Andrew Dunstan <andrew@dunslane.net> writes:
> Here's a little optimization for the parallel restore code, that
> inhibits reopening the archive file unless the worker will actually need
> to read from the file (i.e. a data member). It seems to work OK on both
> Linux and Windows, and I propose to apply it in a day or two.
I think you should close the file immediately at fork if you're not
going to reopen it --- otherwise it's a foot-gun waiting to fire.
IOW, not this, but something more like
if (te->section == SECTION_DATA) (AH->ReopenPtr) (AH);else (AH->ClosePtr) (AH);
... worker task ...
if (te->section == SECTION_DATA) (AH->ClosePtr) (AH);
> I've seen a recent error that suggests we are clobbering memory
> somewhere in the parallel code, as well as Olivier Prennant's reported
> error that suggests the same thing, although I'm blessed if I can see
> where it might be. Maybe some more eyeballs on the code would help.
Can you put together even a weakly reproducible test case? Something
that only fails every tenth or hundredth time would still help.
regards, tom lane