The one caveat is that the memory rise only happens when using a HashAggregate query plan (included in the gist), which I can't find a way to get Postgres to reliably use.
If you need it, I could probably find another test case.
— Theron
On Thu, Aug 25, 2016 at 5:27 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
Hm, well the only thing there that looks even slightly out of the ordinary is the amount of free space in TopMemoryContext itself:
TopMemoryContext: 3525712 total in 432 blocks; 3444272 free (12654 chunks); 81440 used
Normally, TopMemoryContext doesn't get to more than a few hundred K, and in the cases I've seen where it does, it's usually been because of leaky coding that was allocating stuff there and never cleaning it up. But you've got no more than the typical amount of space still allocated there, which seems to kill the "leak in TopMemoryContext" theory. And in any case there is nowhere near 100MB accounted for by the whole dump.
Are you using any other PLs besides plpgsql? We've seen cases where bloat occurred within plpython or plperl, and wasn't visible in this dump because those languages don't use PG's memory management code. Or maybe some nonstandard extension?
If not that, then I'd have to speculate that the query you're running is triggering some bug or otherwise pathological behavior. Can you put together a self-contained test case?