Details
-
Improvement
-
Status: Closed
-
Major
-
Resolution: Won't Fix
-
2.2.0
-
None
-
None
Description
Feature of adaptive query execution is a good way to avoid generating too many small files on HDFS, like mentioned in SPARK-16188.
When feature of adaptive query execution is enabled, all shuffles will be coordinated. The drawbacks:
1. It's hard to balance the num of reducers(this decides the processing speed) and file size on HDFS
2. It generates some unnecessary shuffles(https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/exchange/EnsureRequirements.scala#L101)
3. It generates lots of jobs, which have extra cost for scheduling.
We can add a config and enable adaptive query execution only for the last shuffle.
