MLOps Problem List

Reading time: ~40 min | Interview relevance: Critical | Roles: MLOps Engineer, ML Platform Engineer, ML Infrastructure Engineer, Production ML Engineer

It is 2 AM and you get paged: a production model's latency spiked from 50ms to 800ms, recommendations are returning stale results, and the data pipeline feeding features has been silently failing for 6 hours. Your engineering manager asks you: "How do we make sure this never happens again?" If you can design the monitoring, alerting, and reliability infrastructure to answer that question, you are an MLOps engineer.

MLOps sits at the intersection of DevOps, data engineering, and machine learning. This list of 45 problems covers the full scope: CI/CD for ML, model serving, monitoring, pipeline orchestration, and infrastructure at scale.

MLOps Interview Structure

Round	Duration	What They Test	Weight
System Design	45-60 min	ML infrastructure architecture	30-35%
Coding	45-60 min	Python, infra scripting, pipeline code	20-25%
ML Operations Knowledge	45-60 min	Tooling, best practices, failure modes	20-25%
DevOps / Cloud	30-45 min	Kubernetes, Docker, cloud services, IaC	15-20%
Behavioral	30-45 min	Incident response, cross-team collaboration	10%

:::tip The MLOps Mindset MLOps is not about building the best model. It is about making sure the best model runs reliably, scales efficiently, and can be updated safely. Think like an SRE who specializes in ML systems. :::

Section 1: CI/CD for ML (8 Problems)

#	Problem	Difficulty	Time	Key Concept	Why It Matters	Company Tags
1	Design a CI/CD Pipeline for ML Model Deployment	Medium	35 min	Testing stages, validation gates, rollback strategy	The foundational MLOps system design problem	FAANG, Unicorns
2	Implement Automated Model Validation Tests	Medium	25 min	Performance thresholds, data checks, regression tests	Models need testing just like software	All
3	Design a Canary Deployment Strategy for ML Models	Medium	30 min	Traffic splitting, metric monitoring, auto-rollback	Safe production rollout of model changes	Google, Meta, Uber
4	Implement a Model Registry with Versioning	Medium	30 min	Model metadata, lineage tracking, artifact storage	Track what is in production and how it got there	All
5	Design a Feature Branch Workflow for ML Experiments	Medium	25 min	Experiment isolation, reproducibility, merge strategy	ML development needs version control discipline	FAANG, Unicorns
6	Implement Automated Data Validation in a Pipeline	Medium	25 min	Schema validation, distribution checks, anomaly detection	Bad data is the #1 cause of ML system failures	Google, Uber, Databricks
7	Design a Blue-Green Deployment for Model Serving	Medium	25 min	Zero-downtime deployment, instant rollback	Minimize risk during model updates	All
8	Build a Reproducibility System for ML Experiments	Hard	35 min	Code versioning, data versioning, environment capture	"But it worked on my machine" is not acceptable	AI Labs, FAANG

:::warning CI/CD for ML is Not Just CI/CD for Software ML pipelines have additional dimensions: data versioning, model artifact management, and performance-based validation gates. A model can pass all unit tests and still be a terrible model. Always include data validation and model performance checks. :::

Section 2: Model Serving & Inference (8 Problems)

#	Problem	Difficulty	Time	Key Concept	Why It Matters	Company Tags
9	Design a Low-Latency Model Serving System	Hard	40 min	Model optimization, batching, caching, load balancing	Serving is where MLOps meets users	FAANG, AI Labs
10	Implement Model A/B Testing Infrastructure	Medium	30 min	Traffic routing, metric collection, statistical analysis	Every model change needs experimental validation	FAANG, Big Tech
11	Design a Multi-Model Serving Architecture	Hard	35 min	Model registry integration, routing, resource isolation	Production systems serve many models simultaneously	Google, Meta, Amazon
12	Optimize Inference Latency by 10x	Hard	35 min	Quantization, distillation, ONNX, TensorRT, batching	Latency optimization is a core MLOps skill	FAANG, AI Labs
13	Design an Auto-Scaling Strategy for Model Endpoints	Medium	30 min	Horizontal scaling, request queuing, cold start mitigation	Traffic patterns are bursty; scaling must match	All
14	Implement a Feature Serving Layer with Consistent Online/Offline Features	Hard	40 min	Feature store architecture, point-in-time correctness	Online/offline feature skew causes silent failures	Uber, Airbnb, Databricks
15	Design a GPU Resource Management System	Hard	35 min	GPU sharing, scheduling, quota management	GPUs are expensive; efficient utilization matters	FAANG, AI Labs
16	Build a Model Caching Strategy	Medium	25 min	Result caching, embedding caching, cache invalidation	Caching reduces cost and latency dramatically	All

Section 3: Monitoring & Observability (8 Problems)

#	Problem	Difficulty	Time	Key Concept	Why It Matters	Company Tags
17	Design an ML Model Monitoring Dashboard	Medium	35 min	Metrics hierarchy, alerting thresholds, visualization	You cannot fix what you cannot see	All
18	Implement Data Drift Detection	Medium	25 min	PSI, KS test, feature distribution monitoring	Data changes cause model degradation	All
19	Implement Model Performance Degradation Detection	Medium	30 min	Delayed labels, proxy metrics, statistical process control	Catch model issues before users notice	All
20	Design an Alerting Strategy for ML Systems	Medium	25 min	Alert hierarchy, severity levels, runbook integration	Too many alerts = alert fatigue = missed incidents	FAANG, Big Tech
21	Implement Prediction Logging and Audit Trail	Medium	25 min	Structured logging, sampling strategy, storage optimization	Debugging production issues requires prediction history	All
22	Design a Root Cause Analysis System for Model Failures	Hard	35 min	Dependency tracking, bisection, automated diagnostics	Fast RCA = fast recovery	Google, Meta, Uber
23	Monitor Fairness Metrics in Production	Hard	30 min	Demographic parity, equalized odds, disparate impact	Fairness monitoring is a regulatory requirement	FAANG, Fintech
24	Design Cost Monitoring for ML Infrastructure	Medium	25 min	Per-model cost attribution, budget alerts, optimization recommendations	ML infrastructure costs can spiral without visibility	All

:::tip The Three Pillars of ML Monitoring

1. Data quality -- Is the input data what the model expects?
2. Model performance -- Is the model still making good predictions?
3. System health -- Is the infrastructure performing reliably?

Most MLOps failures happen because teams monitor only system health and ignore data quality and model performance. :::

Section 4: Pipeline Orchestration & Data (7 Problems)

#	Problem	Difficulty	Time	Key Concept	Why It Matters	Company Tags
25	Design an ML Training Pipeline with Airflow/Kubeflow	Medium	35 min	DAG design, retry logic, dependency management	Orchestrated pipelines are the backbone of ML systems	All
26	Implement an Automated Retraining Pipeline	Medium	30 min	Trigger mechanisms, data freshness, validation gates	Models need regular retraining to stay current	All
27	Design a Data Versioning Strategy for ML	Medium	25 min	DVC, Delta Lake, snapshot management	Reproducibility requires data versioning	All
28	Handle Data Pipeline Failures Gracefully	Medium	25 min	Retry logic, dead letter queues, backfill strategies	Pipelines fail; recovery must be automated	All
29	Design a Feature Computation Pipeline	Hard	35 min	Batch vs. streaming features, backfill, consistency	Features are the lifeblood of ML models	Uber, Airbnb, Databricks
30	Implement Pipeline Idempotency and Exactly-Once Processing	Hard	30 min	Idempotent operations, deduplication, checkpointing	Non-idempotent pipelines produce incorrect data	All
31	Design a Multi-Environment Pipeline (Dev/Staging/Prod)	Medium	25 min	Environment parity, config management, promotion workflow	Code that works in dev must work in prod	All

Section 5: Infrastructure & Scalability (8 Problems)

#	Problem	Difficulty	Time	Key Concept	Why It Matters	Company Tags
32	Design a Distributed Training Infrastructure	Hard	40 min	Data parallelism, model parallelism, communication patterns	Large models require distributed training	Google, Meta, AI Labs
33	Containerize an ML Application with Docker	Easy	20 min	Dockerfile best practices, layer caching, multi-stage builds	Every ML deployment starts with containerization	All
34	Deploy ML Models on Kubernetes	Medium	30 min	Deployment specs, resource limits, health checks, HPA	K8s is the standard ML deployment platform	All
35	Design an Infrastructure-as-Code Setup for ML	Medium	30 min	Terraform/Pulumi, reproducible infrastructure, state management	Manual infrastructure does not scale	FAANG, Big Tech
36	Optimize Cloud Costs for ML Workloads	Medium	30 min	Spot instances, right-sizing, reserved capacity, scheduling	ML infra is expensive; optimization is high-impact	All
37	Design a Secrets and Configuration Management System for ML	Medium	25 min	Vault, environment variables, config injection, rotation	API keys, model endpoints, and feature flags need management	All
38	Implement Efficient Data Loading for Large-Scale Training	Medium	25 min	Prefetching, parallel loading, data format optimization	Data loading is often the training bottleneck	Google, Meta, AI Labs
39	Design a Multi-Cloud ML Platform	Hard	40 min	Portability, vendor lock-in avoidance, unified API	Business requirements sometimes mandate multi-cloud	Big Tech, Enterprise

Section 6: Incident Response & Reliability (6 Problems)

#	Problem	Difficulty	Time	Key Concept	Why It Matters	Company Tags
40	Triage a Production Model Outage	Medium	25 min	Incident response, severity classification, communication	On-call response is a core MLOps responsibility	All
41	Design a Disaster Recovery Plan for ML Systems	Hard	35 min	Backup strategies, RTO/RPO, failover, data recovery	Plan for the worst; hope for the best	FAANG, Big Tech
42	Implement Graceful Degradation for AI Features	Medium	25 min	Fallback models, cached predictions, feature flags	AI features must fail gracefully, not catastrophically	All
43	Design an SLA Framework for ML Services	Medium	25 min	Availability, latency, accuracy SLAs, error budgets	SLAs create accountability and alignment	FAANG, Big Tech
44	Conduct a Post-Incident Review for an ML System Failure	Medium	25 min	Blameless retrospective, timeline, action items	Learning from failures prevents future incidents	All
45	Design a Chaos Engineering Strategy for ML Systems	Hard	30 min	Failure injection, blast radius, steady-state verification	Test failure modes before they test you	Netflix, Google, Amazon

:::danger The MLOps Reliability Hierarchy

1. Can you deploy a model? (Table stakes)
2. Can you roll back a bad deployment? (Essential)
3. Can you detect when a model is degrading? (Differentiator)
4. Can you automatically recover from failures? (Senior-level)
5. Can you prevent failures before they happen? (Staff-level) :::

4-Week MLOps Study Plan

Week	Focus	Problems	Daily Load
Week 1	CI/CD + Serving	#1-16	2-3 problems/day
Week 2	Monitoring + Pipelines	#17-31	2 problems/day
Week 3	Infrastructure + Reliability	#32-45	2 problems/day
Week 4	Integration + Mock	Full system designs	1 deep design + review/day

Week-by-Week Breakdown

Week 1: CI/CD and Serving

Day 1: #1, #2 (CI/CD pipeline, model validation)
Day 2: #3, #4 (canary deployment, model registry)
Day 3: #5, #6 (feature branches, data validation)
Day 4: #7, #8 (blue-green deployment, reproducibility)
Day 5: #9, #10 (low-latency serving, A/B testing infra)
Day 6: #11, #12 (multi-model serving, latency optimization)
Day 7: #13-16 (auto-scaling, feature serving, GPU management, caching)

Week 2: Monitoring and Pipelines

Day 1: #17, #18 (monitoring dashboard, data drift)
Day 2: #19, #20 (performance degradation, alerting)
Day 3: #21, #22 (prediction logging, root cause analysis)
Day 4: #23, #24 (fairness monitoring, cost monitoring)
Day 5: #25, #26 (training pipeline, automated retraining)
Day 6: #27, #28 (data versioning, pipeline failure handling)
Day 7: #29-31 (feature pipeline, idempotency, multi-environment)

MLOps Tooling Landscape

Know the major tools in each category:

Category	Key Tools	When to Mention
Orchestration	Airflow, Kubeflow Pipelines, Prefect, Dagster	Pipeline design questions
Model Serving	TensorFlow Serving, Triton, Seldon, BentoML, vLLM	Inference and serving
Feature Store	Feast, Tecton, Hopsworks	Feature engineering at scale
Experiment Tracking	MLflow, Weights & Biases, Neptune	Reproducibility
Model Registry	MLflow, Vertex AI Model Registry, SageMaker	Model management
Data Validation	Great Expectations, TensorFlow Data Validation, Pandera	Data quality
Monitoring	Evidently, Whylabs, Arize, Fiddler	Model monitoring
Containerization	Docker, Kubernetes, Helm	Deployment
IaC	Terraform, Pulumi, CloudFormation	Infrastructure
CI/CD	GitHub Actions, GitLab CI, Jenkins, Argo CD	Automation

:::note Tool Knowledge vs. Concepts Interviewers care more about concepts than specific tools. Saying "I would use a feature store with online and offline serving" is better than "I would use Feast." But knowing specific tools signals practical experience. :::

Problem Deep Dive: Design a CI/CD Pipeline for ML

This is the single most asked MLOps interview question. Here is a thorough answer framework:

Pipeline Stages

Key Design Decisions

Decision	Options	Recommendation
Trigger for retraining	Scheduled vs. drift-based vs. manual	Start with scheduled, add drift-based as you mature
Validation threshold	Relative (beat prod by X%) vs. absolute	Both: must beat prod AND meet absolute minimums
Deployment strategy	Blue-green vs. canary vs. shadow	Canary for gradual rollout with rollback
Rollback trigger	Manual vs. automated	Automated for critical metrics, manual for nuanced cases

Difficulty Distribution

Difficulty	Problems	Count
Easy	#33	1
Medium	#1, #2, #3, #4, #5, #6, #7, #10, #13, #16, #17, #18, #19, #20, #21, #24, #25, #26, #27, #28, #31, #34, #35, #36, #37, #38, #40, #42, #43, #44	30
Hard	#8, #9, #11, #12, #14, #15, #22, #23, #29, #30, #32, #39, #41, #45	14

Next Steps

After completing the MLOps problem list:

• MLE Problems to strengthen ML fundamentals
• Data Engineer Problems for deeper data infrastructure skills
• Google-Style Problems since Google heavily tests infrastructure thinking
• Section 15: Role-Specific Prep for the full MLOps preparation path