Open cindyyuanjiang opened 10 months ago
Thanks @viadea for the comment: "There are some hard rules we can consider when should we use BHJ (BroadcastHashJoin) vs SHJ (ShuffledHashJoin) in the GPU mode. Those hard rules can be implemented in our Profiling tool as well. For example:
As a result, we can check the join type and data size on the smaller side to determine if we should promote the BHJ or not by setting a larger spark.sql.adaptive.autoBroadcastJoinThreshold."
Is your feature request related to a problem? Please describe. Follow up issue for PR #688
We need to investigate further on
spark.sql.adaptive.autoBroadcastJoinThreshold
in order to make an accurate recommendation.