=True, offload_lfc_interval_seconds=AUTOOFFLOAD_INTERVAL_SECS)
     else:
         endpoint.start()
 
@@ -146,10 +142,12 @@ def test_lfc_prewarm(neon_simple_env: NeonEnv, method: PrewarmMethod):
     lfc_cur = lfc_conn.cursor()
     prewarm_endpoint(method, client, pg_cur, lfc_state)
 
-    pg_cur.execute("select lfc_value from neon_lfc_stats where lfc_key='file_cache_used_pages'")
+    pg_cur.execute(
+        "select lfc_value from neon.neon_lfc_stats where lfc_key='file_cache_used_pages'"
+    )
     lfc_used_pages = pg_cur.fetchall()[0][0]
     log.info(f"Used LFC size: {lfc_used_pages}")
-    pg_cur.execute("select * from get_prewarm_info()")
+    pg_cur.execute("select * from neon.get_prewarm_info()")
     total, prewarmed, skipped, _ = pg_cur.fetchall()[0]
     log.info(f"Prewarm info: {total=} {prewarmed=} {skipped=}")
     progress = (prewarmed + skipped) * 100 // total
@@ -162,7 +160,6 @@ def test_lfc_prewarm(neon_simple_env: NeonEnv, method: PrewarmMethod):
     lfc_cur.execute("select sum(pk) from t")
     assert lfc_cur.fetchall()[0][0] == n_records * (n_records + 1) / 2
 
-    check_pinned_entries(pg_cur)
     desired = {"status": "completed", "total": total, "prewarmed": prewarmed, "skipped": skipped}
     check_prewarmed(method, client, desired)
 
@@ -191,7 +188,7 @@ def test_lfc_prewarm_under_workload(neon_simple_env: NeonEnv, method: PrewarmMet
 
     pg_conn = endpoint.connect()
     pg_cur = pg_conn.cursor()
-    pg_cur.execute("create extension neon")
+    pg_cur.execute("create schema neon; create extension neon with schema neon")
     pg_cur.execute("CREATE DATABASE lfc")
 
     lfc_conn = endpoint.connect(dbname="lfc")
@@ -243,9 +240,9 @@ def test_lfc_prewarm_under_workload(neon_simple_env: NeonEnv, method: PrewarmMet
     prewarm_thread.start()
 
     def prewarmed():
-        assert n_prewarms > 5
+        assert n_prewarms > 3
 
-    wait_until(prewarmed)
+    wait_until(prewarmed, timeout=40)  # debug builds don't finish in 20s
 
     running = False
     for t in workload_threads:
@@ -256,7 +253,6 @@ def test_lfc_prewarm_under_workload(neon_simple_env: NeonEnv, method: PrewarmMet
     total_balance = lfc_cur.fetchall()[0][0]
     assert total_balance == 0
 
-    check_pinned_entries(pg_cur)
     if method == PrewarmMethod.POSTGRES:
         return
     desired = {

test_runner/regress/test_pageserver_api.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 from typing import TYPE_CHECKING
 
 from fixtures.common_types import Lsn, TenantId, TimelineId
+from fixtures.log_helper import log
 from fixtures.neon_fixtures import (
     DEFAULT_BRANCH_NAME,
     NeonEnv,
@@ -164,3 +165,15 @@ def test_pageserver_http_index_part_force_patch(neon_env_builder: NeonEnvBuilder
             {"rel_size_migration": "legacy"},
         )
         assert client.timeline_detail(tenant_id, timeline_id)["rel_size_migration"] == "legacy"
+
+
+def test_pageserver_get_tenant_visible_size(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.num_pageservers = 1
+    env = neon_env_builder.init_start()
+    env.create_tenant(shard_count=4)
+    env.create_tenant(shard_count=2)
+
+    json = env.pageserver.http_client().list_tenant_visible_size()
+    log.info(f"{json}")
+    # initial tennat + 2 newly created tenants
+    assert len(json) == 7

test_runner/regress/test_replica_promotes.py

@@ -60,7 +60,7 @@ def test_replica_promote(neon_simple_env: NeonEnv, method: PromoteMethod):
 
     with primary.connect() as primary_conn:
         primary_cur = primary_conn.cursor()
-        primary_cur.execute("create extension neon")
+        primary_cur.execute("create schema neon;create extension neon with schema neon")
         primary_cur.execute(
             "create table t(pk bigint GENERATED ALWAYS AS IDENTITY, payload integer)"
         )
@@ -172,7 +172,7 @@ def test_replica_promote_handler_disconnects(neon_simple_env: NeonEnv):
     secondary: Endpoint = env.endpoints.new_replica_start(origin=primary, endpoint_id="secondary")
 
     with primary.connect() as conn, conn.cursor() as cur:
-        cur.execute("create extension neon")
+        cur.execute("create schema neon;create extension neon with schema neon")
         cur.execute("create table t(pk bigint GENERATED ALWAYS AS IDENTITY, payload integer)")
         cur.execute("INSERT INTO t(payload) SELECT generate_series(1, 100)")
         cur.execute("show neon.safekeepers")

workspace_hack/Cargo.toml

@@ -72,6 +72,7 @@ num-rational = { version = "0.4", default-features = false, features = ["num-big
 num-traits = { version = "0.2", features = ["i128", "libm"] }
 once_cell = { version = "1" }
 p256 = { version = "0.13", features = ["jwk"] }
+paracord = { version = "0.1", features = ["serde"] }
 parquet = { version = "53", default-features = false, features = ["zstd"] }
 prost = { version = "0.13", features = ["no-recursion-limit", "prost-derive"] }
 rand = { version = "0.8", features = ["small_rng"] }