Rebalancing vs restarting--which one happens when?

[John Yost <so...@gmail.com>]

Hey Everyone,

We're currently using 0.9.4 and I've seen situations where a small number
of bolts or spouts will fail--usually five or less in a topology of 20
KafkaSpout and 800 bolt executors--and the topology re-balances, and
therefore restarts all workers.  I've seen other situations where failure
of a small number spouts/bolts results in restarts of the same number of
spouts/bolts in new workers instances but no re-balancing of the entire
topology.

Does anyone know what the algorithm is for Storm rebalancing vs. restarting
the workers corresponding to the failed executors? Any insights would be
really appreciated.-

Thanks

--John