Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
Description
At now, if a vertex receives multiple shuffle edges (in a joining case or something like that), DataSkewPolicy will optimize the partitioning of the two edges respectively.
This makes possible that two key-value elements having an identical key are assigned to different tasks. In this case, the result will differ from the expected one.
We need to collect the data metric for the two edges in a single metric aggregation vertex and optimize the two edges at once (like in Dryad paper).
Attachments
Issue Links
- links to