On Wed, May 21, 2025, at 3:50 AM, Laurenz Albe wrote:
> On Tue, 2025-05-20 at 16:58 -0400, Scott Mead wrote:
> > On Wed, May 14, 2025, at 4:06 AM, Laurenz Albe wrote:
> > > On Tue, 2025-05-13 at 17:53 -0400, Scott Mead wrote:
> > > > On Tue, May 13, 2025, at 5:07 PM, Greg Sabino Mullane wrote:
> > > > > On Tue, May 13, 2025 at 4:37 PM Scott Mead <sc...@meads.us> wrote:
> > > > > > I'll open by proposing that we prevent the planner from
> > > > > > automatically
> > > > > > selecting parallel plans by default
> > >
> > > > > > What is the fallout? When a high-volume, low-latency query flips to
> > > > > > parallel execution on a busy system, we end up in a situation where
> > > > > > the database is effectively DDOSing itself with a very high rate of
> > > > > > connection establish and tear-down requests.
> > >
> > > You are painting a bleak picture indeed. I get to see PostgreSQL
> > > databases
> > > in trouble regularly, but I have not seen anything like what you describe.
> > >
> > > With an argument like that, you may as well disable nested loop joins.
> > > I have seen enough cases where disabling nested loop joins, without any
> > > deeper analysis, made very slow queries reasonably fast.
> >
> > My argument is that parallel query should not be allowed to be invoked
> > without
> > user intervention. Yes, nestedloop can have a similar impact, but let's
> > take
> > a look at the breakdown at scale of PQ:
> >
> > [pgbench run that shows that parallel query is bad for throughput]
>
> I think that your experiment is somewhat misleading. Sure, if you
> overload the machine with parallel workers, that will eventually also
> harm the query response time. But many databases out there are not
> overloaded, and the shorter response time that parallel query offers
> makes many users happy.
It's not intended to be misleading, sorry for that. I agree that PQ can have a
positive effect, the point is that our current defaults will very quickly take
a basic workload on a modest (16 CPU box) and quickly swamp it with a
concurrency of 5, which is counter-intuitive, hard to debug, and usually not
desired (again, in the case of a plan that silently invokes parallelism).
FWIW, setting max_parallel_workers_per_gather to 0 by default only disables
automatic PQ selection behind a SIGHUP (or with a user context), users can
easily re-enable it if they think want without having to restart (similar to
parallel_setup_cost, but without the uncertainty).
During my testing, I actually found (again, at concurrency = 5) that the
default max_parallel_workers and max_worker_processes of 8 is not high enough.
If the default max_parallel_workers_per_gather is 0, then we'd be able to to
crank those defaults up (especially max_worker_processes which requires a
restart).
>
> It is well known that what is beneficial for response time is detrimental
> for the overall throughput and vice versa.
It is well-known. What's not is that the postgres defaults will quickly swamp
a machine with parallelism. That's a lesson that many only learn after it's
happened to them. ISTM that the better path is to let someone try to optimize
with parallelism rather than have to fight with it during an emergent event.
IOW: I'd rather know that I'm walking into a marsh with rattlesnakes rather
than find out after I'd been bitten.
> Now parallel query clearly is a feature that is good for response time
> and bad for throughput, but that is not necessarily wrong.
Agreed, I do like and use parallel query. I just don't think it's wise that we
allow that planner to make that decision on a user's behalf when the overhead
is this high and the concurrency behavior falls apart so spectacularly fast.
>
> Essentially, you are arguing that the default configuration should favor
> throughput over response time.
That's one take on it, I'm actually saying that the default configuration
should protect medium-sized systems from unintended behavior that quickly
degrades performance while being very hard to identify and quantify.
>
> > Going back to the original commit which enabled PQ by default[1], it was
> > done so that the feature would be tested during beta. I think it's time
> > that we limit the accidental impact this can have to users by disabling
> > the feature by default.
>
> I disagree.
> My experience is that parallel query often improves the user experience.
> Sure, there are cases where I recommend disabling it, but I think that
> disabling it by default would be a move in the wrong direction.
>
> On the other hand, I have also seen cases where bad estimates trigger
> parallel query by mistake, making queries slower. So I'd support an
> effort to increase the default value for "parallel_setup_cost".
I'm open to discussing a value for parallel_setup_cost that protects users from
runaway here, I just haven't been able to find a value that allows users to be
protected while simultaneously allowing users who want automatic parallel-plan
selection to take advantage of it.
What I've found (and it sounds somewhat similar to what you are saying) is that
if you use parallelism intentionally and design for it (hardware, concurrency
model, etc...) it's very, very powerful. In cases where it 'just kicks in', I
haven't seen an example that makes users happy.
>
> Yours,
> Laurenz Albe
>
--
Scott Mead
Amazon Web Services
sc...@meads.us
