Re: create_unique_path and GEQO

On Fri, Mar 24, 2017 at 11:52 AM, Rushabh Lathia
<rushabh(dot)lathia(at)gmail(dot)com> wrote:
>
>
> On Wed, Mar 22, 2017 at 3:43 PM, Ashutosh Bapat
> <ashutosh(dot)bapat(at)enterprisedb(dot)com> wrote:
>>
>> Hi,
>> In create_unique_path() there's comment
>> /*
>> * We must ensure path struct and subsidiary data are allocated in
>> main
>> * planning context; otherwise GEQO memory management causes trouble.
>> */
>> oldcontext = MemoryContextSwitchTo(root->planner_cxt);
>>
>> pathnode = makeNode(UniquePath);
>>
>> This means that when GEQO resets the memory context, the RelOptInfo
>> for which this path is created and may be set to cheapest_unique_path
>> goes away, the unique path lingers on in the planner context.
>> Shouldn't we instead allocate the path in the same context as the
>> RelOptInfo similar to mark_dummy_rel()?
>>
>
> Do you have test case, which can reproduce the issue you
> explained above?
>

No. It would require some surgery in standard_planner() to measure the
memory consumed in the planner context OR build the code with
SHOW_MEMORY_STATS defined and dump memory context statistics and check
planner context memory usage. I don't think I can produce a testcase
quickly right now. But then, I think the problem is quite apparent
from the code inspection alone, esp. comparing what mark_dummy_rel()
does with what create_unique_path() is doing.

--
Best Wishes,
Ashutosh Bapat
EnterpriseDB Corporation
The Postgres Database Company