Design pluggable backend for adaptive routing state (file → sqlite / redis) #2
Description
Context
Adaptive routing currently uses file-backed state for simplicity.
This works for single-process experimentation but does not scale to
multi-worker or distributed deployments.
Open problem
Design a pluggable state backend that:
- Is concurrency-safe
- Supports reproducibility and debugging
- Can swap between file / sqlite / redis implementations
Questions
- What state guarantees are actually required?
- How should replay/debugging work?
- What tradeoffs exist between performance and determinism?
This is a systems design problem, not a bug.