Domain 4: Serving and Scaling Models¶
Official exam weight: approximately 20%
Official objectives¶
4.1 Serving models¶
- Deploy batch and online inference using appropriate services such as Agent Platform, Model Garden, Cloud Run, and GKE.
- Package and serve models such as PyTorch and XGBoost using prebuilt or custom containers.
- Organize and version models in Gemini Enterprise Agent Platform Model Registry.
- Implement A/B testing and canary deployments to compare model versions.
- Develop inference preprocessing and postprocessing.
4.2 Scaling online model serving¶
- Manage and serve features with Agent Platform Feature Store.
- Deploy models to public and private endpoints.
- Choose CPU, GPU, TPU, or edge hardware.
- Scale the serving backend based on throughput, including Agent Platform Inference and containerized serving.
- Tune models for production training and serving requirements.
Guided labs to build¶
| Status | Lab | Objective |
|---|---|---|
| Planned | Compare batch and online inference | 4.1 |
| Planned | Package a PyTorch or XGBoost model in a custom container | 4.1 |
| Planned | Register and version a model | 4.1 |
| Planned | Deploy a model and validate online predictions | 4.1 |
| Planned | Implement a canary or traffic-split rollout | 4.1 |
| Planned | Compare public and private endpoints | 4.2 |
| Planned | Scale serving capacity for throughput and latency | 4.2 |
Google Skills¶
Open the official PMLE learning path for the current model deployment, prediction, and serving activities.
Exam focus¶
Be ready to choose batch versus online prediction, managed versus containerized serving, and public versus private access. Understand how preprocessing, model versioning, rollout safety, hardware, latency, throughput, and cost affect the design.
📬 DevopsPilot Weekly — Learn DevOps, Cloud & Gen AI the simple way.
👉 Subscribe here