zookeeper-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "bit1129@163.com" <bit1...@163.com>
Subject Re: Re: How does Zookeeper handle this situation
Date Mon, 10 Nov 2014 11:08:38 GMT
Thanks Eason, this makes sense, but the answer leads me for another question.

Say, A is the leader, BCDE are the followers, consider the following scenario:

1.  A sends a proposal  to BCDE for a update. 
2.  BCDE ack this proposal. 
3.  A send the COMMIT request to BCDE.
4.  B commit the request, but CDE dont' because of power off before B commits.

At this point, A and B are alive but there is no majority any more, so the zookeeper service
is actually down. The user notices this and first shut down A and B, and then restart CDEAB
one by one.

Because CDE don't have the latest update commit, is it the case that when A and B startup
,their last committed update will be discarded?

From: Eason Liao
Date: 2014-11-10 18:41
To: user
Subject: Re: How does Zookeeper handle this situation
It's possible that D, E have more recent data, and the data of D, E will be
discarded if they join leader. This is fine as long as the data is not
We consider the transaction is committed only if the transaction is
persisted in the log of majority servers. In this case, if D, E have the
committed data, then at least
one of A, B, C will also have it. So the committed data won't get lost.
Hope this answers your question.
2014-11-10 2:36 GMT-08:00 bit1129@163.com <bit1129@163.com>:
> Hi zookeepers,
> I got a question about leader election during ensemble starts up.   Say, I
> got five servers, ABCDE, and start them one by one.
> According to the election algorithm, ABC forms the majority, and then a
> leader is elected among ABC. When D, E start up, they will follow the
> leader. What if D or E have more recent data among ABCDE?
> D,E's latest data will simply be discard? If this is case, this is a
> critical issue.
> Very likely that I missed something how leader election works during
> startup, any explanation on this? thinks.
> bit1129@163.com
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message