Which vector database actually held up in production?

Vector DBs are easy to demo and hard to run at scale. If you've taken one to production, the community would love to hear the real story.

A few things worth covering:

• Which one did you pick (Pinecone, and the various open-source options are all in the directory) and why?
• Roughly what scale — vectors, queries per second, latency targets?
• What broke or surprised you once real traffic hit?
• Would you choose it again?

War stories welcome — the messy details are the useful part.