MLOps interview questions and answers

CI for ML is automatically testing every change to training code, feature logic, configs, and sometimes data schemas.

CD is the path from a successful training run to a safely released model, often through staging, shadow, or canary.

ML quality depends on data, randomness, and expensive jobs, so tests cannot only assert status code 200.

Notebooks hide execution order, mix exploration with production logic, and are painful to unit test.

Docker packages code, system libraries, and Python dependencies into an image that runs the same on a laptop, CI, and a cluster.

An image is the immutable template of layers

A Dockerfile is the recipe that copies your serving code, installs requirements, and sets the start command such as uvicorn.

latest can silently pull a new OS or CUDA patch and change numerical results or break GPU drivers.

Compose lets you run the model API, Redis, and a fake feature store together with one command.

Kubernetes is a cluster orchestrator that schedules containers, restarts failures, and exposes services.

A Pod is the smallest deployable unit in Kubernetes, usually one container plus optional sidecars sharing network and volumes.

A Deployment declares how many replicas of a Pod template to keep and how to roll out new images.

A Service gives a stable DNS name and load-balances traffic across healthy Pods.

Kubernetes gives rolling deploys, health checks, autoscaling, and isolation across models.

MLflow is an open-source platform for experiment tracking, packaging, and a model registry.

Tracking records experiments and runs: metrics, params, and artifacts.

A run is one execution of training or evaluation, with its own ID, parameters, metrics, tags, and stored files.

Parameters are inputs you chose, such as learning rate or tree depth.

Watch operational health such as latency, errors, and saturation, plus ML health such as input drift, prediction volume, and delayed outcome metrics.

Training accuracy is a historical number on a static split.