MLOps interview questions and answers

Request rate, p95/p99 latency, error rate, model version tag, feature freshness, PSI for top features, and prediction score histograms.

Each online feature should carry an event timestamp or ingestion time.

Hash a stable user id with a salt into buckets so the same user always sees the same model.

Besides the primary profit or default metric, guard approval rate, fairness slices, latency, and operational loss.

Bandits help when the cost of exploring a bad model is high and you want to shift traffic toward winners faster.

Producers drop feature keys on Kafka or SQS, workers pull, fetch online features, score, and write results back.

Use one registered artifact and the same feature definitions

The server waits a few milliseconds to group incoming requests into one GPU or SIMD forward pass.

Use Airflow, Prefect, Dagster, or similar to run extract, validate, transform, and materialize with retries and SLAs.

A contract is a schema plus semantics: types, null rates, allowed ranges, and owners.

Recompute the historical offline table for the affected window, version the new feature, and decide whether to retrain.

Atomic adds, cuDNN autotune, and mixed precision can change reduction order and thus floats.

Store a dataset hash, partition dates, query, and DVC or table snapshot ID with the run.

dvc.yaml declares stages with deps, outs, and commands so DVC can skip unchanged steps using checksums.

Push cache to a shared remote, pull before experiments, and keep .dvc files in the same PR as code that consumes that data.

Start small enough that a disaster is contained, but large enough to see latency and error signal within minutes, often 1 to 5 percent.

Shadow serving sends live features to the new model but does not affect the user

Blue-green stands up a full new environment and switches all traffic at once after checks.

I would check deploy correlation, GC or Python GIL stalls, feature store latency, GC of large batches, and downstream timeouts.

If a shared dependency such as a feature store, GPU, or database is saturated, extra Pods just queue there.