MLOps Interview Questions (Easy)

These questions cover the fundamentals that usually appear in junior-to-mid MLOps, ML platform, or ML infrastructure interviews.

Foundations

1. What is MLOps?

MLOps is the discipline of building, deploying, monitoring, and governing machine learning systems in production. It extends DevOps by treating data, features, and models as first-class artifacts.

2. How is MLOps different from traditional DevOps?

Traditional DevOps mainly manages code and infrastructure. MLOps must also manage training data, model artifacts, experiment metadata, and model quality after deployment.

3. What is the difference between training and inference?

Training is the process of fitting a model on historical data. Inference is the process of using the trained model to generate predictions on new data.

4. What is model serving?

Model serving is the process of exposing a trained model so applications can call it for predictions, usually through an API, batch job, or stream consumer.

5. What is a model artifact?

A model artifact is the saved output of training, such as a serialized model file, along with metadata needed to load and run it.

6. Why is reproducibility important in MLOps?

Because you must be able to trace a production model back to the exact code, data, parameters, and environment that created it. Without that, debugging and compliance become difficult.

7. What is experiment tracking?

Experiment tracking records parameters, metrics, code version, and artifacts for each training run so teams can compare results and reproduce the best model.

8. What is a model registry?

A model registry stores versioned models and their metadata, and often supports lifecycle stages such as staging, approved, and production.

Data And Features

9. Why is data versioning needed in MLOps?

Because changing the dataset can change the model even when the code stays the same. Versioning helps reproduce training runs and compare models fairly.

10. Why not store large datasets directly in Git?

Git is not designed for large binary data and becomes slow and bloated. Tools like DVC store metadata in Git and the actual data in object storage.

11. What is DVC?

DVC, or Data Version Control, is a tool that tracks datasets and ML artifacts while integrating with Git for metadata and remote object storage for data.

12. What is a feature in machine learning?

A feature is an input variable used by a model to make a prediction, such as age, transaction amount, or device type.

13. What is a feature store?

A feature store is a system that manages reusable features for both training and inference, helping keep transformations consistent across environments.

14. What is training-serving skew?

Training-serving skew happens when the feature processing used in production differs from what was used during training, causing incorrect or degraded predictions.

15. What is label leakage?

Label leakage happens when information that would not truly be available at prediction time leaks into training data, making offline evaluation look better than real production behavior.

Deployment And Delivery

16. What is CI, CT, and CD in MLOps?

CI validates code and tests. CT, or continuous training, retrains models when needed. CD deploys model-serving changes safely into production.

17. What is batch inference?

Batch inference runs predictions over large datasets in bulk, often on a schedule. It is common for reporting, recommendations, or nightly scoring jobs.

18. What is online inference?

Online inference returns predictions in real time for a live request, usually with strict latency requirements.

19. What is async inference?

Async inference accepts requests quickly and processes them later, which is useful when predictions take too long for a synchronous API response.

20. What is shadow deployment?

Shadow deployment sends production traffic to a new model without returning its predictions to users. It lets you compare behavior safely before exposure.

21. What is canary deployment for models?

Canary deployment sends a small percentage of production traffic to a new model and increases traffic gradually if metrics remain healthy.

22. What is A/B testing in MLOps?

A/B testing exposes different user groups to different models so you can compare business outcomes such as conversion rate or click-through rate.

23. What is champion-challenger?

It is a setup where the current production model is the champion and a candidate model is the challenger. The challenger is evaluated before full promotion.

Monitoring And Quality

24. What is data drift?

Data drift means the input data distribution in production is changing compared with the training baseline.

25. What is concept drift or model drift?

Concept drift means the relationship between inputs and outputs changes over time, so a model that used to be accurate becomes less useful.

26. Why is monitoring model accuracy harder than monitoring uptime?

Because labels often arrive late or not at all, so you may not know true accuracy in real time. You often need proxy metrics and delayed evaluation.

27. What kinds of metrics do you monitor for a model-serving system?

Latency, throughput, error rate, resource usage, feature null rates, confidence distribution, drift signals, and business KPI impact.

28. Why can a model endpoint be healthy but still be wrong?

Because the infrastructure may be serving requests correctly while the model, features, or input schema are incorrect.

Infrastructure

29. Why are GPUs used in MLOps?

GPUs accelerate parallel numerical operations and are often needed for deep learning training and high-performance inference.

30. Why are GPUs handled differently from CPUs in Kubernetes?

They are limited, expensive, and often require device plugins, node isolation, and special scheduling controls such as taints and tolerations.

31. What is cold start in model serving?

Cold start is the latency hit when a model or runtime is loaded for the first time, especially when large model weights must be loaded into memory or GPU.

Governance

32. Why is lineage important in MLOps?

Lineage tells you which code, data, features, and model version produced a given prediction service. It is essential for debugging, audits, and rollback.

33. Why does MLOps care about PII and compliance?

Because models may be trained on sensitive data, and the platform must enforce access control, auditability, and safe handling of regulated information.

34. What is rollback in MLOps?

Rollback means restoring a previous trusted model, feature pipeline, or serving configuration when the new model degrades production behavior.

35. What is LLMOps?

LLMOps is a specialization of MLOps focused on large language models, prompt versioning, evaluation, retrieval pipelines, inference cost, and safety controls.