[Tarantool-patches] [PATCH v3] wal: introduce limits on simultaneous writes

Konstantin Osipov kostja.osipov at gmail.com
Sat Mar 13 22:14:45 MSK 2021


* Serge Petrenko <sergepetrenko at tarantool.org> [21/03/12 20:14]:
> > * Vladislav Shpilevoy via Tarantool-patches <tarantool-patches at dev.tarantool.org> [21/03/09 23:44]:
> > > Hi! Thanks for the patch!
> > > 
> > > LGTM.
> > > 
> > > > https://github.com/tarantool/tarantool/issues/5536
> > > > https://github.com/tarantool/tarantool/tree/sp/gh-5536-replica-oom
> > I don't understand why you need two parameters, size and len. What
> > difference does it make how many requests are in the queue? The
> > queue is IO and memory bound, not CPU bound, at least not more cpu
> > bound than any other subsystem.
> > 
> > Pushing the problem to the user by forcing them to choose not one,
> > but two configuration values is poor design. How am I supposed to
> > know what is the right value for the option?
> 
> After some consideration, let's only leave size.
> I've force pushed the changes on the branch.
> 
> I've also implemented some kind of a fiber semaphore in a separate commit on
> top.
> I'm not sure I like how it turned out, so feel free to throw it away. Or
> keep it.

As long as you can benchmark it showing that it is has a
reasonable default (it should be pretty straightforward to
benchmark - send a huge xlog over 10Gb link to tarantool running
on a reasonably fresh SSD and find out optimal limit size) I will
be pretty happy with dropping the size.

-- 
Konstantin Osipov, Moscow, Russia


More information about the Tarantool-patches mailing list