On Wed, Dec 17, 2014 at 6:05 PM, Kouhei Kaigai <kai...@ak.jp.nec.com> wrote:
>
> Simon,
>
> Its concept is good to me. I think, the new combined function should be
> responsible to take a state data type as argument and update state object
> of the aggregate function. In other words, combined function performs like
> transition function but can update state object according to the summary
> of multiple rows. Right?
>
> It also needs some enhancement around Aggref/AggrefExprState structure to
> inform which function shall be called on execution time.
> Combined functions are usually no-thank-you. AggrefExprState updates its
> internal state using transition function row-by-row. However, once someone
> push-down aggregate function across table joins, combined functions have
> to be called instead of transition functions.
> I'd like to suggest Aggref has a new flag to introduce this aggregate
> expects
> state object instead of scalar value.
>
> Also, I'd like to suggest one other flag in Aggref not to generate final
> result, and returns state object instead.
>
>
>
So are you proposing not calling transfuncs at all and just use combined
functions?
That sounds counterintuitive to me. I am not able to see why you would want
to avoid transfns totally even for the case of pushing down aggregates that
you mentioned.
>From Simon's example mentioned upthread:
PRE-AGGREGATED PLAN
Aggregate
-> Join
-> PreAggregate (doesn't call finalfn)
-> Scan BaseTable1
-> Scan BaseTable2
finalfn wouldnt be called. Instead, combined function would be responsible
for getting preaggregate results and combining them (unless of course, I am
missing something).
Special casing transition state updating in Aggref seems like a bad idea to
me. I would think that it would be better if we made it more explicit i.e.
add a new node on top that does the combination (it would be primarily
responsible for calling combined function).
Not a good source of inspiration, but seeing how SQL Server does it
(Exchange operator + Stream Aggregate) seems intuitive to me, and having
combination operation as a separate top node might be a cleaner way.
I may be wrong though.
Regards,
Atri
