Interview questions in MLOps & Infrastructure, page 3
Diagnosing poor distributed training scaling
Communication overhead (gradient all-reduce, interconnect), data-loading starvation, load imbalance, and small per-GPU batches; profile with the PyTorch profiler and Nsight.
Robust checkpointing strategy for multi-day training jobs and seamless resumption
Tests production-grade distributed training reliability. Cover async atomic checkpoints, MTBF-based cadence, tiered storage, and recovery drills. Red flag: blocking synchronous writes that ignore silent corruption or straggler finalization.
Design training job submission to a shared Kubernetes cluster
Gateway with artifact caching; namespace quotas; GPU schedulers like Volcano; Prometheus metrics and cost attribution.
Compare Airflow and Kubeflow for ML training pipelines
Tests orchestrator-to-workload fit. Strong answers contrast Airflow's data integration and Python DAGs with Kubeflow's K8s scaling, container reproducibility, and experiment tracking. Red flag: claiming one is always better without stage-specific reasoning.
Design multi-tenant GPU cluster scheduling and preemption policies
Tests ability to design fair GPU scheduling preventing starvation and noisy-neighbor issues. Answer: Kueue for fair-share, namespace quotas with MIG, priority classes with backoff.
Key differences between traditional and ML CI/CD pipelines?
Contrast code deploys with data versioning, model registries, and retraining; note holdout eval.
What automated tests belong in CI before deploying a classification model?
Name data schema checks, performance regression vs baseline, bias audits, and artifact integrity.
What is a model registry's purpose in CI/CD4ML and its CI/CD interaction?
Tests if you see the model registry as the bridge between experiments and production, not just storage. A strong answer explains how CI publishes validated artifacts and CD consumes versioned models. Red flag: calling it a passive file dump without versioning.
What events trigger automatic model retraining beyond code changes?
This tests whether you treat ML pipelines as event-driven systems, not just software CI/CD. A strong answer lists data drift, scheduled cron jobs, production metric degradation, schema changes, and upstream data pipeline completion.
How would you implement shadow deployment and which metrics justify promotion?
Tests zero-impact validation when feedback loops are broken. Mirror traffic to a shadow variant, log predictions, and compare latency, errors, and drift against SLAs. Red flag: calling it A/B testing or claiming live business metrics from unserved responses.
What infrastructure is needed for a Continuous Training pipeline?
Tests event-driven ML system design beyond CI/CD. Strong answers name orchestrators, feature stores, model registries, and validation gates, mapping triggers to retraining and promotion. Red flag: conflating CT with CI/CD or skipping model validation.
Fairness and robustness gates in CI/CD
Sliced fairness metrics across subgroups, robustness checks via perturbation and adversarial sets, all compared to thresholds that fail the build.
How would you design safe, automatic schema evolution in CI?
Tests whether you separate schema evolution from semantic validation. Strong answer: versioned data contracts allowing additive enums, unknown-category model buckets, and automated contract negotiation. Red flag: manual allow-lists or disabling validation.
CI/CD for microservice-based ML systems
Independent per-service pipelines, contract testing to protect interfaces and schemas, and incremental deploys (canary, blue-green); manage data and model contracts, not just code.
Describe the difference between online and batch inference.
Online serves single requests in ms on live endpoints; batch processes data offline with elastic compute.
Where to place feature transformations: client, serving API, or upstream service?
Tests separation of concerns in ML systems. Client causes duplication and skew; serving API couples compute to requests; dedicated service adds a network hop but centralizes logic. Red flag: ignoring training-serving skew.
Design a system for batch scoring millions of customer records daily
Shard jobs via scheduler; partition storage; right-size CPU/GPU on spot; retry.
How would you systematically debug an inference API latency breach?
This tests structured debugging across the full inference stack. A strong answer traces the request path from ingress to GPU, splits TTFT from token-generation latency, inspects queuing and batching, then applies targeted fixes.
Architectural challenges for deploying ML models on resource-constrained edge devices
Tests Edge MLOps architecture under severe constraints. Strong answers hit quantization and delta OTA updates for flaky networks, power-aware scheduling, and closed-loop drift detection.
Describe the difference between online and batch inference.
Online uses autoscaling APIs for millisecond-to-second latency; batch uses scheduled compute for minute-to-hour latency.
We are hiring for this. Every open role lists the topics its interview covers, so you can prepare for the real thing rather than guessing.
See open roles