postgres.git / summary / log / commit / refs

commit    eeed86523bb9cd7064e540f29e057d123e18914a
Author:   Michael Paquier <michael@paquier.xyz>
Date:     Fri Jun 05 23:53:00 2026 +0000

    Lift shutdown assertion in pgstats for WAL senders
    
    Before v17, WAL senders can shut down after the checkpointer.  If a WAL
    sender still has pending statistics when the checkpointer has already
    exited, its shutdown callback may attempt to report those statistics and
    trigger assertions in pgstats.  In that case, the pending statistics are
    lost.
    
    This commit adjusts the assertion handling so that attempts to report
    pending WAL sender statistics after the checkpointer has completed its
    final stats flush are skipped.
    
    Preserving the existing assertion would require backpatching an
    equivalent of 87a6690cc69, ensuring that the checkpointer is always the
    last process to exit.  Such a change would be considerably more invasive
    and risky for stable branches because it alters the shutdown sequence,
    and the consequence is only some loss of stats data for the WAL sender.
    
    This assertion failure was periodically detected in the buildfarm,
    leading to spurious failures.
    
    Reported-by: Alexander Lakhin <exclusion@gmail.com>
    Reviewed-by: Xuneng Zhou <xunengzhou@gmail.com>
    Discussion: https://postgr.es/m/18158-88f667028dbc7e7b@postgresql.org
    Backpatch-through: 15-17


 src/backend/utils/activity/pgstat.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)

diff --git a/src/backend/utils/activity/pgstat.c b/src/backend/utils/activity/pgstat.c
index b66b6dba1bf..45e90cbcbd2 100644
--- a/src/backend/utils/activity/pgstat.c
+++ b/src/backend/utils/activity/pgstat.c
@@ -97,6 +97,7 @@
 #include "lib/dshash.h"
 #include "pgstat.h"
 #include "port/atomics.h"
+#include "replication/walsender.h"
 #include "storage/fd.h"
 #include "storage/ipc.h"
 #include "storage/lwlock.h"
@@ -509,8 +510,12 @@ pgstat_shutdown_hook(int code, Datum arg)
 
 	pgstat_report_stat(true);
 
-	/* there shouldn't be any pending changes left */
-	Assert(dlist_is_empty(&pgStatPending));
+	/*
+	 * There shouldn't be any pending changes left, unless this is a WAL
+	 * sender that would shut down after the checkpointer has flushed the
+	 * stats.
+	 */
+	Assert(dlist_is_empty(&pgStatPending) || am_walsender);
 	dlist_init(&pgStatPending);
 
 	pgstat_detach_shmem();
@@ -600,7 +605,13 @@ pgstat_report_stat(bool force)
 	 * assert that before the checks above, as there is an unconditional
 	 * pgstat_report_stat() call in pgstat_shutdown_hook() - which at least
 	 * the process that ran pgstat_before_server_shutdown() will still call.
+	 *
+	 * WAL senders would be shut down after the checkpointer and may still
+	 * have stats.  Skip them.
 	 */
+	if (pgStatLocal.shmem->is_shutdown && am_walsender)
+		return 0;
+
 	Assert(!pgStatLocal.shmem->is_shutdown);
 
 	now = GetCurrentTransactionStopTimestamp();


[parent: 51f63ba2bf7f]