Description
Tez auto-parallelism currently has some limitation:
1. ShuffledVertexManager only decrease parallelism not increase
2. Pig currently exaggerate parallelism at frontend, ShuffledVertexManager might get initial parallelism way large than actual, that would be costly
Instead of that, we can gradually adjust initial vertex parallelism at runtime once upstream vertexes finishes.
Attachments
Attachments
Issue Links
- is blocked by
-
TEZ-2234 Add API for statistics information - allow vertex managers to get output size per source vertex
- Closed
-
TEZ-2232 Allow setParallelism to be called multiple times before tasks get scheduled
- Closed
-
TEZ-2233 Allow EdgeProperty of an edge to be changed by VertexManager
- Closed
-
TEZ-2242 Refactor ShuffleVertexManager code
- Closed
- is broken by
-
TEZ-2310 Deadlock caused by StateChangeNotifier sending notifications on thread holding locks
- Closed