I agree the behavior is awkward. The setNumWorkers config appears to behave as an upper limit of workers that will be utilized (e.g. setNumWorkers(200) will not fail to deploy when workers < 200), so this can also impact the ORDER of topology deployment required when workers < sum of all setNumWorkers configs for all topologies.
For example, I originally setNumWorkers(200) - artificially high - so that we can scale our worker pool up to 200 without deploying new code. However, when worker pool is 10, and this topology gets deployed first, then no other topologies can be deployed, since all the workers are assigned to this single topology. Now, while this is awkward IMHO (I would rather have some additional knobs, like “assign a percentage of workers” or “assign a topology priority”), we can get around this by using rebalance and setting the number of workers to reflect the current state of the cluster. Since you MUST rebalance to increase/decrease the number of threads (which would typically go hand in hand with changing the number of workers, I think), then its not really extra work. At some point it would be nice to have some pluggable logic that controls the number of executors dynamically, so that when a worker is added, the number of executors can be altered programmatically, instead of requiring manual intervention. Thanks Tyson On Nov 5, 2014, at 9:05 AM, Dan DeCapria, CivicScience <[email protected]<mailto:[email protected]>> wrote: Hi Nathan, Sounds like I need to just bite-the-bullet and manually define the number of workers for each topology considering all topologies that will be running concurrently. The secondary process you mentioned is interesting wrt using Thrift to query the worker utilization and then auto-balance all topologies during runtime - I'll have to look into that process further. Thanks for you help, -Dan On Wed, Nov 5, 2014 at 11:50 AM, Nathan Leung <[email protected]<mailto:[email protected]>> wrote: It doesn't make sense to automatically balance worker load because you can get yourself into strange situations (e.g. 3 workers and each topology requests 4, or 4 workers and > 4 topologies, etc). It would be nice if the UI or the logs gave a better indication that there were not enough workers to go around though. You could write something that reads cluster information from the nimbus over thrift and rebalances all topologies as necessary, but I think it would be better to just make sure that you have enough workers available (or even better, more than enough workers) to satisfy the needs of your applications. On Tue, Nov 4, 2014 at 4:45 PM, Dan DeCapria, CivicScience <[email protected]<mailto:[email protected]>> wrote: Use Case: I have a production storm cluster running with six workers. Currently topology A is active and consuming all six workers via conf.setNumWorkers(6). Now, launching Topology B with six workers (again via conf.setNumWorkers(6)) states the topology is active, but currently there are no available workers on the cluster for Topology B to use (as Topology A has claimed them all already) and hence Topology B is doing nothing. I believe this is due to storm requiring a priori that the sum of all topology's workers requested <= cluster worker capacity. I am wondering why the sum of all topology workers is not normalizing the allocations for the worker pool when capacity is exceeded and auto-adjusting as new topologies come and go? Meaning, from the use case, since both topologies requested the same count of six workers, and given the finite capacity of the cluster at six actual workers, the implemented normalized proportion of the cluster resources for each topology would be 50% split - such that Topology A gets three actual workers and Topology B gets three actual workers as well. How would I go about implementing a dynamic re-allocation of topology workers based on the proportion of expected workers given the cluster's finite worker capacity? An idea would be allocating workers as desired for all topologies until the cluster capacity is reached, at which point the actual number of workers desired becomes a normalized proportion allocation model over the cluster. Many thanks, -Dan
