We ran into problems setting event time timers per-element in the Python
SDK. Pipeline progress would stall.
Turns out, although the Python SDK does not expose the timer output
timestamp feature to the user, it sets the timer output timestamp to the
current input timestamp of an element.
This will lead to holding back the watermark until the timer fires (the
Flink Runner respects the timer output timestamp when advancing the
output watermark). We had set the fire timestamp to a timestamp so far
in the future, that pipeline progress would completely stall for
downstream transforms, due to the held back watermark.
Considering that this feature is not even exposed to the user in the
Python SDK, I think we should set the default output timestamp to the
fire timestamp, and not to the input timestamp. This is also how timer
work in the Java SDK.
Let me know what you think.
-Max
PR: https://github.com/apache/beam/pull/12531
