MLA-C01 Practice Question 25: Experiment Tracking MLflow
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 3 · question 5 of 10
Six ML engineers each run dozens of training jobs a week, some from SageMaker Studio notebooks and some from scripts on their laptops. Hyperparameters and metrics are pasted into a shared spreadsheet, which is already out of date and does not say which S3 artifact produced which score. The team wants one central store for run parameters, metrics, and artifacts, a UI for comparing runs side by side, and a path to register the winning model in the SageMaker Model Registry. They do not want to patch or scale any servers.
Show answer and explanation
Answer:
Fully managed MLflow on SageMaker AI is built for exactly this: the tracking server's compute and backend metadata store are hosted and operated by SageMaker AI, artifacts go to an S3 bucket in your account, and any MLflow client, a Studio notebook or a script anywhere, logs to it with the server ARN. The MLflow UI compares parameters and metrics across runs, and models kept in the MLflow Model Registry can be registered automatically into the SageMaker Model Registry (option B).
- A is wrong because CloudWatch stores numeric time series, not the parameters, tags, and model artifacts that define a run. Dashboards plot metrics over time rather than comparing runs as rows. - C is wrong only because of the constraint: self-hosted MLflow gives the same API but leaves the team patching the EC2 instance, sizing the RDS database, and handling backups and availability. - D is wrong because it is a tracking system the team would have to build and maintain, no run comparison UI, no artifact lineage, and nothing linking a JSON file to a registered model version.
Reference: Accelerate generative AI development using managed MLflow on Amazon SageMaker AI