What aspect of AI infrastructure design is MOST critical for ensuring high availability of production...

The correct answer is A because high availability depends on redundancy, failover, orchestration, and scaling rather than manual recovery. NVIDIA Run:ai documentation says the goal of high availability is “to reduce downtime and eliminate single points of failure” by using Kubernetes best practices with NVIDIA Run:ai configuration. It identifies two fundamental strategies: using multiple system nodes for critical services and replicating core and third-party services so workloads are distributed and single points of failure are reduced. NVIDIA also states that if a component fails on one node, requests can be routed to another instance.

This aligns exactly with automated failover orchestration and elastic scaling across redundant nodes. NVIDIA’s NIM autoscaling guidance also describes using Kubernetes Horizontal Pod Autoscaling with NVIDIA NIM microservices to automatically scale services up and down based on metrics, which supports production availability and capacity management.

Why the other options are incorrect: Custom GPU driver builds do not ensure service continuity during node failure. Dataset expansion is related to model training quality, not production high availability. Manual GPU restarts and ad hoc redeployment are reactive and unreliable, while production AI services require automated orchestration and redundancy.