[DNM] matrix run test suite with legacy & tiered compaction

We do it as a part of more complicated tests like test_compute_restarts, but
let's have a simple test as well.

.github/workflows/build_and_test.yml

@@ -432,8 +432,9 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        build_type: [ debug, release ]
-        pg_version: [ v14, v15, v16 ]
+        build_type: [ debug ]
+        pg_version: [ v15 ]
+        pageserver_compaction_algorithm_kind: [ "legacy", "tiered" ]
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -461,6 +462,9 @@ jobs:
           PAGESERVER_GET_VECTORED_IMPL: vectored
           PAGESERVER_GET_IMPL: vectored
           PAGESERVER_VALIDATE_VEC_GET: true
+          PAGESERVER_DEFAULT_TENANT_CONFIG_COMPACTION_ALGORITHM: 'kind="${{ matrix.pageserver_compaction_algorithm_kind }}"'
+          # catch the tests that override `tenant_config` as a whole without specifying the compaction algorithm `kind`
+          NEON_PAGESERVER_PANIC_ON_UNSPECIFIED_COMPACTION_ALGORITHM: true
 
       # Temporary disable this step until we figure out why it's so flaky
       # Ref https://github.com/neondatabase/neon/issues/4540

libs/pageserver_api/src/models.rs

@@ -451,6 +451,8 @@ impl EvictionPolicy {
 )]
 #[strum(serialize_all = "kebab-case")]
 pub enum CompactionAlgorithm {
+    #[strum(disabled)]
+    NotSpecified,
     Legacy,
     Tiered,
 }

libs/utils/src/failpoint_support.rs

@@ -9,6 +9,33 @@ use serde::{Deserialize, Serialize};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
 
+/// Declare a failpoint that can use the `pause` failpoint action.
+/// We don't want to block the executor thread, hence, spawn_blocking + await.
+#[macro_export]
+macro_rules! pausable_failpoint {
+    ($name:literal) => {
+        if cfg!(feature = "testing") {
+            tokio::task::spawn_blocking({
+                let current = tracing::Span::current();
+                move || {
+                    let _entered = current.entered();
+                    tracing::info!("at failpoint {}", $name);
+                    fail::fail_point!($name);
+                }
+            })
+            .await
+            .expect("spawn_blocking");
+        }
+    };
+    ($name:literal, $cond:expr) => {
+        if cfg!(feature = "testing") {
+            if $cond {
+                pausable_failpoint!($name)
+            }
+        }
+    };
+}
+
 /// use with fail::cfg("$name", "return(2000)")
 ///
 /// The effect is similar to a "sleep(2000)" action, i.e. we sleep for the

pageserver/src/config.rs

@@ -5,7 +5,7 @@
 //! See also `settings.md` for better description on every parameter.
 
 use anyhow::{anyhow, bail, ensure, Context, Result};
-use pageserver_api::shard::TenantShardId;
+use pageserver_api::{models::CompactionAlgorithm, shard::TenantShardId};
 use remote_storage::{RemotePath, RemoteStorageConfig};
 use serde;
 use serde::de::IntoDeserializer;
@@ -15,7 +15,7 @@ use utils::crashsafe::path_with_suffix_extension;
 use utils::id::ConnectionId;
 use utils::logging::SecretString;
 
-use once_cell::sync::OnceCell;
+use once_cell::sync::{Lazy, OnceCell};
 use reqwest::Url;
 use std::num::NonZeroUsize;
 use std::str::FromStr;
@@ -1067,6 +1067,19 @@ impl PageServerConf {
 
         conf.default_tenant_conf = t_conf.merge(TenantConf::default());
 
+        {
+            const VAR_NAME: &str = "NEON_PAGESERVER_PANIC_ON_UNSPECIFIED_COMPACTION_ALGORITHM";
+            static VAR: Lazy<Option<bool>> = Lazy::new(|| utils::env::var(VAR_NAME));
+            if VAR.unwrap_or(false)
+                && conf.default_tenant_conf.compaction_algorithm.kind
+                    == CompactionAlgorithm::NotSpecified
+            {
+                panic!(
+                        "Unspecified compaction algorithm in default tenant configuration. \
+                        Set the algorithm explicitly in the pageserver.toml's `tenant_config` field or unset the environment variable {VAR_NAME}");
+            }
+        }
+
         Ok(conf)
     }
 

pageserver/src/tenant.rs

@@ -42,6 +42,7 @@ use utils::completion;
 use utils::crashsafe::path_with_suffix_extension;
 use utils::failpoint_support;
 use utils::fs_ext;
+use utils::pausable_failpoint;
 use utils::sync::gate::Gate;
 use utils::sync::gate::GateGuard;
 use utils::timeout::timeout_cancellable;
@@ -122,32 +123,6 @@ use utils::{
     lsn::{Lsn, RecordLsn},
 };
 
-/// Declare a failpoint that can use the `pause` failpoint action.
-/// We don't want to block the executor thread, hence, spawn_blocking + await.
-macro_rules! pausable_failpoint {
-    ($name:literal) => {
-        if cfg!(feature = "testing") {
-            tokio::task::spawn_blocking({
-                let current = tracing::Span::current();
-                move || {
-                    let _entered = current.entered();
-                    tracing::info!("at failpoint {}", $name);
-                    fail::fail_point!($name);
-                }
-            })
-            .await
-            .expect("spawn_blocking");
-        }
-    };
-    ($name:literal, $cond:expr) => {
-        if cfg!(feature = "testing") {
-            if $cond {
-                pausable_failpoint!($name)
-            }
-        }
-    };
-}
-
 pub mod blob_io;
 pub mod block_io;
 pub mod vectored_blob_io;

pageserver/src/tenant/config.rs

@@ -40,8 +40,6 @@ pub mod defaults {
 
     pub const DEFAULT_COMPACTION_PERIOD: &str = "20 s";
     pub const DEFAULT_COMPACTION_THRESHOLD: usize = 10;
-    pub const DEFAULT_COMPACTION_ALGORITHM: super::CompactionAlgorithm =
-        super::CompactionAlgorithm::Legacy;
 
     pub const DEFAULT_GC_HORIZON: u64 = 64 * 1024 * 1024;
 
@@ -554,7 +552,13 @@ impl Default for TenantConf {
                 .expect("cannot parse default compaction period"),
             compaction_threshold: DEFAULT_COMPACTION_THRESHOLD,
             compaction_algorithm: CompactionAlgorithmSettings {
-                kind: DEFAULT_COMPACTION_ALGORITHM,
+                kind: if cfg!(test) {
+                    // Rust tests rely on a valid implicit default (TODO: fix this)
+                    CompactionAlgorithm::Legacy
+                } else {
+                    // Python tests are subject to NotSpecified handling
+                    CompactionAlgorithm::NotSpecified
+                },
             },
             gc_horizon: DEFAULT_GC_HORIZON,
             gc_period: humantime::parse_duration(DEFAULT_GC_PERIOD)

pageserver/src/tenant/delete.rs

@@ -8,7 +8,7 @@ use tokio::sync::OwnedMutexGuard;
 use tokio_util::sync::CancellationToken;
 use tracing::{error, instrument, Instrument};
 
-use utils::{backoff, completion, crashsafe, fs_ext, id::TimelineId};
+use utils::{backoff, completion, crashsafe, fs_ext, id::TimelineId, pausable_failpoint};
 
 use crate::{
     config::PageServerConf,

pageserver/src/tenant/remote_timeline_client.rs

@@ -197,6 +197,7 @@ pub(crate) use upload::upload_initdb_dir;
 use utils::backoff::{
     self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
 };
+use utils::pausable_failpoint;
 
 use std::collections::{HashMap, VecDeque};
 use std::sync::atomic::{AtomicU32, Ordering};

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -9,7 +9,7 @@ use std::time::SystemTime;
 use tokio::fs::{self, File};
 use tokio::io::AsyncSeekExt;
 use tokio_util::sync::CancellationToken;
-use utils::backoff;
+use utils::{backoff, pausable_failpoint};
 
 use super::Generation;
 use crate::tenant::remote_timeline_client::{

pageserver/src/tenant/tasks.rs

@@ -17,7 +17,7 @@ use crate::tenant::{Tenant, TenantState};
 use rand::Rng;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
-use utils::{backoff, completion};
+use utils::{backoff, completion, pausable_failpoint};
 
 static CONCURRENT_BACKGROUND_TASKS: once_cell::sync::Lazy<tokio::sync::Semaphore> =
     once_cell::sync::Lazy::new(|| {

pageserver/src/tenant/timeline.rs

@@ -41,7 +41,7 @@ use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::{
     bin_ser::BeSer,
-    fs_ext,
+    fs_ext, pausable_failpoint,
     sync::gate::{Gate, GateGuard},
     vec_map::VecMap,
 };
@@ -1700,6 +1700,9 @@ impl Timeline {
         }
 
         match self.get_compaction_algorithm_settings().kind {
+            CompactionAlgorithm::NotSpecified => {
+                unreachable!("should panic earlier when we construct the default tenant conf")
+            }
             CompactionAlgorithm::Tiered => self.compact_tiered(cancel, ctx).await,
             CompactionAlgorithm::Legacy => self.compact_legacy(cancel, flags, ctx).await,
         }

pageserver/src/tenant/timeline/delete.rs

@@ -7,7 +7,7 @@ use anyhow::Context;
 use pageserver_api::{models::TimelineState, shard::TenantShardId};
 use tokio::sync::OwnedMutexGuard;
 use tracing::{error, info, instrument, Instrument};
-use utils::{crashsafe, fs_ext, id::TimelineId};
+use utils::{crashsafe, fs_ext, id::TimelineId, pausable_failpoint};
 
 use crate::{
     config::PageServerConf,

pgxn/neon/libpagestore.c

@@ -125,13 +125,6 @@ typedef struct
 	 *	- WL_EXIT_ON_PM_DEATH.
 	 */
 	WaitEventSet   *wes_read;
-	/*---
-	 * WaitEventSet containing:
-	 *	- WL_SOCKET_WRITABLE on 'conn'
-	 *	- WL_LATCH_SET on MyLatch, and
-	 *	- WL_EXIT_ON_PM_DEATH.
-	 */
-	WaitEventSet   *wes_write;
 } PageServer;
 
 static PageServer page_servers[MAX_SHARDS];
@@ -336,11 +329,6 @@ CLEANUP_AND_DISCONNECT(PageServer *shard)
 		FreeWaitEventSet(shard->wes_read);
 		shard->wes_read = NULL;
 	}
-	if (shard->wes_write)
-	{
-		FreeWaitEventSet(shard->wes_write);
-		shard->wes_write = NULL;
-	}
 	if (shard->conn)
 	{
 		PQfinish(shard->conn);
@@ -436,22 +424,6 @@ pageserver_connect(shardno_t shard_no, int elevel)
 			return false;
 		}
 
-		shard->wes_read = CreateWaitEventSet(TopMemoryContext, 3);
-		AddWaitEventToSet(shard->wes_read, WL_LATCH_SET, PGINVALID_SOCKET,
-						  MyLatch, NULL);
-		AddWaitEventToSet(shard->wes_read, WL_EXIT_ON_PM_DEATH, PGINVALID_SOCKET,
-						  NULL, NULL);
-		AddWaitEventToSet(shard->wes_read, WL_SOCKET_READABLE, PQsocket(shard->conn), NULL, NULL);
-
-		shard->wes_write = CreateWaitEventSet(TopMemoryContext, 3);
-		AddWaitEventToSet(shard->wes_write, WL_LATCH_SET, PGINVALID_SOCKET,
-						  MyLatch, NULL);
-		AddWaitEventToSet(shard->wes_write, WL_EXIT_ON_PM_DEATH, PGINVALID_SOCKET,
-						  NULL, NULL);
-		AddWaitEventToSet(shard->wes_write, WL_SOCKET_READABLE | WL_SOCKET_WRITEABLE,
-						  PQsocket(shard->conn),
-						  NULL, NULL);
-
 		shard->state = PS_Connecting_Startup;
 		/* fallthrough */
 	}
@@ -460,13 +432,12 @@ pageserver_connect(shardno_t shard_no, int elevel)
 		char	   *pagestream_query;
 		int			ps_send_query_ret;
 		bool		connected = false;
-
+		int poll_result = PGRES_POLLING_WRITING;
 		neon_shard_log(shard_no, DEBUG5, "Connection state: Connecting_Startup");
 
 		do
 		{
 			WaitEvent	event;
-			int			poll_result = PQconnectPoll(shard->conn);
 
 			switch (poll_result)
 			{
@@ -497,25 +468,45 @@ pageserver_connect(shardno_t shard_no, int elevel)
 				}
 			case PGRES_POLLING_READING:
 				/* Sleep until there's something to do */
-				(void) WaitEventSetWait(shard->wes_read, -1L, &event, 1,
-										PG_WAIT_EXTENSION);
-				ResetLatch(MyLatch);
-
-				/* query cancellation, backend shutdown */
-				CHECK_FOR_INTERRUPTS();
-
+				while (true)
+				{
+					int rc = WaitLatchOrSocket(MyLatch,
+											   WL_EXIT_ON_PM_DEATH | WL_LATCH_SET | WL_SOCKET_READABLE,
+											   PQsocket(shard->conn),
+											   0,
+											   PG_WAIT_EXTENSION);
+					elog(DEBUG5, "PGRES_POLLING_READING=>%d", rc);
+					if (rc & WL_LATCH_SET)
+					{
+						ResetLatch(MyLatch);
+						/* query cancellation, backend shutdown */
+						CHECK_FOR_INTERRUPTS();
+					}
+					if (rc & WL_SOCKET_READABLE)
+						break;
+				}
 				/* PQconnectPoll() handles the socket polling state updates */
 
 				break;
 			case PGRES_POLLING_WRITING:
 				/* Sleep until there's something to do */
-				(void) WaitEventSetWait(shard->wes_write, -1L, &event, 1,
-										PG_WAIT_EXTENSION);
-				ResetLatch(MyLatch);
-
-				/* query cancellation, backend shutdown */
-				CHECK_FOR_INTERRUPTS();
-
+				while (true)
+				{
+					int rc = WaitLatchOrSocket(MyLatch,
+											   WL_EXIT_ON_PM_DEATH | WL_LATCH_SET | WL_SOCKET_WRITEABLE,
+											   PQsocket(shard->conn),
+											   0,
+											   PG_WAIT_EXTENSION);
+					elog(DEBUG5, "PGRES_POLLING_WRITING=>%d", rc);
+					if (rc & WL_LATCH_SET)
+					{
+						ResetLatch(MyLatch);
+						/* query cancellation, backend shutdown */
+						CHECK_FOR_INTERRUPTS();
+					}
+					if (rc & WL_SOCKET_WRITEABLE)
+						break;
+				}
 				/* PQconnectPoll() handles the socket polling state updates */
 
 				break;
@@ -524,12 +515,22 @@ pageserver_connect(shardno_t shard_no, int elevel)
 				connected = true;
 				break;
 			}
+			poll_result = PQconnectPoll(shard->conn);
+			elog(DEBUG5, "PQconnectPoll=>%d", poll_result);
 		}
 		while (!connected);
 
 		/* No more polling needed; connection succeeded */
 		shard->last_connect_time = GetCurrentTimestamp();
 
+		shard->wes_read = CreateWaitEventSet(TopMemoryContext, 3);
+		AddWaitEventToSet(shard->wes_read, WL_LATCH_SET, PGINVALID_SOCKET,
+						  MyLatch, NULL);
+		AddWaitEventToSet(shard->wes_read, WL_EXIT_ON_PM_DEATH, PGINVALID_SOCKET,
+						  NULL, NULL);
+		AddWaitEventToSet(shard->wes_read, WL_SOCKET_READABLE, PQsocket(shard->conn), NULL, NULL);
+
+
 		switch (neon_protocol_version)
 		{
 		case 2:

pgxn/neon/pagestore_smgr.c

@@ -584,9 +584,9 @@ prefetch_read(PrefetchRequest *slot)
 		slot->response != NULL ||
 		slot->my_ring_index != MyPState->ring_receive)
 		neon_shard_log(slot->shard_no, ERROR,
-					   "Incorrect prefetch read: status=%d response=%llx my=%llu receive=%llu",
-					   slot->status, (size_t) (void *) slot->response,
-					   slot->my_ring_index, MyPState->ring_receive);
+					   "Incorrect prefetch read: status=%d response=%p my=%lu receive=%lu",
+					   slot->status, slot->response,
+					   (long)slot->my_ring_index, (long)MyPState->ring_receive);
 
 	old = MemoryContextSwitchTo(MyPState->errctx);
 	response = (NeonResponse *) page_server->receive(slot->shard_no);
@@ -606,8 +606,8 @@ prefetch_read(PrefetchRequest *slot)
 	else
 	{
 		neon_shard_log(slot->shard_no, WARNING,
-					   "No response from reading prefetch entry %llu: %u/%u/%u.%u block %u. This can be caused by a concurrent disconnect",
-					   slot->my_ring_index,
+					   "No response from reading prefetch entry %lu: %u/%u/%u.%u block %u. This can be caused by a concurrent disconnect",
+					   (long)slot->my_ring_index,
 					   RelFileInfoFmt(BufTagGetNRelFileInfo(slot->buftag)),
 					   slot->buftag.forkNum, slot->buftag.blockNum);
 		return false;

proxy/src/bin/proxy.rs

@@ -356,7 +356,7 @@ async fn main() -> anyhow::Result<()> {
 
     let cancel_map = CancelMap::default();
 
-    let redis_publisher = match &redis_notifications_client {
+    let redis_publisher = match &regional_redis_client {
         Some(redis_publisher) => Some(Arc::new(Mutex::new(RedisPublisherClient::new(
             redis_publisher.clone(),
             args.region.clone(),

safekeeper/src/http/routes.rs

@@ -287,6 +287,26 @@ async fn timeline_files_handler(request: Request<Body>) -> Result<Response<Body>
         .map_err(|e| ApiError::InternalServerError(e.into()))
 }
 
+/// Force persist control file and remove old WAL.
+async fn timeline_checkpoint_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
+    check_permission(&request, None)?;
+
+    let ttid = TenantTimelineId::new(
+        parse_request_param(&request, "tenant_id")?,
+        parse_request_param(&request, "timeline_id")?,
+    );
+
+    let tli = GlobalTimelines::get(ttid)?;
+    tli.maybe_persist_control_file(true)
+        .await
+        .map_err(ApiError::InternalServerError)?;
+    tli.remove_old_wal()
+        .await
+        .map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, ())
+}
+
 /// Deactivates the timeline and removes its data directory.
 async fn timeline_delete_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
     let ttid = TenantTimelineId::new(
@@ -553,6 +573,10 @@ pub fn make_router(conf: SafeKeeperConf) -> RouterBuilder<hyper::Body, ApiError>
             "/v1/tenant/:tenant_id/timeline/:timeline_id/control_file",
             |r| request_span(r, patch_control_file_handler),
         )
+        .post(
+            "/v1/tenant/:tenant_id/timeline/:timeline_id/checkpoint",
+            |r| request_span(r, timeline_checkpoint_handler),
+        )
         // for tests
         .post("/v1/record_safekeeper_info/:tenant_id/:timeline_id", |r| {
             request_span(r, record_safekeeper_info)

safekeeper/src/pull_timeline.rs

@@ -11,6 +11,7 @@ use tracing::info;
 use utils::{
     id::{TenantId, TenantTimelineId, TimelineId},
     lsn::Lsn,
+    pausable_failpoint,
 };
 
 use crate::{
@@ -162,6 +163,8 @@ async fn pull_timeline(status: TimelineStatus, host: String) -> Result<Response>
     filenames.remove(control_file_index);
     filenames.insert(0, "safekeeper.control".to_string());
 
+    pausable_failpoint!("sk-pull-timeline-after-list-pausable");
+
     info!(
         "downloading {} files from safekeeper {}",
         filenames.len(),
@@ -183,6 +186,13 @@ async fn pull_timeline(status: TimelineStatus, host: String) -> Result<Response>
 
         let mut file = tokio::fs::File::create(&file_path).await?;
         let mut response = client.get(&http_url).send().await?;
+        if response.status() != reqwest::StatusCode::OK {
+            bail!(
+                "pulling file {} failed: status is {}",
+                filename,
+                response.status()
+            );
+        }
         while let Some(chunk) = response.chunk().await? {
             file.write_all(&chunk).await?;
             file.flush().await?;

safekeeper/src/remove_wal.rs

@@ -15,7 +15,7 @@ pub async fn task_main(_conf: SafeKeeperConf) -> anyhow::Result<()> {
         for tli in &tlis {
             let ttid = tli.ttid;
             async {
-                if let Err(e) = tli.maybe_persist_control_file().await {
+                if let Err(e) = tli.maybe_persist_control_file(false).await {
                     warn!("failed to persist control file: {e}");
                 }
                 if let Err(e) = tli.remove_old_wal().await {

safekeeper/src/safekeeper.rs

@@ -827,9 +827,9 @@ where
 
     /// Persist control file if there is something to save and enough time
     /// passed after the last save.
-    pub async fn maybe_persist_inmem_control_file(&mut self) -> Result<bool> {
+    pub async fn maybe_persist_inmem_control_file(&mut self, force: bool) -> Result<bool> {
         const CF_SAVE_INTERVAL: Duration = Duration::from_secs(300);
-        if self.state.pers.last_persist_at().elapsed() < CF_SAVE_INTERVAL {
+        if !force && self.state.pers.last_persist_at().elapsed() < CF_SAVE_INTERVAL {
             return Ok(false);
         }
         let need_persist = self.state.inmem.commit_lsn > self.state.commit_lsn

safekeeper/src/timeline.rs

@@ -821,9 +821,9 @@ impl Timeline {
     /// passed after the last save. This helps to keep remote_consistent_lsn up
     /// to date so that storage nodes restart doesn't cause many pageserver ->
     /// safekeeper reconnections.
-    pub async fn maybe_persist_control_file(self: &Arc<Self>) -> Result<()> {
+    pub async fn maybe_persist_control_file(self: &Arc<Self>, force: bool) -> Result<()> {
         let mut guard = self.write_shared_state().await;
-        let changed = guard.sk.maybe_persist_inmem_control_file().await?;
+        let changed = guard.sk.maybe_persist_inmem_control_file(force).await?;
         guard.skip_update = !changed;
         Ok(())
     }

safekeeper/src/timeline_manager.rs

@@ -106,7 +106,7 @@ pub async fn main_task(
 
             if !is_active {
                 // TODO: maybe use tokio::spawn?
-                if let Err(e) = tli.maybe_persist_control_file().await {
+                if let Err(e) = tli.maybe_persist_control_file(false).await {
                     warn!("control file save in update_status failed: {:?}", e);
                 }
             }

test_runner/fixtures/common_types.py

@@ -5,6 +5,8 @@ from typing import Any, Type, TypeVar, Union
 
 T = TypeVar("T", bound="Id")
 
+DEFAULT_WAL_SEG_SIZE = 16 * 1024 * 1024
+
 
 @total_ordering
 class Lsn:
@@ -67,6 +69,9 @@ class Lsn:
     def as_int(self) -> int:
         return self.lsn_int
 
+    def segment_lsn(self, seg_sz: int = DEFAULT_WAL_SEG_SIZE) -> "Lsn":
+        return Lsn(self.lsn_int - (self.lsn_int % seg_sz))
+
 
 @dataclass(frozen=True)
 class Key:

test_runner/fixtures/neon_fixtures.py

@@ -3771,7 +3771,7 @@ class SafekeeperPort:
 
 
 @dataclass
-class Safekeeper:
+class Safekeeper(LogUtils):
     """An object representing a running safekeeper daemon."""
 
     env: NeonEnv
@@ -3779,6 +3779,13 @@ class Safekeeper:
     id: int
     running: bool = False
 
+    def __init__(self, env: NeonEnv, port: SafekeeperPort, id: int, running: bool = False):
+        self.env = env
+        self.port = port
+        self.id = id
+        self.running = running
+        self.logfile = Path(self.data_dir) / f"safekeeper-{id}.log"
+
     def start(self, extra_opts: Optional[List[str]] = None) -> "Safekeeper":
         assert self.running is False
         self.env.neon_cli.safekeeper_start(self.id, extra_opts=extra_opts)
@@ -3839,11 +3846,38 @@ class Safekeeper:
             port=self.port.http, auth_token=auth_token, is_testing_enabled=is_testing_enabled
         )
 
-    def data_dir(self) -> str:
-        return os.path.join(self.env.repo_dir, "safekeepers", f"sk{self.id}")
+    def get_timeline_start_lsn(self, tenant_id: TenantId, timeline_id: TimelineId) -> Lsn:
+        timeline_status = self.http_client().timeline_status(tenant_id, timeline_id)
+        timeline_start_lsn = timeline_status.timeline_start_lsn
+        log.info(f"sk {self.id} timeline start LSN: {timeline_start_lsn}")
+        return timeline_start_lsn
 
-    def timeline_dir(self, tenant_id, timeline_id) -> str:
-        return os.path.join(self.data_dir(), str(tenant_id), str(timeline_id))
+    def get_flush_lsn(self, tenant_id: TenantId, timeline_id: TimelineId) -> Lsn:
+        timeline_status = self.http_client().timeline_status(tenant_id, timeline_id)
+        flush_lsn = timeline_status.flush_lsn
+        log.info(f"sk {self.id} flush LSN: {flush_lsn}")
+        return flush_lsn
+
+    def pull_timeline(
+        self, srcs: list[Safekeeper], tenant_id: TenantId, timeline_id: TimelineId
+    ) -> Dict[str, Any]:
+        """
+        pull_timeline from srcs to self.
+        """
+        src_https = [f"http://localhost:{sk.port.http}" for sk in srcs]
+        res = self.http_client().pull_timeline(
+            {"tenant_id": str(tenant_id), "timeline_id": str(timeline_id), "http_hosts": src_https}
+        )
+        src_ids = [sk.id for sk in srcs]
+        log.info(f"finished pulling timeline from {src_ids} to {self.id}")
+        return res
+
+    @property
+    def data_dir(self) -> Path:
+        return self.env.repo_dir / "safekeepers" / f"sk{self.id}"
+
+    def timeline_dir(self, tenant_id, timeline_id) -> Path:
+        return self.data_dir / str(tenant_id) / str(timeline_id)
 
     def list_segments(self, tenant_id, timeline_id) -> List[str]:
         """
@@ -3856,6 +3890,35 @@ class Safekeeper:
         segments.sort()
         return segments
 
+    def checkpoint_up_to(self, tenant_id: TenantId, timeline_id: TimelineId, lsn: Lsn):
+        """
+        Assuming pageserver(s) uploaded to s3 up to `lsn`,
+        1) wait for remote_consistent_lsn and wal_backup_lsn on safekeeper to reach it.
+        2) checkpoint timeline on safekeeper, which should remove WAL before this LSN.
+        """
+        cli = self.http_client()
+
+        def are_lsns_advanced():
+            stat = cli.timeline_status(tenant_id, timeline_id)
+            log.info(
+                f"waiting for remote_consistent_lsn and backup_lsn on sk {self.id} to reach {lsn}, currently remote_consistent_lsn={stat.remote_consistent_lsn}, backup_lsn={stat.backup_lsn}"
+            )
+            assert stat.remote_consistent_lsn >= lsn and stat.backup_lsn >= lsn.segment_lsn()
+
+        # xxx: max wait is long because we might be waiting for reconnection from
+        # pageserver to this safekeeper
+        wait_until(30, 1, are_lsns_advanced)
+        cli.checkpoint(tenant_id, timeline_id)
+
+    def wait_until_paused(self, failpoint: str):
+        msg = f"at failpoint {failpoint}"
+
+        def paused():
+            log.info(f"waiting for hitting failpoint {failpoint}")
+            self.assert_log_contains(msg)
+
+        wait_until(20, 0.5, paused)
+
 
 class S3Scrubber:
     def __init__(self, env: NeonEnvBuilder, log_dir: Optional[Path] = None):

test_runner/fixtures/safekeeper/http.py

@@ -177,6 +177,13 @@ class SafekeeperHttpClient(requests.Session):
         )
         res.raise_for_status()
 
+    def checkpoint(self, tenant_id: TenantId, timeline_id: TimelineId):
+        res = self.post(
+            f"http://localhost:{self.port}/v1/tenant/{tenant_id}/timeline/{timeline_id}/checkpoint",
+            json={},
+        )
+        res.raise_for_status()
+
     # only_local doesn't remove segments in the remote storage.
     def timeline_delete(
         self, tenant_id: TenantId, timeline_id: TimelineId, only_local: bool = False

test_runner/fixtures/utils.py

@@ -196,7 +196,7 @@ def query_scalar(cur: cursor, query: str) -> Any:
 
 
 # Traverse directory to get total size.
-def get_dir_size(path: str) -> int:
+def get_dir_size(path: Path) -> int:
     """Return size in bytes."""
     totalbytes = 0
     for root, _dirs, files in os.walk(path):
@@ -560,3 +560,25 @@ def assert_pageserver_backups_equal(left: Path, right: Path, skip_files: Set[str
 
     elapsed = time.time() - started_at
     log.info(f"assert_pageserver_backups_equal completed in {elapsed}s")
+
+
+class PropagatingThread(threading.Thread):
+    _target: Any
+    _args: Any
+    _kwargs: Any
+    """
+    Simple Thread wrapper with join() propagating the possible exception in the thread.
+    """
+
+    def run(self):
+        self.exc = None
+        try:
+            self.ret = self._target(*self._args, **self._kwargs)
+        except BaseException as e:
+            self.exc = e
+
+    def join(self, timeout=None):
+        super(PropagatingThread, self).join(timeout)
+        if self.exc:
+            raise self.exc
+        return self.ret

test_runner/regress/test_wal_acceptor.py

@@ -23,7 +23,6 @@ from fixtures.log_helper import log
 from fixtures.metrics import parse_metrics
 from fixtures.neon_fixtures import (
     Endpoint,
-    NeonEnv,
     NeonEnvBuilder,
     NeonPageserver,
     PgBin,
@@ -48,7 +47,7 @@ from fixtures.remote_storage import (
 )
 from fixtures.safekeeper.http import SafekeeperHttpClient
 from fixtures.safekeeper.utils import are_walreceivers_absent
-from fixtures.utils import get_dir_size, query_scalar, start_in_background
+from fixtures.utils import PropagatingThread, get_dir_size, query_scalar, start_in_background
 
 
 def wait_lsn_force_checkpoint(
@@ -360,7 +359,7 @@ def test_wal_removal(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
 
     # We will wait for first segment removal. Make sure they exist for starter.
     first_segments = [
-        os.path.join(sk.data_dir(), str(tenant_id), str(timeline_id), "000000010000000000000001")
+        sk.timeline_dir(tenant_id, timeline_id) / "000000010000000000000001"
         for sk in env.safekeepers
     ]
     assert all(os.path.exists(p) for p in first_segments)
@@ -445,7 +444,7 @@ def is_flush_lsn_caught_up(sk: Safekeeper, tenant_id: TenantId, timeline_id: Tim
 def is_wal_trimmed(sk: Safekeeper, tenant_id: TenantId, timeline_id: TimelineId, target_size_mb):
     http_cli = sk.http_client()
     tli_status = http_cli.timeline_status(tenant_id, timeline_id)
-    sk_wal_size = get_dir_size(os.path.join(sk.data_dir(), str(tenant_id), str(timeline_id)))
+    sk_wal_size = get_dir_size(sk.timeline_dir(tenant_id, timeline_id))
     sk_wal_size_mb = sk_wal_size / 1024 / 1024
     log.info(f"Safekeeper id={sk.id} wal_size={sk_wal_size_mb:.2f}MB status={tli_status}")
     return sk_wal_size_mb <= target_size_mb
@@ -591,10 +590,10 @@ def test_s3_wal_replay(neon_env_builder: NeonEnvBuilder):
 
     # save the last (partial) file to put it back after recreation; others will be fetched from s3
     sk = env.safekeepers[0]
-    tli_dir = Path(sk.data_dir()) / str(tenant_id) / str(timeline_id)
+    tli_dir = Path(sk.data_dir) / str(tenant_id) / str(timeline_id)
     f_partial = Path([f for f in os.listdir(tli_dir) if f.endswith(".partial")][0])
     f_partial_path = tli_dir / f_partial
-    f_partial_saved = Path(sk.data_dir()) / f_partial.name
+    f_partial_saved = Path(sk.data_dir) / f_partial.name
     f_partial_path.rename(f_partial_saved)
 
     pg_version = sk.http_client().timeline_status(tenant_id, timeline_id).pg_version
@@ -616,7 +615,7 @@ def test_s3_wal_replay(neon_env_builder: NeonEnvBuilder):
         cli = sk.http_client()
         cli.timeline_create(tenant_id, timeline_id, pg_version, last_lsn)
         f_partial_path = (
-            Path(sk.data_dir()) / str(tenant_id) / str(timeline_id) / f_partial_saved.name
+            Path(sk.data_dir) / str(tenant_id) / str(timeline_id) / f_partial_saved.name
         )
         shutil.copy(f_partial_saved, f_partial_path)
 
@@ -1132,8 +1131,8 @@ def cmp_sk_wal(sks: List[Safekeeper], tenant_id: TenantId, timeline_id: Timeline
         )
 
         for f in mismatch:
-            f1 = os.path.join(sk0.timeline_dir(tenant_id, timeline_id), f)
-            f2 = os.path.join(sk.timeline_dir(tenant_id, timeline_id), f)
+            f1 = sk0.timeline_dir(tenant_id, timeline_id) / f
+            f2 = sk.timeline_dir(tenant_id, timeline_id) / f
             stdout_filename = f"{f2}.filediff"
 
             with open(stdout_filename, "w") as stdout_f:
@@ -1631,7 +1630,7 @@ def test_delete_force(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
             with conn.cursor() as cur:
                 cur.execute("CREATE TABLE t(key int primary key)")
     sk = env.safekeepers[0]
-    sk_data_dir = Path(sk.data_dir())
+    sk_data_dir = sk.data_dir
     if not auth_enabled:
         sk_http = sk.http_client()
         sk_http_other = sk_http
@@ -1724,9 +1723,6 @@ def test_delete_force(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
 
 
 def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
-    def safekeepers_guc(env: NeonEnv, sk_names: List[int]) -> str:
-        return ",".join([f"localhost:{sk.port.pg}" for sk in env.safekeepers if sk.id in sk_names])
-
     def execute_payload(endpoint: Endpoint):
         with closing(endpoint.connect()) as conn:
             with conn.cursor() as cur:
@@ -1812,6 +1808,65 @@ def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
     show_statuses(env.safekeepers, tenant_id, timeline_id)
 
 
+# Test pull_timeline while concurrently gc'ing WAL on safekeeper:
+# 1) Start pull_timeline, listing files to fetch.
+# 2) Write segment, do gc.
+# 3) Finish pull_timeline.
+# 4) Do some write, verify integrity with timeline_digest.
+# Expected to fail while holding off WAL gc plus fetching commit_lsn WAL
+# segment is not implemented.
+@pytest.mark.xfail
+def test_pull_timeline_gc(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.num_safekeepers = 3
+    neon_env_builder.enable_safekeeper_remote_storage(default_remote_storage())
+    env = neon_env_builder.init_start()
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
+
+    (src_sk, dst_sk) = (env.safekeepers[0], env.safekeepers[2])
+
+    log.info("use only first 2 safekeepers, 3rd will be seeded")
+    endpoint = env.endpoints.create("main")
+    endpoint.active_safekeepers = [1, 2]
+    endpoint.start()
+    endpoint.safe_psql("create table t(key int, value text)")
+    endpoint.safe_psql("insert into t select generate_series(1, 1000), 'pear'")
+
+    src_flush_lsn = src_sk.get_flush_lsn(tenant_id, timeline_id)
+    log.info(f"flush_lsn on src before pull_timeline: {src_flush_lsn}")
+
+    dst_http = dst_sk.http_client()
+    # run pull_timeline which will halt before downloading files
+    dst_http.configure_failpoints(("sk-pull-timeline-after-list-pausable", "pause"))
+    pt_handle = PropagatingThread(
+        target=dst_sk.pull_timeline, args=([src_sk], tenant_id, timeline_id)
+    )
+    pt_handle.start()
+    dst_sk.wait_until_paused("sk-pull-timeline-after-list-pausable")
+
+    # ensure segment exists
+    endpoint.safe_psql("insert into t select generate_series(1, 180000), 'papaya'")
+    lsn = last_flush_lsn_upload(env, endpoint, tenant_id, timeline_id)
+    assert lsn > Lsn("0/2000000")
+    # Checkpoint timeline beyond lsn.
+    src_sk.checkpoint_up_to(tenant_id, timeline_id, lsn)
+    first_segment_p = src_sk.timeline_dir(tenant_id, timeline_id) / "000000010000000000000001"
+    log.info(f"first segment exist={os.path.exists(first_segment_p)}")
+
+    dst_http.configure_failpoints(("sk-pull-timeline-after-list-pausable", "off"))
+    pt_handle.join()
+
+    timeline_start_lsn = src_sk.get_timeline_start_lsn(tenant_id, timeline_id)
+    dst_flush_lsn = dst_sk.get_flush_lsn(tenant_id, timeline_id)
+    log.info(f"flush_lsn on dst after pull_timeline: {dst_flush_lsn}")
+    assert dst_flush_lsn >= src_flush_lsn
+    digests = [
+        sk.http_client().timeline_digest(tenant_id, timeline_id, timeline_start_lsn, dst_flush_lsn)
+        for sk in [src_sk, dst_sk]
+    ]
+    assert digests[0] == digests[1], f"digest on src is {digests[0]} but on dst is {digests[1]}"
+
+
 # In this test we check for excessive START_REPLICATION and START_WAL_PUSH queries
 # when compute is active, but there are no writes to the timeline. In that case
 # pageserver should maintain a single connection to safekeeper and don't attempt

test_runner/regress/test_wal_acceptor_async.py

@@ -531,6 +531,64 @@ def test_recovery_uncommitted(neon_env_builder: NeonEnvBuilder):
     asyncio.run(run_recovery_uncommitted(env))
 
 
+async def run_wal_truncation(env: NeonEnv):
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
+
+    (sk1, sk2, sk3) = env.safekeepers
+
+    ep = env.endpoints.create_start("main")
+    ep.safe_psql("create table t (key int, value text)")
+    ep.safe_psql("insert into t select generate_series(1, 100), 'payload'")
+
+    # insert with only one sk3 up to create tail of flushed but not committed WAL on it
+    sk1.stop()
+    sk2.stop()
+    conn = await ep.connect_async()
+    # query should hang, so execute in separate task
+    bg_query = asyncio.create_task(
+        conn.execute("insert into t select generate_series(1, 180000), 'Papaya'")
+    )
+    sleep_sec = 2
+    await asyncio.sleep(sleep_sec)
+    # it must still be not finished
+    assert not bg_query.done()
+    # note: destoy will kill compute_ctl, preventing it waiting for hanging sync-safekeepers.
+    ep.stop_and_destroy()
+
+    # stop sk3 as well
+    sk3.stop()
+
+    # now start sk1 and sk2 and make them commit something
+    sk1.start()
+    sk2.start()
+    ep = env.endpoints.create_start(
+        "main",
+    )
+    ep.safe_psql("insert into t select generate_series(1, 200), 'payload'")
+
+    # start sk3 and wait for it to catch up
+    sk3.start()
+    flush_lsn = Lsn(ep.safe_psql_scalar("SELECT pg_current_wal_flush_lsn()"))
+    await wait_for_lsn(sk3, tenant_id, timeline_id, flush_lsn)
+
+    timeline_start_lsn = sk1.get_timeline_start_lsn(tenant_id, timeline_id)
+    digests = [
+        sk.http_client().timeline_digest(tenant_id, timeline_id, timeline_start_lsn, flush_lsn)
+        for sk in [sk1, sk2]
+    ]
+    assert digests[0] == digests[1], f"digest on sk1 is {digests[0]} but on sk3 is {digests[1]}"
+
+
+# Simple deterministic test creating tail of WAL on safekeeper which is
+# truncated when majority without this sk elects walproposer starting earlier.
+def test_wal_truncation(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.num_safekeepers = 3
+    env = neon_env_builder.init_start()
+
+    asyncio.run(run_wal_truncation(env))
+
+
 async def run_segment_init_failure(env: NeonEnv):
     env.neon_cli.create_branch("test_segment_init_failure")
     ep = env.endpoints.create_start("test_segment_init_failure")