MLOps interview questions and answers

Multi-model serving saves memory and operational overhead but couples failure domains and version rollouts.

Primary model, then a previous champion, then a popularity heuristic, each with timeouts.

Event-driven sensors on table partitions or a data-aware orchestrator, not a fixed cron that races the ETL.

Versioned training data with access logs, documented features, fairness slices, human approval, and a registry that cannot be silently overwritten.

AuthN/Z per client, service catalogs, rate quotas, and deprecation windows.

Check deploy and feature freshness, compare current versus previous prediction histograms, and verify the online store.

Spot or preemptible GPUs with checkpointing, dataset sampling for iteration, cached DVC stages, and deleting failed-run artifacts while keeping successful hashes.

Build when many teams share the same pain and you can staff on-call

Time from merged code to a safe production model, rollback time, percent of models with drift monitors, and reproducibility of last quarter's launches.

Join on entity and event time, diff column by column, and check timezone, window closure, and online TTL.

Lock a true holdout in the platform that training jobs cannot read, and only the evaluation service can score it after registration.

If schemas match, move the alias and let Pods reload or roll the Deployment to the old image.

It can adapt faster but makes reproducibility, rollback, and debugging much harder because the live weights keep moving.

Use idempotent upserts keyed by entity and window, or exactly-once sinks where the stream system supports them.

When traffic is heterogeneous, a random 5 percent may miss a high-risk tenant or a region with different latency.

Track score distributions and selection rates by sensitive slices that you are allowed to use, plus delayed outcome gaps when labels arrive.

Checkpoint to object storage every N steps, use a resume flag, and prefer a queue that restarts the Job on the same PVC or a fresh node with the checkpoint.

Accept they are different decisioning layers with a documented feature gap, or delay the real-time decision to a review queue when the extra features matter.

Dev uses sampled data and mock stores

Config debt is dozens of undocumented flags that change behavior, as called out in hidden technical debt work.