Check relation size for gertpage only in case of error

## Problem

We observe image compaction errors after gc-compaction finishes
compacting below the gc_cutoff. This is because `repartition` returns an
LSN below the gc horizon as we (likely) determined that `distance <=
self.repartition_threshold`.

I think it's better to keep the current behavior of when to trigger
compaction but we should skip image compaction if the returned LSN is
below the gc horizon.

## Summary of changes

If the repartition returns an invalid LSN, skip image compaction.

Signed-off-by: Alex Chi Z <chi@neon.tech>

Cargo.lock

@@ -1303,6 +1303,7 @@ dependencies = [
  "futures",
  "http 1.1.0",
  "indexmap 2.0.1",
+ "itertools 0.10.5",
  "jsonwebtoken",
  "metrics",
  "nix 0.27.1",

compute_tools/Cargo.toml

@@ -28,6 +28,7 @@ flate2.workspace = true
 futures.workspace = true
 http.workspace = true
 indexmap.workspace = true
+itertools.workspace = true
 jsonwebtoken.workspace = true
 metrics.workspace = true
 nix.workspace = true

compute_tools/src/compute.rs

@@ -11,6 +11,7 @@ use compute_api::spec::{
 use futures::StreamExt;
 use futures::future::join_all;
 use futures::stream::FuturesUnordered;
+use itertools::Itertools;
 use nix::sys::signal::{Signal, kill};
 use nix::unistd::Pid;
 use once_cell::sync::Lazy;
@@ -18,7 +19,7 @@ use postgres;
 use postgres::NoTls;
 use postgres::error::SqlState;
 use remote_storage::{DownloadError, RemotePath};
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::net::SocketAddr;
 use std::os::unix::fs::{PermissionsExt, symlink};
 use std::path::Path;
@@ -1995,23 +1996,40 @@ LIMIT 100",
         tokio::spawn(conn);
 
         // TODO: support other types of grants apart from schemas?
-        let query = format!(
-            "GRANT {} ON SCHEMA {} TO {}",
-            privileges
-                .iter()
-                // should not be quoted as it's part of the command.
-                // is already sanitized so it's ok
-                .map(|p| p.as_str())
-                .collect::<Vec<&'static str>>()
-                .join(", "),
-            // quote the schema and role name as identifiers to sanitize them.
-            schema_name.pg_quote(),
-            role_name.pg_quote(),
-        );
-        db_client
-            .simple_query(&query)
+
+        // check the role grants first - to gracefully handle read-replicas.
+        let select = "SELECT privilege_type
+            FROM pg_namespace
+                JOIN LATERAL (SELECT * FROM aclexplode(nspacl) AS x) acl ON true
+                JOIN pg_user users ON acl.grantee = users.usesysid
+            WHERE users.usename = $1
+                AND nspname = $2";
+        let rows = db_client
+            .query(select, &[role_name, schema_name])
             .await
-            .with_context(|| format!("Failed to execute query: {}", query))?;
+            .with_context(|| format!("Failed to execute query: {select}"))?;
+
+        let already_granted: HashSet<String> = rows.into_iter().map(|row| row.get(0)).collect();
+
+        let grants = privileges
+            .iter()
+            .filter(|p| !already_granted.contains(p.as_str()))
+            // should not be quoted as it's part of the command.
+            // is already sanitized so it's ok
+            .map(|p| p.as_str())
+            .join(", ");
+
+        if !grants.is_empty() {
+            // quote the schema and role name as identifiers to sanitize them.
+            let schema_name = schema_name.pg_quote();
+            let role_name = role_name.pg_quote();
+
+            let query = format!("GRANT {grants} ON SCHEMA {schema_name} TO {role_name}",);
+            db_client
+                .simple_query(&query)
+                .await
+                .with_context(|| format!("Failed to execute query: {}", query))?;
+        }
 
         Ok(())
     }

libs/pageserver_api/src/key.rs

@@ -561,6 +561,21 @@ pub fn rel_block_to_key(rel: RelTag, blknum: BlockNumber) -> Key {
     }
 }
 
+#[inline(always)]
+pub fn key_to_rel_tag(key: Key) -> RelTag {
+    RelTag {
+        spcnode: key.field2,
+        dbnode: key.field3,
+        relnode: key.field4,
+        forknum: key.field5,
+    }
+}
+
+#[inline(always)]
+pub fn key_to_blknum(key: Key) -> BlockNumber {
+    key.field6
+}
+
 #[inline(always)]
 pub fn rel_size_to_key(rel: RelTag) -> Key {
     Key {

pageserver/src/pgdatadir_mapping.rs

@@ -16,9 +16,9 @@ use bytes::{Buf, Bytes, BytesMut};
 use enum_map::Enum;
 use pageserver_api::key::{
     AUX_FILES_KEY, CHECKPOINT_KEY, CONTROLFILE_KEY, CompactKey, DBDIR_KEY, Key, RelDirExists,
-    TWOPHASEDIR_KEY, dbdir_key_range, rel_block_to_key, rel_dir_to_key, rel_key_range,
-    rel_size_to_key, rel_tag_sparse_key, rel_tag_sparse_key_range, relmap_file_key,
-    repl_origin_key, repl_origin_key_range, slru_block_to_key, slru_dir_to_key,
+    TWOPHASEDIR_KEY, dbdir_key_range, key_to_blknum, key_to_rel_tag, rel_block_to_key,
+    rel_dir_to_key, rel_key_range, rel_size_to_key, rel_tag_sparse_key, rel_tag_sparse_key_range,
+    relmap_file_key, repl_origin_key, repl_origin_key_range, slru_block_to_key, slru_dir_to_key,
     slru_segment_key_range, slru_segment_size_to_key, twophase_file_key, twophase_key_range,
 };
 use pageserver_api::keyspace::{KeySpaceRandomAccum, SparseKeySpace};
@@ -259,7 +259,7 @@ impl Timeline {
         let mut result = Vec::with_capacity(pages.len());
         let result_slots = result.spare_capacity_mut();
 
-        let mut keys_slots: HashMap<Key, smallvec::SmallVec<[(usize, RequestContext); 1]>> =
+        let mut keys_slots: HashMap<Key, smallvec::SmallVec<[(usize, Lsn, RequestContext); 1]>> =
             HashMap::with_capacity(pages.len());
 
         let mut req_keyspaces: HashMap<Lsn, KeySpaceRandomAccum> =
@@ -275,43 +275,6 @@ impl Timeline {
                 continue;
             }
 
-            let nblocks = {
-                let ctx = RequestContextBuilder::from(&ctx)
-                    .perf_span(|crnt_perf_span| {
-                        info_span!(
-                            target: PERF_TRACE_TARGET,
-                            parent: crnt_perf_span,
-                            "GET_REL_SIZE",
-                            reltag=%tag,
-                            lsn=%lsn,
-                        )
-                    })
-                    .attached_child();
-
-                match self
-                    .get_rel_size(*tag, Version::Lsn(lsn), &ctx)
-                    .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
-                    .await
-                {
-                    Ok(nblocks) => nblocks,
-                    Err(err) => {
-                        result_slots[response_slot_idx].write(Err(err));
-                        slots_filled += 1;
-                        continue;
-                    }
-                }
-            };
-
-            if *blknum >= nblocks {
-                debug!(
-                    "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
-                    tag, blknum, lsn, nblocks
-                );
-                result_slots[response_slot_idx].write(Ok(ZERO_PAGE.clone()));
-                slots_filled += 1;
-                continue;
-            }
-
             let key = rel_block_to_key(*tag, *blknum);
 
             let ctx = RequestContextBuilder::from(&ctx)
@@ -326,7 +289,7 @@ impl Timeline {
                 .attached_child();
 
             let key_slots = keys_slots.entry(key).or_default();
-            key_slots.push((response_slot_idx, ctx));
+            key_slots.push((response_slot_idx, lsn, ctx));
 
             let acc = req_keyspaces.entry(lsn).or_default();
             acc.add_key(key);
@@ -347,42 +310,95 @@ impl Timeline {
             Ok(results) => {
                 for (key, res) in results {
                     let mut key_slots = keys_slots.remove(&key).unwrap().into_iter();
-                    let (first_slot, first_req_ctx) = key_slots.next().unwrap();
 
-                    for (slot, req_ctx) in key_slots {
-                        let clone = match &res {
-                            Ok(buf) => Ok(buf.clone()),
-                            Err(err) => Err(match err {
-                                PageReconstructError::Cancelled => PageReconstructError::Cancelled,
+                    // Try to check if error is caused by access beyond end of relation
+                    match &res {
+                        Err(err) => {
+                            let tag = key_to_rel_tag(key);
+                            let blknum = key_to_blknum(key);
+                            let mut first_error_slot: Option<usize> = None;
+                            for (slot, lsn, req_ctx) in key_slots {
+                                // Check relation size only in case of error
+                                let relsize_ctx = RequestContextBuilder::from(&ctx)
+                                    .perf_span(|crnt_perf_span| {
+                                        info_span!(
+                                         target: PERF_TRACE_TARGET,
+                                         parent: crnt_perf_span,
+                                         "GET_REL_SIZE",
+                                         reltag=%tag,
+                                         lsn=%lsn,
+                                        )
+                                    })
+                                    .attached_child();
 
-                                x @ PageReconstructError::Other(_)
-                                | x @ PageReconstructError::AncestorLsnTimeout(_)
-                                | x @ PageReconstructError::WalRedo(_)
-                                | x @ PageReconstructError::MissingKey(_) => {
-                                    PageReconstructError::Other(anyhow::anyhow!(
-                                        "there was more than one request for this key in the batch, error logged once: {x:?}"
-                                    ))
+                                if let Ok(nblocks) = self
+                                    .get_rel_size(tag, Version::Lsn(lsn), &relsize_ctx)
+                                    .maybe_perf_instrument(&ctx, |crnt_perf_span| {
+                                        crnt_perf_span.clone()
+                                    })
+                                    .await
+                                {
+                                    if blknum >= nblocks {
+                                        debug!(
+                                            "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
+                                            tag, blknum, lsn, nblocks
+                                        );
+                                        result_slots[slot].write(Ok(ZERO_PAGE.clone()));
+                                        slots_filled += 1;
+                                        continue;
+                                    }
                                 }
-                            }),
-                        };
+                                if first_error_slot.is_none() {
+                                    first_error_slot = Some(slot);
+                                } else {
+                                    let err = match err {
+                                        PageReconstructError::Cancelled => {
+                                            PageReconstructError::Cancelled
+                                        }
 
-                        result_slots[slot].write(clone);
-                        // There is no standardized way to express that the batched span followed from N request spans.
-                        // So, abuse the system and mark the request contexts as follows_from the batch span, so we get
-                        // some linkage in our trace viewer. It allows us to answer: which GET_VECTORED did this GET_PAGE wait for.
-                        req_ctx.perf_follows_from(ctx);
-                        slots_filled += 1;
+                                        x @ PageReconstructError::Other(_)
+                                        | x @ PageReconstructError::AncestorLsnTimeout(_)
+                                        | x @ PageReconstructError::WalRedo(_)
+                                        | x @ PageReconstructError::MissingKey(_) => {
+                                            PageReconstructError::Other(anyhow::anyhow!(
+                                                "there was more than one request for this key in the batch, error logged once: {x:?}"
+                                            ))
+                                        }
+                                    };
+                                    result_slots[slot].write(Err(err));
+                                };
+                                // There is no standardized way to express that the batched span followed from N request spans.
+                                // So, abuse the system and mark the request contexts as follows_from the batch span, so we get
+                                // some linkage in our trace viewer. It allows us to answer: which GET_VECTORED did this GET_PAGE wait for.
+                                req_ctx.perf_follows_from(ctx);
+                                slots_filled += 1;
+                            }
+                            if let Some(slot) = first_error_slot {
+                                result_slots[slot].write(res);
+                            }
+                        }
+                        Ok(buf) => {
+                            let (first_slot, _first_lsn, first_req_ctx) = key_slots.next().unwrap();
+
+                            for (slot, _lsn, req_ctx) in key_slots {
+                                result_slots[slot].write(Ok(buf.clone()));
+                                // There is no standardized way to express that the batched span followed from N request spans.
+                                // So, abuse the system and mark the request contexts as follows_from the batch span, so we get
+                                // some linkage in our trace viewer. It allows us to answer: which GET_VECTORED did this GET_PAGE wait for.
+                                req_ctx.perf_follows_from(ctx);
+                                slots_filled += 1;
+                            }
+                            result_slots[first_slot].write(res);
+                            first_req_ctx.perf_follows_from(ctx);
+                            slots_filled += 1;
+                        }
                     }
-
-                    result_slots[first_slot].write(res);
-                    first_req_ctx.perf_follows_from(ctx);
-                    slots_filled += 1;
                 }
             }
             Err(err) => {
                 // this cannot really happen because get_vectored only errors globally on invalid LSN or too large batch size
                 // (We enforce the max batch size outside of this function, in the code that constructs the batch request.)
-                for (slot, req_ctx) in keys_slots.values().flatten() {
+                for (slot, _lsn, req_ctx) in keys_slots.values().flatten() {
                     // this whole `match` is a lot like `From<GetVectoredError> for PageReconstructError`
                     // but without taking ownership of the GetVectoredError
                     let err = match &err {
@@ -488,8 +504,6 @@ impl Timeline {
         let mut buf = version.get(self, key, ctx).await?;
         let nblocks = buf.get_u32_le();
 
-        self.update_cached_rel_size(tag, version.get_lsn(), nblocks);
-
         Ok(nblocks)
     }
 
@@ -1343,32 +1357,6 @@ impl Timeline {
         None
     }
 
-    /// Update cached relation size if there is no more recent update
-    pub fn update_cached_rel_size(&self, tag: RelTag, lsn: Lsn, nblocks: BlockNumber) {
-        let mut rel_size_cache = self.rel_size_cache.write().unwrap();
-
-        if lsn < rel_size_cache.complete_as_of {
-            // Do not cache old values. It's safe to cache the size on read, as long as
-            // the read was at an LSN since we started the WAL ingestion. Reasoning: we
-            // never evict values from the cache, so if the relation size changed after
-            // 'lsn', the new value is already in the cache.
-            return;
-        }
-
-        match rel_size_cache.map.entry(tag) {
-            hash_map::Entry::Occupied(mut entry) => {
-                let cached_lsn = entry.get_mut();
-                if lsn >= cached_lsn.0 {
-                    *cached_lsn = (lsn, nblocks);
-                }
-            }
-            hash_map::Entry::Vacant(entry) => {
-                entry.insert((lsn, nblocks));
-                RELSIZE_CACHE_ENTRIES.inc();
-            }
-        }
-    }
-
     /// Store cached relation size
     pub fn set_cached_rel_size(&self, tag: RelTag, lsn: Lsn, nblocks: BlockNumber) {
         let mut rel_size_cache = self.rel_size_cache.write().unwrap();

pageserver/src/tenant/timeline.rs

@@ -199,11 +199,8 @@ pub struct TimelineResources {
 
 /// The relation size cache caches relation sizes at the end of the timeline. It speeds up WAL
 /// ingestion considerably, because WAL ingestion needs to check on most records if the record
-/// implicitly extends the relation.  At startup, `complete_as_of` is initialized to the current end
-/// of the timeline (disk_consistent_lsn).  It's used on reads of relation sizes to check if the
-/// value can be used to also update the cache, see [`Timeline::update_cached_rel_size`].
+/// implicitly extends the relation.
 pub(crate) struct RelSizeCache {
-    pub(crate) complete_as_of: Lsn,
     pub(crate) map: HashMap<RelTag, (Lsn, BlockNumber)>,
 }
 
@@ -2970,7 +2967,6 @@ impl Timeline {
 
                 last_received_wal: Mutex::new(None),
                 rel_size_cache: RwLock::new(RelSizeCache {
-                    complete_as_of: disk_consistent_lsn,
                     map: HashMap::new(),
                 }),
 

pageserver/src/tenant/timeline/compaction.rs

@@ -1277,6 +1277,8 @@ impl Timeline {
             return Ok(CompactionOutcome::YieldForL0);
         }
 
+        let gc_cutoff = *self.applied_gc_cutoff_lsn.read();
+
         // 2. Repartition and create image layers if necessary
         match self
             .repartition(
@@ -1287,7 +1289,7 @@ impl Timeline {
             )
             .await
         {
-            Ok(((dense_partitioning, sparse_partitioning), lsn)) => {
+            Ok(((dense_partitioning, sparse_partitioning), lsn)) if lsn >= gc_cutoff => {
                 // Disables access_stats updates, so that the files we read remain candidates for eviction after we're done with them
                 let image_ctx = RequestContextBuilder::from(ctx)
                     .access_stats_behavior(AccessStatsBehavior::Skip)
@@ -1341,6 +1343,10 @@ impl Timeline {
                 }
             }
 
+            Ok(_) => {
+                info!("skipping repartitioning due to image compaction LSN being below GC cutoff");
+            }
+
             // Suppress errors when cancelled.
             Err(_) if self.cancel.is_cancelled() => {}
             Err(err) if err.is_cancel() => {}
@@ -3606,6 +3612,13 @@ impl Timeline {
                     last_key = Some(key);
                 }
                 accumulated_values.push((key, lsn, val));
+
+                if accumulated_values.len() >= 65536 {
+                    // Assume all of them are images, that would be 512MB of data in memory for a single key.
+                    return Err(CompactionError::Other(anyhow!(
+                        "too many values for a single key, giving up gc-compaction"
+                    )));
+                }
             } else {
                 let last_key: &mut Key = last_key.as_mut().unwrap();
                 stat.on_unique_key_visited(); // TODO: adjust statistics for partial compaction

pgxn/neon/communicator.c

@@ -425,15 +425,12 @@ compact_prefetch_buffers(void)
  * point inside and outside PostgreSQL.
  *
  * This still does throw errors when it receives malformed responses from PS.
- *
- * When we're not called from CHECK_FOR_INTERRUPTS (indicated by
- * IsHandlingInterrupts) we also report we've ended prefetch receive work,
- * just in case state tracking was lost due to an error in the sync getPage
- * response code.
  */
 void
-communicator_prefetch_pump_state(bool IsHandlingInterrupts)
+communicator_prefetch_pump_state(void)
 {
+	START_PREFETCH_RECEIVE_WORK();
+
 	while (MyPState->ring_receive != MyPState->ring_flush)
 	{
 		NeonResponse   *response;
@@ -482,9 +479,7 @@ communicator_prefetch_pump_state(bool IsHandlingInterrupts)
 		}
 	}
 
-	/* We never pump the prefetch state while handling other pages */
-	if (!IsHandlingInterrupts)
-		END_PREFETCH_RECEIVE_WORK();
+	END_PREFETCH_RECEIVE_WORK();
 
 	communicator_reconfigure_timeout_if_needed();
 }
@@ -672,9 +667,10 @@ prefetch_wait_for(uint64 ring_index)
 
 	Assert(MyPState->ring_unused > ring_index);
 
+	START_PREFETCH_RECEIVE_WORK();
+
 	while (MyPState->ring_receive <= ring_index)
 	{
-		START_PREFETCH_RECEIVE_WORK();
 		entry = GetPrfSlot(MyPState->ring_receive);
 
 		Assert(entry->status == PRFS_REQUESTED);
@@ -683,17 +679,18 @@ prefetch_wait_for(uint64 ring_index)
 			result = false;
 			break;
 		}
-
-		END_PREFETCH_RECEIVE_WORK();
 		CHECK_FOR_INTERRUPTS();
 	}
+
 	if (result)
 	{
 		/* Check that slot is actually received (srver can be disconnected in prefetch_pump_state called from CHECK_FOR_INTERRUPTS */
 		PrefetchRequest *slot = GetPrfSlot(ring_index);
-		return slot->status == PRFS_RECEIVED;
+		result = slot->status == PRFS_RECEIVED;
 	}
-	return false;
+	END_PREFETCH_RECEIVE_WORK();
+
+	return result;
 ;
 }
 
@@ -720,6 +717,7 @@ prefetch_read(PrefetchRequest *slot)
 	Assert(slot->status == PRFS_REQUESTED);
 	Assert(slot->response == NULL);
 	Assert(slot->my_ring_index == MyPState->ring_receive);
+	Assert(readpage_reentrant_guard);
 
 	if (slot->status != PRFS_REQUESTED ||
 		slot->response != NULL ||
@@ -802,6 +800,7 @@ communicator_prefetch_receive(BufferTag tag)
 	PrfHashEntry *entry;
 	PrefetchRequest hashkey;
 
+	Assert(readpage_reentrant_guard);
 	hashkey.buftag = tag;
 	entry = prfh_lookup(MyPState->prf_hash, &hashkey);
 	if (entry != NULL && prefetch_wait_for(entry->slot->my_ring_index))
@@ -821,8 +820,12 @@ communicator_prefetch_receive(BufferTag tag)
 void
 prefetch_on_ps_disconnect(void)
 {
+	bool save_readpage_reentrant_guard = readpage_reentrant_guard;
 	MyPState->ring_flush = MyPState->ring_unused;
 
+	/* Prohibit callig of prefetch_pump_state */
+	START_PREFETCH_RECEIVE_WORK();
+
 	while (MyPState->ring_receive < MyPState->ring_unused)
 	{
 		PrefetchRequest *slot;
@@ -851,6 +854,9 @@ prefetch_on_ps_disconnect(void)
 		MyNeonCounters->getpage_prefetch_discards_total += 1;
 	}
 
+	/* Restore guard */
+	readpage_reentrant_guard = save_readpage_reentrant_guard;
+
 	/*
 	 * We can have gone into retry due to network error, so update stats with
 	 * the latest available
@@ -2509,7 +2515,7 @@ communicator_processinterrupts(void)
 	if (timeout_signaled)
 	{
 		if (!readpage_reentrant_guard && readahead_getpage_pull_timeout_ms > 0)
-			communicator_prefetch_pump_state(true);
+			communicator_prefetch_pump_state();
 
 		timeout_signaled = false;
 		communicator_reconfigure_timeout_if_needed();

pgxn/neon/communicator.h

@@ -44,7 +44,7 @@ extern int communicator_read_slru_segment(SlruKind kind, int64 segno,
 										  void *buffer);
 
 extern void communicator_reconfigure_timeout_if_needed(void);
-extern void communicator_prefetch_pump_state(bool IsHandlingInterrupts);
+extern void communicator_prefetch_pump_state(void);
 
 
 #endif

pgxn/neon/pagestore_smgr.c

@@ -1179,7 +1179,7 @@ neon_prefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 		blocknum += iterblocks;
 	}
 
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 	return false;
 }
@@ -1218,7 +1218,7 @@ neon_prefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum)
 
 	communicator_prefetch_register_bufferv(tag, NULL, 1, NULL);
 
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 	return false;
 }
@@ -1262,7 +1262,7 @@ neon_writeback(SMgrRelation reln, ForkNumber forknum,
 	 */
 	neon_log(SmgrTrace, "writeback noop");
 
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 #ifdef DEBUG_COMPARE_LOCAL
 	if (IS_LOCAL_REL(reln))
@@ -1315,7 +1315,7 @@ neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer
 	}
 
 	/* Try to read PS results if they are available */
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 	neon_get_request_lsns(InfoFromSMgrRel(reln), forkNum, blkno, &request_lsns, 1);
 
@@ -1339,7 +1339,7 @@ neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer
 	/*
 	 * Try to receive prefetch results once again just to make sure we don't leave the smgr code while the OS might still have buffered bytes.
 	 */
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 #ifdef DEBUG_COMPARE_LOCAL
 	if (forkNum == MAIN_FORKNUM && IS_LOCAL_REL(reln))
@@ -1449,7 +1449,7 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 				 nblocks, PG_IOV_MAX);
 
 	/* Try to read PS results if they are available */
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 	neon_get_request_lsns(InfoFromSMgrRel(reln), forknum, blocknum,
 						  request_lsns, nblocks);
@@ -1480,7 +1480,7 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 	/*
 	 * Try to receive prefetch results once again just to make sure we don't leave the smgr code while the OS might still have buffered bytes.
 	 */
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 #ifdef DEBUG_COMPARE_LOCAL
 	if (forknum == MAIN_FORKNUM && IS_LOCAL_REL(reln))
@@ -1665,7 +1665,7 @@ neon_write(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, const vo
 
 	lfc_write(InfoFromSMgrRel(reln), forknum, blocknum, buffer);
 
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 #ifdef DEBUG_COMPARE_LOCAL
 	if (IS_LOCAL_REL(reln))
@@ -1727,7 +1727,7 @@ neon_writev(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
 
 	lfc_writev(InfoFromSMgrRel(reln), forknum, blkno, buffers, nblocks);
 
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 #ifdef DEBUG_COMPARE_LOCAL
 	if (IS_LOCAL_REL(reln))
@@ -1902,7 +1902,7 @@ neon_immedsync(SMgrRelation reln, ForkNumber forknum)
 
 	neon_log(SmgrTrace, "[NEON_SMGR] immedsync noop");
 
-	communicator_prefetch_pump_state(false);
+	communicator_prefetch_pump_state();
 
 #ifdef DEBUG_COMPARE_LOCAL
 	if (IS_LOCAL_REL(reln))

storage_controller/src/service.rs

@@ -3886,10 +3886,10 @@ impl Service {
 
             None
         } else if safekeepers {
-            // Note that we do not support creating the timeline on the safekeepers
-            // for imported timelines. The `start_lsn` of the timeline is not known
-            // until the import finshes.
-            // https://github.com/neondatabase/neon/issues/11569
+            // Note that for imported timelines, we do not create the timeline on the safekeepers
+            // straight away. Instead, we do it once the import finalized such that we know what
+            // start LSN to provide for the safekeepers. This is done in
+            // [`Self::finalize_timeline_import`].
             let res = self
                 .tenant_timeline_create_safekeepers(tenant_id, &timeline_info)
                 .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
@@ -3966,11 +3966,22 @@ impl Service {
                 let active = self.timeline_active_on_all_shards(&import).await?;
 
                 match active {
-                    true => {
+                    Some(timeline_info) => {
                         tracing::info!("Timeline became active on all shards");
+
+                        if self.config.timelines_onto_safekeepers {
+                            // Now that we know the start LSN of this timeline, create it on the
+                            // safekeepers.
+                            self.tenant_timeline_create_safekeepers_until_success(
+                                import.tenant_id,
+                                timeline_info,
+                            )
+                            .await?;
+                        }
+
                         break;
                     }
-                    false => {
+                    None => {
                         tracing::info!("Timeline not active on all shards yet");
 
                         tokio::select! {
@@ -4004,9 +4015,6 @@ impl Service {
             .range_mut(TenantShardId::tenant_range(import.tenant_id))
             .for_each(|(_id, shard)| shard.importing = TimelineImportState::Idle);
 
-        // TODO(vlad): Timeline creations in import mode do not return a correct initdb lsn,
-        // so we can't create the timeline on the safekeepers. Fix by moving creation here.
-        // https://github.com/neondatabase/neon/issues/11569
         tracing::info!(%import_failed, "Timeline import complete");
 
         Ok(())
@@ -4021,10 +4029,16 @@ impl Service {
         .await;
     }
 
+    /// If the timeline is active on all shards, returns the [`TimelineInfo`]
+    /// collected from shard 0.
+    ///
+    /// An error is returned if the shard layout has changed during the import.
+    /// This is guarded against within the storage controller and the pageserver,
+    /// and, therefore, unexpected.
     async fn timeline_active_on_all_shards(
         self: &Arc<Self>,
         import: &TimelineImport,
-    ) -> anyhow::Result<bool> {
+    ) -> anyhow::Result<Option<TimelineInfo>> {
         let targets = {
             let locked = self.inner.read().unwrap();
             let mut targets = Vec::new();
@@ -4048,13 +4062,17 @@ impl Service {
                         .expect("Pageservers may not be deleted while referenced");
                     targets.push((*tenant_shard_id, node.clone()));
                 } else {
-                    return Ok(false);
+                    return Ok(None);
                 }
             }
 
             targets
         };
 
+        if targets.is_empty() {
+            anyhow::bail!("No shards found to finalize import for");
+        }
+
         let results = self
             .tenant_for_shards_api(
                 targets,
@@ -4070,10 +4088,17 @@ impl Service {
             )
             .await;
 
-        Ok(results.into_iter().all(|res| match res {
+        let all_active = results.iter().all(|res| match res {
             Ok(info) => info.state == TimelineState::Active,
             Err(_) => false,
-        }))
+        });
+
+        if all_active {
+            // Both unwraps are validated above
+            Ok(Some(results.into_iter().next().unwrap().unwrap()))
+        } else {
+            Ok(None)
+        }
     }
 
     pub(crate) async fn tenant_timeline_archival_config(

storage_controller/src/service/safekeeper_service.rs

@@ -323,6 +323,42 @@ impl Service {
         })
     }
 
+    pub(crate) async fn tenant_timeline_create_safekeepers_until_success(
+        self: &Arc<Self>,
+        tenant_id: TenantId,
+        timeline_info: TimelineInfo,
+    ) -> anyhow::Result<()> {
+        const BACKOFF: Duration = Duration::from_secs(5);
+
+        loop {
+            if self.cancel.is_cancelled() {
+                anyhow::bail!("Shut down requested while finalizing import");
+            }
+
+            let res = self
+                .tenant_timeline_create_safekeepers(tenant_id, &timeline_info)
+                .await;
+
+            match res {
+                Ok(_) => {
+                    tracing::info!("Timeline created on safekeepers");
+                    break;
+                }
+                Err(err) => {
+                    tracing::error!("Failed to create timeline on safekeepers: {err}");
+                    tokio::select! {
+                        _ = self.cancel.cancelled() => {
+                            anyhow::bail!("Shut down requested while finalizing import");
+                        },
+                        _ = tokio::time::sleep(BACKOFF) => {}
+                    };
+                }
+            }
+        }
+
+        Ok(())
+    }
+
     /// Directly insert the timeline into the database without reconciling it with safekeepers.
     ///
     /// Useful if the timeline already exists on the specified safekeepers,

test_runner/fixtures/neon_fixtures.py

@@ -4613,7 +4613,10 @@ class EndpointFactory:
         return self
 
     def new_replica(
-        self, origin: Endpoint, endpoint_id: str, config_lines: list[str] | None = None
+        self,
+        origin: Endpoint,
+        endpoint_id: str | None = None,
+        config_lines: list[str] | None = None,
     ):
         branch_name = origin.branch_name
         assert origin in self.endpoints
@@ -4629,7 +4632,10 @@ class EndpointFactory:
         )
 
     def new_replica_start(
-        self, origin: Endpoint, endpoint_id: str, config_lines: list[str] | None = None
+        self,
+        origin: Endpoint,
+        endpoint_id: str | None = None,
+        config_lines: list[str] | None = None,
     ):
         branch_name = origin.branch_name
         assert origin in self.endpoints

test_runner/regress/test_import_pgdata.py

@@ -24,6 +24,7 @@ from fixtures.utils import (
     skip_in_debug_build,
     wait_until,
 )
+from fixtures.workload import Workload
 from mypy_boto3_kms import KMSClient
 from mypy_boto3_kms.type_defs import EncryptResponseTypeDef
 from mypy_boto3_s3 import S3Client
@@ -97,6 +98,10 @@ def test_pgdata_import_smoke(
         f"http://{cplane_mgmt_api_server.host}:{cplane_mgmt_api_server.port}/storage/api/v1/"
     )
 
+    if neon_env_builder.storage_controller_config is None:
+        neon_env_builder.storage_controller_config = {}
+    neon_env_builder.storage_controller_config["timelines_onto_safekeepers"] = True
+
     env = neon_env_builder.init_start()
 
     # The test needs LocalFs support, which is only built in testing mode.
@@ -286,34 +291,28 @@ def test_pgdata_import_smoke(
     #
     # validate that we can write
     #
-    rw_endpoint = env.endpoints.create_start(
-        branch_name=import_branch_name,
-        endpoint_id="rw",
-        tenant_id=tenant_id,
-        config_lines=ep_config,
-    )
-    rw_endpoint.safe_psql("create table othertable(values text)")
-    rw_lsn = Lsn(rw_endpoint.safe_psql_scalar("select pg_current_wal_flush_lsn()"))
+    workload = Workload(env, tenant_id, timeline_id, branch_name=import_branch_name)
+    workload.init()
+    workload.write_rows(64)
+    workload.validate()
 
-    # TODO: consider using `class Workload` here
-    # to do compaction and whatnot?
+    rw_lsn = Lsn(workload.endpoint().safe_psql_scalar("select pg_current_wal_flush_lsn()"))
 
     #
     # validate that we can branch (important use case)
     #
 
     # ... at the tip
-    _ = env.create_branch(
+    child_timeline_id = env.create_branch(
         new_branch_name="br-tip",
         ancestor_branch_name=import_branch_name,
         tenant_id=tenant_id,
         ancestor_start_lsn=rw_lsn,
     )
-    br_tip_endpoint = env.endpoints.create_start(
-        branch_name="br-tip", endpoint_id="br-tip-ro", tenant_id=tenant_id, config_lines=ep_config
-    )
-    validate_vanilla_equivalence(br_tip_endpoint)
-    br_tip_endpoint.safe_psql("select * from othertable")
+    child_workload = workload.branch(timeline_id=child_timeline_id, branch_name="br-tip")
+    child_workload.validate()
+
+    validate_vanilla_equivalence(child_workload.endpoint())
 
     # ... at the initdb lsn
     _ = env.create_branch(
@@ -330,7 +329,7 @@ def test_pgdata_import_smoke(
     )
     validate_vanilla_equivalence(br_initdb_endpoint)
     with pytest.raises(psycopg2.errors.UndefinedTable):
-        br_initdb_endpoint.safe_psql("select * from othertable")
+        br_initdb_endpoint.safe_psql(f"select * from {workload.table}")
 
 
 @run_only_on_default_postgres(reason="PG version is irrelevant here")

test_runner/regress/test_role_grants.py

@@ -39,3 +39,10 @@ def test_role_grants(neon_simple_env: NeonEnv):
         res = cur.fetchall()
 
         assert res == [(1,)], "select should not succeed"
+
+    # confirm that replicas can also ensure the grants are correctly set.
+    replica = env.endpoints.new_replica_start(endpoint)
+    replica_client = replica.http_client()
+    replica_client.set_role_grants(
+        "test_role_grants", "test_role", "test_schema", ["CREATE", "USAGE"]
+    )

vendor/revisions.json

@@ -1,18 +1,18 @@
 {
   "v17": [
     "17.4",
-    "b763ab54b98d232a0959371ab1d07f06ed77c49e"
+    "0d59c91c1a23e667f1d1169d5f040b3fa0a0ab44"
   ],
   "v16": [
     "16.8",
-    "05ddf212e2e07b788b5c8b88bdcf98630941f6ae"
+    "d72d76f2cdee4194dd052ce099e9784aca7c794a"
   ],
   "v15": [
     "15.12",
-    "b838c8969b7c63f3e637a769656f5f36793b797c"
+    "72f83df76c61ce18d81bd371f0afd2a43d59c052"
   ],
   "v14": [
     "14.17",
-    "108856a4ae76be285b04497a0ed08fcbe60ddbe9"
+    "06b405bc982fd53522689aa4acbfd9c44b7993cf"
   ]
 }