On Tue, 02 Feb 2021 at 07:12, Jeff Janes <jeff.janes@gmail.com> wrote:
> On Mon, Feb 1, 2021 at 6:09 PM Jeff Janes <jeff.janes@gmail.com> wrote:
>
>>
>>
>> create or replace function gibberish(int) returns text language SQL as $_$
>> select left(string_agg(md5(random()::text),$$$$),$1) from
>> generate_series(0,$1/32) $_$;
>>
>> create table j1 as select x, md5(random()::text) as t11, gibberish(1500)
>> as t12 from generate_series(1,20e6) f(x);
>>
>
> I should have added, found it on HEAD, verified it also in 12.5.
>
Here's my analysis:
1) In the explain_ExecutorEnd(), it will create a ExplainState on SQL function
memory context, which is a long-lived, cause the memory grow up.
/*
* Switch to context in which the fcache lives. This ensures that our
* tuplestore etc will have sufficient lifetime. The sub-executor is
* responsible for deleting per-tuple information. (XXX in the case of a
* long-lived FmgrInfo, this policy represents more memory leakage, but
* it's not entirely clear where to keep stuff instead.)
*/
oldcontext = MemoryContextSwitchTo(fcache->fcontext);
2) I try to call pfree() to release ExplainState memory, however, it does not
make sence, I do not know why this does not work? So I try to create it in
queryDesc->estate->es_query_cxt memory context like queryDesc->totaltime, and
it works.
Attached fix the memory leakage in auto_explain. Any thoughts?
--
Regrads,
Japin Li.
ChengDu WenWu Information Technology Co.,Ltd.
