[tarantool-patches] Re: [PATCH rfc] schema: add possibility to find and throw away dead replicas

[Vladimir Davydov]

On Thu, Sep 27, 2018 at 01:27:23PM +0300, Olga Arkhangelskaia wrote:
> 
> 
> 27/09/2018 12:04, Vladimir Davydov пишет:
> > On Thu, Sep 27, 2018 at 10:37:57AM +0300, Olga Arkhangelskaia wrote:
> > > 
> > > 26/09/2018 17:46, Vladimir Davydov пишет:
> > > > On Fri, Sep 21, 2018 at 09:25:03PM +0300, Olga Arkhangelskaia wrote:
> > > > > Adds possibility to get list of alive replicas in a replicaset,
> > > > > prune from box.space_cluster those who is not considered as alive,
> > > > > and if one has doubts see state of replicaset.
> > > > > 
> > > > > Replica is considered alive if it is just added, its status after
> > > > > timeout period is not stopped or disconnected. However it it has both
> > > > > roles (master and replica) we consider such instance dead only if its
> > > > > upstream and downstream status is stopped or disconnected.
> > > > > 
> > > > > If replica is considered dead we can prune its uuid from _cluster space.
> > > > > If one not sure if the replica is dead or is there is any activity on it
> > > > > it is possible to list replicas with its role, status and lsn
> > > > > statistics.
> > > > > 
> > > > > If you have some ideas how else we can/should decide whether replica is dead
> > > > > please share.
> > > > > 
> > > > > Closes #3110
> > > > > ---
> > > > > 
> > > > > https://github.com/tarantool/tarantool/issues/3110
> > > > > https://github.com/tarantool/tarantool/tree/OKriw/gh-3110-prune-dead-replica-from-replicaset-1.10
> > > > A documentation request with the new API description is missing.
> > > > Tests don't pass on Travis CI.
> > > > 
> > > > Regarding the code:
> > > > 
> > > >    1. Why do you add a function that lists *alive* replicas? The issue
> > > >       author didn't ask for that. He asked for a script that would delete
> > > >       dead replicas from the _cluster system space. We might want to add a
> > > >       function that would list *dead* replicas so that he/she could check
> > > >       what replicas would be deleted (aka "dry run"), but it doesn't make
> > > >       sense to list alive replicas.
> > > It is easy to change, but as I understood we need to throw away replica.
> > How does it contradict what I said?
> It does not. Just explain why I did it in such a way.

Talked to Georgy. We agreed that it's worth adding three methods:

  prune_replicas - delete replicas given in arguments
  list_dead_replicas - returns a table with UUIDs of supposedly dead replicas
  prune_dead_replicas - prune_replicas(list_dead_replics())

(Function names might need some polishing)

> > 
> > > >    2. Dead replica detection is utterly ridiculuous: the functions sleeps
> > > >       for the given amount of time and then deletes inactive replicas.
> > > >       As a user, I'd want to have an ability to delete replicas that have
> > > >       been inactive for, say, a day. Does this mean that I have to wait
> > > >       for a whole day before this function completes? Obviously, no.
> > > >       I guess tarantool core should keep track of the time each replica
> > > >       was active
> > > So we need changes in core code? About lastt time of activity, what do you
> > > mean? Lasn change, vclock, status?
> > Real time when the replica was last active.
> I still did not get "active". Is it when writes occure at the specified
> replica?
> And last time it got updates from master?

I think we should maintain two times. When the relay was last active and
when the applier was last active.

> I think I will write you in privat to discuss.
> 
> > 
> > > If replica is dead for long perion of time we can see its status. And as I
> > > undestand we have heartbeat to monitor the connecion, so if there is
> > > problems with it - we see status.
> > Disconnected status only means that the replica is not available
> > right now. We want to delete replicas that haven't been active for
> > the specified amount of time, say a day or even a week.
> 
> Good point. I will try to do.
> > 
> > BTW, forgot to mention: this function should probably be defined in
> > box.ctl.
> 
> Why ctl?

Because ctl namespace is meant to store helper functions.