Which deployment characteristic is MOST important for highly available enterprise AI services?

High availability (HA) is a core principle of enterprise-grade services, ensuring continuous operation and minimizing downtime. The most effective strategy for achieving HA is deploying redundant instances of the service across multiple, isolated failure domains (e.g., different physical servers, racks, or cloud availability zones). If one domain fails due to hardware, power, or network issues, traffic is automatically redirected to healthy instances in other domains, ensuring the service remains accessible to users.

B. Single host with no backup: This configuration creates a single point of failure (SPOF), which is the antithesis of high availability.

C. Manual recovery only: High availability systems rely on automated failover mechanisms to ensure rapid recovery, as manual intervention is too slow for enterprise service level agreements.

D. No autoscaling policy: While related to performance, a lack of autoscaling can lead to availability issues under load, but redundancy across failure domains is the more fundamental HA principle.

1. NVIDIA. (2023). NVIDIA Triton Inference Server Deployment Guide. Section on "High Availability and Scalability." The guide explicitly recommends running multiple Triton instances behind a load balancer and using orchestration systems like Kubernetes to manage replicas for HA.

2. NVIDIA AI Enterprise Documentation. (2024). High Availability for the NVIDIA AI Enterprise Control Plane. This document details the use of redundant nodes and etcd clusters to prevent single points of failure, a direct application of the principle.

3. Patterson, D. A., & Hennessy, J. L. (2017). Computer Organization and Design RISC-V Edition. Morgan Kaufmann. Chapter 6, "Parallel Processors from Client to Cloud," discusses dependability and availability, defining it via redundancy (Section 6.9).