MLOps interview questions and answers

Prediction logging stores request features, model version, and outputs so you can debug, compute delayed metrics, and retrain.

Alerts notify on-call when SLOs break, such as p99 latency, drift PSI, or a drop in prediction volume.

You split live traffic between a control model and a challenger and compare business and quality metrics.

Control is usually the current production model.

Offline data can be biased, delayed, or missing feedback loops that exist in production.

Batch inference scores a large set of records on a schedule, writing results to a table or file.

Online inference answers a request within a tight latency budget, often tens of milliseconds, using live features.

Choose batch when predictions can be precomputed, traffic is huge, or features are only available daily.

A scoring job is a scheduled or triggered pipeline that loads a registered model, reads a batch of features, and writes predictions.

A feature pipeline is the automated path from raw events to training tables and serving stores.

If train uses log(amount + 1) and serve uses raw amount, the model sees a different world and quality collapses.

Reproducibility means you can recreate a model and its metrics from recorded code, data, configs, and environment.

Seeds control weight init, data shuffles, and some augmentations so two runs are comparable.

Pinning locks package versions in requirements.txt or a lockfile so numpy or CUDA updates do not silently change results.

DVC is Git-friendly version control for datasets, models, and pipelines.

Git is designed for small text diffs, not multi-gigabyte Parquet dumps.

dvc add starts tracking a file or folder, moves the content to the DVC cache, and writes a small .dvc metadata file you commit to Git.

A remote is the shared storage backend, typically S3, GCS, or Azure Blob, where cached data and models are pushed.

A canary sends a small slice of live traffic to a new model version while the rest stay on the champion.

A full cutover exposes every user to unknown regressions immediately.