This sounds similar to a non streaming problem we had at WMF. We ingest all event data from Kafka into HDFS/Hive and partition the Hive tables in hourly directories. If there are no events in a Kafka topic for a given hour, we have no way of knowing if the hour has been ingested successfully. For all we know, the upstream producer pipeline might be broken.
We solved this by emitting artificial 'canary' events into each topic multiple times an hour. The canary events producer uses the same code pathways and services that (most) of our normal event producers do. Then, when ingesting into Hive, we filter out the canary events. The ingestion code has work to do and can mark an hour as complete, but still end up writing no events to it. Perhaps you could do the same? Always emit artificial events, and filter them out in your windowing code? The window should still fire since it will always have events, even if you don't use them? On Mon, May 9, 2022 at 8:55 AM Shilpa Shankar <sshan...@bandwidth.com> wrote: > Hello, > We are building a flink use case where we are consuming from a kafka topic > and performing aggregations and generating alerts based on average, max, > min thresholds. We also need to notify the users when there are 0 events in > a Tumbling Event Time Windows. We are having trouble coming up with a > solution to do the same. The options we considered are below, please let us > know if there are other ideas we haven't looked into. > > [1] Querable State : Save the keys in each of the Process Window > Functions. Query the state from an external application and alert when a > key is missing after the 20min time interval has expired. We see Queryable > state feature is being deprecated in the future. We do not want to go down > this path when we already know there is an EOL for it. > > [2] Use Processing Time Windows : Using Processing time instead of Event > time would have been an option if our downstream applications would send > out events in real time. Maintenances of the downstream applications, > delays etc would result in a lot of data loss which is undesirable. > > Flink version : 1.14.3 > > Thanks, > Shilpa >
