Re: Curing plpgsql's memory leaks for statement-lifespan values

On Sun, Jul 24, 2016 at 9:17 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Amit Kapila <amit.kapila16@gmail.com> writes:
>
>> Wouldn't it be better, if each nested sub-block (which is having an
>> exception) has a separate "SPI Proc", "SPI Exec" contexts which would
>> be destroyed at sub-block end (either commit or rollback)?
>
> AFAICS that would just confuse matters.  In the first place, plpgsql
> variable values are not subtransaction-local, so they'd have to live in
> the outermost SPI Proc context anyway.  In the second place, spi.c
> contains a whole lot of assumptions that actions like saving a plan are
> tied to the current SPI Proc/Exec contexts, so SPI-using plpgsql
> statements that were nested inside a BEGIN/EXCEPT would probably break:
> state they expect to remain valid from one execution to the next would
> disappear.
>

I think for all such usage, we can always switch to top level SPI
Proc/Exec context. To do so, we might need to invent a notion of
something like Sub SPI Proc/Exec contexts and that sounds quite
tricky.

>> In short, why do you think it is better to create a new context rather
>> than using "SPI Exec"?
>
> "SPI Exec" has the same problem as the eval_econtext: there are already
> points at which it will be reset, and those can't necessarily be delayed
> till end of statement.  In particular, _SPI_end_call will delete whatever
> is in that context.
>

That's right and I think it might not be even feasible to do so,
mainly because that is done in exposed SPI calls.  I have checked some
other usage of SPI, it seems plperl is handling some similar problems
either via creating temporary work context or by using
PG_TRY/PG_CATCH.  I think it might be better if, whatever we are
trying here could be of help for other similar usage.


-- 
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com