MLOps interview questions for freshers

MLOps applies DevOps ideas to machine learning so data, training code, and models can be versioned, tested, and released reliably.

Training is where you fit a model on historical data, often with GPUs, large batches, and long-running jobs.

Training jobs are bursty, expensive, and allowed to fail and retry, while serving must stay up with tight SLAs.

A model artifact is the packaged output of training: weights, preprocessing objects, signatures, and metadata needed to run inference.

Inference is running a trained model on new inputs to produce predictions, scores, or embeddings.

A model registry is a catalog of trained models with versions, stages, metrics, and lineage back to data and code.

Each training run can change accuracy, latency, and bias, so you need a unique version to roll forward or back.

A champion alias is a movable pointer such as production that always references the currently live model version.

Promotion is usually a gated step involving the ML owner, sometimes a reviewer, and automated quality checks, not a silent overwrite.

A feature store is a system that computes, stores, and serves features consistently for training and inference.

Ad hoc notebook features drift from production code, get duplicated, and are hard to share.

An entity is the object you join features on, such as user_id, device_id, or merchant_id.

Feature freshness is how recently a feature value was updated relative to the prediction time.

Data drift is a change in the distribution of model inputs compared with the data the model was trained on.

A credit model trained on pre-festival spending will see very different transaction amounts during Diwali.

Ground-truth labels often arrive days or weeks late, so accuracy can be blind in the short term.

PSI compares how a feature's bins have shifted between a reference window and a current window.

Concept drift is when the relationship between features and the target changes, even if input distributions look similar.

Data drift is a change in P(X), the feature distribution.

A UPI fraud model may still see similar transaction amounts, but fraudsters change mule patterns after a new bank rule.