On 02/10/17 18:59, Petr Jelinek wrote: >> >> Now fix the trigger function: >> CREATE OR REPLACE FUNCTION replication_trigger_proc() RETURNS TRIGGER AS $$ >> BEGIN >> RETURN NEW; >> END $$ LANGUAGE plpgsql; >> >> And manually perform at master two updates inside one transaction: >> >> postgres=# begin; >> BEGIN >> postgres=# update pgbench_accounts set abalance=abalance+1 where aid=26; >> UPDATE 1 >> postgres=# update pgbench_accounts set abalance=abalance-1 where aid=26; >> UPDATE 1 >> postgres=# commit; >> <hangs> >> >> and in replica log we see: >> 2017-10-02 18:40:26.094 MSK [2954] LOG: logical replication apply >> worker for subscription "sub" has started >> 2017-10-02 18:40:26.101 MSK [2954] ERROR: attempted to lock invisible >> tuple >> 2017-10-02 18:40:26.102 MSK [2882] LOG: worker process: logical >> replication worker for subscription 16403 (PID 2954) exited with exit >> code 1 >> >> Error happens in trigger.c: >> >> #3 0x000000000069bddb in GetTupleForTrigger (estate=0x2e36b50, >> epqstate=0x7ffc4420eda0, relinfo=0x2dcfe90, tid=0x2dd08ac, >> lockmode=LockTupleNoKeyExclusive, newSlot=0x7ffc4420ec40) at >> trigger.c:3103 >> #4 0x000000000069b259 in ExecBRUpdateTriggers (estate=0x2e36b50, >> epqstate=0x7ffc4420eda0, relinfo=0x2dcfe90, tupleid=0x2dd08ac, >> fdw_trigtuple=0x0, slot=0x2dd0240) at trigger.c:2748 >> #5 0x00000000006d2395 in ExecSimpleRelationUpdate (estate=0x2e36b50, >> epqstate=0x7ffc4420eda0, searchslot=0x2dd0358, slot=0x2dd0240) >> at execReplication.c:461 >> #6 0x0000000000820894 in apply_handle_update (s=0x7ffc442163b0) at >> worker.c:736 > > We have locked the same tuple in RelationFindReplTupleByIndex() just > before this gets called and didn't get the same error. I guess we do > something wrong with snapshots. Will need to investigate more. >
Okay, so it's not snapshot. It's the fact that we don't set the es_output_cid in replication worker which GetTupleForTrigger is using when locking the tuple. Attached one-liner fixes it.
This seems like a clear-cut bug with a simple fix.
Lets get this committed, so we don't lose it. The rest of the thread is going off into the weeds a bit issues unrelated to the original problem.