Hey! Do you intend to use RocksDB similarly to how Flink is using it? (One RocksDB instance per task slot) If not, I would argue that you will reach lower performances for multiple reasons:
• Network: Instead of fetching in the taskmanager managed memory, you will need some hops to reach your rocksdb instance elsewhere.
• Granularity: In a KeyedStream, each RocksDB instance will handle a smaller range of keys, resulting in better writes, and caching. If you only have a single large instance of RocksDB, you will have continuous flushes and background compactions.
To conclude, maybe you can use a remote RocksDB, but I'm not sure if you should 😉
Those are my thoughts, if anyone can confirm, I would appreciate.