MLOps interview questions and answers

Replay production-like payloads, including feature store hits and batch size mix, not a tiny tensor of zeros.

REST/JSON is simple to curl and debug

They are Kubernetes-native serving layers that roll out models, scale, and sometimes canary using custom resources.

Kubeflow is closer to Kubernetes-native ML steps, artifacts, and retries on GPU Jobs.

Make steps idempotent, checkpoint weights, and retry from the last successful stage rather than restarting a six-hour fit.

Require tags for project, owner, and git SHA, expire old artifacts, and delete failed smoke runs.

A kill switch routes traffic to a previous model, a heuristic, or a default deny/allow without a full cluster rollback.

A model card documents intended use, data, metrics, limitations, and ethical notes for a version.

Minimize fields, encrypt, tokenize identifiers, set retention, and restrict access.

Clock skew in timestamps, different timezone handling, online features that are not backfilled, and batch aggregations that include the label window.

You track the histogram of scores or class rates over time and compare to a baseline.

TTL expires stale keys so you do not serve last month's session counts as if they were live.

Continuous training automatically retrains on new data when triggers fire.

Precompute when the candidate set is large and latency is tight, then serve a lookup.

Data paths, feature lists, hyperparameters, seeds, and compute settings should be declarative YAML or Hydra configs.

Automated comparison of canary versus baseline metrics with statistical tests, not a human squinting at graphs.

If you wait too long to fill a batch, every request pays that wait, inflating p99.

Compiled runtimes can cut latency and GPU memory versus eager Python.

A producer adds a new enum or changes amount from int to float and serving fills NaNs.

Label delay is the time until the true outcome is known, such as 30-day default.