On 12.12.2014 22:13, Robert Haas wrote:
> On Fri, Dec 12, 2014 at 11:50 AM, Tomas Vondra <tv@fuzzy.cz> wrote:
>> On 12.12.2014 14:19, Robert Haas wrote:
>>> On Thu, Dec 11, 2014 at 5:46 PM, Tomas Vondra <tv@fuzzy.cz> wrote:
>>>
>>>> Regarding the "sufficiently small" - considering today's hardware, we're
>>>> probably talking about gigabytes. On machines with significant memory
>>>> pressure (forcing the temporary files to disk), it might be much lower,
>>>> of course. Of course, it also depends on kernel settings (e.g.
>>>> dirty_bytes/dirty_background_bytes).
>>>
>>> Well, this is sort of one of the problems with work_mem. When we
>>> switch to a tape sort, or a tape-based materialize, we're probably far
>>> from out of memory. But trying to set work_mem to the amount of
>>> memory we have can easily result in a memory overrun if a load spike
>>> causes lots of people to do it all at the same time. So we have to
>>> set work_mem conservatively, but then the costing doesn't really come
>>> out right. We could add some more costing parameters to try to model
>>> this, but it's not obvious how to get it right.
>>
>> Ummm, I don't think that's what I proposed. What I had in mind was a
>> flag "the batches are likely to stay in page cache". Because when it is
>> likely, batching is probably faster (compared to increased load factor).
>
> How will you know whether to set the flag?
I don't know. I just wanted to make it clear that I'm not suggesting
messing with work_mem (increasing it or whatewer). Or maybe I got your
comments about memory overrun etc. wrong - now that I read it again,
maybe it's meant just as an example of how difficult problem it is?
regards
Tomas
