Github user xccui commented on the issue:
https://github.com/apache/flink/pull/5094
Hi @hequn8128, as @fhueske explained, watermarks should not define the
semantics of the join. We just emit join results, in spite of whether they are
late and the join function should not care about the lateness. You can take
"holding back watermarks" as a special watermark assigner, based on which,
there could be out-of-ordered data (join results) being taken as late.
Currently, we just generates watermarks with a static delay that equals to the
larger relative offset. In the future, we should dynamically decide whether a
join result is late according to the real cache size.
---
