MLOps Quiz

Operationalizing machine learning — pipelines, versioning, and the practices behind reliable ML in production.

This category currently has 100 questions in the SERVBG quiz bank. Below are a few sample questions — the full interactive quiz shuffles through the whole set with instant scoring.

Sample questions

In an ML pipeline, what is the primary architectural reason for separating the training and serving environments rather than using a single unified codebase?

A unified codebase makes it impossible to use GPUs for training while using CPUs for inference, so the split is purely hardware-driven.
Training code must be written in Python while serving code must be compiled to a lower-level language for performance.
Regulatory compliance mandates separate environments in all jurisdictions that govern machine learning systems.
Training and serving have different latency, throughput, and resource requirements; unified codebases create hidden coupling that causes silent skew between offline metrics and online behavior.
Model weights cannot be serialized across different runtime environments without data corruption.

A feature store serving both online (low-latency) and offline (batch training) use cases must guarantee point-in-time correctness. Which consistency challenge is unique to the offline retrieval path?

Offline retrieval requires stronger encryption than online paths because training data is more sensitive than inference inputs.
Offline retrieval must account for time-zone offsets introduced by distributed data warehouses spanning multiple regions.
Offline retrieval is only consistent if features are recomputed from raw events on every training run rather than cached.
Offline retrieval must reconstruct the exact feature values available at the label timestamp to avoid future leakage, requiring time-travel queries against a versioned historical store.
Offline retrieval must handle network partitions that do not affect the online path because batch jobs run on separate infrastructure.

Which data versioning strategy is most suitable for large binary datasets (>100 GB of parquet files) in a DVC-managed pipeline where multiple team members commit daily?

Compress all parquet files to tar.gz before committing so that git handles binary diffs more efficiently than DVC.
Rely on S3 object versioning alone without DVC so that the git history stays clean of any data-related commits.
Store binary files directly in the git repository using Git LFS pointer files without any external remote to simplify the toolchain.
Use DVC only for model artifacts; version raw datasets with a dedicated SQL schema version table in the training database.
Use DVC with a remote storage backend (S3, GCS) and content-addressed caching; commit only the small .dvc pointer files to git while binary blobs are stored remotely.

Your training pipeline uses Great Expectations for schema validation. After a upstream data team changes a nullable integer column to a required float, your suite passes silently. What is the most likely root cause?

Great Expectations does not validate numeric type changes between int and float because it treats all numeric types as equivalent by default.
Great Expectations silently passes when the column name matches but logs a deprecation warning visible only in the verbose run log.
The expectation for that column was defined with mostly=1.0 tolerance or the column type check used expect_column_values_to_be_of_type with a type alias that Python coerces implicitly.
The expectation suite was configured with store_backend_defaults which bypasses type-level validation for columnar formats.
The schema change was applied after the checkpoint ran, so the validation executed against the old snapshot before the migration committed.

In training data curation for a classification model, which practice most directly reduces the risk of systematic label noise without collecting new labels?

Use stratified k-fold cross-validation to detect label noise, as inconsistent fold performance always indicates mislabeled samples.
Remove all examples where model confidence is below 0.5 during a pilot training run to eliminate low-quality training points.
Upsample minority-class examples using SMOTE to balance the dataset before training, which corrects label noise through synthetic interpolation.
Replace all labels with soft pseudo-labels generated by an ensemble of pre-trained models available on Hugging Face.
Apply confident learning (e.g., via cleanlab) to identify likely mislabeled examples by estimating the joint distribution between noisy observed labels and latent true labels.

Start the full quiz ›