[Tarantool-patches] [PATCH 1/1] replication: check rs uuid on subscribe process

[Serge Petrenko]

02.06.2021 00:34, Vladislav Shpilevoy пишет:
> Hi! Thanks for the review!

Thanks for the fixes!
>
>>> The UUID ignorance on subscribe decode was introduced here:
>>> https://github.com/tarantool/tarantool/commit/7f8cbde3555084ad6c41f137aec4faba4648c705#diff-fc276b44b551b4eac3431c9433d4bc881790ddd7df76226d7579f80da7798f6e
>>>
>>> And I don't understand why. Maybe I miss something? The tests have
>>> passed. Sergey, do you remember why was it needed really?
>>>
>>> Replicaset UUID mismatch definitely means the node can't connect.
>>> It is not related to whether it is anonymous or not. Because it
>>> has nothing to do with _cluster.
>> Hi! Thanks for the patch!
>>
>> That change was meant to help anonymous replicas fetch data from
>> multiple clusters. There was such an idea back then.
>> It never got implemented though, and I doubt it will.
> Wow, I am not sure how it is supposed to work at all. Starting from
> the problem that in different RS you have conflicting replica_id, and
> you won't be able to keep track of the changes properly. You simply can't
> pack the same replica_id from 2 different replicasets into one replica_id
> in your own journal.
>
> It does not matter if you are anon or not - the changes in the given
> replicasets are generated by non-anon nodes and their replica IDs will
> clash.

I guess you could assign the same replica_id to all the changes coming from
one cluster. Just assign lsns in the order of arrival, or something.
Or you could maintain a vclock per  each cluster, assign the same replica_id
to each change coming from this cluster, and assign vclock_sum as the 
change's
lsn.


>
>> So the idea was that replica should check the replicaset UUID itself.
>> It doesn't work now obviously. And looks like it hadn't worked before
>> you introduced the ER_TOO_EARLY_SUBSCRIBE error.
>> The test for checking uuid on replica side didn't catch this problem
>> because the replica was already registered on master in it.
>>
>> Long story short, I'm ok with this change, but now you should remove
>> unnecessary replicaset uuid checks from replica side (in applier_subscribe).
> If it is not critical for you, I would leave as many checks as possible
> on both sides. It has nothing to do with perf but protects from unexpected
> situations like this one. Let me know if you want it deleted anyway.

Ok, I'm not against it.

>> And looks like replication/gh-3704-misc-replica-checks-cluster-id test is
>> obsolete now.
> Yes, looks like so. I deleted it on the branch.


Sorry, one more nit. I failed to notice this earlier, but there's an 
obsolete
comment in box_process_subscribe().

It goes like
"Tarantool > 2.1.1 master doesn't check that replica has
the same cluster uuid".

LGTM, once you delete the comment.

-- 
Serge Petrenko