Thread: MergeAppend could consider sorting cheapest child path

Hi.

Now when planner finds suitable pathkeys in 
generate_orderedappend_paths(), it uses them, even if explicit sort of 
the cheapest child path could be more efficient.

We encountered this issue on partitioned table with two indexes, where 
one is suitable for sorting, and another is good for selecting data. 
MergeAppend was generated
with subpaths doing index scan on suitably ordered index and filtering a 
lot of data.
The suggested fix allows MergeAppend to consider sorting on cheapest 
childrel total path as an alternative.

-- 
Best regards,
Alexander Pyhalov,
Postgres Professional

Is this still being considered?

---------------------------------------------------------------------------

On Tue, Jun 18, 2024 at 07:45:09PM +0300, Alexander Pyhalov wrote:
> Hi.
> 
> Now when planner finds suitable pathkeys in generate_orderedappend_paths(),
> it uses them, even if explicit sort of the cheapest child path could be more
> efficient.
> 
> We encountered this issue on partitioned table with two indexes, where one
> is suitable for sorting, and another is good for selecting data. MergeAppend
> was generated
> with subpaths doing index scan on suitably ordered index and filtering a lot
> of data.
> The suggested fix allows MergeAppend to consider sorting on cheapest
> childrel total path as an alternative.
> 
> -- 
> Best regards,
> Alexander Pyhalov,
> Postgres Professional

> From d5eb3d326d83e2ca47c17552fcc6fd27b6b98d4e Mon Sep 17 00:00:00 2001
> From: Alexander Pyhalov <a.pyhalov@postgrespro.ru>
> Date: Tue, 18 Jun 2024 15:56:18 +0300
> Subject: [PATCH] MergeAppend could consider using sorted best path.
> 
> This helps when index with suitable pathkeys is not
> good for filtering data.
> ---
>  .../postgres_fdw/expected/postgres_fdw.out    |  6 +-
>  src/backend/optimizer/path/allpaths.c         | 21 +++++
>  src/test/regress/expected/inherit.out         | 45 +++++++++-
>  src/test/regress/expected/partition_join.out  | 87 +++++++++++--------
>  src/test/regress/expected/union.out           |  6 +-
>  src/test/regress/sql/inherit.sql              | 17 ++++
>  6 files changed, 141 insertions(+), 41 deletions(-)
> 
> diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
> index ea566d50341..687591e4a97 100644
> --- a/contrib/postgres_fdw/expected/postgres_fdw.out
> +++ b/contrib/postgres_fdw/expected/postgres_fdw.out
> @@ -10074,13 +10074,15 @@ SELECT t1.a, t2.b FROM fprt1 t1 INNER JOIN fprt2 t2 ON (t1.a = t2.b) WHERE t1.a
>     ->  Nested Loop
>           Join Filter: (t1.a = t2.b)
>           ->  Append
> -               ->  Foreign Scan on ftprt1_p1 t1_1
> +               ->  Sort
> +                     Sort Key: t1_1.a
> +                     ->  Foreign Scan on ftprt1_p1 t1_1
>                 ->  Foreign Scan on ftprt1_p2 t1_2
>           ->  Materialize
>                 ->  Append
>                       ->  Foreign Scan on ftprt2_p1 t2_1
>                       ->  Foreign Scan on ftprt2_p2 t2_2
> -(10 rows)
> +(12 rows)
>  
>  SELECT t1.a, t2.b FROM fprt1 t1 INNER JOIN fprt2 t2 ON (t1.a = t2.b) WHERE t1.a % 25 = 0 ORDER BY 1,2 FOR UPDATE OF
t1;
>    a  |  b  
> diff --git a/src/backend/optimizer/path/allpaths.c b/src/backend/optimizer/path/allpaths.c
> index 4895cee9944..827bc469269 100644
> --- a/src/backend/optimizer/path/allpaths.c
> +++ b/src/backend/optimizer/path/allpaths.c
> @@ -1845,6 +1845,27 @@ generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
>                  /* Assert we do have an unparameterized path for this child */
>                  Assert(cheapest_total->param_info == NULL);
>              }
> +            else
> +            {
> +                /*
> +                 * Even if we found necessary pathkeys, using unsorted path
> +                 * can be more efficient.
> +                 */
> +                Path        sort_path;
> +
> +                cost_sort(&sort_path,
> +                          root,
> +                          pathkeys,
> +                          childrel->cheapest_total_path->total_cost,
> +                          childrel->cheapest_total_path->rows,
> +                          childrel->cheapest_total_path->pathtarget->width,
> +                          0.0,
> +                          work_mem,
> +                          -1.0 /* need all tuples to sort them */ );
> +
> +                if (compare_path_costs(&sort_path, cheapest_total, TOTAL_COST) < 0)
> +                    cheapest_total = childrel->cheapest_total_path;
> +            }
>  
>              /*
>               * When building a fractional path, determine a cheapest
> diff --git a/src/test/regress/expected/inherit.out b/src/test/regress/expected/inherit.out
> index ad732134148..16e78c8d2ff 100644
> --- a/src/test/regress/expected/inherit.out
> +++ b/src/test/regress/expected/inherit.out
> @@ -1555,6 +1555,7 @@ insert into matest2 (name) values ('Test 4');
>  insert into matest3 (name) values ('Test 5');
>  insert into matest3 (name) values ('Test 6');
>  set enable_indexscan = off;  -- force use of seqscan/sort, so no merge
> +set enable_sort = off; -- avoid sorting below MergeAppend
>  explain (verbose, costs off) select * from matest0 order by 1-id;
>                           QUERY PLAN                         
>  ------------------------------------------------------------
> @@ -1608,6 +1609,7 @@ select min(1-id) from matest0;
>  (1 row)
>  
>  reset enable_indexscan;
> +reset enable_sort;
>  set enable_seqscan = off;  -- plan with fewest seqscans should be merge
>  set enable_parallel_append = off; -- Don't let parallel-append interfere
>  explain (verbose, costs off) select * from matest0 order by 1-id;
> @@ -1702,7 +1704,9 @@ order by t1.b limit 10;
>           Merge Cond: (t1.b = t2.b)
>           ->  Merge Append
>                 Sort Key: t1.b
> -               ->  Index Scan using matest0i on matest0 t1_1
> +               ->  Sort
> +                     Sort Key: t1_1.b
> +                     ->  Seq Scan on matest0 t1_1
>                 ->  Index Scan using matest1i on matest1 t1_2
>           ->  Materialize
>                 ->  Merge Append
> @@ -1711,7 +1715,7 @@ order by t1.b limit 10;
>                             Filter: (c = d)
>                       ->  Index Scan using matest1i on matest1 t2_2
>                             Filter: (c = d)
> -(14 rows)
> +(16 rows)
>  
>  reset enable_nestloop;
>  drop table matest0 cascade;
> @@ -2663,6 +2667,43 @@ explain (costs off) select * from mcrparted where a = 10 order by a, abs(b), c;
>  
>  reset enable_bitmapscan;
>  drop table mcrparted;
> +-- Check that sort path can be used by MergeAppend even when there are suitable pathkeys
> +create table hash_parted (i int, j int, k int) partition by hash(i);
> +create table hash_parted_1 partition of hash_parted for values with (modulus 4, remainder 0);
> +create table hash_parted_2 partition of hash_parted for values with (modulus 4, remainder 1);
> +create table hash_parted_3 partition of hash_parted for values with (modulus 4, remainder 2);
> +create table hash_parted_4 partition of hash_parted for values with (modulus 4, remainder 3);
> +--create table hash_parted_5 partition of hash_parted for values with (modulus 6, remainder 4);
> +--create table hash_parted_6 partition of hash_parted for values with (modulus 6, remainder 5);
> +create index on hash_parted(j);
> +create index on hash_parted(k);
> +insert into hash_parted select i, i, i from generate_series(1,10000) i;
> +analyze hash_parted;
> +explain (costs off) select * from hash_parted where k<100 order by j limit 100;
> +                               QUERY PLAN                                
> +-------------------------------------------------------------------------
> + Limit
> +   ->  Merge Append
> +         Sort Key: hash_parted.j
> +         ->  Sort
> +               Sort Key: hash_parted_1.j
> +               ->  Index Scan using hash_parted_1_k_idx on hash_parted_1
> +                     Index Cond: (k < 100)
> +         ->  Sort
> +               Sort Key: hash_parted_2.j
> +               ->  Index Scan using hash_parted_2_k_idx on hash_parted_2
> +                     Index Cond: (k < 100)
> +         ->  Sort
> +               Sort Key: hash_parted_3.j
> +               ->  Index Scan using hash_parted_3_k_idx on hash_parted_3
> +                     Index Cond: (k < 100)
> +         ->  Sort
> +               Sort Key: hash_parted_4.j
> +               ->  Index Scan using hash_parted_4_k_idx on hash_parted_4
> +                     Index Cond: (k < 100)
> +(19 rows)
> +
> +drop table hash_parted;
>  -- Ensure LIST partitions allow an Append to be used instead of a MergeAppend
>  create table bool_lp (b bool) partition by list(b);
>  create table bool_lp_true partition of bool_lp for values in(true);
> diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
> index 6d07f86b9bc..80d480d33d5 100644
> --- a/src/test/regress/expected/partition_join.out
> +++ b/src/test/regress/expected/partition_join.out
> @@ -1309,28 +1309,32 @@ SELECT t1.a, t1.c, t2.b, t2.c, t3.a + t3.b, t3.c FROM (prt1 t1 LEFT JOIN prt2 t2
>  -- This should generate a partitionwise join, but currently fails to
>  EXPLAIN (COSTS OFF)
>  SELECT t1.a, t2.b FROM (SELECT * FROM prt1 WHERE a < 450) t1 LEFT JOIN (SELECT * FROM prt2 WHERE b > 250) t2 ON t1.a
=t2.b WHERE t1.b = 0 ORDER BY t1.a, t2.b;
 
> -                        QUERY PLAN                         
> ------------------------------------------------------------
> - Incremental Sort
> +                           QUERY PLAN                            
> +-----------------------------------------------------------------
> + Sort
>     Sort Key: prt1.a, prt2.b
> -   Presorted Key: prt1.a
> -   ->  Merge Left Join
> -         Merge Cond: (prt1.a = prt2.b)
> -         ->  Sort
> -               Sort Key: prt1.a
> -               ->  Append
> -                     ->  Seq Scan on prt1_p1 prt1_1
> -                           Filter: ((a < 450) AND (b = 0))
> -                     ->  Seq Scan on prt1_p2 prt1_2
> -                           Filter: ((a < 450) AND (b = 0))
> -         ->  Sort
> -               Sort Key: prt2.b
> -               ->  Append
> +   ->  Merge Right Join
> +         Merge Cond: (prt2.b = prt1.a)
> +         ->  Append
> +               ->  Sort
> +                     Sort Key: prt2_1.b
>                       ->  Seq Scan on prt2_p2 prt2_1
>                             Filter: (b > 250)
> +               ->  Sort
> +                     Sort Key: prt2_2.b
>                       ->  Seq Scan on prt2_p3 prt2_2
>                             Filter: (b > 250)
> -(19 rows)
> +         ->  Materialize
> +               ->  Append
> +                     ->  Sort
> +                           Sort Key: prt1_1.a
> +                           ->  Seq Scan on prt1_p1 prt1_1
> +                                 Filter: ((a < 450) AND (b = 0))
> +                     ->  Sort
> +                           Sort Key: prt1_2.a
> +                           ->  Seq Scan on prt1_p2 prt1_2
> +                                 Filter: ((a < 450) AND (b = 0))
> +(23 rows)
>  
>  SELECT t1.a, t2.b FROM (SELECT * FROM prt1 WHERE a < 450) t1 LEFT JOIN (SELECT * FROM prt2 WHERE b > 250) t2 ON t1.a
=t2.b WHERE t1.b = 0 ORDER BY t1.a, t2.b;
 
>    a  |  b  
> @@ -1350,25 +1354,33 @@ SELECT t1.a, t2.b FROM (SELECT * FROM prt1 WHERE a < 450) t1 LEFT JOIN (SELECT *
>  -- partitionwise join does not apply
>  EXPLAIN (COSTS OFF)
>  SELECT t1.a, t2.b FROM prt1 t1, prt2 t2 WHERE t1::text = t2::text AND t1.a = t2.b ORDER BY t1.a;
> -                                       QUERY PLAN                                        
> ------------------------------------------------------------------------------------------
> +                            QUERY PLAN                            
> +------------------------------------------------------------------
>   Merge Join
> -   Merge Cond: ((t1.a = t2.b) AND (((((t1.*)::prt1))::text) = ((((t2.*)::prt2))::text)))
> -   ->  Sort
> -         Sort Key: t1.a, ((((t1.*)::prt1))::text)
> -         ->  Result
> -               ->  Append
> -                     ->  Seq Scan on prt1_p1 t1_1
> -                     ->  Seq Scan on prt1_p2 t1_2
> -                     ->  Seq Scan on prt1_p3 t1_3
> -   ->  Sort
> -         Sort Key: t2.b, ((((t2.*)::prt2))::text)
> -         ->  Result
> -               ->  Append
> +   Merge Cond: (t1.a = t2.b)
> +   Join Filter: ((((t2.*)::prt2))::text = (((t1.*)::prt1))::text)
> +   ->  Append
> +         ->  Sort
> +               Sort Key: t1_1.a
> +               ->  Seq Scan on prt1_p1 t1_1
> +         ->  Sort
> +               Sort Key: t1_2.a
> +               ->  Seq Scan on prt1_p2 t1_2
> +         ->  Sort
> +               Sort Key: t1_3.a
> +               ->  Seq Scan on prt1_p3 t1_3
> +   ->  Materialize
> +         ->  Append
> +               ->  Sort
> +                     Sort Key: t2_1.b
>                       ->  Seq Scan on prt2_p1 t2_1
> +               ->  Sort
> +                     Sort Key: t2_2.b
>                       ->  Seq Scan on prt2_p2 t2_2
> +               ->  Sort
> +                     Sort Key: t2_3.b
>                       ->  Seq Scan on prt2_p3 t2_3
> -(16 rows)
> +(24 rows)
>  
>  SELECT t1.a, t2.b FROM prt1 t1, prt2 t2 WHERE t1::text = t2::text AND t1.a = t2.b ORDER BY t1.a;
>   a  | b  
> @@ -4906,21 +4918,26 @@ EXPLAIN (COSTS OFF)
>  SELECT t1.*, t2.* FROM alpha t1 INNER JOIN beta t2 ON (t1.a = t2.a AND t1.b = t2.b) WHERE t1.b >= 125 AND t1.b < 225
ORDERBY t1.a, t1.b;
 
>                               QUERY PLAN                             
>  --------------------------------------------------------------------
> - Sort
> + Merge Append
>     Sort Key: t1.a, t1.b
> -   ->  Append
> +   ->  Sort
> +         Sort Key: t1_1.a, t1_1.b
>           ->  Hash Join
>                 Hash Cond: ((t1_1.a = t2_1.a) AND (t1_1.b = t2_1.b))
>                 ->  Seq Scan on alpha_neg_p1 t1_1
>                       Filter: ((b >= 125) AND (b < 225))
>                 ->  Hash
>                       ->  Seq Scan on beta_neg_p1 t2_1
> +   ->  Sort
> +         Sort Key: t1_2.a, t1_2.b
>           ->  Hash Join
>                 Hash Cond: ((t2_2.a = t1_2.a) AND (t2_2.b = t1_2.b))
>                 ->  Seq Scan on beta_neg_p2 t2_2
>                 ->  Hash
>                       ->  Seq Scan on alpha_neg_p2 t1_2
>                             Filter: ((b >= 125) AND (b < 225))
> +   ->  Sort
> +         Sort Key: t1_4.a, t1_4.b
>           ->  Hash Join
>                 Hash Cond: ((t2_4.a = t1_4.a) AND (t2_4.b = t1_4.b))
>                 ->  Append
> @@ -4935,7 +4952,7 @@ SELECT t1.*, t2.* FROM alpha t1 INNER JOIN beta t2 ON (t1.a = t2.a AND t1.b = t2
>                                   Filter: ((b >= 125) AND (b < 225))
>                             ->  Seq Scan on alpha_pos_p3 t1_6
>                                   Filter: ((b >= 125) AND (b < 225))
> -(29 rows)
> +(34 rows)
>  
>  SELECT t1.*, t2.* FROM alpha t1 INNER JOIN beta t2 ON (t1.a = t2.a AND t1.b = t2.b) WHERE t1.b >= 125 AND t1.b < 225
ORDERBY t1.a, t1.b;
 
>   a  |  b  |  c   | a  |  b  |  c   
> diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
> index 0fd0e1c38b3..4c1c173d8e6 100644
> --- a/src/test/regress/expected/union.out
> +++ b/src/test/regress/expected/union.out
> @@ -1195,12 +1195,14 @@ select event_id
>  ----------------------------------------------------------
>   Merge Append
>     Sort Key: events.event_id
> -   ->  Index Scan using events_pkey on events
> +   ->  Sort
> +         Sort Key: events.event_id
> +         ->  Seq Scan on events
>     ->  Sort
>           Sort Key: events_1.event_id
>           ->  Seq Scan on events_child events_1
>     ->  Index Scan using other_events_pkey on other_events
> -(7 rows)
> +(9 rows)
>  
>  drop table events_child, events, other_events;
>  reset enable_indexonlyscan;
> diff --git a/src/test/regress/sql/inherit.sql b/src/test/regress/sql/inherit.sql
> index e3bcfdb181e..5331e49283f 100644
> --- a/src/test/regress/sql/inherit.sql
> +++ b/src/test/regress/sql/inherit.sql
> @@ -586,11 +586,13 @@ insert into matest3 (name) values ('Test 5');
>  insert into matest3 (name) values ('Test 6');
>  
>  set enable_indexscan = off;  -- force use of seqscan/sort, so no merge
> +set enable_sort = off; -- avoid sorting below MergeAppend
>  explain (verbose, costs off) select * from matest0 order by 1-id;
>  select * from matest0 order by 1-id;
>  explain (verbose, costs off) select min(1-id) from matest0;
>  select min(1-id) from matest0;
>  reset enable_indexscan;
> +reset enable_sort;
>  
>  set enable_seqscan = off;  -- plan with fewest seqscans should be merge
>  set enable_parallel_append = off; -- Don't let parallel-append interfere
> @@ -976,6 +978,21 @@ reset enable_bitmapscan;
>  
>  drop table mcrparted;
>  
> +-- Check that sort path can be used by MergeAppend even when there are suitable pathkeys
> +create table hash_parted (i int, j int, k int) partition by hash(i);
> +create table hash_parted_1 partition of hash_parted for values with (modulus 4, remainder 0);
> +create table hash_parted_2 partition of hash_parted for values with (modulus 4, remainder 1);
> +create table hash_parted_3 partition of hash_parted for values with (modulus 4, remainder 2);
> +create table hash_parted_4 partition of hash_parted for values with (modulus 4, remainder 3);
> +--create table hash_parted_5 partition of hash_parted for values with (modulus 6, remainder 4);
> +--create table hash_parted_6 partition of hash_parted for values with (modulus 6, remainder 5);
> +create index on hash_parted(j);
> +create index on hash_parted(k);
> +insert into hash_parted select i, i, i from generate_series(1,10000) i;
> +analyze hash_parted;
> +explain (costs off) select * from hash_parted where k<100 order by j limit 100;
> +drop table hash_parted;
> +
>  -- Ensure LIST partitions allow an Append to be used instead of a MergeAppend
>  create table bool_lp (b bool) partition by list(b);
>  create table bool_lp_true partition of bool_lp for values in(true);
> -- 
> 2.34.1
> 


-- 
  Bruce Momjian  <bruce@momjian.us>        https://momjian.us
  EDB                                      https://enterprisedb.com

  When a patient asks the doctor, "Am I going to die?", he means 
  "Am I going to die soon?"

Bruce Momjian <bruce@momjian.us> writes:

> Is this still being considered?

I'd +1 on this feature. I guess this would be more useful on parallel
case, where the Sort can be pushed down to parallel worker, and in the
distributed database case, where the Sort can be pushed down to multiple
nodes, at the result, the leader just do the merge works.

At the high level implementaion, sorting *cheapest* child path looks
doesn't add too much overhead on the planning effort. 

-- 
Best Regards
Andy Fan

Andy Fan писал(а) 2024-10-17 03:33:
> Bruce Momjian <bruce@momjian.us> writes:
> 
>> Is this still being considered?
> 
> I'd +1 on this feature. I guess this would be more useful on parallel
> case, where the Sort can be pushed down to parallel worker, and in the
> distributed database case, where the Sort can be pushed down to 
> multiple
> nodes, at the result, the leader just do the merge works.
> 
> At the high level implementaion, sorting *cheapest* child path looks
> doesn't add too much overhead on the planning effort.

Hi.

I've updated patch. One more interesting case which we found - when 
fractional path is selected, it still can be more expensive than sorted 
cheapest total path (as we look only on indexes whith necessary 
pathkeys, not on indexes which allow efficiently fetch data).
So far couldn't find artificial example, but we've seen inadequate index 
selection due to this issue - instead of using index suited for filters 
in where, index, suitable for sorting was selected as one having the 
cheapest fractional cost.
-- 
Best regards,
Alexander Pyhalov,
Postgres Professional

On 3/28/25 09:19, Alexander Pyhalov wrote:
> Andy Fan писал(а) 2024-10-17 03:33:
> I've updated patch. One more interesting case which we found - when 
> fractional path is selected, it still can be more expensive than sorted 
> cheapest total path (as we look only on indexes whith necessary 
> pathkeys, not on indexes which allow efficiently fetch data).
> So far couldn't find artificial example, but we've seen inadequate index 
> selection due to this issue - instead of using index suited for filters 
> in where, index, suitable for sorting was selected as one having the 
> cheapest fractional cost.
I think it is necessary to generalise the approach a little.

Each MergeAppend subpath candidate that fits pathkeys should be compared 
to the overall-optimal path + explicit Sort node. Let's label this 
two-node composition as base_path. There are three cases exist: 
startup-optimal, total-optimal and fractional-optimal.
In the startup case, base_path should use cheapest_startup_path, the 
total-optimal case - cheapest_total_path and for a fractional case, we 
may employ the get_cheapest_fractional_path routine to detect proper 
base_path.

It may provide a more effective plan either in full, fractional and 
partial scan cases:
1. The Sort node may be pushed down to subpaths under a parallel or 
async Append.
2. When a minor set of subpaths doesn't have a proper index, and it is 
profitable to sort them instead of switching to plain Append.

In general, analysing the regression tests changed by this code, I see 
that the cost model prefers a series of small sortings instead of a 
single massive one. May be it will show some profit for execution time.

I am not afraid of any palpable planning overhead here because we just 
do cheap cost estimation and comparison operations that don't need 
additional memory allocations. The caller is responsible for building a 
proper Sort node if this method is chosen as optimal.

In the attachment, see the patch written according to the idea. There 
are I introduced two new routines:
get_cheapest_path_for_pathkeys_ext
get_cheapest_fractional_path_for_pathkeys_ext

I have designed the code that way to reduce duplicated code in the 
generate_orderedappend_paths routine. But the main point is that I 
envision these new routines may be reused elsewhere.

This feature looks сlose to the one we discussed before [1]. So, let me 
CC the people from the previous conversation to the discussion.

[1] 
https://www.postgresql.org/message-id/flat/CAN-LCVPxnWB39CUBTgOQ9O7Dd8DrA_tpT1EY3LNVnUuvAX1NjA%40mail.gmail.com

-- 
regards, Andrei Lepikhov

Andrei Lepikhov писал(а) 2025-04-24 16:01:
> On 3/28/25 09:19, Alexander Pyhalov wrote:
>> Andy Fan писал(а) 2024-10-17 03:33:
>> I've updated patch. One more interesting case which we found - when 
>> fractional path is selected, it still can be more expensive than 
>> sorted cheapest total path (as we look only on indexes whith necessary 
>> pathkeys, not on indexes which allow efficiently fetch data).
>> So far couldn't find artificial example, but we've seen inadequate 
>> index selection due to this issue - instead of using index suited for 
>> filters in where, index, suitable for sorting was selected as one 
>> having the cheapest fractional cost.
> I think it is necessary to generalise the approach a little.
> 
> Each MergeAppend subpath candidate that fits pathkeys should be 
> compared to the overall-optimal path + explicit Sort node. Let's label 
> this two-node composition as base_path. There are three cases exist: 
> startup-optimal, total-optimal and fractional-optimal.
> In the startup case, base_path should use cheapest_startup_path, the 
> total-optimal case - cheapest_total_path and for a fractional case, we 
> may employ the get_cheapest_fractional_path routine to detect proper 
> base_path.
> 
> It may provide a more effective plan either in full, fractional and 
> partial scan cases:
> 1. The Sort node may be pushed down to subpaths under a parallel or 
> async Append.
> 2. When a minor set of subpaths doesn't have a proper index, and it is 
> profitable to sort them instead of switching to plain Append.
> 
> In general, analysing the regression tests changed by this code, I see 
> that the cost model prefers a series of small sortings instead of a 
> single massive one. May be it will show some profit for execution time.
> 
> I am not afraid of any palpable planning overhead here because we just 
> do cheap cost estimation and comparison operations that don't need 
> additional memory allocations. The caller is responsible for building a 
> proper Sort node if this method is chosen as optimal.
> 
> In the attachment, see the patch written according to the idea. There 
> are I introduced two new routines:
> get_cheapest_path_for_pathkeys_ext
> get_cheapest_fractional_path_for_pathkeys_ext

Hi. I'm a bit confused that 
get_cheapest_fractional_path_for_pathkeys_ext() looks only on sorting 
cheapest fractional path, and get_cheapest_path_for_pathkeys_ext() in 
STARTUP_COST case looks only on sorting cheapest_startup_path.
Usually, sorted cheapest_total_path will be cheaper than sorted 
fractional/startup path at least by startup cost (as after sorting it 
includes total_cost of input path). But we ignore this case when 
selecting cheapest_startup and cheapest_fractional subpaths. As result 
selected cheapest_startup and cheapest_fractional can be not cheapest 
for startup or selecting a fraction of rows.

Consider the partition with the following access paths:

1) cheapest_startup without required pathkeys:
   startup_cost: 0.42
   total_cost: 4004

2) some index_path  with required pathkeys:
   startup_cost: 6.6
   total_cost: 2000

3) cheapest_total_path:
   startup_cost: 0.42
   total_cost: 3.48

Here, when selecting cheapest startup subpath we'll compare costs of 
index path (2) and sorted cheapest_startup (1), but will ignore sorted 
cheapest_total_path (3), despite the fact that it really can be the 
cheapest startup path, providing required sorting order.

-- 
Best regards,
Alexander Pyhalov,
Postgres Professional

On 4/25/25 11:16, Alexander Pyhalov wrote:
> Andrei Lepikhov писал(а) 2025-04-24 16:01:
>> On 3/28/25 09:19, Alexander Pyhalov wrote:
>> In the attachment, see the patch written according to the idea. There 
>> are I introduced two new routines:
>> get_cheapest_path_for_pathkeys_ext
>> get_cheapest_fractional_path_for_pathkeys_ext
> 
> Hi. I'm a bit confused that 
Thanks for the participation!

> get_cheapest_fractional_path_for_pathkeys_ext() looks only on sorting 
> cheapest fractional path, and get_cheapest_path_for_pathkeys_ext() in 
> STARTUP_COST case looks only on sorting cheapest_startup_path.
At first, at the moment, I don't understand why we calculate the 
cheapest_startup path at all. In my opinion, after commit 6b94e7a [1, 
2], the min-fractional path totally covers the case. I began this 
discussion in [3] - maybe we need to scrutinise that issue beforehand.

Looking into the min-fractional-path + Sort, we propose a path for the 
case when extracting minor part of tuples with sorting later is cheaper 
than doing a massive job of non-selective index scan. You also may 
imagine the case of a JOIN as a subpath: non-sorted, highly selective 
parameterised NestLoop may be way more optimal than MergeJoin, which 
fits the pathkeys.

> Usually, sorted cheapest_total_path will be cheaper than sorted 
> fractional/startup path at least by startup cost (as after sorting it 
> includes total_cost of input path). But we ignore this case when 
> selecting cheapest_startup and cheapest_fractional subpaths. As result 
> selected cheapest_startup and cheapest_fractional can be not cheapest 
> for startup or selecting a fraction of rows.
I don't know what you mean by that. The cheapest_total_path is 
considered when we chose optimal cheapest_total path. The same works for 
the fractional path - get_cheapest_fractional_path gives us the most 
optimal fractional path and probes cheapest_total_path too.
As above, not sure about min-startup case for now. I can imagine 
MergeAppend over sophisticated subquery: non-sorted includes highly 
parameterised JOINs and the alternative (with pathkeys) includes 
HashJoin, drastically increasing startup cost. It is only a theory, of 
course. So, lets discover how min-startup works.

At the end, when the sorted path already estimated, we each time compare 
it with previously selected pathkeys-cheapest. So, if the sorted path is 
worse, we end up with the original path and don't lose anything.

[1] 
https://www.postgresql.org/message-id/e8f9ec90-546d-e948-acce-0525f3e92773%40enterprisedb.com
[2] 
https://www.postgresql.org/message-id/1581042da8044e71ada2d6e3a51bf7bb%40index.de
[3] 
https://www.postgresql.org/message-id/f0206ef2-6b5a-4d07-8770-cfa7cd30f685@gmail.com

-- 
regards, Andrei Lepikhov

On 4/25/25 17:13, Andrei Lepikhov wrote:
> On 4/25/25 11:16, Alexander Pyhalov wrote:
>> Usually, sorted cheapest_total_path will be cheaper than sorted 
>> fractional/startup path at least by startup cost (as after sorting it 
>> includes total_cost of input path). But we ignore this case when 
>> selecting cheapest_startup and cheapest_fractional subpaths. As result 
>> selected cheapest_startup and cheapest_fractional can be not cheapest 
>> for startup or selecting a fraction of rows.
> I don't know what you mean by that. The cheapest_total_path is 
> considered when we chose optimal cheapest_total path. The same works for 
> the fractional path - get_cheapest_fractional_path gives us the most 
> optimal fractional path and probes cheapest_total_path too.
> As above, not sure about min-startup case for now. I can imagine 
> MergeAppend over sophisticated subquery: non-sorted includes highly 
> parameterised JOINs and the alternative (with pathkeys) includes 
> HashJoin, drastically increasing startup cost. It is only a theory, of 
> course. So, lets discover how min-startup works.
After a second thought I have caught your idea. I agree that for a 
fractional path it have no sense to choose any other path except a 
cheapest total one.
There are the modified patch in the attachment.

Also, to be more objective, I propose to use examples in argumentation - 
something like in attached test2.sql script.

-- 
regards, Andrei Lepikhov

Andrei Lepikhov писал(а) 2025-04-29 16:52:
> On 4/25/25 17:13, Andrei Lepikhov wrote:
>> On 4/25/25 11:16, Alexander Pyhalov wrote:
>>> Usually, sorted cheapest_total_path will be cheaper than sorted 
>>> fractional/startup path at least by startup cost (as after sorting it 
>>> includes total_cost of input path). But we ignore this case when 
>>> selecting cheapest_startup and cheapest_fractional subpaths. As 
>>> result selected cheapest_startup and cheapest_fractional can be not 
>>> cheapest for startup or selecting a fraction of rows.
>> I don't know what you mean by that. The cheapest_total_path is 
>> considered when we chose optimal cheapest_total path. The same works 
>> for the fractional path - get_cheapest_fractional_path gives us the 
>> most optimal fractional path and probes cheapest_total_path too.
>> As above, not sure about min-startup case for now. I can imagine 
>> MergeAppend over sophisticated subquery: non-sorted includes highly 
>> parameterised JOINs and the alternative (with pathkeys) includes 
>> HashJoin, drastically increasing startup cost. It is only a theory, of 
>> course. So, lets discover how min-startup works.
> After a second thought I have caught your idea. I agree that for a 
> fractional path it have no sense to choose any other path except a 
> cheapest total one.
> There are the modified patch in the attachment.
> 
> Also, to be more objective, I propose to use examples in argumentation 
> - something like in attached test2.sql script.

Hi.
I've looked through new patch and found minor inconsistencies in 
get_cheapest_path_for_pathkeys_ext() and 
get_cheapest_fractional_path_for_pathkeys_ext().

In get_cheapest_fractional_path_for_pathkeys_ext() we check that 
base_path is not NULL
        path = get_cheapest_fractional_path_for_pathkeys(rel->pathlist, 
pathkeys,
                                                                          
                                required_outer, fraction);

        base_path = rel->cheapest_total_path;

        /* Stop here if the path doesn't satisfy necessary conditions */
        if (!base_path || !bms_is_subset(PATH_REQ_OUTER(base_path), 
required_outer))
                return path;

But it seems, base_path can't be NULL (as add_paths_to_append_rel() is 
called after set_rel_pathlist() for childrels).
However,  path can. Can we do these two functions 
get_cheapest_path_for_pathkeys_ext() and 
get_cheapest_fractional_path_for_pathkeys_ext()
more similar?

Also we check base_path for required_outer and require_parallel_safe, 
but if cheapest path for pathkeys is NULL, these checks are not 
performed. Luckily, they seen to be no-op anyway due to 
cheapest_total->param_info == NULL and function arguments being NULL 
(required_outer) and false (require_parallel_safe). Should we do 
something about this? Don't know, perhaps, remove these misleading 
arguments?

Now, if we return cheapest_total_path from 
get_cheapest_fractional_path_for_pathkeys_ext() if cheapest paths for 
pathkeys don't exist, do the following lines


                                 /*
                                  * If we found no path with matching 
pathkeys, use the
                                  * cheapest total path instead.
                                  *
                                  * XXX We might consider partially 
sorted paths too (with an
                                  * incremental sort on top). But we'd 
have to build all the
                                  * incremental paths, do the costing 
etc.
                                  */
                                 if (!cheapest_fractional)
                                         cheapest_fractional = 
cheapest_total;

become no-op? And we do return non-null path from  
get_cheapest_fractional_path_for_pathkeys_ext(), as it seems we return 
either cheapest_total_path or cheapest fractional path from 
get_cheapest_fractional_path_for_pathkeys_ext().

-- 
Best regards,
Alexander Pyhalov,
Postgres Professional