[tarantool-patches] Re: [PATCH 03/10] vinyl: move vylog recovery to vylog thread

[Vladimir Davydov]

On Fri, Jun 07, 2019 at 04:39:54PM +0300, Konstantin Osipov wrote:
> * Vladimir Davydov <vdavydov.dev at gmail.com> [19/06/06 13:24]:
> > > > We used coio, because vylog was written from a WAL thread, which
> > > > shouldn't be used for such a heavy operation as vylog recovery.
> > > > Now, we can move it to the dedicated vylog thread. This allows
> > > > us to simplify rotation logic as well: now most of work is done
> > > > from the same function (vy_log_rotate_f) executed by vylog thread,
> > > > not scattered between coio and WAL, as it used to be.
> > > 
> > > Why do we need to lock out the scheduler while rotating the log in
> > > the first place? 
> > 
> > We rotate vylog by first reading the old vylog and forming a recovery
> > context, then dumping this recovery context to the new vylog. If a new
> > record appears in the old vylog in between, it will be missing in the
> > new vylog. That's why we lock out writers.
> 
> We have two layers of abstractions intermixed here. During
> snapshotting, when we really rotate the vylog, no DDL can happen,
> it's locked out. So no one can take the problematic latch
> anyway.

Except compaction, which isn't locked out by checkpointing.

> So there is, strictly speaking, no problem at all. But
> since we're using a low level latch, and not a centralized
> mechanism to lock out writers, we wouldn't know.
> 
> One option could be to append the writes to vylog which happen
> during checkpointing to the vylog buffer, and not flush them to
> the vylog file which is about-to-become-obsolete.

We must flush those records to disk, otherwise we risk loosing data.

> 
> Anyway, I keep thinking that if you want to kill a latch, there is
> a dozen of ways of killing it, not an own thread.

What's so wrong about the new thread? Could you please give some insight
why we should avoid introducing a separate thread for vylog at all costs
at this point?