hadoop-zookeeper-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Henry Robinson <he...@cloudera.com>
Subject Re: ZOOKEEPER-107 - Allow dynamic changes to server cluster membership
Date Mon, 03 May 2010 17:29:28 GMT
Hi Vishal -

That's right - design, not implementation!

I'd encourage you to share a design document once you feel you understand
exactly what's required. This is probably going to be complex patch and
reviewers will need a study guide :)

cheers,
Henry

On 3 May 2010 10:26, Vishal Kher <vishalmlst@gmail.com> wrote:

> Hi Henry,
>
> Thanks for the info. I will spend some more time to understand the issues
> before starting with the implementation. I will let you know if I have any
> questions (which I am sure I will).
>
> Just to clarify, "by solved issue" you mean from design perspective and not
> from implementation right?
> Regards,
> -Vishal
> On Mon, May 3, 2010 at 1:16 PM, Henry Robinson <henry@cloudera.com> wrote:
>
> > Hi Vishal -
> >
> > Great that you're interested in contributing! This would be a really neat
> > feature to get into ZK.
> >
> > The documentation that exists is essentially all on the JIRA. I had a
> patch
> > that 'worked' but was nowhere near commit-ready. I'm trying to dig it up,
> > but it appears it may have gone to the great bit-bucket in the sky. Trunk
> > has moved sufficiently that a new patch would be required anyhow.
> >
> > There were two main difficulties with this issue. The first is changing
> the
> > voting protocol to cope with changes in views. Since proposals are
> > pipelined, the leader needs to keep track of what the view was that
> should
> > vote for a proposal. IIRC, the other subtlety is making sure that when a
> > view change is proposed, a quorum of votes is received from both the
> > outgoing view and the incoming one. Otherwise it's possible to transition
> > to
> > a 'dead' view in which no progress can be made.
> >
> > The second is to figure out the metadata management - how do we 'find'
> > ZooKeeper servers if the ensemble may have moved onto a completely
> separate
> > set of machines? That is, if the original ensemble was on A, B, C and the
> > current ensemble is D, E, F - where do we look to find where the ensemble
> > is
> > located?
> >
> > The first is a solved issue, the second is more a matter of taste than
> > designing distributed protocols.
> >
> > Really happy to help with this issue - I'd love to see it get
> resurrected.
> >
> > cheers,
> > Henry
> >
> > On 3 May 2010 07:25, Vishal K <vishalmlst@gmail.com> wrote:
> >
> > > Hi Henry,
> > >
> > > I just commented on the Jira. I would be happy to contribute.
> > > Please advise on the current status and next steps. Thanks.
> > >
> > > Regards,
> > > -Vishal
> > >
> >
> >
> >
> > --
> > Henry Robinson
> > Software Engineer
> > Cloudera
> > 415-994-6679
> >
>



-- 
Henry Robinson
Software Engineer
Cloudera
415-994-6679

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message