Currently we don't have any TTL for inferences, so it will grow unbounded. We should allow users to configure TTL for inferences and delete old inferences. We should decide whether to retain inferences with feedback attached.