MLA-C01 Practice Question 1: SageMaker Processing Job Preprocessing
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 1 · question 1 of 10
A data science team runs a pandas and scikit-learn feature engineering script by hand inside an Amazon SageMaker Studio notebook. The raw dataset in Amazon S3 has grown to roughly 200 GB and the notebook now runs out of memory. The team wants the script to run on managed compute that is provisioned on demand and released when the script finishes, reads and writes S3, and can later be reused as a step in an automated ML workflow, with as few changes to the script as possible.
Which approach meets these requirements?
Show answer and explanation
Answer:
SageMaker Processing exists for exactly this job: pre- and post-processing, feature engineering, and model evaluation on fully managed infrastructure. SageMaker AI launches the instances, copies the input data from S3 into the container, runs your script, writes the outputs back to S3, and then releases the compute; you pay only for the duration of the job. The built-in scikit-learn container already has pandas and scikit-learn installed, so an existing script usually needs only its input and output paths changed. A Processing job also maps directly onto a ProcessingStep in SageMaker Pipelines later.
Option A is a stopgap, not a solution. A bigger notebook instance runs continuously and bills continuously, it still requires a human to launch the job, and it cannot be dropped into an automated pipeline.
Option B abuses the training job API. Training jobs are designed to emit a model artifact to /opt/ml/model; using one purely to transform data forces you to package the dataset as a model tarball and gives you no clean way to produce multiple named output channels.
Option D does not fit the workload. Lambda has a 15-minute maximum execution time, limited ephemeral storage and memory, and is a poor fit for a 200 GB batch transformation; the same memory pressure that broke the notebook would break the function.
Reference: Data transformation workloads with SageMaker Processing