Message-ID: <bdb4da6e-4432-4cf3-8ec9-622e5fc64d83@gmail.com>
Date: Mon, 11 Nov 2024 16:41:01 +0700
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: Performance of Query 4 on TPC-DS Benchmark
To: Ba Jinsheng <bajinsheng@u.nus.edu>,
 "pgsql-performance@lists.postgresql.org"
 <pgsql-performance@lists.postgresql.org>
References: 
 <SEZPR06MB6494F6A2837995BDD4E0BF9A8A5F2@SEZPR06MB6494.apcprd06.prod.outlook.com>
Content-Language: en-AU
From: Andrei Lepikhov <lepihov@gmail.com>
In-Reply-To: 
 <SEZPR06MB6494F6A2837995BDD4E0BF9A8A5F2@SEZPR06MB6494.apcprd06.prod.outlook.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Archived-At: 
 <https://www.postgresql.org/message-id/bdb4da6e-4432-4cf3-8ec9-622e5fc64d83%40gmail.com>
Precedence: bulk

On 11/11/24 02:35, Ba Jinsheng wrote:
> Hi all,
> 
> Please see this case:
> 
> 
> Query 4 on TPC-DS benchmark:
Thank you for interesting example!
Looking into explains I see two sortings:
->  Sort  (cost=794037.94..794037.95 rows=1 width=132)
    (actual time=3024403.310..3024403.313 rows=8 loops=1)
->  Sort  (cost=794033.93..794033.94 rows=1 width=132)
    (actual time=8068.869..8068.872 rows=8 loops=1)

Almost the same cost and different execution time. So, I think, the core 
of the problem in accuracy of selectivity estimation.
In this specific example I see lots of composite scan filters:
- ((sale_type = 'w'::text) AND (dyear = 2002))
- ((year_total > '0'::numeric) AND (sale_type = 'w'::text) AND (dyear = 
2001))
- ((year_total > '0'::numeric) AND (sale_type = 's'::text) AND (dyear = 
2001))

It is all the time a challenge for PostgreSQL to estimate such a filter 
because of absent information on joint column distribution.
Can you research this way by building extended statistics on these 
clauses? It could move the plan to the more optimal direction.

-- 
regards, Andrei Lepikhov