Thread: PoC: Partial sort

Hi,

Cool stuff.

On 2013-12-14 13:59:02 +0400, Alexander Korotkov wrote:
> Currently when we need to get ordered result from table we have to choose
> one of two approaches: get results from index in exact order we need or do
> sort of tuples. However, it could be useful to mix both methods: get
> results from index in order which partially meets our requirements and do
> rest of work from heap.

>
------------------------------------------------------------------------------------------------------------------------------------------
>  Limit  (cost=69214.06..69214.08 rows=10 width=16) (actual
> time=0.097..0.099 rows=10 loops=1)
>    ->  Sort  (cost=69214.06..71714.06 rows=1000000 width=16) (actual
> time=0.096..0.097 rows=10 loops=1)
>          Sort Key: v1, v2
>          Sort Method: top-N heapsort  Memory: 25kB
>          ->  Index Scan using test_v1_idx on test  (cost=0.42..47604.42
> rows=1000000 width=16) (actual time=0.017..0.066 rows=56 loops=1)
>  Total runtime: 0.125 ms
> (6 rows)

Is that actually all that beneficial when sorting with a bog standard
qsort() since that doesn't generally benefit from data being pre-sorted?
I think we might need to switch to a different algorithm to really
benefit from mostly pre-sorted input.

> *partial-knn-1.patch*
> 
> KNN-GiST provides ability to get ordered results from index, but this order
> is based only on index information. For instance, GiST index contains
> bounding rectangles for polygons, and we can't get exact distance to
> polygon from index (similar situation is in PostGIS). In attached patch,
> GiST distance method can set recheck flag (similar to consistent method).
> This flag means that distance method returned lower bound of distance and
> we should recheck it from heap.

> See an example.
> 
> create table test as (select id, polygon(3+(random()*10)::int,
> circle(point(random(), random()), 0.0003 + random()*0.001)) as p from
> generate_series(1,1000000) id);
> create index test_idx on test using gist (p);
> 
> We can get results ordered by distance from polygon to point.
> 
> postgres=# select id, p <-> point(0.5,0.5) from test order by p <->
> point(0.5,0.5) limit 10;

>
----------------------------------------------------------------------------------------------------------------------------------
>  Limit  (cost=0.29..1.86 rows=10 width=36) (actual time=0.180..0.230
> rows=10 loops=1)
>    ->  Index Scan using test_idx on test  (cost=0.29..157672.29
> rows=1000000 width=36) (actual time=0.179..0.228 rows=10 loops=1)
>          Order By: (p <-> '(0.5,0.5)'::point)
>  Total runtime: 0.305 ms
> (4 rows)

Rechecking from the heap means adding a sort node though, which I don't
see here? Or am I misunderstanding something?
Greetings,

Andres Freund

-- Andres Freund                       http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services

On 14/12/13 12:54, Andres Freund wrote:
> On 2013-12-14 13:59:02 +0400, Alexander Korotkov wrote:
>> Currently when we need to get ordered result from table we have to choose
>> one of two approaches: get results from index in exact order we need or do
>> sort of tuples. However, it could be useful to mix both methods: get
>> results from index in order which partially meets our requirements and do
>> rest of work from heap.
>
>>
------------------------------------------------------------------------------------------------------------------------------------------
>>   Limit  (cost=69214.06..69214.08 rows=10 width=16) (actual
>> time=0.097..0.099 rows=10 loops=1)
>>     ->  Sort  (cost=69214.06..71714.06 rows=1000000 width=16) (actual
>> time=0.096..0.097 rows=10 loops=1)
>>           Sort Key: v1, v2
>>           Sort Method: top-N heapsort  Memory: 25kB
>>           ->  Index Scan using test_v1_idx on test  (cost=0.42..47604.42
>> rows=1000000 width=16) (actual time=0.017..0.066 rows=56 loops=1)
>>   Total runtime: 0.125 ms
>> (6 rows)
>
> Is that actually all that beneficial when sorting with a bog standard
> qsort() since that doesn't generally benefit from data being pre-sorted?
> I think we might need to switch to a different algorithm to really
> benefit from mostly pre-sorted input.

Eg:  http://www.postgresql.org/message-id/5291467E.6070807@wizmail.org

Maybe Alexander and I should bash our heads together.

-- 
Cheers,   Jeremy

On Sat, Dec 14, 2013 at 06:21:18PM +0400, Alexander Korotkov wrote:
> > Is that actually all that beneficial when sorting with a bog standard
> > qsort() since that doesn't generally benefit from data being pre-sorted?
> > I think we might need to switch to a different algorithm to really
> > benefit from mostly pre-sorted input.
> >
>
> In this patch I don't do full sort of dataset. For instance, index returns
> data ordered by first column and we need to order them also by second
> column. Then this node sorts groups (assumed to be small) where values of
> the first column are same by value of second column. And with limit clause
> only required number of such groups will be processed. But, I don't think
> we should expect pre-sorted values of second column inside a group.

Nice. I imagine this would be mostly beneficial for fast-start plans,
since you no longer need to sort the whole table prior to returning the
first tuple.

Reduced memory usage might be a factor, especially for large sorts
where you otherwise might need to spool to disk.

You can now use an index on (a) to improve sorting for (a,b).

Cost of sorting n groups of size l goes from O(nl log nl) to just O(nl
log l), useful for large n.

Minor comments:

I find cmpTuple a bad name. That's what it's doing but perhaps
cmpSkipColumns would be clearer.

I think it's worthwhile adding a seperate path for the skipCols = 0
case, to avoid extra copies.

Have a nice day,
--
Martijn van Oosterhout   <kleptog@svana.org>   http://svana.org/kleptog/
> He who writes carelessly confesses thereby at the very outset that he does
> not attach much importance to his own thoughts.  -- Arthur Schopenhauer

Hi,

> > >  Limit  (cost=69214.06..69214.08 rows=10 width=16) (actual
> > > time=0.097..0.099 rows=10 loops=1)
> > >    ->  Sort  (cost=69214.06..71714.06 rows=1000000 width=16) (actual
> > > time=0.096..0.097 rows=10 loops=1)
> > >          Sort Key: v1, v2
> > >          Sort Method: top-N heapsort  Memory: 25kB
> > >          ->  Index Scan using test_v1_idx on test  (cost=0.42..47604.42
> > > rows=1000000 width=16) (actual time=0.017..0.066 rows=56 loops=1)
> > >  Total runtime: 0.125 ms
> > > (6 rows)
> >
> > Is that actually all that beneficial when sorting with a bog standard
> > qsort() since that doesn't generally benefit from data being pre-sorted?
> > I think we might need to switch to a different algorithm to really
> > benefit from mostly pre-sorted input.
> >
> 
> In this patch I don't do full sort of dataset. For instance, index returns
> data ordered by first column and we need to order them also by second
> column.

Ah, that makes sense.

> But, I don't think we should expect pre-sorted values of second column
> inside a group.

Yes, if you do it that way, there doesn't seem to any need to assume
that any more than we usually do.

I think you should make the explain output reflect the fact that we're
assuming v1 is presorted and just sorting v2. I'd be happy enough with:
Sort Key: v1, v2
Partial Sort: v2
or even just
"Partial Sort Key: [v1,] v2"
but I am sure others disagree.

> > > *partial-knn-1.patch*

> > Rechecking from the heap means adding a sort node though, which I don't
> > see here? Or am I misunderstanding something?

> KNN-GiST contain RB-tree of scanned items. In this patch item is rechecked
> inside GiST and reinserted into same RB-tree. It appears to be much easier
> implementation for PoC and also looks very efficient. I'm not sure what is
> actually right design for it. This is what I like to discuss.

I don't have enough clue about gist to say wether it's the right design,
but it doesn't look wrong to my eyes. It'd probably be useful to export
the knowledge that we are rechecking and how often that happens to the
outside.
While I didn't really look into the patch, I noticed in passing that you
pass a all_dead variable to heap_hot_search_buffer without using the
result - just pass NULL instead, that performs a bit less work.

Greetings,

Andres Freund

-- Andres Freund                       http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services

On 12/14/2013 10:59 AM, Alexander Korotkov wrote:
> This patch allows to use index for order-by if order-by clause and index
> has non-empty common prefix. So, index gives right ordering for first n
> order-by columns. In order to provide right order for rest m columns,
> sort node is inserted. This sort node sorts groups of tuples where
> values of first n order-by columns are equal.

I recently looked at the same problem. I see that you solved the 
rescanning problem by simply forcing the sort to be redone on 
ExecReScanSort if you have done a partial sort.

My idea for a solution was to modify tuplesort to allow storing the 
already sorted keys in either memtuples or the sort result file, but 
setting a field so it does not sort thee already sorted tuples again. 
This would allow the rescan to work as it used to, but I am unsure how 
clean or ugly this code would be. Was this something you considered?

-- 
Andreas Karlsson

On 12/18/2013 01:02 PM, Alexander Korotkov wrote:
>     My idea for a solution was to modify tuplesort to allow storing the
>     already sorted keys in either memtuples or the sort result file, but
>     setting a field so it does not sort thee already sorted tuples
>     again. This would allow the rescan to work as it used to, but I am
>     unsure how clean or ugly this code would be. Was this something you
>     considered?
>
>
> I'm not sure. I believe that best answer depends on particular
> parameter: how much memory we've for sort, how expensive is underlying
> node and how it performs rescan, how big are groups in partial sort.

Yes, if one does not need a rescan your solution will use less memory 
and about the same amount of CPU (if the tuplesort does not spill to 
disk). While if we keep all the already sorted tuples in the tuplesort 
rescans will be cheap but more memory will be used with an increased 
chance of spilling to disk.

-- 
Andreas Karlsson

On Sun, Dec 22, 2013 at 07:38:05PM +0400, Alexander Korotkov wrote:
> Hi!
>
> Next revision. It expected to do better work with optimizer. It introduces
> presorted_keys argument of cost_sort function which represent number of
> keys already sorted in Path. Then this function uses estimate_num_groups to
> estimate number of groups with different values of presorted keys and
> assumes that dataset is uniformly divided by
> groups. get_cheapest_fractional_path_for_pathkeys tries to select the path
> matching most part of path keys.
> You can see it's working pretty good on single table queries.

Nice work! The plans look good and the calculated costs seem sane also.

I suppose the problem with the joins is generating the pathkeys?

Have a nice day,
--
Martijn van Oosterhout   <kleptog@svana.org>   http://svana.org/kleptog/
> He who writes carelessly confesses thereby at the very outset that he does
> not attach much importance to his own thoughts.  -- Arthur Schopenhauer

On 12/22/2013 04:38 PM, Alexander Korotkov wrote:
> postgres=# explain analyze select * from test order by v1, id limit 10;
>                                                                    QUERY
> PLAN
>
-----------------------------------------------------------------------------------------------------------------------------------------------
>   Limit  (cost=11441.77..11442.18 rows=10 width=12) (actual
> time=79.980..79.982 rows=10 loops=1)
>     ->  Partial sort  (cost=11441.77..53140.44 rows=1000000 width=12)
> (actual time=79.978..79.978 rows=10 loops=1)
>           Sort Key: v1, id
>           Presorted Key: v1
>           Sort Method: top-N heapsort  Memory: 25kB
>           ->  Index Scan using test_v1_idx on test  (cost=0.42..47038.83
> rows=1000000 width=12) (actual time=0.031..38.275 rows=100213 loops=1)
>   Total runtime: 81.786 ms
> (7 rows)

Have you thought about how do you plan to print which sort method and 
how much memory was used? Several different sort methods may have been 
use in the query. Should the largest amount of memory/disk be printed?

> However, work with joins needs more improvements.

That would be really nice to have, but the patch seems useful even 
without the improvements to joins.

-- 
Andreas Karlsson

On 12/29/2013 08:24 AM, David Rowley wrote:
> If it was possible to devise some way to reuse any
> previous tuplesortstate perhaps just inventing a reset method which
> clears out tuples, then we could see performance exceed the standard
> seqscan -> sort. The code the way it is seems to lookup the sort
> functions from the syscache for each group then allocate some sort
> space, so quite a bit of time is also spent in palloc0() and pfree()
>
> If it was not possible to do this then maybe adding a cost to the number
> of sort groups would be better so that the optimization is skipped if
> there are too many sort groups.

It should be possible. I have hacked a quick proof of concept for
reusing the tuplesort state. Can you try it and see if the performance
regression is fixed by this?

One thing which have to be fixed with my patch is that we probably want
to close the tuplesort once we have returned the last tuple from ExecSort().

I have attached my patch and the incremental patch on Alexander's patch.

--
Andreas Karlsson

David Rowley escribió:

> I was about to test it tonight, but I'm having trouble getting the patch to
> compile... I'm really wondering which compiler you are using as it seems
> you're declaring your variables in some strange places.. See nodeSort.c
> line 101. These variables are declared after there has been an if statement
> in the same scope. That's not valid in C. (The patch did however apply
> without any complaints).

AFAIR C99 allows mixed declarations and code.  Visual Studio only
implements C89 though, which is why it fails to compile there.

-- 
Álvaro Herrera                http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services

On 12/28/2013 04:51 PM, Alexander Korotkov wrote:
> I've compiled it with clang. Yeah, there was mixed declarations. I've
> rechecked it with gcc, now it gives no warnings. I didn't try it with
> visual studio, but I hope it will be OK.

I looked at this version of the patch and noticed a possibility for 
improvement. You could decrement the bound for the tuplesort after every 
completed sort. Otherwise the optimizations for small limits wont apply 
to partial sort.

-- 
Andreas Karlsson

On 22/12/13 20:26, Alexander Korotkov wrote:
> On Sat, Dec 14, 2013 at 6:30 PM, Jeremy Harris <jgh@wizmail.org> wrote:
>
>> On 14/12/13 12:54, Andres Freund wrote:
>>> Is that actually all that beneficial when sorting with a bog standard
>>> qsort() since that doesn't generally benefit from data being pre-sorted?
>>> I think we might need to switch to a different algorithm to really
>>> benefit from mostly pre-sorted input.
>>>
>>
>> Eg:  http://www.postgresql.org/message-id/5291467E.6070807@wizmail.org
>>
>> Maybe Alexander and I should bash our heads together.
>
>
> Partial sort patch is mostly optimizer/executor improvement rather than
> improvement of sort algorithm itself.

I finally got as far as understanding Alexander's cleverness, and it
does make the performance advantage (on partially-sorted input) of the
merge-sort irrelevant.

There's a slight tradeoff possible between the code complexity of
the chunking code front-ending the sorter and just using the
enhanced sorter.  The chunking does reduce the peak memory usage
quite nicely too.

The implementation of the chunker does O(n) compares using the
keys of the feed-stream index, to identify the chunk boundaries.
Would it be possible to get this information from the Index Scan?
-- 
Cheers,   Jeremy

On 13/01/14 18:01, Alexander Korotkov wrote:
> Thanks. It's included into attached version of patch. As wall as estimation
> improvements, more comments and regression tests fix.

Would it be possible to totally separate the two sets of sort-keys,
only giving the non-index set to the tuplesort?  At present tuplesort
will, when it has a group to sort, make wasted compares on the
indexed set of keys before starting on the non-indexed set.
-- 
Cheers,  Jeremy

On 31/12/13 01:41, Andreas Karlsson wrote:
> On 12/29/2013 08:24 AM, David Rowley wrote:
>> If it was possible to devise some way to reuse any
>> previous tuplesortstate perhaps just inventing a reset method which
>> clears out tuples, then we could see performance exceed the standard
>> seqscan -> sort. The code the way it is seems to lookup the sort
>> functions from the syscache for each group then allocate some sort
>> space, so quite a bit of time is also spent in palloc0() and pfree()
>>
>> If it was not possible to do this then maybe adding a cost to the number
>> of sort groups would be better so that the optimization is skipped if
>> there are too many sort groups.
>
> It should be possible. I have hacked a quick proof of concept for
> reusing the tuplesort state. Can you try it and see if the performance
> regression is fixed by this?
>
> One thing which have to be fixed with my patch is that we probably want
> to close the tuplesort once we have returned the last tuple from
> ExecSort().
>
> I have attached my patch and the incremental patch on Alexander's patch.

How does this work in combination with randomAccess ?
-- 
Thanks,   Jeremy

On Sat, Jan 18, 2014 at 7:22 PM, Marti Raudsepp <marti@juffo.org> wrote:
> Total runtime: 5.418 ms

Oops, shouldn't have rushed this. Clearly the timings should have
tipped me off that it's broken. I didn't notice that cmpSortSkipCols
was re-using tuplesort's sortkeys.

Here's a patch that actually works; I added a new skipKeys attribute
to SortState. I had to extract the SortSupport-creation code from
tuplesort_begin_heap to a new function; but that's fine, because it
was already duplicated in ExecInitMergeAppend too.

I reverted the addition of tuplesort_get_sortkeys, which is not needed now.

Now the timings are:
Unpatched partial sort: 13478.158 ms
Full sort: 6802.063 ms
Patched partial sort: 6618.962 ms

Regards,
Marti

On 01/18/2014 08:13 PM, Jeremy Harris wrote:
> On 31/12/13 01:41, Andreas Karlsson wrote:
>> On 12/29/2013 08:24 AM, David Rowley wrote:
>>> If it was possible to devise some way to reuse any
>>> previous tuplesortstate perhaps just inventing a reset method which
>>> clears out tuples, then we could see performance exceed the standard
>>> seqscan -> sort. The code the way it is seems to lookup the sort
>>> functions from the syscache for each group then allocate some sort
>>> space, so quite a bit of time is also spent in palloc0() and pfree()
>>>
>>> If it was not possible to do this then maybe adding a cost to the number
>>> of sort groups would be better so that the optimization is skipped if
>>> there are too many sort groups.
>>
>> It should be possible. I have hacked a quick proof of concept for
>> reusing the tuplesort state. Can you try it and see if the performance
>> regression is fixed by this?
>>
>> One thing which have to be fixed with my patch is that we probably want
>> to close the tuplesort once we have returned the last tuple from
>> ExecSort().
>>
>> I have attached my patch and the incremental patch on Alexander's patch.
>
> How does this work in combination with randomAccess ?

As far as I can tell randomAccess was broken by the partial sort patch 
even before my change since it would not iterate over multiple 
tuplesorts anyway.

Alexander: Is this true or am I missing something?

-- 
Andreas Karlsson

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

On Mon, Jan 20, 2014 at 2:43 PM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> Another changes in this version of patch:
> 1) Applied patch to don't compare skipCols in tuplesort by Marti Raudsepp
> 2) Adjusting sort bound after processing buckets.

Hi,

Here's a patch with some whitespace and coding style fixes for
partial-sort-6.patch

I tried to understand the mergejoin regression, but this code still
looks like Chinese to me. Can anyone else have a look at it?

Test case: http://www.postgresql.org/message-id/CABRT9RDd-P2RLRdHsMq8rCOB46k4a5O+bGz_up2bRGeeH4R6oQ@mail.gmail.com
Original report:
http://www.postgresql.org/message-id/CABRT9RCLLUyJ=bkeB132aVA_mVNx5==LvVvQMvUqDguFZtW+cg@mail.gmail.com

Regards,
Marti

On Mon, Jan 27, 2014 at 9:26 PM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> For now, I have attempt to fix extra columns in mergejoin problem. It would
> be nice if you test it.

Yes, it solves the test cases I was trying with, thanks.

> 1) With enable_partialsort = off all mergejoin logic should behave as
> without partial sort patch.
> 2) With partial sort patch get_cheapest_fractional_path_for_pathkeys
> function is much more expensive to execute. With enable_partialsort = off it
> should be as cheap as without partial sort patch.

When it comes to planning time, I really don't think you should
bother. The planner enable_* settings are meant for troubleshooting,
debugging and learning about the planner. You should not expect people
to disable them in a production setting. It's not worth complicating
the code for that rare case.

This is stated in the documentation
(http://www.postgresql.org/docs/current/static/runtime-config-query.html)
and repeatedly on the mailing lists.

But some benchmarks of planning performance are certainly warranted.

Regards,
Marti

On Tue, Jan 28, 2014 at 7:51 AM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> I didn't test it, but I worry that overhead might be high.
> If it's true then it could be like constraint_exclusion option which id off
> by default because of planning overhead.

I see, that makes sense.

I will try to find the time to run some benchmarks in the coming few days.

Regards,
Marti

On Wed, Feb 5, 2014 at 6:58 PM, Marti Raudsepp <marti@juffo.org> wrote:
> I ran some synthetic benchmarks with single-column inner joins between 5
> tables, with indexes on both joined columns, using only EXPLAIN (so
> measuring planning time, not execution) in 9 scenarios to excercise
> different code paths. According to these measurements, the overhead ranges
> between 1.0 and 4.5% depending on the scenario.

Hmm, sounds a little steep.  Why is it so expensive?  I'm probably
missing something here, because I would have thought that planner
support for partial sorts would consist mostly of considering the same
sorts we consider today, but with the costs reduced by the batching.
Changing the cost estimation that way can't be that much more
expensive than what we're already doing, so the overhead should be
minimal.  What the patch is actually doing seems to be something quite
a bit more invasive than that, but I'm not sure what it is exactly, or
why.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Thu, Feb 6, 2014 at 5:31 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> Hmm, sounds a little steep.  Why is it so expensive?  I'm probably
> missing something here, because I would have thought that planner
> support for partial sorts would consist mostly of considering the same
> sorts we consider today, but with the costs reduced by the batching.

I guess it's because the patch undoes some optimizations in the
mergejoin planner wrt caching merge clauses and adds a whole lot of
code to find_mergeclauses_for_pathkeys. In other code paths the
overhead does seem to be negligible.

Notice the removal of:
/* Select the right mergeclauses, if we didn't already */
/** Avoid rebuilding clause list if we already made one;* saves memory in big join trees...*/

Regards,
Marti

On Thu, Feb 6, 2014 at 3:39 AM, Marti Raudsepp <marti@juffo.org> wrote:
> On Thu, Feb 6, 2014 at 5:31 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>> Hmm, sounds a little steep.  Why is it so expensive?  I'm probably
>> missing something here, because I would have thought that planner
>> support for partial sorts would consist mostly of considering the same
>> sorts we consider today, but with the costs reduced by the batching.
>
> I guess it's because the patch undoes some optimizations in the
> mergejoin planner wrt caching merge clauses and adds a whole lot of
> code to find_mergeclauses_for_pathkeys. In other code paths the
> overhead does seem to be negligible.
>
> Notice the removal of:
> /* Select the right mergeclauses, if we didn't already */
> /*
>  * Avoid rebuilding clause list if we already made one;
>  * saves memory in big join trees...
>  */

Yeah, I noticed that.  My feeling is that those optimizations got put
in there because someone found them to be important, so I'm skeptical
about removing them.  It may be that having the capability to do a
partial sort makes it seem worth spending more CPU looking for merge
joins, but I'd vote for making any such change a separate patch.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Thu, Feb 6, 2014 at 9:15 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> It may be that having the capability to do a
> partial sort makes it seem worth spending more CPU looking for merge
> joins, but I'd vote for making any such change a separate patch.

Agreed.

Alexander, should I work on splitting up the patch in two, or do you
want to do it yourself?

Should I merge my coding style and enable_partialsort patches while at
it, or do you still have reservations about those?

Regards,
Marti

Robert Haas <robertmhaas@gmail.com> writes:
> On Thu, Feb 6, 2014 at 3:39 AM, Marti Raudsepp <marti@juffo.org> wrote:
>> I guess it's because the patch undoes some optimizations in the
>> mergejoin planner wrt caching merge clauses and adds a whole lot of
>> code to find_mergeclauses_for_pathkeys. In other code paths the
>> overhead does seem to be negligible.

> Yeah, I noticed that.  My feeling is that those optimizations got put
> in there because someone found them to be important, so I'm skeptical
> about removing them.

I put them in, and yeah they are important.  Even with those, and even
with the rather arbitrary heuristic restrictions that joinpath.c puts on
what mergeclause lists to consider, the existing planner spends a whole
lot of effort on mergejoins --- possibly disproportionate to their actual
value.  I think that any patch that removes those optimizations is not
going to fly.  If anything, it'd be better to reduce the number of
mergejoins considered even further, because a lot of the possible plans
are not usefully different.

It's already the case that we expect indxpath.c to predict the useful
orderings (by reference to query_pathkeys and available mergejoin clauses)
and generate suitable paths, rather than trying to identify the orderings
at join time.  Can't that approach be extended to cover this technique?

In any case, the bottom line is that we don't want this patch to cause
the planner to consider large numbers of new but useless sort orderings.
        regards, tom lane

On Sun, Feb 9, 2014 at 7:37 PM, Alexander Korotkov <aekorotkov@gmail.com> wrote:
> This is not only place that worry me about planning overhead. See
> get_cheapest_fractional_path_for_pathkeys. I had to estimate number of
> groups for each sorting column in order to get right fractional path.

AFAICT this only happens once per plan and the overhead is O(n) to the
number of pathkeys? I can't get worried about that, but I guess it's
better to test anyway.

PS: You didn't answer my questions about splitting the patch. I guess
I'll have to do that anyway to run the tests.

Regards,
Marti

On Mon, Feb 10, 2014 at 8:59 PM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> Done. Patch is splitted.

Thanks!

I think the 1st patch now has a bug in initial_cost_mergejoin; you
still pass the "presorted_keys" argument to cost_sort, making it
calculate a partial sort cost, but generated plans never use partial
sort. I think 0 should be passed instead. Patch attached, needs to be
applied on top of partial-sort-basic-1 and then reverse-applied on
partial-sort-merge-1.

With partial-sort-basic-1 and this fix on the same test suite, the
planner overhead is now a more manageable 0.5% to 1.3%; one test is
faster by 0.5%. Built with asserts disabled, ran on Intel i5-3570K. In
an effort to reduce variance, I locked the server and pgbench to a
single CPU core (taskset -c 3), but there are still noticeable
run-to-run differences, so these numbers are a bit fuzzy. The faster
result is definitely not a fluke, however; it happens every time.

> On Mon, Feb 10, 2014 at 2:33 PM, Marti Raudsepp <marti@juffo.org> wrote:
>> AFAICT this only happens once per plan and the overhead is O(n) to the
>> number of pathkeys?

I was of course wrong about that, it also adds extra overhead when
iterating over the paths list.

----
Test "taskset -c 3 run2.sh" from
https://github.com/intgr/benchjunk/tree/master/partial_sort

Overhead percentages (between best of each 3 runs):
join5.sql 0.7
star5.sql 0.8
line5.sql 0.5
lim_join5.sql -0.5
lim_star5.sql 1.3
lim_line5.sql 0.5
limord_join5.sql 0.6
limord_star5.sql 0.5
limord_line5.sql 0.7

Raw results:
git 48870dd
join5.sql tps = 509.328070 (excluding connections establishing)
join5.sql tps = 509.772190 (excluding connections establishing)
join5.sql tps = 510.651517 (excluding connections establishing)
star5.sql tps = 499.208698 (excluding connections establishing)
star5.sql tps = 498.200314 (excluding connections establishing)
star5.sql tps = 496.269315 (excluding connections establishing)
line5.sql tps = 797.968831 (excluding connections establishing)
line5.sql tps = 797.011690 (excluding connections establishing)
line5.sql tps = 796.379258 (excluding connections establishing)
lim_join5.sql tps = 394.946024 (excluding connections establishing)
lim_join5.sql tps = 395.417689 (excluding connections establishing)
lim_join5.sql tps = 395.482958 (excluding connections establishing)
lim_star5.sql tps = 423.434393 (excluding connections establishing)
lim_star5.sql tps = 423.774305 (excluding connections establishing)
lim_star5.sql tps = 424.386099 (excluding connections establishing)
lim_line5.sql tps = 733.007330 (excluding connections establishing)
lim_line5.sql tps = 731.794731 (excluding connections establishing)
lim_line5.sql tps = 732.356280 (excluding connections establishing)
limord_join5.sql tps = 385.317921 (excluding connections establishing)
limord_join5.sql tps = 385.915870 (excluding connections establishing)
limord_join5.sql tps = 384.747848 (excluding connections establishing)
limord_star5.sql tps = 417.992615 (excluding connections establishing)
limord_star5.sql tps = 416.944685 (excluding connections establishing)
limord_star5.sql tps = 418.262647 (excluding connections establishing)
limord_line5.sql tps = 708.979203 (excluding connections establishing)
limord_line5.sql tps = 710.926866 (excluding connections establishing)
limord_line5.sql tps = 710.928907 (excluding connections establishing)

48870dd + partial-sort-basic-1.patch.gz + fix-cost_sort.patch
join5.sql tps = 505.488181 (excluding connections establishing)
join5.sql tps = 507.222759 (excluding connections establishing)
join5.sql tps = 506.549654 (excluding connections establishing)
star5.sql tps = 495.432915 (excluding connections establishing)
star5.sql tps = 494.906793 (excluding connections establishing)
star5.sql tps = 492.623808 (excluding connections establishing)
line5.sql tps = 789.315968 (excluding connections establishing)
line5.sql tps = 793.875456 (excluding connections establishing)
line5.sql tps = 790.545990 (excluding connections establishing)
lim_join5.sql tps = 396.956732 (excluding connections establishing)
lim_join5.sql tps = 397.515213 (excluding connections establishing)
lim_join5.sql tps = 397.578669 (excluding connections establishing)
lim_star5.sql tps = 417.459963 (excluding connections establishing)
lim_star5.sql tps = 418.024803 (excluding connections establishing)
lim_star5.sql tps = 418.830234 (excluding connections establishing)
lim_line5.sql tps = 729.186915 (excluding connections establishing)
lim_line5.sql tps = 726.288788 (excluding connections establishing)
lim_line5.sql tps = 728.123296 (excluding connections establishing)
limord_join5.sql tps = 383.484767 (excluding connections establishing)
limord_join5.sql tps = 383.021960 (excluding connections establishing)
limord_join5.sql tps = 383.722051 (excluding connections establishing)
limord_star5.sql tps = 414.138460 (excluding connections establishing)
limord_star5.sql tps = 414.063766 (excluding connections establishing)
limord_star5.sql tps = 416.130110 (excluding connections establishing)
limord_line5.sql tps = 706.002589 (excluding connections establishing)
limord_line5.sql tps = 705.632796 (excluding connections establishing)
limord_line5.sql tps = 704.991305 (excluding connections establishing)

Regards,
Marti

On Wed, Feb 12, 2014 at 11:54 PM, Marti Raudsepp <marti@juffo.org> wrote:
> With partial-sort-basic-1 and this fix on the same test suite, the
> planner overhead is now a more manageable 0.5% to 1.3%; one test is
> faster by 0.5%.

Ping, Robert or anyone, does this overhead seem bearable or is that
still too much?

Do these numbers look conclusive enough or should I run more tests?

> I think the 1st patch now has a bug in initial_cost_mergejoin; you
> still pass the "presorted_keys" argument to cost_sort, making it
> calculate a partial sort cost

Ping, Alexander?

Regards,
Marti

On Wed, Feb 19, 2014 at 1:39 PM, Marti Raudsepp <marti@juffo.org> wrote:
> On Wed, Feb 12, 2014 at 11:54 PM, Marti Raudsepp <marti@juffo.org> wrote:
>> With partial-sort-basic-1 and this fix on the same test suite, the
>> planner overhead is now a more manageable 0.5% to 1.3%; one test is
>> faster by 0.5%.
>
> Ping, Robert or anyone, does this overhead seem bearable or is that
> still too much?
>
> Do these numbers look conclusive enough or should I run more tests?

Tom should really be the one to comment on this, I think.  I read
through the patch quickly and it looks much less scary than the early
versions, but it's not obvious to me whether the remaining overhead is
enough to worry about.  I'd need to spend more time studying it to
form a really sound opinion on that topic, and unfortunately I don't
have that time right now.

I think it'd be interesting to try to determine specifically where
that overhead is coming from.  Pick the test case where it's the worst
(1.3%) and do a "perf" with and without the patch and look at the
difference in the call graph.  It's possible we could have changes on
that order of magnitude just from more or less fortuitous code layout
decisions as code shifts around, but it's also possible that there's a
real effect there we should think harder about.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Mon, Feb 10, 2014 at 10:59 AM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> Done. Patch is splitted.

I took a quick look at this.

Have you thought about making your new cmpSortSkipCols() function not
use real comparisons? Since in the circumstances in which this
optimization is expected to be effective (e.g. your original example)
we can also expect a relatively low cardinality for the first n
indexed attributes (again, as in your original example), in general
when cmpSortSkipCols() is called there is a high chance that it will
return true. If any pair of tuples (logically adjacent tuples fed in
to cmpSortSkipCols() by an index scan in logical order) are not fully
equal (i.e. their leading, indexed attributes are not equal) then we
don't care about the details -- we just know that a new sort grouping
is required.

The idea here is that you can get away with simple binary equality
comparisons, as we do when considering HOT-safety. Of course, you
might find that two bitwise unequal values are equal according to
their ordinary B-Tree support function 1 comparator (e.g. two numerics
that differ only in their display scale). AFAICT this should be okay,
since that just means that you have smaller sort groupings than
strictly necessary. I'm not sure if that's worth it to more or less
duplicate heap_tuple_attr_equals() to save a "mere" n expensive
comparisons, but it's something to think about (actually, there are
probably less than even n comparisons in practice because there'll be
a limit).

A similar idea appears in my SortSupport for text ("Poor man's
normalized key"/strxfrm()) patch. A poor man's key comparison didn't
work out, and there may be further differences that aren't captured in
the special simple key representation, so we need to do a "proper
comparison" to figure it out for sure. However, within the sortsupport
routine comparator, we know that we're being called in this context,
as a tie-breaker for a poor man's normalized key comparison that
returned 0, and so are optimistic about the two datums being fully
equal. An optimistic memcmp() is attempted before a strcoll() here if
the lengths also match.

I have not actually added special hints so that we're optimistic about
keys being equal in other places (places that have nothing to do with
the general idea of poor man's normalized keys), but that might not be
a bad idea. Actually, it might not be a bad idea to just always have
varstr_cmp() attempt a memcmp() first when two texts have equal
length, no matter how it's called.

-- 
Peter Geoghegan

On Fri, Sep 12, 2014 at 2:19 PM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> Actually, higher cardinality skip columns is better. Sorting of smaller
> groups is faster than sorting larger groups of same size. Also, with smaller
> groups you achieve limit more accurate (in average), i.e. sort smaller
> amount of total rows.

Higher cardinality leading key columns are better for what? Do you
mean they're better in that those cases are more sympathetic to your
patch, or better in the general sense that they'll perform better for
the user? The first example query, that you chose to demonstrate your
patch had a leading, indexed column (column "v1") with much lower
cardinality/n_distinct than the column that had to be sorted on
(column "v2"). That was what my comments were based on.

When this feature is used, there will often be a significantly lower
cardinality in the leading, indexed column (as in your example).
Otherwise, the user might well have been happy to just order on the
indexed/leading column alone. That isn't definitely true, but it's
often true.

>> I'm not sure if that's worth it to more or less
>> duplicate heap_tuple_attr_equals() to save a "mere" n expensive
>> comparisons, but it's something to think about (actually, there are
>> probably less than even n comparisons in practice because there'll be
>> a limit).
>
> Not correct. Smaller groups are not OK. Imagine that two representations of
> same skip column value exists. Index may return them in any order, even
> change them one by one. In this case sorting on other column never takes
> place, while it should.

I think I explained this badly - it wouldn't be okay to make the
grouping smaller, but as you say we could tie-break with a proper
B-Tree support function 1 comparison to make sure we really had
reached the end of our grouping.

FWIW I want all bttextcmp()/varstr_cmp() comparisons to try a memcmp()
first, so the use of the equality operator with text in mind that you
mention may soon not be useful at all. The evidence suggests that
memcmp() is so much cheaper than special preparatory NUL-termination +
strcoll() that we should always try it first when sorting text, even
when we have no good reason to think it will work. The memcmp() is
virtually free. And so, you see why it might be worth thinking about
this when we already have reasonable confidence that many comparisons
will indicate that datums are equal. Other datatypes will have
expensive "real" comparators, not just text or numeric.

-- 
Peter Geoghegan

Some quick comments on partial-sort-basic-2.patch:

> *** a/src/include/utils/tuplesort.h
> --- b/src/include/utils/tuplesort.h
> ***************
> *** 24,29 ****
> --- 24,30 ----
>   #include "executor/tuptable.h"
>   #include "fmgr.h"
>   #include "utils/relcache.h"
> + #include "utils/sortsupport.h"

Why include sortsupport.h here?

I would like to see more comments, especially within ExecSort(). The
structure of that routine is quite unclear.

I don't really like this MakeSortSupportKeys() stuff within ExecSort():

> !       /* Support structures for cmpSortSkipCols - already sorted columns */
> !       if (skipCols)
> !           node->skipKeys = MakeSortSupportKeys(skipCols,
> !                                                plannode->sortColIdx,
> !                                                plannode->sortOperators,
> !                                                plannode->collations,
> !                                                plannode->nullsFirst);
>
> +       /* Only pass on remaining columns that are unsorted */
>         tuplesortstate = tuplesort_begin_heap(tupDesc,
> !                                             plannode->numCols - skipCols,
> !                                             &(plannode->sortColIdx[skipCols]),
> !                                             &(plannode->sortOperators[skipCols]),
> !                                             &(plannode->collations[skipCols]),
> !                                             &(plannode->nullsFirst[skipCols]),
>                                               work_mem,
>                                               node->randomAccess);

You're calling the new function MakeSortSupportKeys() (which
encapsulates setting up sortsupport state for all attributes) twice;
first, to populate the skip keys (the indexed attribute(s)), and
second, when tuplesort_begin_heap() is called, so that there is state
for unindexed sort groups that must be manually sorted. That doesn't
seem great.

I think we might be better off if a tuplesort function was called
shortly after tuplesort_begin_heap() is called. How top-n heap sorts
work is something that largely lives in tuplesort's head. Today, we
call tuplesort_set_bound() to hint to tuplesort "By the way, this is a
top-n heap sort applicable sort". I think that with this patch, we
should then hint (where applicable) "by the way, you won't actually be
required to sort those first n indexed attributes; rather, you can
expect to scan those in logical order. You may work the rest out
yourself, and may be clever about exploiting the sorted-ness of the
first few columns". The idea of managing a bunch of tiny sorts from
with ExecSort(), and calling the new function tuplesort_reset() seems
questionable. tuplesortstate is supposed to be private/opaque to
nodeSort.c, and the current design strains that.

I'd like to keep nodeSort.c simple. I think it's pretty clear that the
guts of this do not belong within ExecSort(), in any case. Also, the
additions there should be much better commented, wherever they finally
end up.

In this struct:

> *** a/src/include/nodes/execnodes.h
> --- b/src/include/nodes/execnodes.h
> *************** typedef struct SortState
> *** 1670,1678 ****
> --- 1670,1682 ----
>      bool        bounded;        /* is the result set bounded? */
>      int64       bound;          /* if bounded, how many tuples are needed */
>      bool        sort_Done;      /* sort completed yet? */
> +   bool        finished;       /* fetching tuples from outer node
> +                                  is finished ? */
>      bool        bounded_Done;   /* value of bounded we did the sort with */
>      int64       bound_Done;     /* value of bound we did the sort with */
>      void       *tuplesortstate; /* private state of tuplesort.c */
> +   SortSupport skipKeys;       /* columns already sorted in input */
> +   HeapTuple   prev;           /* previous tuple from outer node */
>   } SortState;

I think you should be clearer about the scope and duration of fields
like "finished", if this really belongs here. In general, there should
be some high-level comments about how the feature added by the patch
fits together, which there isn't right now.

That's all I have for now.

-- 
Peter Geoghegan

On 09/15/2014 01:58 PM, Alexander Korotkov wrote:
> On Sun, Sep 14, 2014 at 9:32 AM, Peter Geoghegan <pg@heroku.com
> <mailto:pg@heroku.com>> wrote:
>
>     I think we might be better off if a tuplesort function was called
>     shortly after tuplesort_begin_heap() is called. How top-n heap sorts
>     work is something that largely lives in tuplesort's head. Today, we
>     call tuplesort_set_bound() to hint to tuplesort "By the way, this is a
>     top-n heap sort applicable sort". I think that with this patch, we
>     should then hint (where applicable) "by the way, you won't actually be
>     required to sort those first n indexed attributes; rather, you can
>     expect to scan those in logical order. You may work the rest out
>     yourself, and may be clever about exploiting the sorted-ness of the
>     first few columns". The idea of managing a bunch of tiny sorts from
>     with ExecSort(), and calling the new function tuplesort_reset() seems
>     questionable. tuplesortstate is supposed to be private/opaque to
>     nodeSort.c, and the current design strains that.
>
>     I'd like to keep nodeSort.c simple. I think it's pretty clear that the
>     guts of this do not belong within ExecSort(), in any case. Also, the
>     additions there should be much better commented, wherever they finally
>     end up.
>
>
> As I understand, you propose to incapsulate partial sort algorithm into
> tuplesort. However, in this case we anyway need some significant change
> of its interface: let tuplesort decide when it's able to return tuple.
> Otherwise, we would miss significant part of LIMIT clause optimization.
> tuplesort_set_bound() can't solve all the cases. There could be other
> planner nodes between the partial sort and LIMIT.

Hi,

Are you planning to work on this patch for 9.6?

I generally agree with Peter that the changes to the sorting probably 
belong in the tuplesort code rather than in the executor. This way it 
should also be theoretically possible to support mark/restore.

Andreas

On Sun, Jun 7, 2015 at 8:10 AM, Andreas Karlsson <andreas@proxel.se> wrote:
> Are you planning to work on this patch for 9.6?

FWIW I hope so. It's a nice patch.

-- 
Peter Geoghegan

On Tue, Oct 20, 2015 at 4:17 AM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> Planner regression is fixed in the attached version of patch. It appears
> that get_cheapest_fractional_path_for_pathkeys() behaved wrong when no
> ordering is required.

I don't see an entry in the CF app for this. This seems like something
I should review, though.

-- 
Peter Geoghegan

On Tue, Oct 20, 2015 at 4:17 AM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> Planner regression is fixed in the attached version of patch. It appears
> that get_cheapest_fractional_path_for_pathkeys() behaved wrong when no
> ordering is required.

I took a look at this. My remarks are not comprehensive, but are just
what I noticed having looked at this for the first time in well over a
year.

Before anything else, I would like to emphasize that I think that this
is pretty important work; it's not just a "nice to have". I'm very
glad you picked it up, because we need it. In the real world, there
will be *lots* of cases that this helps.

Explain output
-------------------

A query like your original test query looks like this for me:

postgres=# explain analyze select * from test order by v1, v2 limit 100;
               QUERY
 
PLAN

--------------------------------------------------------------------------------------------------------------------------------------------Limit
(cost=429.54..434.53 rows=100 width=16) (actual
 
time=15125.819..22414.038 rows=100 loops=1)  ->  Partial sort  (cost=429.54..50295.52 rows=1000000 width=16)
(actual time=15125.799..22413.297 rows=100 loops=1)        Sort Key: v1, v2        Presorted Key: v1        Sort
Method:top-N heapsort  Memory: 27kB        ->  Index Scan using test_v1_idx on test
 
(cost=0.42..47604.43 rows=1000000 width=16) (actual time=1.663..13.066
rows=151 loops=1)Planning time: 0.948 msExecution time: 22414.895 ms
(8 rows)

I thought about it for a while, and decided that you have the basic
shape of the explain output right here. I see where you are going by
having the sort node drive things.

I don't think the node should be called "Partial sort", though. I
think that this is better presented as just a "Sort" node with a
presorted key.

I think it might be a good idea to also have a "Sort Groups: 2" field
above. That illustrates that you are in fact performing 2 small sorts
per group, which is the reality. As you said, it's good to have this
be high, because then the sort operations don't need to do too many
comparisons, which could be expensive.

Sort Method
----------------

Even thought the explain analyze above shows "top-N heapsort" as its
sort method, that isn't really true. I actually ran this through a
debugger, which is why the above plan took so long to execute, in case
you wondered. I saw that in practice the first sort executed for the
first group uses a quicksort, while only the second sort (needed for
the 2 and last group in this example) used a top-N heapsort.

Is it really worth using a top-N heapsort to avoid sorting the last
little bit of tuples in the last group? Maybe we should never push
down to an individual sort operation (we have one
tuplesort_performsort() per group) that it should be bounded. Our
quicksort falls back on insertion sort in the event of only having 7
elements (at that level of recursion), so having this almost always
use quicksort may be no bad thing.

Even if you don't like that, the "Sort Method" shown above is just
misleading. I wonder, also, if you need to be more careful about
whether or not "Memory" is really the high watermark, as opposed to
the memory used by the last sort operation of many. There could be
many large tuples in one grouping, for example. Note that the current
code will not show any "Memory" in explain analyze for cases that have
memory freed before execution is done, which this is not consistent
with. Maybe that's not so important. Unsure.

trace_sort output shows that these queries often use a large number of
tiny individual sorts. Maybe that's okay, or maybe we should make it
clearer that the sorts are related. I now use trace_sort a lot.

Abbreviated Keys
-----------------------

It could be very bad for performance that the first non-presorted key
uses abbreviated keys. There needs to be a way to tell tuplesort to
not waste its time with them, just as there currently is for bounded
(top-N heapsort) sorts. They're almost certainly the wrong way to go,
unless you have huge groups (where partial sorting is unlikely to win
in the first place).

Other issues in executor
--------------------------------

This is sort of an optimizer issue, but code lives in execAmi.c.
Assert is redundant here:

+               case T_Sort:
+                       /* We shouldn't reach here without having plan node */
+                       Assert(node);
+                       /* With skipCols sort node holds only last bucket */
+                       if (node && ((Sort *)node)->skipCols == 0)
+                               return true;
+                       else
+                               return false;

I don't like that you've added a Plan node argument to
ExecMaterializesOutput() in this function, too.

There is similar assert/pointer test code within
ExecSupportsBackwardScan() and ExecSupportsMarkRestore(). In general,
I have concerns about the way the determination of a sort's ability to
do stuff like be scanned backwards is now made dynamic, which this new
code demonstrates:
       /*
+        * skipCols can't be used with either EXEC_FLAG_REWIND,
EXEC_FLAG_BACKWARD
+        * or EXEC_FLAG_MARK, because we hold only current bucket in
+        * tuplesortstate.
+        */
+       Assert(node->skipCols == 0 || (eflags & (EXEC_FLAG_REWIND |
+                 EXEC_FLAG_BACKWARD |
+                 EXEC_FLAG_MARK)) == 0);
+

I need to think some more about this general issue.

Misc. issues
----------------

_readSort() needs READ_INT_FIELD().  _outSort() similarly needs
WRITE_INT_FIELD(). You've mostly missed this stuff.

Please be more careful about this. It's always a good idea to run the
regression tests with "#define COPY_PARSE_PLAN_TREES" from time to
time, which tends to highlight these problems.

tuplesort.h should not include sortsupport.h. It's a modularity
violation, and besides which is unnecessary. Similarly, pathkeys.c
should not include optimizer/cost.h.

What is this?

+               if (inner_cheapest_total &&
inner_cheapest_total->pathtype == T_Sort)
+                       elog(ERROR, "Sort");

Optimizer
-------------

I am not an expert on the optimizer, but I do have some feedback.

* cost_sort() needs way way more comments.  Doesn't even mention
indexes. Not worth commenting further on until I know what it's
*supposed* to do, though.

* pathkeys_useful_for_ordering() now looks like a private convenience
wrapper for the new public function pathkeys_common().  I think that
comments should make this quite clear.

* compare_bifractional_path_costs() should live beside
compare_fractional_path_costs() and be public, I think. The existing
compare_fractional_path_costs() also only has a small number of
possible clients, and is still not static.

* Think it's not okay that there are new arguments, such as the
"tuples" argument for get_cheapest_fractional_path_for_pathkeys().

It seems a bad sign, design-wise, that a new argument of "PlannerInfo
*root" was added at end, for the narrow purpose of passing stuff to
estimate number of groups for the benefit of this patch.  ISTM that
grouping_planner() caller should do the
work itself as and when it alone needs to.

* New loop within get_cheapest_fractional_path_for_pathkeys() requires
far more explanation.

Explain theory behind derivation of compare_bifractional_path_costs()
fraction arguments, please. I think there might be simple heuristics
like this elsewhere in the optimizer or selfuncs.c, but you need to
share why you did things that way in the code.

* Within planner.c, "partial_sort_path" variable name in
grouping_planner() might be called something else.

Its purpose isn't clear. Also, the way that you mix path costs from
the new get_cheapest_fractional_path_for_pathkeys() into the new
cost_sort() needs to be explained in detail (as I already said,
cost_sort() is currently very under-documented).

Obviously the optimizer part of this needs the most work -- no
surprises there. I wonder if we cover all useful cases? I haven't yet
got around to using "#define OPTIMIZER_DEBUG" to see what's really
going on, which seems essential to understanding what is really
happening, but it looks like merge append paths can currently use the
optimization, whereas unique paths cannot. Have you thought about
that?

That's all I have for now...

-- 
Peter Geoghegan

Hi,

On 10/20/2015 01:17 PM, Alexander Korotkov wrote:
> On Fri, Oct 16, 2015 at 7:11 PM, Alexander Korotkov
> <aekorotkov@gmail.com <mailto:aekorotkov@gmail.com>> wrote:
>
>     On Sun, Jun 7, 2015 at 11:01 PM, Peter Geoghegan <pg@heroku.com
>     <mailto:pg@heroku.com>> wrote:
>
>         On Sun, Jun 7, 2015 at 8:10 AM, Andreas Karlsson
>         <andreas@proxel.se <mailto:andreas@proxel.se>> wrote:
>         > Are you planning to work on this patch for 9.6?
>
>         FWIW I hope so. It's a nice patch.
>
>
>     I'm trying to to whisk dust. Rebased version of patch is attached.
>     This patch isn't passing regression tests because of plan changes.
>     I'm not yet sure about those changes: why they happens and are they
>     really regression?
>     Since I'm not very familiar with planning of INSERT ON CONFLICT and
>     RLS, any help is appreciated.
>
>
> Planner regression is fixed in the attached version of patch. It appears
> that get_cheapest_fractional_path_for_pathkeys() behaved wrong when no
> ordering is required.
>

Alexander, are you working on this patch? I'd like to look at the patch, 
but the last available version (v4) no longer applies - there's plenty 
of bitrot. Do you plan to send an updated / rebased version?


The main thing I'm particularly interested in is how much is this 
coupled with the Sort node, and whether it's possible to feed partially 
sorted tuples into other nodes.

I'm particularly thinking about Hash Aggregate, because the partial sort 
allows to keep only the "current group" in a hash table, making it much 
more memory efficient / faster. What do you think?

regards

--
Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On Sat, Jan 23, 2016 at 4:07 AM, Tomas Vondra
<tomas.vondra@2ndquadrant.com> wrote:
> The main thing I'm particularly interested in is how much is this coupled
> with the Sort node, and whether it's possible to feed partially sorted
> tuples into other nodes.

That's cool, but I'm particularly interested in seeing Alexander get
back to this because it's an important project on its own. We should
really have this.

-- 
Peter Geoghegan

Alexander Korotkov wrote:
> I'm sorry that I didn't found time for this yet. I'm certainly planning to
> get back to this in near future. The attached version is just rebased
> without any optimization.

Great to have a new version -- there seems to be a lot of interest in
this patch.  I'm moving this one to the next commitfest, thanks.

-- 
Álvaro Herrera                http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On Sun, Jan 31, 2016 at 4:16 AM, Alvaro Herrera
<alvherre@2ndquadrant.com> wrote:
> Great to have a new version -- there seems to be a lot of interest in
> this patch.  I'm moving this one to the next commitfest, thanks.

I am signed up to review this patch.

I was very surprised to see it in "Needs Review" state in the CF app
(Alexander just rebased the patch, and didn't do anything with the CF
app entry). Once again, this seems to have happened just because
Alvaro moved the patch to the next CF.

I've marked it "Waiting on Author" once more. Hopefully the CF app
will be fixed soon, so moving a patch to the next commitfest no longer
clobbers its state.

-- 
Peter Geoghegan

Hi,

On Tue, Mar 1, 2016 at 7:06 AM, Alexander Korotkov <aekorotkov@gmail.com> wrote:
> I finally went over your review.

I'll respond to your points here. Note that I'm reviewing
"partial-sort-basic-7.patch", which you sent on March 13. I respond
here because this is where you answered my questions (I had no
feedback on "partial-sort-basic-6.patch", which didn't use the new
upper planner pathification stuff, unlike this latest version).

> On Wed, Nov 4, 2015 at 4:44 AM, Peter Geoghegan <pg@heroku.com> wrote:
>>
>> Explain output
>> -------------------

>> I think it might be a good idea to also have a "Sort Groups: 2" field
>> above. That illustrates that you are in fact performing 2 small sorts
>> per group, which is the reality. As you said, it's good to have this
>> be high, because then the sort operations don't need to do too many
>> comparisons, which could be expensive.
>
>
> I agree with your notes. In the attached version of path explain output was
> revised as you proposed.

Cool.

>> Sort Method
>> ----------------
>>
>> Even thought the explain analyze above shows "top-N heapsort" as its
>> sort method, that isn't really true. I actually ran this through a
>> debugger, which is why the above plan took so long to execute, in case
>> you wondered. I saw that in practice the first sort executed for the
>> first group uses a quicksort, while only the second sort (needed for
>> the 2 and last group in this example) used a top-N heapsort.

> With partial sort we run multiple sorts in the same node. Ideally, we need
> to provide some aggregated information over runs.
> This situation looks very similar to subplan which is called multiple times.
> I checked how it works for now.

Noticed this in nodeSort.c:

+       if (node->tuplesortstate != NULL)
+       {
+               tuplesort_reset((Tuplesortstate *) node->tuplesortstate);
+               node->groupsCount++;
+       }
+       else
+       {
+               /* Support structures for cmpSortSkipCols - already
sorted columns */
+               if (skipCols)
+                       prepareSkipCols(plannode, node);

+               /*
+                * Only pass on remaining columns that are unsorted.
Skip abbreviated
+                * keys usage for partial sort.  We unlikely will have
huge groups
+                * with partial sort.  Therefore usage of abbreviated
keys would be
+                * likely a waste of time.
+                */               tuplesortstate = tuplesort_begin_heap(tupDesc,

You should comment on which case is which, and put common case (no
skip cols) first. Similarly, the ExecSort() for(;;) should put the
common (non-partial) case first, which it does, but then the "first
tuple in partial sort" case first, then the "second or subsequent
partial sort" case last.

More comments here, please:

+typedef struct SkipKeyData
+{
+ FunctionCallInfoData fcinfo;
+ FmgrInfo flinfo;
+ OffsetNumber attno;
+} SkipKeyData;

(What's SkipKeyData?)

Also want comments for new SortState fields. SortState.prev is a
palloc()'d copy of tuple, which should be directly noted, as it is for
similar aggregate cases, etc.

Should you be more aggressive about freeing memory allocated for
SortState.prev tuples?

The new function cmpSortSkipCols() should say "Special case for
NULL-vs-NULL, else use standard comparison", or something. "Lets
pretend NULL is a value for implementation convenience" cases are
considered the exception, and are always noted as the exception.

> In the case of subplan explain analyze gives us just information about last
> subplan run. This makes me uneasy. From one side, it's probably OK that
> partial sort behaves like subplan while showing information just about last
> sort run. From the other side, we need some better solution for that in
> general case.

I see what you mean, but I wasn't so much complaining about that, as
complaining about the simple fact that we use a top-N heap sort *at
all*. This feels like the "limit" case is playing with partial sort
sub-sorts in a way that it shouldn't.

I see you have code like this to make this work:

+       /*
+        * Adjust bound_Done with number of tuples we've actually sorted.
+        */
+       if (node->bounded)
+       {
+               if (node->finished)
+                       node->bound_Done = node->bound;
+               else
+                       node->bound_Done = Min(node->bound,
node->bound_Done + nTuples);

But, why bother? Why not simply prevent tuplesort.c from ever using
the top-N heapsort method when it is called from nodeSort.c for a
partial sort (probably in the planner)?

Why, at a high level, does it make sense to pass down a limit to *any*
sort operation that makes up a partial sort, even the last? This seems
to be adding complexity without a benefit. A big advantage of top-N
heapsorts is that much less memory could be used, but this patch
already has the memory allocated that belonged to previous performsort
calls (mostly -- certainly has all those tuplesort.c memtuples
throughout, a major user of memory overall).  It's not going to be
very good at preventing work, except almost by accident because we
happen to have a limit up to just past the beginning of a smaller
partial sort group. I'd rather use quicksort, which is very versatile.
Top-N sorts make sense when sorting itself is the bottleneck, which it
probably won't be for a partial sort (that's the whole point).

If the sort method was very likely the same for every performsort
(quicksort), which it otherwise would be, then I'd care way way less
that that could be a bit misleading in EXPLAIN ANALYZE output, because
typically the last one would be "close enough". Although, this isn't
quite like your SubPlan example, because the Sort node isn't reported
as e.g. "SubPlan 1" by EXPLAIN.

I think that this has bugs for external sorts:

+void
+tuplesort_reset(Tuplesortstate *state)
+{
+       int i;
+
+       if (state->tapeset)
+               LogicalTapeSetClose(state->tapeset);
+
+       for (i = 0; i < state->memtupcount; i++)
+               free_sort_tuple(state, state->memtuples + i);
+
+       state->status = TSS_INITIAL;
+       state->memtupcount = 0;
+       state->boundUsed = false;
+       state->tapeset = NULL;
+       state->currentRun = 0;
+       state->result_tape = -1;
+       state->bounded = false;
+}

It's not okay to reset like this, especially not after the recent
commit 0011c0091, which could make this code unacceptably leak memory.
I realize that we really should never use an external sort here, but,
as you know, this is not the point.

So, I want to suggest that you use the regular code to destroy and
recreate a tuplesort in this case. Now, obviously that has some
significant disadvantages -- you want to reuse everything in the
common case when each sort is tiny. But you can still do that for that
very common case.

I think you need to use sortcontext memory context here on general
principle, even if current usage isn't broken by that.

Even if you get this right for external sorts once, it will break
again without anyone noticing until it's too late. Better to not rely
on it staying in sync, and find a way of having the standard
tuplesort.c initialization begin again.

Even though these free_sort_tuple() calls are still needed, you might
also call "MemoryContextReset(state->tuplecontext)" at the end. That
might prevent palloc() fragmentation when groups are of wildly
different sizes. Just an idea.

>> I don't like that you've added a Plan node argument to
>> ExecMaterializesOutput() in this function, too.
>
>
> I don't like this too. But I didn't find better solution without significant
> rework of planner.
> However, "Upper planner pathification" by Tom Lane seems to have such
> rework. It's likely sort becomes separate path node there.
> Then ExecMaterializesOutput could read parameters of path node.

A tuplesort may be randomAccess, or !randomAccess, as the caller
wishes. It's good for performance if the caller does not want
randomAccess, because then we can do our final merge on-the-fly if
it's an external sort, which helps a lot.

How is this different? ExecMaterializesOutput() seems to be about
whether or not the plan *could* materialize its output in principle,
even though you might well want to make it not do so in specific
cases. So, it's not so much that the new argument is ugly; rather, I
worry that it's wrong to make ExecMaterializesOutput() give a more
specific answer than anticipated by current callers.

Is the difference basically just that a partial sort could be
enormously faster, whereas a !randomAccess conventional sort is nice
to have, but not worth e.g. changing cost_sort() to account for?

You might just make a new function, ExecPlanMaterializesOutput(),
instead. That would call ExecMaterializesOutput() for non-Sort cases.

>> Optimizer
>> -------------
>>

>> * cost_sort() needs way way more comments.  Doesn't even mention
>> indexes. Not worth commenting further on until I know what it's
>> *supposed* to do, though.
>
>
> I've added some comments.

Looking at cost_sort() now, it's a bit clearer. I think that you
should make sure that everything is costed as a quicksort, though, if
you accept that we should try and make every small sort done by the
partial sort a quicksort. Which I think is a good idea. The common
case is that groups are small, but the qsort() insertion sort will be
very very fast for that case.

>> * New loop within get_cheapest_fractional_path_for_pathkeys() requires
>> far more explanation.
>>
>> Explain theory behind derivation of compare_bifractional_path_costs()
>> fraction arguments, please. I think there might be simple heuristics
>> like this elsewhere in the optimizer or selfuncs.c, but you need to
>> share why you did things that way in the code.
>
>
> Idea is that since partial sort fetches data per group then it would require
> fetching more data than fully presorted path.

I think I get it.

>> * Within planner.c, "partial_sort_path" variable name in
>> grouping_planner() might be called something else.
>>
>> Its purpose isn't clear. Also, the way that you mix path costs from
>> the new get_cheapest_fractional_path_for_pathkeys() into the new
>> cost_sort() needs to be explained in detail (as I already said,
>> cost_sort() is currently very under-documented).
>
>
> I've tried to make it more clear. partial_sort_path is renamed to
> presorted_path.

> Unique paths occasionally can use this optimization.

> But it depends on attribute order. I could work out this case, but I would
> prefer some simple case to commit before. I already throw merge join
> optimization away for the sake of simplicity.

I think that was the right decision under our time constraints.
However, I suggest noting that this should happen for unique paths in
the future, say within create_unique_path().

Other notes:

This looks like an old change you missed:

- * compare_path_fractional_costs
+ * compare_fractional_path_costs

All in all, this looks significantly better. Thanks for your work on
this. Sorry for the delay in my response, and that my review was
relatively rushed, but I'm rather busy at the moment with fighting
fires.

-- 
Peter Geoghegan

Hi Alexander,

On 3/23/16 8:39 PM, Peter Geoghegan wrote:

> This looks like an old change you missed:
>
> - * compare_path_fractional_costs
> + * compare_fractional_path_costs
>
> All in all, this looks significantly better. Thanks for your work on
> this. Sorry for the delay in my response, and that my review was
> relatively rushed, but I'm rather busy at the moment with fighting
> fires.

It looks like a new patch is required before this can be marked "ready 
for committer".  Will you have that ready soon?

Thanks,
-- 
-David
david@pgmasters.net

On Wed, Mar 30, 2016 at 8:02 AM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> Hmm... I'm not completely agree with that. In typical usage partial sort
> should definitely use quicksort.  However, fallback to other sort methods is
> very useful.  Decision of partial sort usage is made by planner.  But
> planner makes mistakes.  For example, our HashAggregate is purely in-memory.
> In the case of planner mistake it causes OOM.  I met such situation in
> production and not once.  This is why I'd like partial sort to have graceful
> degradation for such cases.

I think that this should be moved to the next CF, unless a committer
wants to pick it up today.


-- 
Peter Geoghegan

On Tue, Sep 13, 2016 at 5:32 PM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> On Fri, Apr 8, 2016 at 10:09 PM, Peter Geoghegan <pg@heroku.com> wrote:
>>
>> On Wed, Mar 30, 2016 at 8:02 AM, Alexander Korotkov
>> <aekorotkov@gmail.com> wrote:
>> > Hmm... I'm not completely agree with that. In typical usage partial sort
>> > should definitely use quicksort.  However, fallback to other sort
>> > methods is
>> > very useful.  Decision of partial sort usage is made by planner.  But
>> > planner makes mistakes.  For example, our HashAggregate is purely
>> > in-memory.
>> > In the case of planner mistake it causes OOM.  I met such situation in
>> > production and not once.  This is why I'd like partial sort to have
>> > graceful
>> > degradation for such cases.
>>
>> I think that this should be moved to the next CF, unless a committer
>> wants to pick it up today.
>
>
> Patch was rebased to current master.

Applies on HEAD at e8bdee27 and passes make-check, now I am seeing
zero documentation so it is a bit hard to see what this patch is
achieving without reading the thread.

$ git diff master --check
src/backend/optimizer/prep/prepunion.c:967: trailing whitespace.
+   cost_sort(&sorted_p, root, NIL, 0,
src/backend/utils/sort/tuplesort.c:1244: trailing whitespace.
+ * tuplesort_updatemax

+ * Returns true if the plan node isautomatically materializes its output
Typo here.

Still, this has received to reviews, so moved to next CF.
-- 
Michael

On Tue, Sep 13, 2016 at 4:32 AM, Alexander Korotkov
<aekorotkov@gmail.com> wrote:
> On Fri, Apr 8, 2016 at 10:09 PM, Peter Geoghegan <pg@heroku.com> wrote:
>> On Wed, Mar 30, 2016 at 8:02 AM, Alexander Korotkov
>> <aekorotkov@gmail.com> wrote:
>> > Hmm... I'm not completely agree with that. In typical usage partial sort
>> > should definitely use quicksort.  However, fallback to other sort
>> > methods is
>> > very useful.  Decision of partial sort usage is made by planner.  But
>> > planner makes mistakes.  For example, our HashAggregate is purely
>> > in-memory.
>> > In the case of planner mistake it causes OOM.  I met such situation in
>> > production and not once.  This is why I'd like partial sort to have
>> > graceful
>> > degradation for such cases.
>>
>> I think that this should be moved to the next CF, unless a committer
>> wants to pick it up today.
>
> Patch was rebased to current master.

Just a few quick observations on this...

It strikes me that the API contract change in ExecMaterializesOutput
is pretty undesirable.  I think it would be better to have a new
executor node for this node rather than overloading the existing
"Sort" node, sharing code where possible of course.  The fact that
this would distinguish them more clearly in an EXPLAIN plan seems
good, too.  "Partial Sort" is the obvious thing, but there might be
even better alternatives -- maybe "Incremental Sort" or something like
that?  Because it's not partially sorting the data, it's making data
that already has some sort order have a more rigorous sort order.

I think that it will probably be pretty common to have queries where
the data is sorted by (mostly_unique_col) and we want to get it sorted
by (mostly_unique_col, disambiguation_col).  In such cases I wonder if
we'll incur a lot of overhead by feeding single tuples to the
tuplesort stuff and performing lots of 1-item sorts.  Not sure if that
case needs any special optimization.

I also think that the "HeapTuple prev" bit in SortState is probably
not the right way of doing things.  I think that should use an
additional TupleTableSlot rather than a HeapTuple.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Mon, Nov 21, 2016 at 11:04 PM, Haribabu Kommi
<kommi.haribabu@gmail.com> wrote:
> you assigned as reviewer to the current patch in the 11-2016 commitfest.
> But you haven't shared your review yet in this commitfest on the latest
> patch posted by the author. If you don't have any comments on the patch,
> please move the patch into "ready for committer" state to get committer's
> attention. This will help us in smoother operation of commitfest.

Sorry for the delay on this.

I agree with Robert's remarks today on TupleTableSlot, and would like
to see a revision that does that.

-- 
Peter Geoghegan

On Mon, Dec 5, 2016 at 2:04 PM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote:
>
>
> On Fri, Dec 2, 2016 at 4:05 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>>
>> On Tue, Sep 13, 2016 at 4:32 AM, Alexander Korotkov
>> <aekorotkov@gmail.com> wrote:
>> > On Fri, Apr 8, 2016 at 10:09 PM, Peter Geoghegan <pg@heroku.com> wrote:
>> >> On Wed, Mar 30, 2016 at 8:02 AM, Alexander Korotkov
>> >> <aekorotkov@gmail.com> wrote:
>> >> > Hmm... I'm not completely agree with that. In typical usage partial
>> >> > sort
>> >> > should definitely use quicksort.  However, fallback to other sort
>> >> > methods is
>> >> > very useful.  Decision of partial sort usage is made by planner.  But
>> >> > planner makes mistakes.  For example, our HashAggregate is purely
>> >> > in-memory.
>> >> > In the case of planner mistake it causes OOM.  I met such situation
>> >> > in
>> >> > production and not once.  This is why I'd like partial sort to have
>> >> > graceful
>> >> > degradation for such cases.
>> >>
>> >> I think that this should be moved to the next CF, unless a committer
>> >> wants to pick it up today.
>> >
>> > Patch was rebased to current master.
>>
>> Just a few quick observations on this...
>>
>> It strikes me that the API contract change in ExecMaterializesOutput
>> is pretty undesirable.  I think it would be better to have a new
>> executor node for this node rather than overloading the existing
>> "Sort" node, sharing code where possible of course.  The fact that
>> this would distinguish them more clearly in an EXPLAIN plan seems
>> good, too.  "Partial Sort" is the obvious thing, but there might be
>> even better alternatives -- maybe "Incremental Sort" or something like
>> that?  Because it's not partially sorting the data, it's making data
>> that already has some sort order have a more rigorous sort order.
>>
>> I think that it will probably be pretty common to have queries where
>> the data is sorted by (mostly_unique_col) and we want to get it sorted
>> by (mostly_unique_col, disambiguation_col).  In such cases I wonder if
>> we'll incur a lot of overhead by feeding single tuples to the
>> tuplesort stuff and performing lots of 1-item sorts.  Not sure if that
>> case needs any special optimization.
>>
>> I also think that the "HeapTuple prev" bit in SortState is probably
>> not the right way of doing things.  I think that should use an
>> additional TupleTableSlot rather than a HeapTuple.
>>
>
> The feedback from the reviewer has received at the end of commitfest.
> So Moved to next CF with "waiting on author" status.

This patch is on its 6th commit fest now. As the thread has died and
as feedback has been provided but not answered I am marking it as
returned with feedback.
-- 
Michael