MLOps Engineer interview question
What is training-serving skew?
Training-serving skew means production features or preprocessing differ from training in definition, code, timing or distribution. Prevent it with shared transformations, point-in-time joins, versioned schemas, parity tests and monitoring.
Direct answer
What a strong answer covers
How it works
Skew arises whenever the inputs a model sees in production are computed differently from the ones it trained on: a feature implemented twice in SQL for training and in Python for serving, a default value that differs, a timestamp that means event time in one place and processing time in the other, or a distribution that shifts because the serving population differs from the training sample. The model runs and returns numbers, but its accuracy silently degrades. Prevention is structural: one implementation of each transformation shared by both paths, point-in-time joins for training data, versioned feature schemas, and a parity test that scores the same records through both paths and compares.
Example
A fraud model trained on a feature defined as transactions in the last 24 hours computed by a batch job at midnight, while the serving path computed it as a rolling 24-hour window in real time; the two disagreed for most of the day and production recall was 20 points below validation. Moving both to a shared feature definition materialised through the same code and adding a daily parity check on 1,000 sampled records closed the gap.
Common mistakes
- Reimplementing feature logic in the serving language and assuming it matches.
- Training on features computed at the end of the day for predictions made in the morning.
- Never comparing production feature values with training values for the same records.
A useful answer structure
Answer from your own work.
Do not memorise a script: use only examples you can explain honestly in follow-up questions.
Open the desktop workspace and rehearse a clear answer in your own words.