2010/12/10 Tom Lane <tgl@sss.pgh.pa.us>:
> I wrote:
>> We're throwing away one tuple at a time as we advance forward through
>> the tuplestore, and moving 100000+ tuple pointers each time. Ugh.
>> This code was all right when written, because (IIRC) the mergejoin
>> case was actually the only caller. But it's not all right for
>> WindowAgg's less-predictable usage patterns.
>
>> I thought for a bit about changing things around so that the first-used
>> tuple slot isn't necessarily state->memtuples[0], but just like the
>> comment says, that complicates a lot of other logic. And there isn't
>> any easy place to reclaim the wasted slots later.
>
>> What seems like the best bet is to put in a heuristic to make
>> tuplestore_trim simply not do anything until nremove reaches some
>> reasonably large amount, perhaps 10% of the number of stored tuples.
>> This wastes up to 10% of the alloted memory, but that seems tolerable.
>
> On reflection I think just not doing anything isn't a very good idea.
> The problem with that is that a mis-coded caller could try to fetch
> tuples that it had already told the tuplestore could be trimmed away;
> and this would work, most of the time, until you got unlucky and the
> trim operation had actually deleted them. I think it's pretty important
> for bug-catching purposes that the tuplestore enforce that those tuples
> are not available anymore.
I see it's too late now that you've committed it, but it seems there
was another way to avoid it by not trimming from percent_rank()
individually. Once the whole partition is fit to the memory, you don't
need to trim it since it never grows. The trimming logic is for
something like moving aggregates and (simple) rank(), which grows
tuplestore content as it advances. percent_rank() doesn't seem to
match the optimization.
Regards,
--
Hitoshi Harada