We used multiple regionservers on the same host for a long time. Our use case was a little different.
We had a table (1 table on a cluster) that was 99% reads (random gets) and 1% writes. The table was tall and skinny and we needed really low latency gets. To keep everything in the block cache and to minimize GC time we had 5 regionservers per machine instead of one RS with a huge heap. This worked really well for us, deployment was a bit of a pain as a lot of the start / stop scripts wont work out of the box (or at least didn't at the time). Once bucket cache came around we just reverted back to one regionserver per machine as it was easier to handle operationally. But other than the operational pain, it seemed to work fine for us. On Wed, Aug 24, 2016 at 2:09 AM, Ted Yu <yuzhih...@gmail.com> wrote: > One downside is that when the machine crashes, you would lose more than one > region server. > > You may need to subclass RSGroupBasedLoadBalancer so that balancing > decision fits your requirement. > > Cheers > > On Tue, Aug 23, 2016 at 4:53 PM, GuangYang <yguan...@outlook.com> wrote: > > > Hello,We are recently exploring running multiple region servers on the > > same host, some of the potential benefits:1. smaller region server > > instance, thus potential less contention for better performance?2. We are > > running region server groups, thus we can run region server groups with > > different workloads on the same group of hardware. > > We plan to go with cgroups to isolate region servers on the same host, > > mainly because we have had good experience using cgroups within other > > hadoop projects. > > Has anyone played around with running multiple RS on the same host and > > would like to share the experience? > > Thanks,Guang >
