Hi,
On 2025-02-11 18:45:13 +0100, Jelte Fennema-Nio wrote:
> On Tue, 11 Feb 2025 at 17:19, Andres Freund <andres@anarazel.de> wrote:
> > Yes, at least initially:
>
> Ah, then I understand your point of view much better. Still I think we
> could easily frame it as: If you enable io_uring, you also get these
> additional fancy stats.
> Also afaict the items don't have to mean that
>
> > 1) it's not enabled on the kernel level everywhere
>
> Is that really a common thing to do? From a quick internet search, I
> can "only" find that Google does this. (Google is definitely a big
> cloud player, so I don't want to suggest that that is not important,
> but if that's really the only one still the bulk of systems would have
> io_uring support)
RHEL had it disabled for quite a while, not sure if that's still the case.
> > 2) it requires an optional build dependency
>
> What build dependency is this?
Liburing.
> In any case, can't we choose the default at build time based on the
> available build dependencies? And if we cannot, I think we could always add
> an "auto" default that would mean the best available AIO implementation
> (where io_uring is better than bgworkers).
We could, but because of 3) I don't want to do that right now.
> > 3) it requires tuning the file descriptor ulimit, unless we can convince Tom
> > that it's ok to do that ourselves
>
> I think we should just do this, given the reasoning in the blog[1]
> from the systemd author I linked in the AIO thread. I agree that a
> response/explicit approval from Tom would be nice though.
I think it's the right path, but that's a fight to fight after AIO has been
merged, not before.
Greetings,
Andres Freund
