Re: [bug?] Missed parallel safety checks, and wrong parallel safety

On Tue, Jun 15, 2021 at 7:31 PM Robert Haas <robertmhaas@gmail.com> wrote:
>
> On Tue, Jun 15, 2021 at 7:05 AM Amit Kapila <amit.kapila16@gmail.com> wrote:
> > Okay, but I think if we go with your suggested model where whenever
> > there is a change in parallel-safety of any function, we need to send
> > the new invalidation then I think it won't matter whether the function
> > is associated with index expression, check constraint in the table, or
> > is used in any other way.
>
> No, it will still matter, because I'm proposing that the
> parallel-safety of functions that we only access via operator classes
> will not even be checked.
>

I am not very clear on what exactly you have in your mind in this
regard. I understand that while computing parallel-safety for a rel we
don't need to consider functions that we only access via operator
class but how do we distinguish such functions during Alter Function?
Is there a simple way to deduce that this is an operator class
function so don't register invalidation for it? Shall we check it via
pg_depend?

>
> > We can additionally check the
> > parallel safety of partitions when we are trying to insert into a
> > particular partition and error out if we detect any parallel-unsafe
> > clause and we are in parallel-mode. So, in this case, we won't be
> > completely relying on the users. Users can either change the parallel
> > safe option of the table or remove/change the parallel-unsafe clause
> > after error. The new invalidation message as we are discussing would
> > invalidate the parallel-safety for individual partitions but not the
> > root partition (partitioned table itself). For root partition, we will
> > rely on information specified by the user.
>
> Yeah, that may be the best we can do. Just to be clear, I think we
> would want to check whether the relation is still parallel-safe at the
> start of the operation, but not have a run-time check at each function
> call.
>

Agreed, that is what I also had in mind.

> > I am not sure if we have a simple way to check the parallel safety of
> > partitioned tables. In some way, we need to rely on user either (a) by
> > providing an option to specify whether parallel Inserts (and/or other
> > DMLs) can be performed, or (b) by providing a guc and/or rel option
> > which indicate that we can check the parallel-safety of all the
> > partitions. Yet another option that I don't like could be to
> > parallelize inserts on non-partitioned tables.
>
> If we figure out a way to check the partitions automatically that
> actually works, we don't need a switch for it; we can (and should)
> just do it that way all the time. But if we can't come up with a
> correct algorithm for that, then we'll need to add some kind of option
> where the user declares whether it's OK.
>

Yeah, so let us think for some more time and see if we can come up
with something better for partitions, otherwise, we can sort out
things further in this direction.

-- 
With Regards,
Amit Kapila.