AWS Certified Machine Learning – SpecialtyMachine Learning Implementation and OperationsHard
A company is using a SageMaker real-time endpoint for a critical application. They need to minimize the cold start time for the endpoint when scaling up or after updates. The model is large, and loading it into memory takes a significant amount of time. Which SageMaker endpoint configuration parameter can directly address this issue by pre-loading the model onto instances?
- AAutoScalingPolicy
- BProductionVariant.InitialInstanceCount
- CProductionVariant.ModelDataDownloadTimeoutInSeconds
- DProductionVariant.ContainerStartupHealthCheckTimeoutInSeconds
Show answer & explanationAnswer & explanation
Correct answer: D. ProductionVariant.ContainerStartupHealthCheckTimeoutInSeconds
The `ContainerStartupHealthCheckTimeoutInSeconds` parameter defines how long SageMaker waits for the container to become healthy. For large models, increasing this timeout allows more time for the model to load into memory before the health check fails, effectively reducing perceived cold start by preventing premature restarts.
Why the other options are wrong
- A. AutoScalingPolicy defines how instances scale, but doesn't directly address cold start time for individual instances.
- B. InitialInstanceCount sets the initial number of instances, but doesn't prevent cold start on *new* instances when scaling up.
- C. ModelDataDownloadTimeoutInSeconds specifies how long the instance has to download the model artifact. While related, it doesn't cover the *loading into memory* part after download.
ContainerStartupHealthCheckTimeoutInSeconds
A SageMaker endpoint configuration parameter that specifies the maximum time (in seconds) for a container to pass its health check after startup.
- Crucial for large models that take time to load into memory.
- Increasing it can prevent premature container restarts due to model loading times.
- Part of `ProductionVariant` configuration.
Memory trick: Health check timeout determines if a cold instance gets a warm welcome or a cold restart.