I have a hot_standby system and use it to bear the load of various reporting
queries that take 15-60 minutes each. In an effort to avoid long pauses in
recovery, I set a vacuum_defer_cleanup_age constituting roughly three hours of
the master's transactions. Even so, I kept seeing recovery pause for the
duration of a long-running query. In each case, the culprit record was an
XLOG_BTREE_DELETE arising from on-the-fly deletion of an index tuple. The
attached test script demonstrates the behavior (on HEAD); the index tuple
reclamation conflicts with a concurrent "SELECT pg_sleep(600)" on the standby.
Since this inserting transaction aborts, HeapTupleSatisfiesVacuum reports
HEAPTUPLE_DEAD independent of vacuum_defer_cleanup_age. We go ahead and remove
the index tuples. On the standby, btree_xlog_delete_get_latestRemovedXid does
not regard the inserting-transaction outcome, so btree_redo proceeds to conflict
with snapshots having visibility over that transaction. Could we correctly
improve this by teaching btree_xlog_delete_get_latestRemovedXid to ignore tuples
of aborted transactions and tuples inserted and deleted within one transaction?
Thanks,
nm