Keeping keys in a state for a very long time (keys expiry unknown)

I have a use case which I think might be a good addition to the pipelines patterns:

 
beam (java sdk) reads two kind of records from data stream like Kafka:
 
1. Records of type A containing key and corresponding metadata. 
2. Records of type B containing the same key, but no metadata. Beam then needs to fill metadata for records of type B  by doing a lookup for metadata using keys received in records of type A. 
 
Idea is to save metadata or rather state for keys received in records of type A and then do a lookup when records of type B are received.
 Beam's "@State" construct  can be used here, however, problem is that we don't know when keys should expire. I don't think keeping a global window will be a good idea as there could be many keys (may be millions over a period of time) to be saved in a state.
 
One possible solution as suggested by Reza Ardeshir Rokni (rarokni@gmail.com):
 
We can maintain a state in a large fixed window (1 day or so), so that GC can happen within a window bound. After window expire, save the metadata values in an external DB like BigQuery. If we get a record with same key in a new window looking for this metadata, fetch the metadata for that key from external DB and save it in window's state again.
 
 
 
 

Imported from Jira [BEAM-10019](https://issues.apache.org/jira/browse/BEAM-10019). Original Jira may contain additional context.
Reported by: mohilkhare.