[Tarantool-patches] [RFC 4/4] qsync: allow to specify replication_synchro_quorum as a formula

[Serge Petrenko]

20.11.2020 15:01, Cyrill Gorcunov пишет:
> On Fri, Nov 20, 2020 at 01:50:50PM +0300, Serge Petrenko wrote:
>>> @@ -4216,6 +4217,7 @@ unregister_replica(struct trigger *trigger, void * /* event */)
>>>    	struct replica *replica = replica_by_uuid(&old_uuid);
>>>    	assert(replica != NULL);
>>>    	replica_clear_id(replica);
>>> +	replication_on_cluster_update();
>>>    	return 0;
>>>    }
>> We usually perform all the work related to replica register/unregister
>> directly in replica_set_id
>> and replica_clear_id.
>>
>> Besides, these are the places where replicaset.registered_count is updated,
>> so it'd be nice to call replication_on_cluster_update from there.
> Sure.
>
>>> --- a/src/box/box.cc
>>> +++ b/src/box/box.cc
>>> @@ -634,14 +634,6 @@ box_check_replication_synchro_quorum(void)
>>>    		 */
>>>    		int value = replication_synchro_quorum;
>>>    		quorum = eval_replication_synchro_quorum(value);
>>> Pass replicaset.registered_count instead of replication_synchro_quorum here.
> Wait, this is bootstrap, replicaset.registered_count is 0 at this moment, no?

Hm, I didn't think of this.

Yes, this is  either bootstrap or reconfiguration.
Well, if we have a max(1, value)  guard nothing bad should happen even if we
pass 0 on bootstrap. The formula will be reevaluated each time
a replica is registered anyway. Even when the node registers itself.

Anyway, passing replication_synchro_quorum as parameter to evaluate quorum
is even more random.


>
>>>    }
>>> +/**
>>> + * Evaluate the new synchro quorum number when replica
>>> + * get registered/unregistered and the quorum depends on
>>> + * their amount via formula in config.
>>> + */
>>> +void
>>> +replication_on_cluster_update(void)
>>> +{
>>> +	if (!replication_synchro_quorum_eval)
>>> +		return;
>>> +
>>> +	/*
>>> +	 * Account only registered replicas when evaluating
>>> +	 * quorum number from a fromula present in config.
>>> +	 */
>>> +	int value = replicaset.registered_count - replicaset.anon_count;
>>
>> registered_count stands for 'normal' replica count, so no need to subtract
>> anon_count from it.
> Yeah, thanks!
>
>>> +
>>> +	say_info("replication: evaluated quorum is %d", quorum);
>>
>> quorum -> replication_synchro_quorum
> ok

-- 
Serge Petrenko