Обсуждение: 9.1rc1: TRAP: FailedAssertion("!(item_width > 0)", File: "costsize.c", Line: 3274)

I am developing against 9.1beta3 and encountered this problem. I repeated it against rc1 (or
actually, REL9_1_STABLE taken from git today).

The below SQL causes:

TRAP: FailedAssertion("!(item_width > 0)", File: "costsize.c", Line: 3274)

and the session crashes.


select *
from ( select      col1   ,  col2   ,  reviewed from   (values       ( 1111, 'aweyu', 3)     ,    ( 2222, 'bwetu', 2)
 , ( 3333, 'vwxyz', 1)    )       as r(col1,col2,col3) left join (               select 1 as reviewed, c2 from ( values
('abcde')) as t1(c2)     union all select 0 as reviewed, c2 from ( values ('vwxyz') ) as t2(c2) )    as up_all on
up_all.c2= r.col2 order by col3
 
) as f
;


To simplify I used 'values' for the actual tables that I am using, but the above still causes the
FailedAssertion.

Removing the outer select avoids the error and
removing order by avoids the error


Erik Rijkers

On Mon, Aug 22, 2011 at 4:16 PM, Erik Rijkers <er@xs4all.nl> wrote:
> I am developing against 9.1beta3 and encountered this problem. I repeated it against rc1 (or
> actually, REL9_1_STABLE taken from git today).
>
> The below SQL causes:
>
> TRAP: FailedAssertion("!(item_width > 0)", File: "costsize.c", Line: 3274)
>
> and the session crashes.
>
> select *
> from (
>  select
>       col1
>    ,  col2
>    ,  reviewed
>  from
>    (values
>        ( 1111, 'aweyu', 3)
>      , ( 2222, 'bwetu', 2)
>      , ( 3333, 'vwxyz', 1)
>     )
>        as r(col1,col2,col3)
>  left join (
>                select 1 as reviewed, c2 from ( values ('abcde') ) as t1(c2)
>      union all select 0 as reviewed, c2 from ( values ('vwxyz') ) as t2(c2)
>  )
>     as up_all on up_all.c2 = r.col2
>  order by col3
> )
>  as f
> ;
>
>
> To simplify I used 'values' for the actual tables that I am using, but the above still causes the
> FailedAssertion.
>
> Removing the outer select avoids the error and
> removing order by avoids the error

I can reproduce this on master but not on 9.0.  I suspect the problem
was introduced by this commit:

commit 0f61d4dd1b4f95832dcd81c9688dac56fd6b5687
Author: Tom Lane <tgl@sss.pgh.pa.us>
Date:   Fri Nov 19 17:31:50 2010 -0500
   Improve relation width estimation for subqueries.

...but I haven't confirmed that yet.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Robert Haas <robertmhaas@gmail.com> writes:
> On Mon, Aug 22, 2011 at 4:16 PM, Erik Rijkers <er@xs4all.nl> wrote:
>> The below SQL causes:
>> TRAP: FailedAssertion("!(item_width > 0)", File: "costsize.c", Line: 3274)

> I can reproduce this on master but not on 9.0.  I suspect the problem
> was introduced by this commit:
> commit 0f61d4dd1b4f95832dcd81c9688dac56fd6b5687

Well, that's a pretty safe bet considering the Assert is in code that
didn't exist before that commit ;-).  The question is why the per-column
width estimate hasn't been set for that column.  Looking at the
coredump, I see that the width *did* get set for the subquery's other
output column, which makes it even weirder.  Too tired to dig further
right now though.
        regards, tom lane

"Erik Rijkers" <er@xs4all.nl> writes:
> The below SQL causes:
> TRAP: FailedAssertion("!(item_width > 0)", File: "costsize.c", Line: 3274)

Thanks, I've applied a patch for this.
        regards, tom lane