[Tarantool-patches] [PATCH rfc 0/1] replication: stop resetting existing connections

[Konstantin Osipov]

* Olga Arkhangelskaia <arkholga at tarantool.org> [20/04/02 19:07]:
> 02.04.2020 16:47, Konstantin Osipov пишет:
> > * Olga Arkhangelskaia <arkholga at tarantool.org> [20/04/02 16:30]:
> > 
> > Could you please explain the idea of the patch?
> > 
> > As far I can see, you do some strcmp magic to understand if the
> > uri is the same.
> Yes.
> > 
> > Why not use instance uuids?
> 
> I did not think of uuids in such a way. But  after reading your description
> it
> 
> seems good idea (As I understand  even if we have 3301 and localhost:3301
> they will have the same uuid).
> 
> But I am not sure about appliers that should be removed.
> 
> If there is no UUIDS of existing applier in new appliers pool it should be
> removed?
> 
> And we still need to do all this work even if we just changed order in the
> config.

We could perhaps optimize things based on some heuristics, like
old and new uri match 100%, but first we need to settle on the
main algorithm. 

E.g. one variation could be that, given appliers are ref-counted,
we do not create a new applier for a 100% match, but simply add it
directly to the new uuid -> applier hash. 

Please suggest an edit to this:
> 
> > 0) Introduce uuid -> applier hash.
> > 1) Create new appliers for all entries in replication.
> > 2) Establish connections, but do not issue SUBSCRIBE yet.
> > 3) Look at connection server uuids. If this uuid is already
> >     connected to, then (easy) quietly close the new connection,
> >     and add existing connection to the new instance of uuid hash.
> > 4) Add the new connection to the new instance of uuid hash.
> > If there is a duplicate in the *new* instance of the uuid hash,
> > produce "duplicate connection" error. Otherwise remove/delete all
> > appliers that are left in the old uuid hash.
> > 
> > Perhaps we need applier reference counting to make this work
> > smoothly

-- 
Konstantin Osipov, Moscow, Russia