The replace_branch incurs an expensive IO overhead for most operations in the normal code path. For HDFS, it is a namenode access, and for object storage, it is a separate billing. And for object storage, every IO cost 10ms+. It is very costly.
This is difficult to accept, and if replace_branch is not as useful, we should remove this operation.
Purpose
The replace_branch incurs an expensive IO overhead for most operations in the normal code path. For HDFS, it is a namenode access, and for object storage, it is a separate billing. And for object storage, every IO cost 10ms+. It is very costly.
This is difficult to accept, and if replace_branch is not as useful, we should remove this operation.
Detail is here: https://lists.apache.org/thread/q46clxx38fz7n1xw0sgscmcslo3qrp5c
Tests
API and Format
Documentation