Horton

Re: POC, WIP: OR-clause support for indexes

Andrei Lepikhov <lepihov@gmail.com>

View thread
From: Andrei Lepikhov <lepihov@gmail.com>
To: Alexander Korotkov <aekorotkov@gmail.com>
Cc: Alena Rybakina <a.rybakina@postgrespro.ru>, Ranier Vilela <ranier.vf@gmail.com>, Alexander Lakhin <exclusion@gmail.com>, jian he <jian.universality@gmail.com>, Robert Haas <robertmhaas@gmail.com>, Nikolay Shaplov <dhyan@nataraj.su>, pgsql-hackers@lists.postgresql.org, pgsql-hackers@postgresql.org, Peter Geoghegan <pg@bowt.ie>, Marcos Pegoraro <marcos@f10.com.br>, teodor@sigaev.ru, Peter Eisentraut <peter@eisentraut.org>
Date: 2025-01-27T08:52:16Z
Lists: pgsql-hackers

Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →

  1. Make group_similar_or_args() reorder clause list as little as possible

  2. Allow usage of match_orclause_to_indexcol() for joins

  3. Skip not SOAP-supported indexes while transforming an OR clause into SAOP

  4. Remove the wrong assertion from match_orclause_to_indexcol()

  5. Teach bitmap path generation about transforming OR-clauses to SAOP's

  6. Transform OR-clauses to SAOP's during index matching

  7. Fix the value of or_to_any_transform_limit in postgresql.conf.sample

  8. Transform OR clauses to ANY expression

  9. MergeAttributes code deduplication

  10. SEARCH and CYCLE clauses

  11. Improve estimation of OR clauses using extended statistics.

  12. Teach btree to handle ScalarArrayOpExpr quals natively.

  13. Revise collation derivation method and expression-tree representation.

  14. Instead of trying to force WHERE clauses into CNF or DNF normal form,

On 1/25/25 12:04, Alexander Korotkov wrote:
> On Wed, Jan 15, 2025 at 10:24 AM Andrei Lepikhov <lepihov@gmail.com> wrote:
>> causes SEGFAULT during index keys evaluation. I haven't dived into it
>> yet, but it seems quite a typical misstep and is not difficult to fix.
> 
> Segfault appears to be caused by a typo.  Patch used parent rinfo
> instead of child rinfo.  Fixed in the attached patch.
Great!
> 
> It appears that your first query also changed a plan after fixing
> this.  Could you, please, provide another example of a regression for
> short-circuit optimization, which is related to this patch?
Yes, it may be caused by the current lazy InitPlan evaluation strategy, 
which would only happen if it was really needed.

Examples:
---------

EXPLAIN (ANALYZE, COSTS OFF, BUFFERS OFF, TIMING OFF)
SELECT * FROM bitmap_split_or t1
WHERE t1.a=2 AND (t1.b=2 OR t1.b = (
    SELECT avg(x) FROM generate_series(1,1e6) AS x)::integer);

without optimisation:

  Index Scan using t_a_b_idx on bitmap_split_or t1 (actual rows=1 loops=1)
    Index Cond: (a = 2)
    Filter: ((b = 2) OR (b = ((InitPlan 1).col1)::integer))
    InitPlan 1
      ->  Aggregate (never executed)
            ->  Function Scan on generate_series x (never executed)
  Planning Time: 0.564 ms
  Execution Time: 0.182 ms

But having it as a part of an array, we forcedly evaluate it for (not 
100% sure) more precise selectivity estimation:

  Index Scan using t_a_b_idx on bitmap_split_or t1
    (actual rows=1 loops=1)
    Index Cond: ((a = 2) AND
                 (b = ANY (ARRAY[2, ((InitPlan 1).col1)::integer])))
    InitPlan 1
      ->  Aggregate (actual rows=1 loops=1)
            ->  Function Scan on generate_series x
                (actual rows=1000000 loops=1)
  Planning Time: 0.927 ms
  Execution Time: 489.933 ms

This also means that if, before the patch, we executed a query 
successfully, after applying the patch, we sometimes may get the error:
'ERROR: more than one row returned by a subquery used as an expression' 
because of early InitPlan evaluation. See the example below:

EXPLAIN (ANALYZE, COSTS OFF)
SELECT * FROM bitmap_split_or t1
WHERE t1.a=2 AND (t1.b=2 OR t1.b = (
    SELECT random() FROM generate_series(1,1e6) AS x)::integer);

  Index Scan using t_a_b_idx on bitmap_split_or t1
    Index Cond: ((a = 2) AND (b = ANY (ARRAY[2, ((InitPlan 
1).col1)::integer])))
    InitPlan 1
      ->  Function Scan on generate_series x

I think optimisation should have never happened and this is another 
issue, isn't it?

-- 
regards, Andrei Lepikhov