What are the trade-offs between different vector databases like Pinecone or FAISS in a production environment?