Question about multi-master deployment

classic Classic list List threaded Threaded
2 messages Options
Reply | Threaded
Open this post in threaded view
|

Question about multi-master deployment

rong111
Luckily to have a place to ask questions. I have been trying to deploy
Alluxio on a cluster for months. I came across several hard problems to
solve.

The first important problem is on high-availability deployment. I deployed
two masters, and a zookeeper server on each Alluxio master/worker node. At a
time, only one master can work, and the other master takes over only when
the leader master fails. My question is: Is there a deployment that multiple
masters can work together to provide metadata service? I think it is quite
important for large-scale cluster with for example thousands of clients. If
thousands of threads try to access IO, may the only leader master becomes a
bottleneck? I wish there is a solution for this problem, but did not find an
answer yet.




--
Sent from: http://alluxio-users.85194.x6.nabble.com/

--
You received this message because you are subscribed to the Google Groups "Alluxio Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: Question about multi-master deployment

Andrew Audibert
Hi Rong,

Alluxio has a single-active master architecture, so the short answer is no. We are considering multi-active in Alluxio 2, but for now only one master at a time can be active, with standby masters shadowing the master state and only starting to serve requests if they win Zookeeper election. The system is designed to minimize the load on the master by delegating all IO operations to workers, and only serving metadata from master. As long as your workloads aren't focused on small files, you should be able to support thousands of clients since clients will spend most of their time communicating with workers instead of the master. The master is designed to be highly concurrent, so you can get more performance out by using more powerful master nodes.

Regarding the cluster setup, consider using either 1 or 3 nodes for Zookeeper. Zookeeper recommends using an odd number of nodes, and I've run into trouble in the past with trying an even number. Alluxio doesn't require Zookeeper to be running on the same nodes as it.

- Andrew

On Tue, Jul 24, 2018 at 6:49 AM rong111 <[hidden email]> wrote:
Luckily to have a place to ask questions. I have been trying to deploy
Alluxio on a cluster for months. I came across several hard problems to
solve.

The first important problem is on high-availability deployment. I deployed
two masters, and a zookeeper server on each Alluxio master/worker node. At a
time, only one master can work, and the other master takes over only when
the leader master fails. My question is: Is there a deployment that multiple
masters can work together to provide metadata service? I think it is quite
important for large-scale cluster with for example thousands of clients. If
thousands of threads try to access IO, may the only leader master becomes a
bottleneck? I wish there is a solution for this problem, but did not find an
answer yet.




--
Sent from: http://alluxio-users.85194.x6.nabble.com/

--
You received this message because you are subscribed to the Google Groups "Alluxio Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
--

--
You received this message because you are subscribed to the Google Groups "Alluxio Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.