Remote sources provide insufficient metadata about relative sizes of splits
Open
core
improvement
java
P3
- Dominant language
- Java
- Stars
- 8.7k
- Forks
- 4.7k
- Avg merge
- 1d 20h
- Merged PRs (30d)
- 196
Description
In the current split protocol, SourceMetadata is reported for the initial parent source. Subsequent splits drop the SourceMetadata. Without this additional information, downstream systems make simplifying assumptions that result in decorrelation between input fraction and the actual fraction of input represented by a task.
This decorrelation of input fraction has cascading negative effects for any system relying on trends in input fraction (eg., Cloud Dataflow's autotuning).
Imported from Jira [BEAM-6819](https://issues.apache.org/jira/browse/BEAM-6819). Original Jira may contain additional context.
Reported by: skp.
Contributor guide
Assessment
This issue has not been assessed yet.