What stream data should my application keep?

That depends on your application, but the minimal set should include UUIDs (or names if using named streams), descriptions of the stream content, the content length, the content type, and the number of replicas expected in the cluster. This will allow quick approximation of the number of bytes in the cluster attributable to your application. This number could be approximated by a script that executed an SCSP INFO command using the countreps=yes query argument on each name or UUID in the application, and then adding the bytes used by each stream. Arriving at the same number through a database query is quicker and easier though. Using Swarm Search, you can also get an approximation of matching streams' content sizes with an enumeration. Swarm Search adds the expected number of replicas to get total size in bytes, so if your stream is over-replicated or under-replicated temporarily, these anomalies will not be detected in a Swarm Search enumeration.