The Hermit Hacker <scrappy@hub.org> writes:
> I think that, based on this, the changes should be back'd out of v6.5.2
> until further testing and analysis can be done.
If we can't find a solution to the inval-too-late problem pronto,
what we can do is comment out the FileUnlink call in mdtruncate.
I don't see a need to back out the other fixes in md.c.
But I think we ought to fix the underlying problem, not this symptom.
What we now see is that after one backend has done something that
requires invalidating a relcache entry, another backend is able
to complete an entire query using the *old* relcache info before it
notices the shared-inval signal. That's got to have bad consequences
for more than just md.c.
regards, tom lane