How are production ML systems typically handling distribution shift over time? [D]

In deployed ML systems, data distribution drift seems unavoidable over longer time horizons. I’m trying to understand what approaches are commonly used in practice: Continuous retraining pipelines (fixed intervals vs trigger-based) Online monitoring for feature or prediction drift Use of shadow models or fallback models in production Human-in-the-loop review for edge cases In most real deployments I’ve seen discussed, retraining strategy seems more operationally constrained than model-related. Curious what