Re: Missing optimization when filters are applied after window functions - Mailing list pgsql-hackers

From Hitoshi Harada
Subject Re: Missing optimization when filters are applied after window functions
Date
Msg-id CAP7QgmmEpBJ8+oVwG5JZJDCa0iYZ+qVvvf=syBk3=mKabfGc2g@mail.gmail.com
Whole thread Raw
In response to Re: Missing optimization when filters are applied after window functions  (Volker Grabsch <vog@notjusthosting.com>)
List pgsql-hackers
On Thu, May 17, 2012 at 7:26 AM, Volker Grabsch <vog@notjusthosting.com> wrote:
> Hitoshi Harada schrieb:
>> On Wed, May 16, 2012 at 12:50 AM, Volker Grabsch <vog@notjusthosting.com> wrote:
>> > I propose the following general optimization: If all window
>> > functions are partitioned by the same first field (here: id),
>> > then any filter on that field should be executed before
>> > WindowAgg. So a query like this:
>>
>> I think that's possible.  Currently the planner doesn't think any
>> qualification from the upper query can be pushed down to a query that
>> has a window function.  It would be possible to let it push down if
>> the expression matches PARTITION BY expression.
>
> Sounds great!
>
>> However, the
>> challenge is that a query may have a number of window functions that
>> have different PARTITION BY expressions.  At the time of pushing down
>> in the planning, it is not obvious which window function comes first.
>
> I'm don't really unterstand what you mean with "which window function
> comes first", because to my understanding, all window functions of
> a query belong to the same level in the query hierarchy. But then,
> my knowledge of PostgreSQL internals isn't very deep, either.

No, you can specify as many window specification as you like.  Say,

SELECT count(*) over (w1), count(*) over (w2)
FROM tbl
WINDOW w1 AS (PARTITION BY x ORDER BY y), w2 AS (PARTITION BY y ORDER BYx);

and in the same query level there are different type of window
specifications.  The code as stands today doesn't have any semantics
about which of w1 or w2 to run first (probably w1 will be run first,
but the query semantics doesn't enforce anything).

>> One idea is to restrict such optimization in only case of single
>> window function, and the other is to make it generalize and cover a
>> lot of cases.
>
> From a practical point of view, the restriction to a single window
> function wouldn't be that bad, although I'd prefer to think about
> the number of different windows rather than number of window functions.
>
> In other words, every optimization that is correct for a single window
> function is also correct for multiple window functions if those use
> all the same window.

Yeah, I mean, multiple windows, not window functions.

>> That said, our planner on window functions has a lot of improvement to
>> be done.  Every kind of optimization I see is what I raised above;
>> they can be done easily by hacking in a small case, or they can be
>> done by generalizing for the most of cases.  My understanding is our
>> project tends to like the latter and it takes a little time but covers
>> more use cases.
>
> I'd also prefer to see a general solution, as this provides less
> room for unpleasant surprises (e.g. "This query is only slightly
> different from the previous one. Why does it take so much longer?").
>
> On the other hand, any small improvement is a big step forward
> regarding window functions.
>
> Unfortunately, I can't voluteer on that, as it is currently
> impossible for me to allocate enough time for this.
>
> However, any pointer to where to look at the source (or in the
> manual) would be of great. Maybe I'll find at least enough time
> to provide a rough proposal, or to improve existing attempts
> to solve this issue.

Look at subquery_is_pushdown_safe in allpath.c.  Here we stop looking
deeper if the upper qualification can be pushed down to the inner
sub-query if the sub-query has any window function expressions.


Thanks,
--
Hitoshi Harada


pgsql-hackers by date:

Previous
From: Joel Jacobson
Date:
Subject: Re: Schema version management
Next
From: Florian Pflug
Date:
Subject: Re: Per-Database Roles