MLOps Interview Playbook

Use this file for MLOps, ML platform, ML infrastructure, inference platform, and model-serving interviews.

What Interviewers Are Really Testing

1. Reproducibility

In MLOps, a deployment is not just code. A strong answer shows that you understand versioning across:

• code

• data

• features

• model artifacts

• configuration

2. Training Versus Inference

Interviewers want to know that you understand the difference between:

• offline training workloads

• batch inference

• real-time inference

• asynchronous scoring

Each one has different latency, cost, autoscaling, and observability requirements.

3. Data And Feature Discipline

Many MLOps failures are not "server down" failures. They are:

• schema mismatch

• feature skew

• stale features

• label leakage

• data drift

You should be able to talk about these as production risks, not just data science details.

4. Safe Model Delivery

A good MLOps engineer knows how to move a model safely from experiment to production:

• experiment tracking

• registry promotion

• validation gates

• shadow deployment

• canary rollout

• rollback

5. Platform And Cost Awareness

Senior MLOps answers should include:

• GPU scheduling

• autoscaling and queueing

• checkpointing long jobs

• spot or preemptible trade-offs

• throughput versus latency

• cost per training run or cost per inference

6. Monitoring Beyond Uptime

In traditional DevOps, 200 OK might mean success. In MLOps, a service can be technically healthy and still wrong. You should discuss:

• drift

• confidence distribution

• feature null rates

• data freshness

• business KPI impact

• delayed ground truth

A Strong MLOps Answer Framework

For most interview questions, use this structure:

1. Identify whether the problem is in data, training, model, serving, or platform.

2. Clarify the artifact versions involved: code, data, features, model, config.

3. Explain how you would validate correctness before optimizing performance.

4. Talk about safe rollout, rollback, and observability.

5. Close with reproducibility and prevention controls.

Example:

I would first separate platform health from model correctness. If the endpoint is healthy but predictions are wrong, I would validate input schema, feature transformation parity, model version, and registry lineage before changing infrastructure.

What You Should Know By Topic

Lifecycle And Artifacts

• code, data, features, model, and metadata as first-class artifacts

• experiment tracking

• registry promotion and approval

• lineage from training run to production endpoint

Data And Feature Management

• data validation gates

• feature stores

• offline versus online features

• training-serving skew

• schema evolution and contracts

Pipelines

• CI for code and tests

• CT for retraining

• CD for model serving

• pipeline orchestration tools such as Kubeflow, Airflow, or MLflow workflows

Serving Patterns

• batch inference

• online inference

• async inference

• A/B testing, shadow, canary, champion-challenger

• REST or gRPC serving

Platform And Infrastructure

• Kubernetes scheduling

• GPU nodes and device plugins

• autoscaling

• model loading and cold start behavior

• storage for datasets, features, and model artifacts

Observability

• service latency, throughput, error rate

• drift and freshness metrics

• confidence score shifts

• online versus offline evaluation

• feedback loops and delayed labels

Security And Governance

• PII handling

• secrets and credentials

• access to datasets and models

• audit trail for promotion and rollback

• approval gates for regulated environments

LLMOps As An Advanced Specialization

For modern MLOps interviews, it also helps to mention:

• prompt and evaluation versioning

• retrieval pipeline quality

• token cost controls

• guardrails and fallback models

• latency and throughput trade-offs for large models

Must-Know Commands And Checks

DVC

• dvc add <path>

• dvc push

• dvc pull

• dvc status

• dvc repro

MLflow

• mlflow ui

• mlflow models serve -m <model_uri>

• run metadata, params, metrics, and model registry stages

Kubernetes And GPU Checks

• kubectl get pods

• kubectl describe pod <name>

• kubectl logs <pod>

• kubectl describe node <name>

• kubectl get events --sort-by=.lastTimestamp

• nvidia-smi

Serving Validation

• curl a sample payload against the inference endpoint

• compare expected features and payload schema

• inspect model version, registry stage, and serving config

High-Value Scenarios To Practice

Wrong Predictions With 200 OK

Mention:

• input schema validation

• feature parity between training and serving

• model version and registry lineage

• stale or missing features

Data Drift Or Concept Drift

Mention:

• statistical comparison against training baseline

• delayed-label problem

• proxy metrics such as confidence or output distribution

• retraining trigger rules

GPU Pods Stuck In Pending

Mention:

• device plugin health

• available GPU capacity

• taints and tolerations

• driver and CUDA compatibility

Latency Spike After A New Model Release

Mention:

• model size or cold start

• batch size and concurrency

• CPU versus GPU inference choice

• canary rollback or traffic shift

Expensive Training Jobs

Mention:

• spot or preemptible workers

• checkpointing

• data locality

• artifact caching

• experiment pruning

Strong Signals In Senior MLOps Answers

• You distinguish platform health from model quality.

• You talk about lineage and reproducibility without being prompted.

• You mention rollout safety and rollback for models, not just services.

• You connect data quality and feature quality to production risk.

• You think about cost, GPU utilization, and throughput, not just model accuracy.

Common Weak Signals

• Treating MLOps as ordinary CI/CD with a notebook attached

• Ignoring data versioning

• Saying the model is fine because the endpoint returned 200

• Recommending retraining without defining a trigger or validation gate

• Ignoring feature skew and delayed labels

Final Revision Checklist

• I can explain the difference between DevOps and MLOps.

• I can explain reproducibility across code, data, features, and model versions.

• I can explain model registry, feature store, experiment tracking, and drift.

• I can discuss batch, online, and async inference trade-offs.

• I can explain safe rollout patterns for new models.

• I can troubleshoot wrong predictions even when infrastructure is healthy.

• I can discuss GPU scheduling, latency, and cost trade-offs at a senior level.