lucene-solr-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aman Tandon <amantandon...@gmail.com>
Subject Re: Help needed to understand zookeeper in solrcloud
Date Sat, 07 Mar 2015 15:02:20 GMT
Thanks svante for clearing the doubt.

With Regards
Aman Tandon

On Thu, Mar 5, 2015 at 2:15 PM, svante karlsson <saka@csi.se> wrote:

> The network will "only" split if you get errors on your network hardware.
> (or fiddle with iptables) Let's say you placed your zookeepers in separate
> racks and someone pulls network cable between them - that will leave you
> with 5 working servers but they can't reach each other. This is "split
> brain scenario".
>
> >Are they guaranteed to split 4/0
> Yes. A node failure will not partition the network.
>
> > any odd number - it could be 21 even
> Since all write a synchronous you don't want to use a too large number of
> zookeepers since that would slow down the cluster. Use a reasonable number
> to reach your SLA. (3 or 5 are common choices)
>
> >and from a single failure you drop to an even number - then there is the
> danger of NOT getting quorum.
> No, se above.
>
> BUT, if you first lose most of  your nodes due to a network partition and
> then lose another due to node failure - then you are out of quorum.
>
>
> /svante
>
>
>
> 2015-03-05 9:29 GMT+01:00 Julian Perry <jules@limitless.co.uk>:
>
> >
> > I start out with 5 zk's.  All good.
> >
> > One zk fails - I'm left with four.  Are they guaranteed
> > to split 4/0 or 3/1 - because if they split 2/2 I'm screwed,
> > right?
> >
> > Surely to start with 5 zk's (or in fact any odd number - it
> > could be 21 even), and from a single failure you drop to an
> > even number - then there is the danger of NOT getting quorum.
> >
> > So ... I can only assume that there is a mechanism in place
> > inside zk to guarantee this cannot happen, right?
> >
> > --
> > Cheers
> > Jules.
> >
> >
> >
> > On 05/03/2015 06:47, svante karlsson wrote:
> >
> >> Yes, as long as it is three (the majority of 5) or more.
> >>
> >> This is why there is no point of having a 4 node cluster. This would
> also
> >> require 3 nodes for majority thus giving it the fault tolerance of a 3
> >> node
> >> cluster but slower and more expensive.
> >>
> >>
> >>
> >> 2015-03-05 7:41 GMT+01:00 Aman Tandon <amantandon.10@gmail.com>:
> >>
> >>  Thanks svante.
> >>>
> >>> What if in the cluster of 5 zookeeper only 1 zookeeper goes down, will
> >>> zookeeper election can occur with 4 / even number of zookeepers alive?
> >>>
> >>> With Regards
> >>> Aman Tandon
> >>>
> >>> On Tue, Mar 3, 2015 at 6:35 PM, svante karlsson <saka@csi.se> wrote:
> >>>
> >>>  synchronous update of state and a requirement of more than half the
> >>>> zookeepers alive (and in sync) this makes it impossible to have a
> "split
> >>>> brain" situation ie when you partition a network and get let's say 3
> >>>>
> >>> alive
> >>>
> >>>> on one side and 2 on the other.
> >>>>
> >>>> In this case the 2 node networks stops serving request since it's not
> in
> >>>> majority.
> >>>>
> >>>>
> >>>>
> >>>>
> >>>>
> >>>>
> >>>>
> >>>>
> >>>> 2015-03-03 13:15 GMT+01:00 Aman Tandon <amantandon.10@gmail.com>:
> >>>>
> >>>>  But how they handle the failure?
> >>>>>
> >>>>> With Regards
> >>>>> Aman Tandon
> >>>>>
> >>>>> On Tue, Mar 3, 2015 at 5:17 PM, O. Klein <klein@octoweb.nl>
wrote:
> >>>>>
> >>>>>  Zookeeper requires a majority of servers to be available. For
> >>>>>>
> >>>>> example:
> >>>
> >>>> Five
> >>>>>
> >>>>>> machines ZooKeeper can handle the failure of two machines. That's
> why
> >>>>>>
> >>>>> odd
> >>>>
> >>>>> numbers are recommended.
> >>>>>>
> >>>>>
>

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