Description
I recently saw a case where a bad node in the cluster produced corrupted shuffle data that caused the codec to throw IllegalArgumentException when trying to fetch. Fetchers currently only handle IOException and InternalError, and any other type of exception will cause the entire task to be torn down. We should consider catching Exception like MapReduce does to be more robust in light of other types of errors coming from the codec and allow retries to occur.
Attachments
Attachments
Issue Links
- relates to
-
TEZ-3196 java.lang.InternalError from decompression codec is fatal to a task during shuffle
- Closed