MLOps interview questions and answers

I would package the scaler, encoder, and feature builder with the model, or better, call the same library from both jobs.

It is acceptable if the feature logic is exported to a portable form, such as a feature store, ONNX, or generated code, not rewritten by hand.

I look at model size, p99 budget, QPS, and cost.

Each version should store Git SHA, DVC or table snapshot IDs, training run ID, metrics, and the serving image digest.

A signature declares expected input names, types, and shapes so serving can reject bad payloads early.

Use namespaced registered models, IAM on who can transition stages, and required reviews on production.

The offline store holds historical feature tables for training, usually warehouse or lake files.

When you build a training row, you must use feature values that were known at that event's timestamp, not future updates.

Feast lets you declare feature views, materialize to an online store, and retrieve vectors for training or serving.

A tabular feature store serves named numeric or categorical columns keyed by entity.

I would snapshot reference histograms from the training window, then compare live daily distributions with PSI or KS per feature.

If the model is well calibrated on the new region of X, or if the shift is a known campaign you already encoded as a feature, retraining can wait.

Track frequency of each level, the share of unknown or new levels, and entropy of the distribution.

Use proxy or leading indicators, such as chargeback rates that arrive earlier, and monitor residual errors on a delayed labeled sample.

If the model’s scores change who gets a loan, future training data is selected by the previous model.

Unit tests for transforms, schema tests on sample data, a tiny training job that must beat a dummy baseline, and a serving smoke test that loads the artifact.

After training, an evaluation step compares the candidate to the current champion on a frozen eval set and required slices.

A golden set is a versioned batch of inputs with expected scores or labels used as a regression test.

Inject them from a secret manager at runtime, never commit them, and scope IAM to read-only training tables.

Pin the CUDA base image digest, driver-compatible tags, and Python lockfile, and record nvidia-smi output in the run.