Data availability means required data can be fetched reliably for training and remain available during runtime. The goal is not merely storing data, but making necessary sources accessible through a practical command, call, or pipeline. section 6.2.1
Treat availability as an operational checklist:
- Make combined data sources fetchable through a repeatable process rather than manual assembly. section 6.2.1
- Check that the data used for training can also be obtained during inference; missing serving data produces incomplete or inaccurate predictions. section 6.2.1 section 11.1.1
- Balance storage design against dataset size and growth instead of overengineering small, stable datasets. section 6.2.1
- Test external dependencies, monitor feature performance, and track feature versions and dependencies as data changes. section 11.1.1
From API response to serving features
A pipeline can fetch JSON from a third-party API and store it as CSV or Parquet for training. Later, the serving path must still obtain the necessary source data; otherwise, features may be missing even when the model itself is healthy. section 6.2.1 section 11.1.1
