Question 390
A team has corrected this HPE compute solution issue: A GPU service sends no metrics about queue time or per-request latency despite strict response-time commitments. Which continuing practice would best prevent the same problem?
Explanation
Correct Answer
Define inference service-level observability before production deployment. This recurring assurance measure is tied to the diagnosed risk: Hardware utilization alone does not reveal application inference bottlenecks. For the active incident, the corrective action remains: Instrument endpoint queue, processing latency, throughput and errors with supported monitoring.
Incorrect Answers
Hardware utilization alone does not reveal application inference bottlenecks. Using the proposed safeguard would guard against a separate risk: Internet uplink speed is unrelated to the east-west fabric demanded by distributed AI jobs. Define inference service-level observability before production deployment.
Unpinned environments cause inconsistent dependencies and irreproducible experiments. The selected safeguard helps with that exposure, but it leaves the recurrence risk here unaddressed. Hardware utilization alone does not reveal application inference bottlenecks. Define inference service-level observability before production deployment.
This control protects against a condition in which workload scheduling is distinct from model validation and governance services. The issue to prevent here is: Hardware utilization alone does not reveal application inference bottlenecks. Define inference service-level observability before production deployment.