Keeping models current.
WHAT THE OPTIONS ARE
On a schedule When performance degrades When drift is detected When enough new data accumulates
WHAT SCHEDULED RETRAINING PROVIDES
Simplicity, and predictable freshness.
WHAT IT COSTS
Retraining when unnecessary.
WHAT TRIGGERED RETRAINING PROVIDES
Retraining only when warranted.
WHAT IT REQUIRES
Reliable detection, which is harder than it sounds.
WHAT MOST TEAMS SHOULD DO
Schedule it, at a frequency informed by observed decay.
WHAT TO DECIDE
The training window: all history, or recent data only.
WHAT RECENT-ONLY PROVIDES
Adaptation to current conditions.
WHAT IT COSTS
Forgetting patterns that recur.
WHAT TO NEVER DO
Deploy a retrained model without evaluating it.
WHY
Retraining can produce a worse model, particularly on drifted or broken data.
WHAT TO COMPARE
The new model against the current one, on the same held-out data.
WHAT TO AUTOMATE
The pipeline: data, training, evaluation, comparison.
WHAT TO KEEP MANUAL
The decision to promote, unless the comparison is thoroughly trustworthy.
WHAT TO RECORD
Every trained model, its data and its results.
