Received: from malur.postgresql.org ([217.196.149.56]) by arkaria.postgresql.org with esmtp (Exim 4.80) (envelope-from ) id 1WAPDJ-0005Oe-7N for pgsql-performance@arkaria.postgresql.org; Mon, 03 Feb 2014 19:29:49 +0000 Received: from localhost ([127.0.0.1] helo=postgresql.org) by malur.postgresql.org with smtp (Exim 4.80) (envelope-from ) id 1WAPDI-0006L3-Lx for pgsql-performance@arkaria.postgresql.org; Mon, 03 Feb 2014 19:29:48 +0000 Received: from magus.postgresql.org ([2a02:c0:301:0:ffff::29]) by malur.postgresql.org with esmtp (Exim 4.80) (envelope-from ) id 1WAPDH-0006Kq-Eg for pgsql-performance@postgresql.org; Mon, 03 Feb 2014 19:29:47 +0000 Received: from v124499.home.net.pl ([188.128.182.39]) by magus.postgresql.org with smtp (Exim 4.80) (envelope-from ) id 1WAPDD-0005AR-UN for pgsql-performance@postgresql.org; Mon, 03 Feb 2014 19:29:47 +0000 Received: from 5-226-122-185.ip.netia.com.pl [5.226.122.185] (HELO ArturW7) by serwer1372326.home.pl [188.128.182.39] with SMTP (IdeaSmtpServer v0.80) id 54f62251b01ed7d8; Mon, 3 Feb 2014 20:29:43 +0100 From: =?iso-8859-2?Q?Artur_Zaj=B1c?= To: Subject: Planner estimates and VACUUM/autovacuum Date: Mon, 3 Feb 2014 20:29:34 +0100 Message-ID: <036301cf2116$44b73c10$ce25b430$@ang.com.pl> MIME-Version: 1.0 Content-Type: multipart/alternative; boundary="----=_NextPart_000_0364_01CF211E.A67E3C20" X-Mailer: Microsoft Outlook 15.0 Thread-Index: Ac8hFjwSIpfI4rp6Sc6t8aceTEB6sg== Content-Language: pl X-Pg-Spam-Score: -1.9 (-) List-Archive: List-Help: List-ID: List-Owner: List-Post: List-Subscribe: List-Unsubscribe: X-Mailing-List: pgsql-performance Precedence: bulk Sender: pgsql-performance-owner@postgresql.org This is a multipart message in MIME format. ------=_NextPart_000_0364_01CF211E.A67E3C20 Content-Type: text/plain; charset="iso-8859-2" Content-Transfer-Encoding: 7bit Hi, I am not sure if it is bug or not but I found some strange behaviour. Maybe it is the same as described on http://www.postgresql.org/message-id/14616.1244317490@sss.pgh.pa.us ?). If yes - I'm sorry for the trouble, but I think that my example is more obvious. Tested on PostgreSQL 9.2.4 and 9.2.6. Console 1: BEGIN; DECLARE a CURSOR FOR SELECT * FROM tab; --- Keep cursor open for disallow full vacuum of tab Console 2: SELECT count(*) FROM tab; ---- Result: 3588; select reltuples from pg_class where relname='table'; --- Result: 3588 UPDATE tab SET id=id; UPDATE tab SET id=id; UPDATE tab SET id=id; VACUUM ANALYZE tab; select reltuples from pg_class where relname='table'; --- Result: 3588 Now wait few seconds :) select reltuples from pg_class where relname='table'; --- Result: 12560 VACUUM ANALYZE tab; select reltuples from pg_class where relname='table'; --- Result: 3588 There is 3588 live records and 12560 live+dead records in table. That is strange for me. VACUUM updates pg_class.reltuples differently (only live roiws count) than autovacuum (live and dead rows). Why? Also in planning: explain SELECT id FROM tab; QUERY PLAN ---------------------------------------------------------------------- Seq Scan on tab (cost=0.00..1074.60 rows=12560 width=4) Estimation is done with the use of current pg_class.reltuples value. This value includes dead rows count after autovacuum so estimation is bad, especially in more complex planner tree, for example: Explain SELECT a.id FROM tab AS a JOIN tab AS b USING (id); QUERY PLAN ---------------------------------------------------------------------------- ----------------------------- Nested Loop (cost=0.00..6410.70 rows=12560 width=4) -> Seq Scan on tab a (cost=0.00..1074.60 rows=12560 width=8) -> Index Only Scan using tab_pkey on tab b (cost=0.00..0.41 rows=1 width=4) Index Cond: (id = a.id) PostgreSQL estimates 12560 records in query result. This is wrong estimation if dead tuples are removed during seq scan or index scan (I suppose that it is). I don't think that AUTOVACUUM and VACUUM ANALYZE should behave differently :( -------------------------------------------------------------------------- Artur Zajac ------=_NextPart_000_0364_01CF211E.A67E3C20 Content-Type: text/html; charset="iso-8859-2" Content-Transfer-Encoding: quoted-printable

Hi,

 

I am not sure if it is bug or not = but I found some strange behaviour. Maybe it is the same as described on = http://www.postgresql.org/message-id/14616.1244317490@sss.pgh.pa.us ?). If yes – I’m sorry for the trouble, but I think that = my example is more obvious.

 

Tested on PostgreSQL 9.2.4 and = 9.2.6.

 

Console 1:

BEGIN;

DECLARE a CURSOR FOR SELECT * FROM = tab;

--- = Keep cursor open for disallow full vacuum of tab

 

Console 2:

SELECT count(*) FROM = tab;    

---- Result: = 3588;

select reltuples from pg_class where = relname=3D'table’; 

--- Result: = 3588

UPDATE = tab SET id=3Did;

UPDATE tab SET id=3Did;

UPDATE tab SET = id=3Did;

VACUUM ANALYZE tab;

select reltuples from pg_class = where relname=3D'table’;

--- Result: = 3588

 

Now wait few seconds J

 

select reltuples from pg_class where = relname=3D'table’;

--- Result: = 12560

 

VACUUM ANALYZE tab;

select reltuples from pg_class = where relname=3D'table’;

--- Result: = 3588

 

There is 3588 live records and 12560 live+dead records in = table.

That = is strange for me. VACUUM updates pg_class.reltuples differently (only = live roiws count) than autovacuum (live and dead rows). = Why?

 

Also in planning:

 

explain SELECT id FROM = tab;

          =             &= nbsp;       QUERY = PLAN

------------------------------------------------------------= ----------

Seq Scan on tab  (cost=3D0.00..1074.60 rows=3D12560 = width=3D4)

 

Estimation is done with the use of current = pg_class.reltuples value.  This value includes dead rows count = after autovacuum so estimation is bad, especially in more complex = planner tree, for example:

 

Explain SELECT a.id FROM tab AS a = JOIN tab AS b USING (id);

 

          =             &= nbsp;           &n= bsp;            = QUERY PLAN

------------------------------------------------------------= ---------------------------------------------

Nested Loop  = (cost=3D0.00..6410.70 rows=3D12560 width=3D4)

   ->  Seq Scan = on tab a  (cost=3D0.00..1074.60 rows=3D12560 = width=3D8)

   ->  Index Only Scan using tab_pkey on = tab b  (cost=3D0.00..0.41 rows=3D1 = width=3D4)

         Index = Cond: (id =3D a.id)

 

PostgreSQL estimates 12560 records in query result. This is = wrong estimation if dead tuples are removed during seq scan or index = scan (I suppose that it is).

 

I don’t think that AUTOVACUUM = and VACUUM ANALYZE should behave differently L

 

---------------------------------------= -----------------------------------

Artur Zajac

 

 

------=_NextPart_000_0364_01CF211E.A67E3C20--