Обсуждение: BUG #4170: Rows estimation which are cast from TEXT is inaccurate.

The following bug has been logged online:

Bug reference:      4170
Logged by:          Tashuhito Kasahara
Email address:      kasahara.tatsuhito@oss.ntt.co.jp
PostgreSQL version: 8.3.1
Operating system:   Linux
Description:        Rows estimation which are cast from TEXT  is inaccurate.
Details:

I noticed that rows estimation is not accurate when we cast some datetype to
TEXT.
See the following example. (TEXT -> TIMESTAMP)

============================================================================
====
test=# SELECT count(*) FROM test WHERE t < '2008-05-14 23:55:00';
 count
-------
 86099
(1 row)

test=# EXPLAIN SELECT * FROM test WHERE t < '2008-05-14 23:55:00';
                             QUERY PLAN
--------------------------------------------------------------------
 Seq Scan on test  (cost=0.00..1727.00 rows=85721 width=12)
   Filter: (t < '2008-05-14 23:55:00'::timestamp without time zone)
(2 rows)


test=# EXPLAIN SELECT * FROM test WHERE t < '2008-05-14
23:55:00'::text::timestamp;
                                 QUERY PLAN
----------------------------------------------------------------------------

 Seq Scan on test  (cost=0.00..2209.00 rows=32133 width=12)  <- too little
number of the estimates
   Filter: (t < ('2008-05-14 23:55:00'::text)::timestamp without time zone)
(2 rows)

test=# SELECT count(*) FROM test WHERE t < '2008-05-14 23:55:00';
 count
-------
 86099
(1 row)
============================================================================
====

We can avoid this problem by setting appropriate cast-function.

============================================================================
====
CREATE FUNCTION text2timestamp(text) RETURNS timestamp AS
$$
    SELECT timestamp_in(textout($1), 0, 0);
$$
LANGUAGE sql STRICT STABLE;

CREATE CAST (text AS timestamp) WITH FUNCTION text2timestamp(text) AS
ASSIGNMENT;

test=# EXPLAIN SELECT * FROM test WHERE t < '2008-05-14
23:55:00'::text::timestamp;
                               QUERY PLAN
-------------------------------------------------------------------------
 Seq Scan on test  (cost=0.00..1968.00 rows=85721 width=12)
   Filter: (t < timestamp_in('2008-05-14 23:55:00'::cstring, 0::oid, 0))
(2 rows)
============================================================================
====

I think it's a bug and will be troubled at plan optimization.

Best regards.

"Tashuhito Kasahara" <kasahara.tatsuhito@oss.ntt.co.jp> writes:
> test=# EXPLAIN SELECT * FROM test WHERE t < '2008-05-14
> 23:55:00'::text::timestamp;
>                                  QUERY PLAN
> ----------------------------------------------------------------------------

>  Seq Scan on test  (cost=0.00..2209.00 rows=32133 width=12)  <- too little
> number of the estimates
>    Filter: (t < ('2008-05-14 23:55:00'::text)::timestamp without time zone)
> (2 rows)

Hmm ... as of 8.3 this will generate a CoerceViaIO node, and it looks
like I forgot to teach eval_const_expressions how to simplify those.

            regards, tom lane

Hi.

Tom Lane wrote:
> "Tashuhito Kasahara" <kasahara.tatsuhito@oss.ntt.co.jp> writes:
>> test=# EXPLAIN SELECT * FROM test WHERE t < '2008-05-14
>> 23:55:00'::text::timestamp;
>>                                  QUERY PLAN
>> ----------------------------------------------------------------------------
>
>>  Seq Scan on test  (cost=0.00..2209.00 rows=32133 width=12)  <- too little
>> number of the estimates
>>    Filter: (t < ('2008-05-14 23:55:00'::text)::timestamp without time zone)
>> (2 rows)
>
> Hmm ... as of 8.3 this will generate a CoerceViaIO node, and it looks
> like I forgot to teach eval_const_expressions how to simplify those.
Relevant issues ocurred on PostgreSQL versions 7.4 also.

8.2, 8.1 and 8.0 seemed to be accurate estimates on simple test.


============= 7.4
test=# EXPLAIN ANALYZE SELECT * FROM test WHERE id < '2008-5-14 00:01:00'::text::timestamp;
                                             QUERY PLAN
-----------------------------------------------------------------------------------------------------
 Seq Scan on test  (cost=0.00..224.98 rows=3667 width=8) (actual time=0.026..30.987 rows=60 loops=1)
   Filter: (id < ('2008-5-14 00:01:00'::text)::timestamp without time zone)
 Total runtime: 31.074 ms
(3 rows)

============= 8.2
test=# EXPLAIN ANALYZE SELECT * FROM test WHERE id < '2008-5-14 00:01:00'::text::timestamp;
                                            QUERY PLAN
---------------------------------------------------------------------------------------------------
 Seq Scan on test  (cost=0.00..224.98 rows=55 width=8) (actual time=0.043..13.896 rows=60 loops=1)
   Filter: (id < ('2008-5-14 00:01:00'::text)::timestamp without time zone)
 Total runtime: 13.951 ms
(3 rows)

============= 8.1
test=# EXPLAIN ANALYZE SELECT * FROM test WHERE id < '2008-5-14 00:01:00'::text::timestamp;
                                            QUERY PLAN
---------------------------------------------------------------------------------------------------
 Seq Scan on test  (cost=0.00..224.98 rows=60 width=8) (actual time=0.034..15.409 rows=60 loops=1)
   Filter: (id < ('2008-5-14 00:01:00'::text)::timestamp without time zone)
 Total runtime: 15.464 ms
(3 rows)

============= 8.0
test=# EXPLAIN ANALYZE SELECT * FROM test WHERE id < '2008-5-14 00:01:00'::text::timestamp;
                                            QUERY PLAN
---------------------------------------------------------------------------------------------------
 Seq Scan on test  (cost=0.00..224.98 rows=63 width=8) (actual time=0.041..18.645 rows=60 loops=1)
   Filter: (id < ('2008-5-14 00:01:00'::text)::timestamp without time zone)
 Total runtime: 18.706 ms
(3 rows)
=============

Best regards.

--
Tatsuhito Kasahara
kasahara.tatsuhito@oss.ntt.co.jp