Objective 4.2

PMLE

Scaling online model serving

Objective 4.2 sits in Serving and scaling models, which carries 20% of the ML Engineer exam. The questions below are original, written from the official objective title above, and each explanation cites the Google Cloud page it rests on.

Objective title verbatim from the official objectives. Google Cloud exam page

A worked example

Shown solved, with the whole explanation open: this is what every question here carries.

4-2Serving and scaling models

A team wants GPU accelerated online serving for an XGBoost model exported from training. What does the accelerator rule say?

GPUs are unavailable for that model typeCorrect · your answerCorrect. That runtime never reaches the hardware.
GPUs work with any prebuilt serving containerA prebuilt image for that framework has no device path.
GPUs work once the model is registeredRegistration does not change the serving runtime.
GPUs require at least four replicas deployedReplica count is not what gates accelerator use.

Correct.

Concept

An accelerator is only usable when the serving code knows how to reach it. A framework whose runtime has no device path gains nothing from attached hardware, so the platform refuses the combination outright.

Why A

You can only use GPUs when the model resource is based on a TensorFlow SavedModel, or when you use a custom container designed to take advantage of GPUs. GPUs cannot be used for scikit-learn or XGBoost models.

Source

You can only use GPUs when your Model resource is based on a TensorFlow SavedModel , or when you use a custom container that has been designed to take advantage of GPUs. You can't use GPUs for scikit-learn or XGBoost models.

Configure compute resources for inference, Agent Platform documentation, checked August 2026
#gcp#serving#gpu#frameworks

Now you: objective 4.2 questions

No account needed. The explanation opens when you answer.

Sample question 1 of 3

4-2Serving and scaling models

A team benchmarks a model on a single core machine type and the result looks worse than expected. What does the platform reserve on each replica?

Sample question 2 of 3

4-2Serving and scaling models

A team assumes the cheapest machine type per hour will give the cheapest serving bill. What does the cost guidance point out?

Sample question 3 of 3

4-2Serving and scaling models

An engineer wants to mix two accelerator models behind one deployed model to spread capacity around. What does the rule allow?

Full ML Engineer question bank coming

We’re writing the complete bank from the official objectives right now. Leave your email and we’ll tell you when it ships, nothing else, ever.

Read the sources

These are the official pages the questions above cite. Reading them is studying the objective from the primary source, which is what the explanations point you toward anyway.

More objectives in Serving and scaling models