Re: [HACKERS][BUG] Cache invalidation for queries that contains constof temporary composite type - Mailing list pgsql-hackers

From Maksim Milyutin
Subject Re: [HACKERS][BUG] Cache invalidation for queries that contains constof temporary composite type
Date
Msg-id fd7f8d67-e680-3c4f-d13b-c641dc6599b2@gmail.com
Whole thread Raw
In response to [HACKERS][BUG] Cache invalidation for queries that contains const oftemporary composite type  (Maksim Milyutin <milyutinma@gmail.com>)
Responses Re: [HACKERS][BUG] Cache invalidation for queries that contains constof temporary composite type  (Maksim Milyutin <milyutinma@gmail.com>)
List pgsql-hackers

25.09.17 20:50, Maksim Milyutin wrote:

I have found out the problem when try to sequentially call the function that casts constant to composite type of temporary table that is deleted at each transaction termination (i.e. at each function call completion).


For example, we have the following function 'test':

CREATE OR REPLACE FUNCTION test()
 RETURNS void
 LANGUAGE plpgsql
AS $function$
begin
    create temp table tbl (id int) on commit drop;
    perform json_populate_record(null::tbl, '{ "id": 11 }'::json) as tt;
end;
$function$


Оn the second and subsequent calls we'll observe the following error:

ERROR:  cache lookup failed for type 16392


I investigated the problem and realized that result type of function json_populate_record (funcresulttype field of FuncExpr struct) as well as type of the first null argument (consttype field of Const struct) refer to the invalid composite type related with temporary table 'tbl'. Namely they take a value of oid gotten from the first 'tbl' initialization. The plan of query 'perform json_populate_record(null::tbl, '{ "id": 11 }'::json) as tt' is cached and is not invalidated at each function call. This is because the statement of this query doesn't have any dependency from the 'tbl' relation (relationOids list of CachedPlanSource struct).


Attached patch resolves this problem by adding dependency from relation upon detection Const expression of composite type of that relation


Updated patchset contains more transparent definition of composite type for constant node and regression test for described above buggy case.

-- 
Regards,
Maksim Milyutin
Attachment

pgsql-hackers by date:

Previous
From: Tom Lane
Date:
Subject: Re: [HACKERS] v10 pg_ctl compatibility
Next
From: David Rowley
Date:
Subject: Re: [HACKERS] path toward faster partition pruning