Hi,
I'm not aware of any performance tests for AWS specifically, more so, I'm
not aware of any cloud based performance tests regarding your question, so
I'm going to speak general ZK deployment.
tickTime - regulates how often a heartbeat is sent or connection times out
(it's the base unit, not the exact time). This has no direct effect to sync
time.
initLimit - initLimit * tickTime equals the time allowed for followers to
connect and sync with Leader. This will not speed up your sync times. If
you have lot of data stored in ZK, you might wan't to increase this,
allowing more time to sync.
syncLimit - Pretty similar to initLimit, minus the connection time. This is
for every sync operation, so your follower might be dropped after a while.
Again, not much to do with speeding up the sync time.
These settings are mainly for timeouts. I would touch these if I start
seeing followers getting dropped due to timeout.
There isn't really a way I'm aware of that would speed up sync time, other
than to keep data stored in ZK minimal, and keep your jute.maxbuffer fairly
small, default is 1MB I think, and you shouldn't go above a few MBs.
Regards,
Norbert
On Mon, Mar 11, 2019 at 1:49 PM Behroz Sikander wrote:
> This seems to be a straight forward question :). Anyone?
>
> On Fri, Mar 8, 2019 at 10:32 AM Behroz Sikander
> wrote:
>
> > Hello,
> > Currently, I have a Spark cluster which uses 3-node zookeeper underneath
> > for leader election. I want to reuse the zookeeper cluster for storing
> some
> > configuration information and traffic in zookeeper will increase. I want
> > the cluster to become synced as early as possible.
> >
> >
> > What are the recommended configuration settings for this clusters
> assuming
> > that I am running on AWS?
> >
> > The following values are the ones I am using now.
> >
> > tickTime=4000
> > initLimit=30
> > syncLimit=15
> >
> > Any reasoning on why specific values would work best would also be
> helpful.
> >
> > Regards,
> > Behroz Sikander
> >
> >
>
