Graceful Handling of Topology Kill

[jimhe]

I'm wondering if there are any good approaches to ensure that tuples are flushed to HDFS when the topology is killed? Under normal operation, the bolt flushes according to the sync policy, but it acks every tuple even if they haven't been yet flushed. If the topology is shut down (storm issues kill -9 during a normal shut down) or if the bolt worker dies, the tuples left in its buffer are left untracked.