Online vs Offline Training

Your recommendation model was trained on last month's data. New trends emerged. New users appeared. The model is already stale — predictions degrade by the hour. Do you re-train nightly (offline), continuously update weights as events stream in (online), or something in between? Different answers cost orders of magnitude in infrastructure and produce dramatically different freshness/correctness tradeoffs.

Most production ML systems are offline-trained, online-served. A small but important class genuinely needs online learning. Knowing which is which is the interview-grade insight.

Despite all the hype, ~95% of production ML is periodic offline training. Reasons:

• Reproducibility. Re-train on a fixed snapshot → identical weights. Online updates make every model state ephemeral, hard to debug.
• Hyperparameter stability. Tuning a model is an offline experiment. Online updates with the wrong learning rate destroy what you trained.
• Feature monitoring. Offline pipelines run validation, distribution-shift alerts, A/B comparisons. Online updates skip all that.
• Rollback. Offline = checkpoint per training run. Roll back instantly. Online = "we lost the state from 3 hours ago" might be impossible.
• Sufficient freshness. Daily-trained model + real-time features (via feature store) is often as good as continuous training and dramatically simpler.

Online learning earns its complexity in three scenarios:

1. Multi-armed bandits (recommendation exploration). Show ad A or ad B? Don't wait a day to learn — every impression provides signal. Thompson sampling or UCB updates per click.

2. Drift-heavy problems. Twitter trending — what was hot 2 hours ago is dead. Fashion recs in fast-fashion ecommerce. News recommendation.

3. Cold-start personalization. A new user arrives. Static model has no signal. Bandit-style online updates from their first 5 actions converge fast.

Architecture: keep a base model trained offline (the "warm prior"). Apply online deltas on top — usually as a simple linear model or shallow correction layer. If online state is lost, fall back to base model. The two layers compose.

• Training-serving skew. Solved by sharing feature transformation code via a feature store.
• Distribution drift. The world the model was trained on isn't the world it serves. Monitor input feature distributions vs training distributions; alert on KL divergence above threshold.
• Label leakage. Training data accidentally contains future signal. Always use point-in-time joins.
• Ground truth lag. "Did the user buy?" answer arrives hours/days later. Online models must handle delayed labels.
• Catastrophic forgetting (online learning specific). Sustained updates on a new distribution erase old knowledge. Mitigate with replay buffers or elastic weight consolidation.

Recommendation algorithm uses offline training + online feature freshness. News feed retrains ranker daily but applies fresh engagement features online. Leaderboard's scoring weights rarely change — pure static.