On 07/27/2012 10:31 AM, Craig Ringer wrote:
> Ouch. Sure looks like a leak to me, yeah.
... but it turns out I'm not thinking very straight. I forgot to check
the size of your `shared_buffers' or `work_mem' and forgot to get you to
report `free -m' output after each run to measure _real_ memory use.
During testing I did got a backend crash when running an INSERT - which
I didn't expect given that I have only 20MB of work_mem and 256MB of
shared_buffers. I was surprised by that as I would not have expected
that query to require a huge gob of RAM. I didn't dig much further as
I'm on a swapless system with overcommit enabled ( 'cos the Java VM
does't work with overcommit off ) which isn't exactly a recommended Pg
configuration.
--
Craig Ringer