Description
The HA Dispatch implementations catch IOException, and initiate failover logic regardless of the type of error the exception represents. For instance, some IOExceptions indicate interrupted data transfer while others represent connection errors.
This distinction is especially important for PUT and POST requests, for which InputStreamEntity is used for the content. InputStreamEntity is a non-repeatable entity type, making the results of subsequent attempts unreliable.
We should probably only failover / retry on connection-related IOException types:
- java.net.SocketException
- java.net.UnknownHostException
And return an error response to the client for other IOException types. Maybe it makes sense to consider the HTTP method to make this decision (e.g., retry GET requests, but not PUT or POST).
The affected dispatch implementations includes at least:
- org.apache.knox.gateway.ha.dispatch.DefaultHaDispatch
- org.apache.knox.gateway.ha.dispatch.AtlasApiHaDispatch
- org.apache.knox.gateway.ha.dispatch.AtlasHaDispatch
- org.apache.knox.gateway.dispatch.NiFiHaDispatch
- org.apache.knox.gateway.hdfs.dispatch.AbstractHdfsHaDispatch
If retry is configured, but we won't retry, then perhaps java.net.HttpRetryException should be thrown.