MLA-C01 Practice Question 27: Inference Pipeline Preprocessing
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 3 · question 7 of 10
A scikit-learn preprocessor fitted during training (imputation values, a scaler, and a category encoder) is currently re-implemented inside three client applications, which transform each record before calling the model endpoint. The three implementations have drifted apart and one of them now produces a different encoding than training did. The team wants a single place that applies the exact fitted preprocessing artifacts and then the model, keeps added latency to a minimum, and does not introduce another endpoint or job for the clients to call.
Show answer and explanation
Answer:
An inference pipeline is a single SageMaker AI model made of two to fifteen containers that run in sequence. SageMaker AI installs all of them on the same EC2 instances in the endpoint, so the response of the preprocessing container is passed straight to the model container with no network hop. That reuses the data transformers built for training, the fix for the drift, behind the one endpoint the clients already call, and it costs nothing beyond the instances the endpoint already runs (option A).
- B is wrong because the preprocessing logic is still separate from the model artifact and can drift from it again. It also adds a Lambda invocation plus an extra network hop to every request. - C is wrong because it gives clients two endpoints to call in order, adds a full network round-trip per request, and doubles the hosting bill for the same traffic. - D is wrong because processing jobs are batch compute: each one provisions instances and takes minutes to start, which is unusable in a per-request path.
Reference: Inference pipelines in Amazon SageMaker AI