Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
Description
BucketWriter has several race conditions that came up during my performance testing over the weekend. One issue that caused data loss was the lack of atomic close() and open() semantics related to the "retry" mechanism after the abort() call in HDFSEventSink.process().
Another issue is the lack of clearly delineated responsibilities for calling open(), flush(), close(), etc. For example, HDFSEventSink.start() calls open(), HDFSEventSink.process() calls and abort() which calls open(), and BucketWriter.append() also calls close() and open().
There is another race condition related to the JVM shutdown hooks, which cause .tmp files not to be renamed.
These APIs need to be refactored and their responsibilities need to be clarified.
Attachments
Attachments
Issue Links
- is duplicated by
-
FLUME-1163 HDFSEventSink leaves .tmp files in place when Flume is stopped
- Closed
-
FLUME-1173 HDFSEventSink can leave orphaned .tmp files
- Closed
- links to