zookeeper ensemble with hbase (versus single node)

[Koert Kuipers <ko...@tresata.com>]

in various places i have read that in production one should really have a
zookeeper ensemble (with an odd number of members) as opposed to a single
zookeeper. this was also mentioned for example in the base documentation.
however, if your cluster has a single machine with namenode, secondary
namenode and hbase master, what is the benefit of having a zookeeper
cluster? if that one machine goes down your hbase isn't doing anything
anyhow, so why even bother running a zookeeper ensemble in this case? why
not just use a single zookeeper?

are the performance benefits a reason to run a zookeeper ensemble even in
this scenario? (faster reads perhaps)

Re: zookeeper ensemble with hbase (versus single node)

[Mahadev Konar <ma...@hortonworks.com>]

Koert,
 I think this is a more appropriate question for the HBase mailing
list. HBase uses ZooKeeper for Master failover. It is possible that if
the Master is down and namenode/others are up, HBase can failiover to
the new master. You have a good point of having ZK running on
Master/NN/SNN, so you'd have to avoid running ZK on those machines.
Hope that helps.

mahadev

On Fri, Nov 11, 2011 at 7:15 AM, Koert Kuipers <ko...@tresata.com> wrote:
> in various places i have read that in production one should really have a
> zookeeper ensemble (with an odd number of members) as opposed to a single
> zookeeper. this was also mentioned for example in the base documentation.
> however, if your cluster has a single machine with namenode, secondary
> namenode and hbase master, what is the benefit of having a zookeeper
> cluster? if that one machine goes down your hbase isn't doing anything
> anyhow, so why even bother running a zookeeper ensemble in this case? why
> not just use a single zookeeper?
>
> are the performance benefits a reason to run a zookeeper ensemble even in
> this scenario? (faster reads perhaps)
>

Re: zookeeper ensemble with hbase (versus single node)

[Ted Dunning <te...@gmail.com>]

Yes.  It is nice to still have a valid indicator of what is up and down
after a critical node goes down.

This is independent of hbase.  If you have a system with a critical
function monitored by ZK, then you would like that monitoring to survive
the loss of that single node.

On Fri, Nov 11, 2011 at 10:15 AM, Koert Kuipers <ko...@tresata.com> wrote:

> are the performance benefits a reason to run a zookeeper ensemble even in
> this scenario? (faster reads perhaps)
>