AI-300 Practice Question 64: Readiness Probe Slow Model Load
Microsoft Machine Learning Operations Engineer Associate (AI-300) · Test 7 · question 4 of 10
A large model needs about 10 minutes to load inside its container. When deployed to a managed online endpoint with the default probe settings, the deployment fails and the logs show the container being restarted before it finishes loading. Running the same container manually works. What should the team change?
Show answer and explanation
Answer:
Managed online deployments use liveness and readiness probes, with settings such as initial_delay, period, timeout and failure_threshold. If the container needs longer than the probes allow, the liveness probe gives up and the container is restarted, and the readiness probe keeps it out of traffic. Raising the probe delay or threshold gives the model time to load (option C).
- A is wrong because this setting limits how long a scoring request may take. It does not affect the startup period during which the probes run. - B is wrong because each new instance would load the same model and be restarted in the same way, so more instances do not help. - D is wrong because authentication controls who can call the endpoint. It has no effect on whether the container starts.
Reference: Managed online deployment YAML schema