How do you pick a vector database for production RAG?

For teams shipping RAG in production, what criteria have actually mattered beyond benchmarks? I am weighing Pinecone, Weaviate, Qdrant, and pgvector. - Operational: backup/restore, multi-tenant isolation, failover SLAs - Retrieval quality: HNSW vs IVF vs disk-based, re-ranking integration - Cost: index build time, storage vs recall trade-offs, hot vs cold tiers - Developer experience: schema flexibility, hybrid search (BM25+dense), filters - Security/compliance: VPC, encryption, audit logs, regional residency What did you learn the hard way (e.g., index parameter choices, memory sizing, write amplification)? Any anti-patterns with chunking and metadata filters that hurt recall? Share real-world numbers if possible.