It seems to be the best solution, solving the problem of ungrounded
wal_begin_checkpoint usage and in the same time providing direct vlock
obtainment. Sent v3 of the patch.

Вторник, 29 октября 2019, 0:10 +03:00 от Konstantin Osipov <kostja.osipov@gmail.com>:

* Vladislav Shpilevoy <v.shpilevoy@tarantool.org> [19/10/29 00:08]:
> It looks wrong, that we do begin_checkpoint() but omit
> commit_checkpoint(). I would call commit too. Probably now
> begin() does not allocate any resources, but it may start
> in future, and we will have a leak here.

This would be wrong, too. Wal uses commit_checkpoint() as a singal
that it can recycle wal logs. This is the code Georgy is moving to
GC (but his patch is not in yet).

One way to fix it would be to add vclock out parameter to
wal_sync().

--
Konstantin Osipov, Moscow, Russia