MLA-C01 Practice Question 14: Managed Spot Checkpointing
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 2 · question 4 of 10
A research team retrains a custom PyTorch model a few times each month. Every run takes about 30 hours on GPU instances and is currently billed at On-Demand rates. The team wants to cut training cost substantially and can tolerate runs finishing later than they do today, but after an interruption a run must not start over from the first epoch. The team does not want to make any long-term spending commitment or manage EC2 capacity itself.
Which configuration meets these requirements?
Show answer and explanation
Answer:
Managed Spot Training runs the job on EC2 Spot capacity that SageMaker AI manages for you, reducing training cost by up to 90% compared with On-Demand. Spot capacity can be interrupted, so AWS recommends checkpointing for anything that does not finish quickly: the script saves checkpoints to a local path, SageMaker AI copies them to the configured S3 location, and when the job restarts it copies them back so training resumes from the last checkpoint instead of from scratch. MaxWaitTimeInSeconds covers both the time spent waiting for Spot capacity and the run itself, which is why it must be at least as long as MaxRuntimeInSeconds.
Option A does not meet the requirement. Whatever the two timeouts are set to, it runs without checkpoints, so an interrupted 30-hour run restarts from the beginning, which the stem forbids.
Option B addresses a different problem. Warm pools reduce the startup latency of back-to-back jobs by keeping instances provisioned; they do nothing to lower the On-Demand hourly rate.
Option D can save money, but it hands the team the capacity management it wants to avoid, and checkpoints on the instance's local storage disappear with the instance when Spot reclaims it.
Reference: Managed Spot Training in Amazon SageMaker AI