Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Plain View
Zookeeper >> mail # user >> Can I recover full zookeeper state from observer?


+
Kuba Lekstan 2013-02-07, 12:24
+
Alexander Shraer 2013-02-07, 21:28
Copy link to this message
-
Re: Can I recover full zookeeper state from observer?
Yea I can reconfigure cluster manually but because I have only observers in
DC2 it's not possible because I don't have transaction logs in DC2, or
maybe it is?

2013/2/7 Alexander Shraer <[EMAIL PROTECTED]>

> Hi,
>
> You can probably do this manually now, but I'll let someone else comment on
> this.
>
> In 3.5.0 release (with ZK-107 in) you will be able to turn observers into
> leader + followers and vise versa by running a simple zookeeper command.
> But unfortunately not in the scenario you described.
>
> In your scenario, there are two possible problems that make it unsafe to
> reconfigure:
> - it possible that the connection between the datacenters is down and not
> DC1, and therefore if DC2 independently decides
> to start leading we may have "split brain" - both quorums will be
> independently processing and committing operations.
> - it is very possible that DC1 is down but managed to commit things before
> it went
> down which are unknown to DC2, and therefore if we activate DC2 we'll loose
> committed state.
>
> You will only be able to reconfigure if a quorum (majority) of the old
> configuration and a quorum of the new configuration are up and connected to
> the current leader.
>
> Alex
>
> On Thu, Feb 7, 2013 at 4:24 AM, Kuba Lekstan <[EMAIL PROTECTED]> wrote:
>
> > Hello
> >
> > I'm considering situation when I have 2 data centers, In 1s DC I'm having
> > leader + followers, in 2nd only observers.
> > In case when my 1st DC goes down, can I somehow reconfigure observers in
> > 2nd DC to leader+followers?
> >
> > From what I found out the observers doesn't get transaction logs from the
> > leader, is there any option I could set to change this behavior?
> >
> > Thanks.
> >
>