> On 18 Jan 2017, at 21:50, kant kodali <kanth...@gmail.com> wrote: > > Anyone has any experience using spark in the banking industry? I have couple > of questions.
> 2. How can I make spark cluster highly available across multi datacenter? Any > pointers? That's not, AFAIK, been a design goal. The communications and scheduling for spark assume that (a) there's low latency between executors and driver, and (b) that data is close enough to any executor that you don't have to get placement right: if you can't schedule work near the data, then running it on other nodes is better than not running work. oh, and failure modes are that of a single cluster: node and rack failures, not a single long-haul connection which could cut the entire cluster in half. If that happens, then all work running on the cluster without the driver is lost: you don't have the failure resliience you wanted, as it means that if the cluster with the driver actually failed, then the other cluster would not be able to take over. Similarly, cluster filesystems tend to assume they are single DC, with its failure modes. life is more complex across two sites. I do know HDFS doesn't handle it, though there are things that do. I would try to come up with a strategy for having separate applications running on the different DCs, with a story for data replication and reconciliation. Even there, though, there'll inevitably be an SPOF. How do you find it? You Wait: it will find you. -steve --------------------------------------------------------------------- To unsubscribe e-mail: user-unsubscr...@spark.apache.org