We have a Dataflow Job copying data from multiple Kinesis streams into Big Query. Recently we have noticed that the watermark on one of the streams frequently gets stuck although data from that stream is still being processed (it progress only when the traffic increases or Dataflow autoscaling feature kicks in).
Looking at the CloudWatch statistics for the affected stream, it has a really low traffic rate - only ~1 event every few minutes . After investigation and consulting the issue with Google's Dataflow Team, it looks like with such small amount of data on the stream, the function calculating the watermark in KinesisReader reports progress incorrectly.
From my initial investigation, I suspect that the issue might be related to usage of MovingFunction in KinesisReader. In the current implementation, it covers 1 minute period of samples, since obtaining the min value flushes stale values, if the traffic is very low the following call to significance check always returns false (as it relies on the number of samples, and most of them were flushed by get() invocation).