Imagine a situation where one of your nodes running a kafka broker and zookeeper node goes down. You now have to contend with two distributed systems that need to do leader election and consensus in the case of a zookeeper ensemble and partition rebalancing/repair in the case of a kafka cluster so I think Jun's point is that when running distributed systems try to isolate them as much as possible from running on the same node to achieve better fault tolerance and high availability.
>From the Kafka docs you can see that a zookeeper cluster does't need to sit on very powerful hardware to be reliable so I believe the suggestion is to run a small independent zookeeper cluster that will be used by kafka and by all means don't hesitate to reuse that zookeeper ensemble for other systems as long as you can guarantee that all the systems using the zk ensemble use some form of znode root to keep their data seperated within the zookeeper znode directory structure. This is an interesting topic and I'd love to hear if anyone else is running their zk alongside their kafka brokers in production? Ray On Sat, Mar 15, 2014 at 10:28 AM, Carlile, Ken <carli...@janelia.hhmi.org>wrote: > I'd rather not purchase dedicated hardware for ZK if I don't absolutely > have to, unless I can use it for multiple clusters (ie Kafka, HBase, other > things that rely on ZK). Would adding more cores help with ZK on the same > machine? Or is that just a waste of cores, considering that it's java under > all of this? > > --Ken > > On Mar 15, 2014, at 12:07 AM, Jun Rao <jun...@gmail.com> wrote: > > > The spec looks reasonable. If you have other machines, it may be better > to > > put ZK on its own machines. > > > > Thanks, > > > > Jun > > > > > > On Fri, Mar 14, 2014 at 10:52 AM, Carlile, Ken < > carli...@janelia.hhmi.org>wrote: > > > >> Hi all, > >> > >> I'm looking at setting up a (small) Kafka cluster for streaming > microscope > >> data to Spark-Streaming. > >> > >> The producer would be a single Windows 7 machine with a 1Gb or 10Gb > >> ethernet connection running http posts from Matlab (this bit is a little > >> fuzzy, and I'm not the user, I'm an admin), the consumer would be 10-60 > (or > >> more) Linux nodes running Spark-Streaming with 10Gb ethernet > connections. > >> Target data rate per the user is <200MB/sec, although I can see this > >> scaling in the future. > >> > >> Based on the documentation, my initial thoughts were as follows: > >> > >> 3 nodes, all running ZK and the broker > >> > >> Dell R620 > >> 2x8 core 2.6GHz Xeon > >> 256GB RAM > >> 8x300GB 15K SAS drives (OS runs on 2, ZK on 1, broker on the last 5) > >> 10Gb ethernet (single port) > >> > >> Do these specs make sense? Am I over or under-speccing in any of the > >> areas? It made sense to me to make the filesystem cache as large as > >> possible, particularly when I'm dealing with a small number of brokers. > >> > >> Thanks, > >> Ken Carlile > >> Senior Unix Engineer, Scientific Computing Systems > >> Janelia Farm Research Campus, HHMI > >> > >
