Hi,
If I remember correctly, I don't think Spark provides watermark value
itself for the current batch to the public API. That said, if you're
dealing with "event time" (and I guess you belong to this case as you worry
about late events), unless you employ a new logical/physical plan to expose
I have a Spark structured streaming based application that performs
window(...) construct followed by aggregation.
This construct discards latecomer events that arrive past the watermark. I
need to be able to detect these late events to handle them out-of-band.
The application maintains a raw