TL;DR: Answer across four planes with rough time allocation (pipelines and platform ~40%, serving fleet ~25%, monitoring and on-call ~20%, paved road for data scientists ~15%), name the orchestrator, registry, and what actually paged you, and close with one manual process you automated away. Ownership verbs beat a tool inventory.
How to approach it
This sounds like small talk. It isn't. The interviewer is triaging: real operator, or someone who trained models in notebooks and added "MLOps" during the job search? Your answer determines whether the next forty minutes are deep technical questions or polite filler. Don't recite a job description, describe a system you ran, with ownership verbs and proportions.
A strong answer
Structure it across four planes, with rough time allocation, because real jobs have a shape:
Pipelines and platform (roughly 40%). "I owned the training and deployment pipelines for N models, ours were SageMaker Pipelines triggered by CodePipeline, with evaluation gates and registry promotion. Day to day that meant extending pipeline steps, fixing broken DAG runs, and reviewing data scientists' PRs against the pipeline templates." Name the orchestrator, the registry, the CI system. Specificity here is the whole game.
Serving fleet (roughly 25%). What was actually in production, how many endpoints, what traffic, what infrastructure. "Six real-time endpoints on EKS behind an internal gateway, peak around 800 requests/sec, plus nightly batch scoring jobs" tells the interviewer your scale in one sentence. They calibrate every later question to this number, so be honest; inflated scale gets exposed in follow-ups.
Monitoring and on-call (roughly 20%). This is the credibility section. Mention the rotation, the dashboards (Prometheus/Grafana, CloudWatch, Model Monitor), and above all what paged you: latency SLO breaches, feature pipeline failures, drift alerts. A candidate who can say "our noisiest alert was schema changes from an upstream team, so I built a data contract check that failed their CI instead of paging us" has just demonstrated more MLOps judgment than any tool list.
Paved road for data scientists (roughly 15%). The part senior interviewers listen for: making other people productive. Templates, a model-onboarding checklist, golden-path Docker images, "you bring the training code, the platform handles deploy and monitoring."
Close with one concrete recent week: "Last month, for example: migrated two models from stage-based MLflow promotion to aliases, debugged a training-serving skew issue in a pricing model, and ran the postmortem for a drift incident." One real week beats ten abstract responsibilities.
| Where the time actually goes | Share of a typical week |
|---|---|
| Pipelines that broke overnight, and why | The largest single slice |
| Making someone else's notebook reproducible and deployable | Large |
| Monitoring, drift checks, and chasing an alert | Moderate |
| Promotions, rollbacks and release mechanics | Moderate |
| Cost and capacity work nobody asked for until the invoice | Small, until it is not |
| Actually training a model | Smallest, and it surprises people |
Answering with the last row first is the tell of someone describing the job title rather than the job. The honest version leads with reliability and reproducibility, because that is what the role exists to provide.
If you are transitioning and have never held the title, do not perform a week you never lived; map the planes onto the system you did run, and name the gap yourself before the interviewer finds it. A DevOps or platform engineer has the pipelines, serving, and on-call planes cold, just for stateless services: say that, then show you know what changes ("my pages were infra pages; the new failure class I'd be learning is the model-quality page, where everything is green and the predictions are wrong, and here is how I'd instrument for it"). A data scientist has the opposite shape: real model-quality stories, no rotation, and the honest version is "I was the person the platform team's paved road was built for, here is what I would have built differently having been its customer." Both of those answers are hirable because they demonstrate the four-plane map itself plus self-awareness about position on it. The unhirable version is the vague composite that claims all four planes at once, because the follow-ups ("what paged you?") locate you in about ninety seconds anyway.
What interviewers probe next
- "What pages did you actually get?", have two ready: one infrastructure (OOM on a GPU pod, queue backup) and one model-quality (drift, skew). Candidates who only have infra pages get tagged as platform-only; only model pages, as never on-call.
- "What part did you build versus inherit?", answer honestly and name the design decision you made on the part you built. Inheriting is fine; pretending isn't.
- "How many models, how many engineers?", they're sizing your claims. "Twelve models, three platform engineers, fifteen data scientists" is a normal, believable shape.
- "What did you automate away?", the strongest close: a manual process you killed, with the before/after ("deploys went from a two-week ticket queue to same-day self-service").
Common mistakes
Answering with a tool inventory ("we used Docker, Kubernetes, MLflow, Airflow..."), tools without ownership verbs signal tutorial knowledge. Describing only model development, which answers a different job title. Claiming end-to-end ownership of everything, at any real company the boundaries are visible, and "I did all of it" reads as "I watched all of it." And skipping on-call entirely: an MLOps engineer who has never been paged hasn't operated anything yet, and interviewers know it.