Thread: Re: [HACKERS] FDW and parallel execution

Hi,

At Sun, 02 Apr 2017 16:30:24 +0300, Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote in <58E0FCF0.2070603@postgrespro.ru>

Hi hackers and personally Robet (you are the best expert in both
areas).
I want to ask one more question concerning parallel execution and FDW.
Below are two plans for the same query (TPC-H Q5): one for normal
tables, another for FDW to vertical representation of the same data.
FDW supports analyze function and is expected to produce the similar
statistic as for original tables.

<big explain>

The plans look very similar, but first one is parallel and second -
not.
My FDW provides implementation for IsForeignScanParallelSafe which
returns true.
I wonder what can prevent optimizer from using parallel plan in this
case?

Parallel execution requires partial paths. It's the work for
GetForeignPaths of your FDW.

Thank you very much for explanation.
But unfortunately I still do not completely understand what kind of queries allow parallel execution with FDW.

Section "FDW Routines for Parallel Execution" of FDW specification says:

A ForeignScan node can, optionally, support parallel execution. A parallel ForeignScan will be executed in multiple processes and should return each row only once across all cooperating processes. To do this, processes can coordinate through fixed size chunks of dynamic shared memory. This shared memory is not guaranteed to be mapped at the same address in every process, so pointers may not be used. The following callbacks are all optional in general, but required if parallel execution is to be supported.

I provided IsForeignScanParallelSafe, EstimateDSMForeignScan, InitializeDSMForeignSca and InitializeWorkerForeignScan in my FDW.
IsForeignScanParallelSafe returns true.
Also in GetForeignPaths function I created path with baserel->consider_parallel == true.
Is it enough or I should do something else?

But unfortunately I failed to find any query: sequential scan, grand aggregation, aggregation with group by, joins... when parallel execution plan is used for this FDW.
Also there are no examples of using this functions in Postgres distributive and I failed to find any such examples in Internet.

Can somebody please clarify my situation with parallel execution and FDW and may be point at some examples?
Thank in advance.

-- 
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company 

Sorry for the too-brief reply.

At Tue, 11 Apr 2017 20:08:46 +0300, Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote in
<94c8692a-f299-b72b-6227-270b8a9ed7ad@postgrespro.ru>
>
> On 04.04.2017 13:29, Kyotaro HORIGUCHI wrote:
> > Hi,
> >
> > At Sun, 02 Apr 2017 16:30:24 +0300, Konstantin Knizhnik
> > <k.knizhnik@postgrespro.ru> wrote in <58E0FCF0.2070603@postgrespro.ru>
> >> My FDW provides implementation for IsForeignScanParallelSafe which
> >> returns true.
> >> I wonder what can prevent optimizer from using parallel plan in this
> >> case?
> > Parallel execution requires partial paths. It's the work for
> > GetForeignPaths of your FDW.
>
> Thank you very much for explanation.
> But unfortunately I still do not completely understand what kind of
> queries allow parallel execution with FDW.

At Tue, 11 Apr 2017 19:20:04 +0200, PostgreSQL - Hans-Jürgen Schönig <postgres@cybertec.at> wrote in
<0c9c101d-0fbb-1e19-f04c-7a6ec577d960@cybertec.at>
> did you check out antonin houska's patches?
> we basically got code, which can do that.

Parallel aggregation is already available. Antonin's patch is
partition-wise aggregation, which boosts the case where partition
key is aggregation key, I suppose. parallel aggregation seems to
be considered when any appropriate partial path is available. (I
haven't tried anything, though.)

set_plain_rel_pathlist() does the work for plain relations so
what we should do in GetForeignPaths would be follows.

- check rel->consider_parallel (won't be requried since the fDW knows that) and rel->lateral_relids.

- If parallel is OK, create a path with create_foreignscan_path in ordinary way then change some parallel related
membersas necessary. 

- Like create_plain_partial_paths(), check certain conditions and finally add_partial_path() the created partial
foreignscan path. 

I haven't really done this, so I might be wrong.

> Section "FDW Routines for Parallel Execution" of FDW specification
> says:
> > A ForeignScan node can, optionally, support parallel execution. A
> > parallel ForeignScan will be executed in multiple processes and should
> > return each row only once across all cooperating processes. To do
> > this, processes can coordinate through fixed size chunks of dynamic
> > shared memory. This shared memory is not guaranteed to be mapped at
> > the same address in every process, so pointers may not be used. The
> > following callbacks are all optional in general, but required if
> > parallel execution is to be supported.
>
> I provided IsForeignScanParallelSafe, EstimateDSMForeignScan,
> InitializeDSMForeignSca and InitializeWorkerForeignScan in my FDW.
> IsForeignScanParallelSafe returns true.
> Also in GetForeignPaths function I created path with
> baserel->consider_parallel == true.
> Is it enough or I should do something else?

Creating partial paths, I think. create_grouping_paths() requires
partial_pathlist in input_rel.

The section is explaning FDW routines specially provided for
parallel execution. But it doesn't seem mentioning "how to run a
parallel execution" as a whole.

> But unfortunately I failed to find any query: sequential scan, grand
> aggregation, aggregation with group by, joins... when parallel
> execution plan is used for this FDW.
> Also there are no examples of using this functions in Postgres
> distributive and I failed to find any such examples in Internet.

Maybe you're the pioneer in this area.

> Can somebody please clarify my situation with parallel execution and
> FDW and may be point at some examples?
> Thank in advance.

regards,

--
Kyotaro Horiguchi
NTT Open Source Software Center