Model governance keeps machine learning systems owned, understandable, auditable, maintainable, and controllable throughout their operational life. It prevents organizations deploying hundreds or thousands of models from losing track of responsibility as systems change or become obsolete. section 9.5.1
Implement a minimum governance record for every production model and its containing system:
- Assign a responsible and empowered system owner within the client organization. section 9.5.1
- Link the model and system artifacts needed for access, understanding, auditing, management, maintenance, and support. section 9.5.1
- Record implementation, changes, reviews, detected problems, responses, withdrawal, and planned deletion of the system and its records. section 9.5.1
- Give owners a way to recognize failure and turn the model off, repair it, or replace it. section 9.5.1
Governance depends on operational evidence. Log events relevant to system operation, including database connections, user logins, data updates, and model decisions; monitor functional behavior as well as nonfunctional performance; and raise alarms when failures emerge or appear imminent. Balance logging density, metric overhead, and retention against application needs and storage cost. section 9.5 section 9.5.1
Detecting domain change
A model predicting music sales from pre-streaming data may fail after streaming becomes widespread. Production monitoring should therefore detect relevant behavioral changes rather than checking only whether the software still runs. section 9.5
