A platform team is taking a single-instance LLM assistant to a multi-region fleet behind a load balancer, and they must keep safety behaviour identical on every replica while supporting a large rise in concurrent users. Which design choices most directly help the safety controls remain both consistent and scalable as the fleet grows? Select TWO.
- AServe the guardrail policy from a shared, centrally versioned configuration source that every replica reads, rather than editing a local file on each instance. Correct
- BHard-code the blocklist and refusal rules into each replica's container image so the safety logic ships immutably with the binary.
- CMake the safety filter stateless so any replica can serve any request and instances can be added or removed without warm-up or session affinity. Correct
- DPin each user to one replica with sticky sessions so that replica can cache that user's prior safety decisions in local memory.
- ERoute every request through one central GPU node that applies all safety checks before forwarding to the worker fleet for generation.
Why A is correct: A single versioned policy source removes per-replica drift, so every instance enforces the same rules even as the fleet scales horizontally.
Why B is wrong: Baking rules into the image looks safe and reproducible, but a policy change then needs a full rebuild and redeploy, so replicas drift until every one is rebuilt.
Why C is correct: Stateless filtering lets the load balancer spread traffic freely and autoscale, which is what keeps enforcement uniform and scalable under rising load.
Why D is wrong: Session affinity seems to aid caching, but it concentrates load and means a stale local policy on one replica keeps affecting the same users instead of being smoothed out.
Why E is wrong: A single chokepoint node appears to guarantee one consistent decision, but it caps total throughput and becomes a single point of failure, defeating the scaling goal.