How to catch potential issues with the new Faiss vector format?

[kaivalnp]

Description

Opening this to discuss how we could catch issues (early on) with the new Faiss vector format (#14178) -- like memory leaks or bugs in Faiss itself, under more rigorous scenarios -- perhaps using long-running (or "monster" tests)

This is important because we're dealing with native code, without fine control over memory allocation and error handling -- and could result in the JVM crashing!

One such example is #14847, where we discovered an issue with attempting to index >2GB of vectors in a single field in a single segment (this issue was more on the codec side and how we passed data over to Faiss instead of the native part, but the above tests would help catch this nonetheless!)

[onestardao]

You're absolutely right about the risks in native FAISS-based vector formats. We've actually documented multiple failure patterns in OpenSearch clusters that stem from semantic bootstrapping mismatches and untracked FAISS state transitions.

In our diagnostic map, this falls under Problem No.14 and No.15, especially when the system silently proceeds with misconfigured engines and later crashes at runtime. (Sometimes even deletes the data directory.)

We’ve built a full test suite and detection framework for these edge cases. Happy to share details or help if you’re exploring structured mitigations.

Let us know if this is useful.