Bruce Momjian <maillist@candle.pha.pa.us> writes:
>> btw, vacuum crash when postmaster starts with -B 1024 option
>> seems fixed now !!!! I tried many times 'vacuum analyze' and never
>> get problem ! Probably this is a bonus of last fixes :-)
> My guess is the catalog changes Tom did fixed it. Vacuum analyze uses
> them quite a bit, and buffer cache size could affect which duplicate was
> picked.
Hmm, that would be an unexpected side benefit, wouldn't it! It could be
true, if vacuum depends on pg_operator entries. That'd explain why the
rest of us couldn't duplicate Oleg's problem: I'll bet no one who tried
had tables containing the data types that had bogus entries. (In fact,
I imagine you need to have some *indexes* on those data types before
you'd see such a problem in vacuum, no?)
It occurs to me that there ought to be a VACUUM ANALYZE somewhere in
the regression suite, probably at the end where it has a whole database
of weird stuff to chew on.
regards, tom lane