Github user revans2 commented on the issue:
https://github.com/apache/storm/pull/2463
@srdo One of the issues I was trying to address with using the array was
both performance and memory usage. Maps are very slow and use lots of memory
compared to an array. Hadoop did a very similar thing as part of their generic
resource work for similar reasons. They tend to do a lot more computation on
their resources looking at fairness etc.
We are fining that the performance of the scheduler loop in general is our
biggest bottleneck to scaling our clusters larger. It directly correlates to
how responsive a cluster is to user requests. So for cases where they want to
kill a topology and launch a new one, if they cannot do it quickly because of
scheduling they are not very happy.
---
