When the first error was reported, we entered AbortTransaction(), proc->xid was changed to InvalidTransactionId in ProcArrayEndTransactionInternal().
After AbortTransaction() is done, we enter into CleanupTransaction(), and error would be reported again. So we enter AbortTransaction() again because
the transaction block state is still TBLOCK_STARTED. MyProc->xid now is InvalidTransactionId, so the assert failed.
Changing the behavior of tuplestore can work for this issue, but I'm not sure if this change will affect other components which depend on BufFile(like nodeMaterial)
At Wed, 21 Feb 2024 12:00:01 +0000, PG Bug reporting form <noreply@postgresql.org> wrote in
> triggers two errors, a warning, and an assertion failure:
> ERROR: temporary file size exceeds temp_file_limit (100kB)
> WARNING: AbortTransaction while in ABORT state
> ERROR: temporary file size exceeds temp_file_limit (100kB)
> server closed the connection unexpectedly
...
> For the second error:
> 2024-02-21 11:40:07.001 UTC|law|regression|65d5e116.13cfcb|ERROR: temporary
> file size exceeds temp_file_limit (100kB)
> 2024-02-21 11:40:07.001 UTC|law|regression|65d5e116.13cfcb|BACKTRACE:
> FileWrite at fd.c:2183:5
> BufFileDumpBuffer at buffile.c:537:18
> BufFileFlush at buffile.c:723:3
> BufFileClose at buffile.c:419:9
> tuplestore_end at tuplestore.c:459:5
> MemoryContextSwitchTo at palloc.h:142:23
> (inlined by) PortalDrop at portalmem.c:587:3
> AtCleanup_Portals at portalmem.c:907:3
Therefore, BufFileClose should not flush the content during error
handling. In the first place, tuplestore doesn't need to flush the
underlying files in _end and _clear. In this case, I would choose to
change the general behavior of tuplestore. The attached PoC patch
fixes the issue for me. It introduces a new "extended" function to
control flushing, avoiding the addition of an unnatural parameter to
BufFileClose. I suspect that it is usable in some other places, but I
haven't checked that.
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center
--