This is a JVM infrastructure optimization, not a recommendation architecture breakthrough. At Netflix scale, most of the field runs GPU-accelerated ANN search (FAISS, ScaNN) for candidate retrieval.
From the sounds of it, they're looking for items that are different from their watch history among some candidate set, which vector dbs aren't optimized for.
First mover disadvantage: you need to invent a way of doing things (ANN) before any decent tools have been built. 10 years later the world is laughing at you for using Java to do vector math.
