On all but the largest clusters, the computing needs of the
master/gc/monitor/tracer can all run on one node and can be co-located with
a zookeeper server.
Strictly speaking, you don't need a tracer. Unless you are running a very
large cluster, one tracer will be enough.
The tracer can be used to identify components that are performing slowly,
so it needs to run all the time.
On Wed, May 1, 2013 at 3:57 PM, Aaron G <[EMAIL PROTECTED]> wrote:
> Generic/best practices question about the 6 services: master, gc,
> tserver, logger, monitor, tracer
> For this example/discussion let's say I have a cluster, with 10 nodes
> (n01-n10)...3 of the nodes running zookeeper
> n01: NameNode, zooKeeper
> n02: SecondaryNameNode, zooKeeper
> n03: JobTracker, zooKeeper
> n04: empty (for now)
> Let's label these storage/compute nodes:
> n05-n10: dataNode, taskTracker
> So, how i thought this could be setup for Accumulo services:
> n04: master & gc
> n05-n10: each runs a tserver & logger
> I think my main questions revolve around the monitor & tracer services and
> where they run:
> 1. Does those need to run on every "compute node"?
> 2. Do you only need one running instance of monitor? Perhaps on the n04?
> Or does it need to run on every tserver as well?
> 3. Do you only need the tracer service running on compute nodes? Or
> everywhere (master & gc included)? Do you only need the tracer for the
> purpose of development of Iterators, Scanners, Writers? They primarily
> there to help with that activity? Or are they useful to have running "all
> the time?"
> Thanks in advance,