You find a 12 percent PSI between logged serving features and training features for the same ids. How do you debug?
PICTURE THIS: DATA SPLIT
Fit on train, tune on val, report on test once.
Simple meaning
Join on entity and event time, diff column by column, and check timezone, window closure, and online TTL.
WHY — Training-Serving Skew instead of guessing?
Why interviewers care about Training-Serving Skew:
people who only read docs from people who shipped.
and tied to MLOps work.
Name the idea, why it exists, then one short example.
End with when you use it and one common pitfall.
STEPS — What happens step by step?
Before you speak the answer, walk the interviewer through these steps:
- 1Join on entity and
event time, diff column by column, and check timezone, window closure, and online TTL.
- 2Compare library versions and
default values for missing keys.
- 3Fix the shared definition,
backfill, and only then retrain if the model saw the wrong world.
- 4Give an example
One tiny concrete case you can say aloud.
- 5Common mistake
What juniors usually get wrong.
- 6Close
When you pick this over the alternative.
EXAMPLE — See it in action
Here's a short line you can speak, broken into clear beats:
Note: Adapt this scaffold to your own project — keep it under 60–90 seconds.
Key takeaway
Join on entity and event time, diff column by column, and check timezone, window closure, and online TTL. Compare library versions and default values for missing keys.