Not recent but quite helpful. If you are planning to use Flink or even if you already use it in production this post is worth the time.
Points 2 and 7 are specially relevant.
Kryo serialization looks like a great solution initially and usually becomes a nightmare in the long term.
Rocksdb is a great piece of engineering. Going deeper into its multiple options and parameters is worth it at least and mandatory in many cases. I’ve stumbled upon (and solved) the same problem (Rocksdb using too much memory) both in Kafka Streams and Flink.