Re: [Tarantool-patches] [PATCH 4/5] vclock: ignore 0th component in comparisons.

["Sergey Petrenko" <sergepetrenko@tarantool.org>]

[-- Attachment #1: Type: text/plain, Size: 2468 bytes --]




>Пятница, 27 декабря 2019, 16:31 +03:00 от Konstantin Osipov <kostja.osipov@gmail.com>:
>
>* Sergey Petrenko < sergepetrenko@tarantool.org > [19/12/27 15:56]:
>> >Четверг, 26 декабря 2019, 8:03 +03:00 от Konstantin Osipov < kostja.osipov@gmail.com >:
>> I couldn't find any code where id 0 is reserved.
>
>It is used in initial join.
Yes, master sends snapshot rows with id 0 and 0 lsn, but this doesn't
interfere with my change, AFAICS.

>
>
>> What do you mean by "the changes of expelled replicas"?
>
>Check the comment in replica_clear_id. Right now when you delete
>replica from _cluster, you keep its slot in vclock. The goal is to
>reuse it.
>
>> However, it's true that vclock comparisons are used in creating snapshots
>> and finding the latest xlog on recovery.
>> So an anonymous replica won't create new snapshots if the only new changes
>> are the one made on the anonymous replica. Some problems with recovery may
>> also exist. I don't know whether it's severe enough, but looks not so good.
>> Thanks for pointing this out!
>> 
>> >
>> >
>> > A much safer bet would be to use a new special id number, like
>> > UINT64_MAX, and not change meaning of  an existing id.
>> 
>> This won't help IMO. We still have cases where this vclock component
>> should be ignored (replication) and cases where it should be taken into
>> account (checkpoint/xlog clock).
>> What about this change? I pushed it to the branch.
>> Also there's no need to fix vclock tests anymore.
>
>It is also a hack. It's best if all of the complexity of an
>anonymous replica resides on it and the master doesn't deal with
>it in any way. Refusing the connection from master with a proper
>error message seems to be simple & reliable way to do it without
>mangling vclock logic.
Both places where vclock_compare ignores 0 component are on
anon replica side. First place is checking whether we are in sync with
master, the second place is finding replicaset leader.

>
>
>The anonymous replica would still have to find  a legal slot in
>vclock for its own changes, but this would be a standard slot. 
If I understand you correctly, this implies some kind of replica id remapping.
Otherwise no one guarantees that a non-anonymous instance with same id
won't be added later. Also we can use anonymous replicas in a cluster with
32 "normal" replicas. Where to find a valid slot then?
>
>
>-- 
>Konstantin Osipov, Moscow, Russia


-- 
Sergey Petrenko

[-- Attachment #2: Type: text/html, Size: 3787 bytes --]