Message-ID: <54F119A6.9080106@agliodbs.com>
Date: Fri, 27 Feb 2015 17:28:06 -0800
From: Josh Berkus <josh@agliodbs.com>
Organization: PostgreSQL Experts Inc.
User-Agent: Mozilla/5.0 (X11; Linux x86_64;
 rv:31.0) Gecko/20100101 Thunderbird/31.4.0
MIME-Version: 1.0
To: pgsql-performance@postgresql.org
CC: Aaron Brashears <abrashears@justin.tv>
Subject: Bad cost estimate with FALSE filter condition
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit
Precedence: bulk
Sender: pgsql-performance-owner@postgresql.org

All:

This got posted to pgsql-bugs, but got no attention there[1], so I'm
sending it to this list.

Test case:

createdb bench
pgbench -i -s bench
\c bench

bench=# explain select * from pgbench_accounts where aid = 2;
                                          QUERY PLAN
---------------------------------------------------------------
 Index Scan using pgbench_accounts_pkey on pgbench_accounts
(cost=0.42..8.44 rows=1 width=97)
   Index Cond: (aid = 2)
(2 rows)

bench=# explain select * from pgbench_accounts where aid = 2 and false;

                                QUERY PLAN
-------------------------------------------------
 Result  (cost=0.00..26394.00 rows=1 width=97)
   One-Time Filter: false
   ->  Seq Scan on pgbench_accounts  (cost=0.00..26394.00 rows=1 width=97)
(3 rows)

This seems like a special case of the "aborted plan cost", that is, when
the planner expects to abort a plan early, it nevertheless returns the
full cost for the non-aborted version of the query, rather than the
working cost, which is based on the abort.

For example:

bench=# create index on pgbench_accounts(bid);
CREATE INDEX
bench=# explain select * from pgbench_accounts where bid = 2;
                                                QUERY PLAN

----------------------------------------------------------------------------------------------------------
 Index Scan using pgbench_accounts_bid_idx on pgbench_accounts
(cost=0.42..4612.10 rows=102667 width=97)
   Index Cond: (bid = 2)
(2 rows)

bench=# explain select * from pgbench_accounts where bid = 2 limit 1;
                                   QUERY PLAN

--------------------------------------------------------------------------------
 Limit  (cost=0.00..0.28 rows=1 width=97)
   ->  Seq Scan on pgbench_accounts  (cost=0.00..28894.00 rows=102667
width=97)
         Filter: (bid = 2)
(3 rows)

So in this case, the top-level node returns a lower cost because the
planner knows that it will find a row with bid=2 fairly quickly in the
seq scan.  But in the WHERE FALSE example, that scan *is* the top-level
node, so the planner returns a fictitious cost for the whole query.

Or is there something else at work here?

[1]
http://www.postgresql.org/message-id/20150225194953.2546.86836@wrigleys.postgresql.org

-- 
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com


-- 
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance