A container packages an application with its dependencies so it runs predictably and efficiently across environments. In machine-learning pipelines, this makes training easier to automate and reproduce. section 11.3.2 section 11.3.3
Containers support practical ML operations in several ways:
- Bundle the required Python libraries with the training code instead of depending on each execution environment. section 11.3.3
- Use prebuilt images for frameworks such as TensorFlow, PyTorch, and XGBoost. section 11.3.3
- Scale demanding training across multiple compute engines, although a single node is sufficient for a small real-estate prediction problem. section 11.3.3
Passing training artifacts
Treat the container as a self-contained machine: local files are not automatically available. Store training, validation, test, and configuration artifacts in Cloud Storage, then pass their URIs through environment variables or job arguments. section 11.3.4
Figure 11.30 Training code interactions with external files
Use the filesystem interaction diagram to see why containerized training must receive external artifacts through explicit paths. section 11.3.4
