Hi,
On Tue, Jan 14, 2014 at 5:49 PM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
>On Tue, Jan 14, 2014 at 12:54 AM, Marti Raudsepp <marti@juffo.org> wrote:
>> I've been trying it out in a few situations. I implemented a new
>> enable_partialsort GUC to make it easier to turn on/off, this way it's a lot
>> easier to test. The attached patch applies on top of partial-sort-5.patch
>
> I though about such option. Generally not because of testing convenience,
> but because of overhead of planning. This way you implement it is quite
> naive :)
I don't understand. I had another look at this and cost_sort still
seems like the best place to implement this, since that's where the
patch decides how many pre-sorted columns to use. Both mergejoin and
simple order-by plans call into it. If enable_partialsort=false then I
skip all pre-sorted options except full sort, making cost_sort behave
pretty much like it did before the patch.
I could change pathkeys_common to return 0, but that seems like a
generic function that shouldn't be tied to partialsort. The old code
paths called pathkeys_contained_in anyway, which has similar
complexity. (Apart for initial_cost_mergejoin, but that doesn't seem
special enough to make an exception for).
Or should I use?: enable_partialsort ? pathkeys_common(...) : 0
> For instance, merge join rely on partial sort which will be
> replaced with simple sort.
Are you saying that enable_partialsort=off should keep
partialsort-based mergejoins enabled?
Or are you saying that merge joins shouldn't use "simple sort" at all?
But merge join was already able to use full Sort nodes before your
patch.
What am I missing?
Regards,
Marti