MLOps interview questions: pipelines, deployment and monitoring

Twelve common MLOps interview questions

1. Describe the ML lifecycle and where MLOps fits.
2. How do you make training reproducible?
3. What is a feature store, and what problem does it solve?
4. What does a model registry do?
5. How is CI/CD for ML different from CI/CD for software?
6. Batch versus online inference — how do you choose?
7. How do you roll out a new model safely?
8. What is data drift versus concept drift?
9. What is training–serving skew?
10. What would you monitor for a production model?
11. When and how do you retrain?
12. What changes for LLM applications?

Reproducibility and versioning

A training run is reproducible when you can rebuild the same model from versioned code, versioned data, a pinned environment and recorded configuration and random seeds. Experiment trackers record parameters, metrics and artefacts per run; data versioning tools or immutable dataset snapshots record exactly what was trained on; a model registry stores approved models with metadata and lineage, and manages stages such as staging and production.

Feature stores and skew

Training–serving skew happens when features are computed differently in training and production — a different code path, a time-window bug, or data that was only available in hindsight (leakage). A feature store addresses this by defining features once and serving the same logic for offline training data and low-latency online lookups, with point-in-time correct joins so training never sees future values.

Serving and rollout

Monitoring

• Service health: latency, throughput, errors, resource use.
• Data quality: missing values, schema changes, out-of-range inputs.
• Data drift: input distributions shift (population stability index, KS tests).
• Concept drift: the relationship between inputs and outcomes changes; visible only when labels arrive.
• Model performance: accuracy or business metrics on labelled feedback, segmented by group.

Tools you may be asked about

Interviewers rarely require a specific stack, but they expect you to map concepts to tools you've used: an experiment tracker and model registry such as MLflow or Weights & Biases, an orchestrator such as Airflow, Kubeflow or a cloud ML pipeline service, a feature store, containerisation with Docker and Kubernetes for serving, and a monitoring setup for both system and model metrics. Describe what each did in your pipeline and what you would change, rather than listing names. See Airflow interview questions and Kubernetes interview questions.

Retraining strategy

Retrain on a schedule when data changes predictably, on triggers when monitoring shows drift or degraded performance, or continuously for fast-moving domains. Whatever the trigger, the retrained model goes through the same validation gates — offline metrics against the current production model, fairness and slice checks — before rollout. Automating retraining without automated validation is a common and costly mistake.

LLMOps

For LLM applications the model is often an API, so operations shift to prompts, retrieval and evaluation: version prompts and their test sets, run evaluation suites on every change, monitor cost per request, latency, refusal and error rates, and track quality through user feedback and sampled reviews. Retrieval indexes need their own freshness monitoring. See RAG interview questions.

Practising

Interviewers follow up on the most specific thing you say, so rehearse the second question as well as the first. CoPilot Interview, a free AI interview assistant, runs MLOps mock interviews with follow-ups and produces a report of likely next-round questions. Prepare one story about a production model problem you diagnosed — skew, drift or a bad rollout — with how you'd prevent it now. See also machine learning engineer interview questions.

Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly. Everything else on this page applies whichever tool you use, or none.

Frequently asked questions

What is MLOps in an interview answer?

The practices and tooling for taking ML models to production and keeping them reliable: reproducible training, versioning, CI/CD and continuous training, serving, safe rollout and monitoring.

What is the difference between data drift and concept drift?

Data drift is a change in input distributions. Concept drift is a change in the relationship between inputs and the target, so the same inputs now mean something different.

What is training–serving skew?

A mismatch between how features are computed in training and in production, which makes production performance worse than offline metrics predicted.

How do you deploy a new model safely?

Validate offline against the current model, run it in shadow mode on live traffic, then canary it to a small share of users and expand while monitoring metrics, with a fast rollback path.