MLA-C01 Practice Question 48: Multi Container Direct Invocation
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 5 · question 8 of 10
An analytics group runs four low-traffic real-time models, each on its own endpoint: a TensorFlow image model, a PyTorch text model, a scikit-learn churn model, and an XGBoost pricing model. Each endpoint averages under 5% CPU utilization. The models are unrelated, each is called independently by a different application, and all four must answer synchronously with low latency at any time of day. The group wants to cut hosting costs by sharing instances.
Which hosting option should the group use?
Show answer and explanation
Answer:
A multi-container endpoint hosts up to 15 containers, which *can use different models and frameworks*, on the same endpoint instances. With direct invocation, the caller names the container to run in TargetContainerHostname, so each application reaches its own model independently while all four share the same always-on instances. That fits four unrelated models on different frameworks, each lightly used, that still need synchronous low-latency answers (option B).
- A is wrong because a multi-model endpoint uses one shared serving container. It is built for many models on the *same* framework, and TensorFlow, PyTorch, scikit-learn, and XGBoost models cannot all be served from one framework container. - C is wrong because an inference pipeline chains containers, so every request passes through all of them in order. These models are unrelated, and each application needs its own model's output. - D is wrong because serverless endpoints cold-start after idle periods, which conflicts with low latency at any time of day for traffic this sparse. It also gives each model separate capacity instead of sharing instances.
Reference: Multi-container endpoints