Re: buildfarm: strange OOM failures on markhor (running CLOBBER_CACHE_RECURSIVELY)

On 18.5.2014 20:49, Tom Lane wrote:
> With both of these things fixed, I'm not seeing any significant memory
> bloat during the first parallel group of the regression tests. I don't
> think I'll have the patience to let it run much further than that
> (the uuid and enum tests are still running after an hour :-().

I've been watching barnacle running the tests with this commit applied,
and so far it seems fine - the memory consumption is hovering around
240MB, and it's running for ~8 hours now without a single OOM.
Previously the first OOM happened within ~3 hours.

I've also included a simple python script that checks amount of memory
used by the postgres processes every minute, and if it's more that 512MB
uses gdb to flush memory context statistics into postmaster log.
So far there was no such case.

Sadly, with 'mite' this resulted in 900+ MB postmaster.log, which then
apparently caused the buildfarm client to be killed by OOM (seems the
client is loading the log into memory, not expecting it to be this
large). That's why there are no results from mite available.

BTW, I see this was applied to HEAD only. Shouldn't this be applied to
the other branches too?

regards
Tomas