Page History

...

The con is that the implementation could be complicated and sensitive since wrapping the StreamRecord with a new field. But I believe it is acceptable since the modification is limited in the SourceTask. On the other hand, the wrapping StreamRecord could help to reduce duplicated extraction and computation of key and keyGroup.

For these two options, there is a common limitation, which is the number of splits should not be larger than the number of key groups. Because finally, we want to build a mapping from a split to a key group to ensure correctness during rescaling.

...