That wouldn't really work. First of all we don't have Nimbus HA yet, so
whichever colo you put nimbus in would be a single point of failure. Even
after nimbus HA all the state is stored in ZK, which uses a quorum to be
functional, so whichever colo had the most ZK nodes in it would be the single
point of failure. If we are doing a true lambda architecture we often don't
have a BCP for storm, as it can be down for a while and the batch will stop
over the results. If we are doing more of a kappa architecture it varies by
the need. But in those cases we will either run hot-hot and pick the result
with the most complete data, or hot-warm and use something else to trigger the
failover.
- Bobby
On Thursday, March 5, 2015 8:37 PM, Rama Ramani <[email protected]>
wrote:
Hello,
Can Storm be deployed in a topology where the nimbus node and data
nodes in a single cluster can span different data centers (for Disaster
recovery scenarios)? Or is there another recommended deployment topology for DR
scenarios?
Regards
Rama
