pg.ddx.io  pgsql-general@postgresql.org mailing list archive  
help / color / mirror / Atom feed
From: Anna B. <terzi@bk.ru>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: pgsql-general@lists.postgresql.org
Subject:  Re[2]: Index-only scan not working when IN clause has 2 or more values
Date: Mon, 28 Nov 2022 16:09:08 +0300
Message-ID: <1669640948.115398878@f420.i.mail.ru> (raw)
In-Reply-To: <4056383.1669390835@sss.pgh.pa.us>
References: <1669301102.998685549@f535.i.mail.ru>
	<4056383.1669390835@sss.pgh.pa.us>


Hi Tom and community,
 
Thank you very much! 
After digging how Postgres planner uses statistics, I have increased table statistics from 100 to 1000. It was enough for planner to use multiple scans of the index and then sort!
(Also I have added dependency extended stats on the three columns as you suggested).
 
Can I ask one more question. I am also testing same index but covering version:
 
create index "ix-transaction-client-trans_dttm-include-division"
    on transaction (client_id,
                    trans_dttm desc)
    include (division_code);
 
Why tuned statistics does not improved it?
 
EXPLAIN (ANALYZE, BUFFERS)
select *
from transaction
where client_id = 123456
  and (trans_dttm between TO_DATE('01.01.2020', 'dd.mm.yyyy') and TO_DATE('31.12.2022', 'dd.mm.yyyy'))
  and (division_code in
       ('not_existing_code1', 'not_existing_code2'))
order by trans_dttm desc
 
"Index Scan using ""ix-transaction-client-trans_dttm-include-division"" on transaction  (cost=0.57..8243559.04 rows=240 width=921) (actual time=23920.988..23920.989 rows=0 loops=1)"
"  Index Cond: ((client_id = '123456'::numeric) AND (trans_dttm >= to_date('01.01.2020'::text, 'dd.mm.yyyy'::text)) AND (trans_dttm <= to_date('31.12.2022'::text, 'dd.mm.yyyy'::text)))"
"  Filter: ((division_code)::text = ANY ('{not_existing_code1,not_existing_code2}'::text[]))"
  Rows Removed by Filter: 10000000
  Buffers: shared hit=8021895 read=2038341
  I/O Timings: read=8902.706
Planning Time: 1.278 ms
Execution Time: 23921.026 ms
 
Yes, I have read about covering indexes in Postgres, about why it has to check rows visibility. But do not understand why Postgres prefers to filter 10000000 table rows instead of filtering in index + using visibility map.
Btw, visibility map is up to date:
relpages, reltuples, relallvisible
23478634, 210520464, 23478634
 
Thank you in advance,
Dmitry
>Пятница, 25 ноября 2022, 18:40 +03:00 от Tom Lane <tgl@sss.pgh.pa.us>:
> 
>=?UTF-8?B?QW5uYSBCLg==?= < terzi@bk.ru > writes:
>> create index "ix-transaction-client-trans_dttm-division"
>> on transaction (client_id,
>> trans_dttm desc,
>> division_code);
>>
>> EXPLAIN (ANALYZE, BUFFERS)
>> select *
>> from transaction
>> where client_id = 123456
>> and (trans_dttm between TO_DATE('01.01.2020', 'dd.mm.yyyy') and TO_DATE('31.12.2022', 'dd.mm.yyyy'))
>> and (division_code in
>> ('not_existing_code1', 'not_existing_code2'))
>> order by trans_dttm desc
>> limit 50 offset 0;
>
>The reason you get a plan like this:
>
>> " -> Index Scan using ""ix-transaction-client-trans_dttm-division"" on transaction (cost=0.57..8350814.66 rows=28072 width=2675) (actual time=703291.834..703291.835 rows=0 loops=1)"
>> " Index Cond: ((client_id = '123456'::numeric) AND (trans_dttm >= to_date('01.01.2020'::text, 'dd.mm.yyyy'::text)) AND (trans_dttm <= to_date('31.12.2022'::text, 'dd.mm.yyyy'::text)))"
>> " Filter: ((division_code)::text = ANY ('{not_existing_code1,not_existing_code2}'::text[]))"
>
>is that if the =ANY clause were an index condition, it would result
>in multiple scans of the index, therefore the output would (in all
>probability) not be sorted in index order. To produce the demanded
>result, the plan would have to read the entire index scan and sort
>its output. The planner estimates that that would be slower than
>what it has done here. In practice it looks like you're reading
>the whole scan output anyway because there are less than 50
>matching rows, but the planner didn't know that.
>
>The problem with =ANY producing unordered output can be dodged if
>the =ANY is on the first index column; but I suppose that does not
>help you here, since making division_code the first index column
>would defeat getting output that's sorted by trans_dttm anyway.
>
>You might try making extended stats on these three columns to see
>if that helps the planner to get a better rowcount estimate.
>If it understood that there were fewer than 50 matching rows,
>it might opt for the use-the-=ANY-and-sort plan type.
>
>regards, tom lane 
 
 
 
 

Message-ID: <1669640948.115398878@f420.i.mail.ru>
Permalink:  ../1669640948.115398878@f420.i.mail.ru/
Also on:    postgresql.org/message-id/1669640948.115398878@f420.i.mail.ru

 · 

reply

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Reply to all the recipients from your mail client; the link fills in
  the headers below: reply via email
  Some clients, webmail especially, drop In-Reply-To and References from
  a link, and the reply then starts a new thread: use the mbox below.

  To: pgsql-general@postgresql.org
  Cc: terzi@bk.ru, tgl@sss.pgh.pa.us
  Subject: Re: Index-only scan not working when IN clause has 2 or more values
  In-Reply-To: <1669640948.115398878@f420.i.mail.ru>
  References: <1669301102.998685549@f535.i.mail.ru> <4056383.1669390835@sss.pgh.pa.us> <1669640948.115398878@f420.i.mail.ru>

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox