Thread: [sqlsmith] ERROR: plan should not reference subplan's variable

Updating master from f8c5855..1bdae16, sqlsmith triggers "failed to
generate plan" errors again.  Below is the smallest query logged so far.

regards,
Andreas

-- ERROR:  plan should not reference subplan's variable
set force_parallel_mode = 'on';
set max_parallel_workers_per_gather = '1';

explain WITH
jennifer_0 AS (select   (select b from public.rtest_v1 limit 1 offset 5)      as c0,
pg_catalog.pg_current_xlog_location()as c1,   sample_0.a as c2,   sample_0.a as c3 from   public.rtest_view4 as
sample_0tablesample system (5.9) where cast(null as bigint) = pg_catalog.hashinet(     cast((select client_addr from
pg_catalog.pg_stat_activitylimit 1 offset 35)        as inet)) limit 76)
 
select   ref_0.sl_name as c0 from   public.shoelace as ref_0 where (cast(null as anyrange) < cast(null as anyrange))
and(EXISTS (     select         39 as c0       from         jennifer_0 as ref_1       where cast(null as real) =
cast(nullas real)       limit 81)) limit 96;
 

Andreas Seltenreich <seltenreich@gmx.de> writes:
> Updating master from f8c5855..1bdae16, sqlsmith triggers "failed to
> generate plan" errors again.  Below is the smallest query logged so far.

Hmm, interesting.  This can be reduced to

set force_parallel_mode = on;

explain
with j1 as (select * from int8_tbl)
select * from int4_tbl where false and EXISTS (select 1 as c0 from j1);

The "plan should not reference subplan's variable" fail seems to be due
to ancient fuzzy thinking in SS_finalize_plan.  When I fix that, I get
a plan like so:
Gather  (cost=0.00..0.00 rows=0 width=0)  Workers Planned: 1  Single Copy: true  ->  Result  (cost=1.05..1.05 rows=0
width=4)       One-Time Filter: false        CTE j1          ->  Seq Scan on int8_tbl  (cost=0.00..1.05 rows=5
width=16)

but if I try to actually execute the query, it crashes at runtime,
apparently because the CTE has not been passed over to the parallel
worker.  Robert, is it expected that CTEs should be parallel-safe?
I'd have thought not.
        regards, tom lane

On Fri, Jul 1, 2016 at 5:29 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Andreas Seltenreich <seltenreich@gmx.de> writes:
>> Updating master from f8c5855..1bdae16, sqlsmith triggers "failed to
>> generate plan" errors again.  Below is the smallest query logged so far.
>
> Hmm, interesting.  This can be reduced to
>
> set force_parallel_mode = on;
>
> explain
> with j1 as (select * from int8_tbl)
> select * from int4_tbl
>   where false and EXISTS (select 1 as c0 from j1);
>
> The "plan should not reference subplan's variable" fail seems to be due
> to ancient fuzzy thinking in SS_finalize_plan.  When I fix that, I get
> a plan like so:
>
>  Gather  (cost=0.00..0.00 rows=0 width=0)
>    Workers Planned: 1
>    Single Copy: true
>    ->  Result  (cost=1.05..1.05 rows=0 width=4)
>          One-Time Filter: false
>          CTE j1
>            ->  Seq Scan on int8_tbl  (cost=0.00..1.05 rows=5 width=16)
>
> but if I try to actually execute the query, it crashes at runtime,
> apparently because the CTE has not been passed over to the parallel
> worker.  Robert, is it expected that CTEs should be parallel-safe?
> I'd have thought not.

Not.  See the RTE_CTE case in set_rel_consider_parallel.

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

Robert Haas <robertmhaas@gmail.com> writes:
> On Fri, Jul 1, 2016 at 5:29 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> ... but if I try to actually execute the query, it crashes at runtime,
>> apparently because the CTE has not been passed over to the parallel
>> worker.  Robert, is it expected that CTEs should be parallel-safe?
>> I'd have thought not.

> Not.  See the RTE_CTE case in set_rel_consider_parallel.

OK; but this is crashing anyway, because after const-simplification the
CTE isn't actually referenced anywhere.  But it still ends up as an
initplan attached to the top plan node.  This may have accidentally
failed to fail before you removed this bit in 5ce5e4a12:
   if (glob->subplans != NULL)       glob->wholePlanParallelSafe = false;

I think in the current code, the equivalent thing is that we need to
consider that a plan node is parallel-unsafe once we've stuck some
initPlans on it (which happens only in createplan.c and so doesn't affect
the parallel_safe state of the source Path).  Will see if that helps.
        regards, tom lane