Affects Version/s: None
Fix Version/s: None
I have identified an issue where the watermark does not advance when using the beam PubSubIO when volumes are very low.
I have created a mini example project to demonstrate the behaviour with a python script for generating messages at different frequencies:
[note: this is in a directory of a Beam fork for corp hoop jumping convenience on my end, it is not intended for merging].
The behaviour is easily replicated if you apply a fixed window triggering after the watermark passes the end of the window.
With this triggering, using both the flink local runner the direct runner, panes will be fired after a long delay (minutes) for low frequencies of messages in pubsub (seconds). The biggest issue is that it seems no panes will ever be emitted if you just send a few events and stop. This is particularly likely trip up people new to Beam.
If I change the triggering to have early firings I get exactly the emitted panes that you would expect.
I can use any variation of early firing triggers and they work as expected.
We believe that the watermark is not advancing when the volume is too low because of the sampling that PubSubIO does to determine it's watermark. It just never has a large enough sample.
This problem occurs in the direct runner and flink runner, but not in the dataflow runner (because dataflow uses it's own PubSubIO because dataflow has access to internal details of pubsub and so doesn't need to do any sampling).
For extra context from the user@ list:
Thanks to your info, I think it is the configuration of MovingFunction  that is the likely culprit, but I don't totally understand why. It is configured like so:
- store 60 seconds of data
- update data every 5 seconds
- require at least 10 messages to be 'significant'
- require messages from at least 2 distinct 5 second update periods to 'significant'
I would expect a rate of 1 message per second to satisfy this. I may have read something wrong.
Have you filed an issue in Jira ?
Not sure that this can be very helpful but I recall a similar issue with KinesisIO   and it was a bug in MovingFunction which was fixed.