Major Description
For WEBHDFS invocations through Knox, if a datanode write fails with an IOException, the currently configured HA dispatch (DefaultHaDispatch) attempts the invocation the configured number of times.
HDFS dev says that it would be appropriate to more simply respond with the error, deferring the decision to retry to the client.
The best resolution for this is to have the datanode interactions use the DefaultDispatch (rather than the DefaultHaDispatch) even when HA is enabled, such that no retry/failover will be attempted for those interactions.
(FYI: the datanode dispatch is configured in the WEBHDFS service definition)