Details
-
Bug
-
Status: Open
-
P3
-
Resolution: Unresolved
-
None
-
None
-
None
Description
SparkUnboundedSource will advance the watermark to the MAX of the watermark of any partition. You can see it at https://github.com/apache/beam/blob/fab12c772d461fc8db4b3c361d38fe2781926fff/runners/spark/src/main/java/org/apache/beam/runners/spark/io/SparkUnboundedSource.java#L204 .
This should be the MIN - this is a combining of watermarks - not advancing. This currently means the watermark moves too quickly and the slowest partition of an unbounded source has elements that are routinely marked late.