2010/10/1 Tom Lane <tgl@sss.pgh.pa.us>:
> Hitoshi Harada <umi.tanuki@gmail.com> writes:
>> 2010/9/26 Pavel Stehule <pavel.stehule@gmail.com>:
>>> This patch needs a few work - can share a compare functionality with
>>> tuplesort.c, but I would to verify a concept now.
>
>> Sorry for delay. I read the patch and it seems the result is sane. For
>> window function calls, I agree that the current tuplesort is not
>> enough to implement median functions and the patch introduces its own
>> memsort mechanism, although memsort has too much copied from
>> tuplesort. It looks to me not so difficult to modify the existing
>> tuplesort to guarantee staying in memory always if an option to do so
>> is specified from caller. I think that option can be used by other
>> cases in the core code.
>
> If this patch tries to force the entire sort to happen in memory,
> it is not committable. What will happen when you get a lot of data?
> You need to be working on a variant that will work anyway, not working
> on an unacceptable lobotomization of the main sort code.
The median function checking a calling context - under window
aggregate uses a just memory sort solution - and under standard
aggregate context it uses a full tuplesort. It bases on request of
window aggregate function - the final function have to be called more
time - and it isn't possible with tuplesort. So as window aggregate it
uses just memory sort limited with work_mem. Other usage is unlimited.
Second option was a block median function under window aggregates.
It this design possible? We cannot use any complex source under window
aggregates, because there isn't any way to unlink it.
Regards
Pavel
>
> regards, tom lane
>