[tarantool-patches] Re: [PATCH v2 1/2] box: added replication_dead/rw_gap options

[Konstantin Osipov <kostja@tarantool.org>]

* Olga Arkhangelskaia <arkholga@tarantool.org> [18/10/23 21:41]:

> >> In scope of gh-3110 we need options that store periods of time,
> >> to be compared with time of last activity of relay and applier.
> >> This patch introduces replication_dead_gap and replication_rw_gap options.
> >>
> >> replication_dead_gap is configured in box.cfg, with default 0 value.
> >> If time that passed from now till last reader/writer activity of given replica
> >> exceeds replication_dead_gap value, replica is suspected to be dead.
> >> replication_dead_gap is measured in hours.
> >>
> >> replication_rw_gap is configured in box.cfg, with default 0 value.
> >> If time difference between last reader activity and last writer activity of
> >> given replica exceeds replication_rw_gap value, replica is suspected to be dead.
> >> replication_rw_gap is measured in hours.
> > Why do we need this if we have heartbeats?
> I used to think that we need some parameters, that can be set by user, 
> to check that replica is not active.
> For example, if replica is not active for XXXX seconds - it is dead. 
> However, I did not think about the idea of passing this parameter as a 
> function argument: list_dead_replicas(XXXX). So I will throw it away.

OK.
> 
> Another question that is worth to discuss - is kind of statistics to use 
> for accusing replica to be dead.
> The is two ways - save time of last write/read by applier and relay. I 
> implemented it, but as Vova pointed out, may be we need to save period 
> of time that replica spends in stopped status. So we decided to do 
> statistics in separate patch set, and implement both way. And than 
> decide. However, may be you have better ideas, etc.

I think unless this statistics is persistent it is of little
value.

> > And with swim on board we will have gossip information about entire replica set?
> I have read about swim, and as I understand it :
> if we have replica set with some topology except full-mesh, we can save 
> dead replicas mask, numbers, etc, (that we obtained using 
> list_dead_replicas on some of replicas), and in the end, after some 
> questioning,  we will definitely  have information about every replica 
> in the set.
> If that what you mean.
> If not, can you be more specific.

We can simply query which replicas are dead according to swim and
correlate this information with relay state. If a replica is dead
according to relay/applier state and it's dead according to swim,
it's dead.

-- 
Konstantin Osipov, Moscow, Russia, +7 903 626 22 32
http://tarantool.io - www.twitter.com/kostja_osipov