MLOps interview questions for freshers

Latency is how long one request takes from arrival to response, often summarized as p50 or p99.

Throughput is how many predictions you complete per second, often QPS.

p99 is the latency below which 99 percent of requests finish

QPS is queries per second, a throughput measure of request rate.

Model serving is the production system that loads a registered model and answers prediction requests.

An endpoint decouples the model from each product service and lets you scale, version, and auth in one place.

Experiment tracking records hyperparameters, metrics, code version, and artifacts for each training attempt.

An artifact store is object storage for models, plots, datasets, and pipeline outputs, addressed by run ID or hash.

Rollback switches serving back to a previously known-good registry version, often by moving the production alias.

Training-serving skew is when the model is trained on features computed one way and served with features computed another way.

MLOps is the practice of deploying, monitoring, and iterating ML systems reliably like DevOps for models.

A store for versioned model artifacts and metadata.

Without versions you cannot reproduce a metric or roll back a bad release.

Batch scores many rows on a schedule

A consistent place to serve features for training and inference.

Storing versioned model artifacts with metadata and stage tags like Staging or Prod.

Batch scores many rows on a schedule

Same runtime in dev and prod, pinned system libs, and easier rollbacks.