MLOps Engineer Interview Questions:
What UK Employers Actually Ask

MLOps interviews are unlike most engineering interviews. They test a combination of infrastructure depth, ML domain knowledge, operational thinking, and system design — and the balance differs significantly from a standard software engineering interview process. Here's what to expect and how to prepare.

The Interview Process at UK Companies

Most UK MLOps interview processes follow four stages:

1. Recruiter screen (30 min): Background, motivation, salary expectations. Nothing technical.
2. Technical phone screen (45–60 min): Infrastructure concepts, Python scripting question, high-level ML pipeline knowledge. Used to filter candidates who lack basic technical foundations.
3. Take-home task (4–8 hours): Often an infrastructure problem — "here's a broken deployment configuration, diagnose and fix it" or "write a CI/CD pipeline for this ML training job". Sometimes a design document exercise.
4. Final interview loop (3–4 rounds in a day): Infrastructure deep-dive, system design, ML concepts, behavioural/culture round.

Infrastructure Questions You Must Be Able to Answer

These appear in almost every MLOps technical screen:

• "Walk me through how Kubernetes schedules a pod. What happens when resources are unavailable?"
• "How would you manage GPU resource allocation for multiple concurrent training jobs on a shared cluster?"
• "What's the difference between a Kubernetes Deployment and a StatefulSet? When would you use each for ML workloads?"
• "How do you manage secrets in a Kubernetes-based ML platform?"
• "Walk me through how you'd structure Terraform code to provision a cloud ML training environment reproducibly."

ML Pipeline Questions

• "How would you design a training pipeline that automatically retrains a model when data quality drops below threshold?"
• "What would you include in a CI/CD pipeline for an ML model? What gates would you add before production deployment?"
• "How do you handle the problem of training-serving skew? What causes it and how do you prevent it?"
• "What does your ideal experiment tracking setup look like? What do you log and why?"

Model Monitoring Questions

• "What's the difference between data drift and concept drift? How would you detect each in production?"
• "What metrics would you monitor for a classification model in production? What thresholds would trigger retraining?"
• "A model that was performing well last month is now performing 8% worse by your key metric. Walk me through how you'd diagnose this."
• "How would you design a monitoring system that doesn't require labelled data to detect model degradation?"

System Design: Design an ML Platform for 50 Data Scientists

This is one of the most common senior MLOps system design questions. Here's how to structure a strong answer:

Start with requirements clarification (2–3 minutes): What workloads? CPU or GPU? Real-time serving or batch? Any compliance requirements? Existing cloud provider?

Identify the key components:

• Compute management: Kubernetes cluster with GPU nodes, resource quotas per team, spot instances for training cost reduction
• Experiment tracking: MLflow (self-hosted for data governance) or W&B, with a defined logging standard
• Data and model versioning: DVC for dataset versioning, MLflow model registry for model artefacts
• Pipeline orchestration: Airflow or Prefect for scheduled jobs, GitHub Actions for CI/CD triggers
• Serving infrastructure: Kubernetes deployments with Horizontal Pod Autoscaler, canary deployments for new model versions
• Monitoring: Evidently AI for data drift, Prometheus/Grafana for infrastructure, custom dashboards for ML metrics

Discuss trade-offs explicitly: "I'd recommend self-hosted MLflow over W&B here because the data governance requirements you mentioned suggest you don't want experiment data leaving your cloud boundary."

End with the phased rollout: Don't try to build everything at once. Suggest a phase 1 (experiment tracking and basic serving), phase 2 (automated retraining), phase 3 (monitoring and governance).

Frequently Asked Questions

How much coding vs infrastructure?

Roughly 30% Python scripting and 70% infrastructure/architecture discussion. The coding round involves pipeline scripts or deployment configurations, not algorithm problems.

What system design questions are most common?

"Design an ML platform for N data scientists" and "How would you architect continuous training for a model that degrades in production?" — both test trade-off reasoning, not tool memorisation.

Do I need to know ML to pass the technical screen?

Yes, at a conceptual level. You need to discuss training pipelines, evaluation metrics, and model lifecycle intelligently. You don't need to implement ML algorithms.

How long is the process at UK companies?

Typically 3–5 weeks: recruiter screen, technical phone screen, take-home task, and final loop of 3–4 rounds.

What separates good from great answers?

Explicit trade-off reasoning and operational awareness. State your choice, explain why you'd make it over alternatives, and address how you'd operate and monitor it in production.