Question

A data science team orchestrates an ML retraining workflow using AWS Step Functions. The workflow includes a SageMaker training job task followed by a model evaluation task. Occasionally the training task fails due to transient resource limits, and when it does, the team wants the workflow to publish a message to an SNS topic notifying operators and then transition to a graceful cleanup state rather than failing the entire execution. Which Step Functions configuration accomplishes this with the least custom code?

Accepted Answer

Add a Catch field to the training task state that routes specified error types to an SNS publish state, which then transitions to the cleanup state.

Answer

Wrap the entire state machine in a Lambda function that uses try/except blocks to invoke SNS and run cleanup logic when the training call raises an exception.

Answer

Add a Retry field to the training task with an unlimited maximum attempt count so the job never reaches a failure state.

Answer

Enable Step Functions execution logging to CloudWatch and create a metric filter that triggers an SNS alarm, then manually run cleanup afterward.