Why architecture matters here
Online learning fails on adversarial input, feedback loops, and lack of rollback. The architecture matters because you need drift detection, clipping, and instant rollback — plus auditing for every model change.
With the pieces mapped, online learning becomes a controlled loop rather than a runaway train.
The architecture: every piece explained
The top strip is the pipeline. Event stream carries user actions + labels. Feature pipeline computes real-time features. Online trainer updates via SGD-style methods. Model registry versions and hosts a shadow.
The middle row is safety. Deployment canaries new models. Safety guards clip updates and validate outputs. Drift detection monitors input + label + prediction distributions. Feedback loop integrates label arrival.
The lower rows are governance. Auditing records changes + provenance. Metrics track online AUC + regret. Ops ensures reproducibility, kill switches, and governance.
End-to-end flow
End-to-end: a ranking model updates every 5 minutes from feedback. New labels stream in; features computed; SGD update. Safety guard clips gradient norms; validates predictions in bounds. Drift detector alarms if input distribution shifts. Auditing logs every checkpoint. If online AUC drops, kill switch reverts to shadow model. Regret metric confirms the online system beats a static baseline.