Github user revans2 commented on the pull request:
https://github.com/apache/storm/pull/838#issuecomment-157057138
@bastiliu There is a possibility of that. If you have a congested network
or if there are other issues around the pacemaker node that it could become a
bottleneck. We don't see this as a silver bullet that will fix all of the
problems, just a step in the right direction. This is similar to what JStorm
has done with the TopologyMaster to offload ZooKeeper, although quite different
in it's design.
If you watch the talk I gave at Hadoop Summing about scaling storm and the
calculations we did
https://www.youtube.com/watch?v=pB9d3tMM__k
http://www.slideshare.net/Hadoop_Summit/from-gust-to-tempest-scaling-storm
Theoretically a single pacemaker node with the current metrics level and
gigabit Ethernet should be able to support about 6000 nodes. The reality is
probably closer to between 2000 and 3000 nodes. With 10GigE we are looking at
a lot more and nimbus is likely to be the bottleneck at that point. But the
real reason for this is a step in the direction of moving metrics out of
zookeeper and into a pluggable time series database backend. I have not looked
deeply into the metrics changes that JStorm has yet, so all of this code might
end up being removed as part of the merger. We mostly want to make sure that
all of the code we write for storm ends up in open source.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
