FAISS or a Real Vector Database?

A recommendations team builds a "similar products" feature: 150,000 product embeddings, rebuilt from scratch once a night by a single batch job, queried read-only by that same process to precompute similar-item lists. A senior engineer proposes standing up Pinecone "so it's production-grade from day one."

1. Would you recommend an in-process ANN library (FAISS, hnswlib) or a dedicated vector database here, and why?
2. Name two concrete requirements that, if added to this system, would flip your recommendation toward a real database.
3. What operational capability is the team implicitly giving up by choosing a library, and why doesn't it matter for this workload?