Thread: BUG #17723: cache lookup failed for type 0

The following bug has been logged on the website:

Bug reference:      17723
Logged by:          Vik Fearing
Email address:      vik@postgresfriends.org
PostgreSQL version: Unsupported/Unknown
Operating system:   Ubuntu
Description:

This query:

WITH RECURSIVE

run(x, y) AS (
  SELECT 0, 0
  UNION ALL
  SELECT x, y FROM run AS r WHERE r.is_cycle
)
CYCLE x, y SET is_cycle USING path

TABLE run
;

in which I mistakenly tried to access the is_cycle column from inside the
wle, provokes the following error:

ERROR:  XX000: cache lookup failed for type 0
LOCATION:  typeOrDomainTypeRelid, parse_type.c:699

Even though I did not need to add that WHERE clause because the CYCLE clause
does it for me, I still should have been able to.  And in any case, I should
not have received an XX000 error.

PG Bug reporting form <noreply@postgresql.org> writes:
> This query:

> WITH RECURSIVE

> run(x, y) AS (
>   SELECT 0, 0
>   UNION ALL
>   SELECT x, y FROM run AS r WHERE r.is_cycle
> )
> CYCLE x, y SET is_cycle USING path

> TABLE run
> ;

> in which I mistakenly tried to access the is_cycle column from inside the
> wle, provokes the following error:

> ERROR:  XX000: cache lookup failed for type 0
> LOCATION:  typeOrDomainTypeRelid, parse_type.c:699

Yeah.  We are calling addRangeTableEntryForCTE inside parse analysis of
the CTE's query, and it's generating a ParseNamespaceItem with p_vartype = 0
because analyzeCTE hasn't yet identified the cycle_mark_type.  That
eventually results in a Var with vartype 0, confusing later parse analysis.

It looks to me like we can just move that part of the code up to before
we recurse to parse_sub_analyze, though.  AFAICS, identification of the
cycle column type needn't (and had better not) depend on any
characteristics of the CTE's query.

            regards, tom lane

On 12/16/22 18:10, Tom Lane wrote:
> PG Bug reporting form <noreply@postgresql.org> writes:
>> This query:
> 
>> WITH RECURSIVE
> 
>> run(x, y) AS (
>>    SELECT 0, 0
>>    UNION ALL
>>    SELECT x, y FROM run AS r WHERE r.is_cycle
>> )
>> CYCLE x, y SET is_cycle USING path
> 
>> TABLE run
>> ;
> 
>> in which I mistakenly tried to access the is_cycle column from inside the
>> wle, provokes the following error:
> 
>> ERROR:  XX000: cache lookup failed for type 0
>> LOCATION:  typeOrDomainTypeRelid, parse_type.c:699
> 
> Yeah.  We are calling addRangeTableEntryForCTE inside parse analysis of
> the CTE's query, and it's generating a ParseNamespaceItem with p_vartype = 0
> because analyzeCTE hasn't yet identified the cycle_mark_type.  That
> eventually results in a Var with vartype 0, confusing later parse analysis.
> 
> It looks to me like we can just move that part of the code up to before
> we recurse to parse_sub_analyze, though.  AFAICS, identification of the
> cycle column type needn't (and had better not) depend on any
> characteristics of the CTE's query.

Indeed, it should only depend on the TO and DEFAULT subclauses (not 
shown here).

Thanks for the fix!
-- 
Vik Fearing