Which statement best defines MLOps?

1. A method used only to choose Machine Learning algorithms
2. A discipline for developing and operating ML systems through collaboration, automation, testing, monitoring, and governance
3. A database used exclusively for storing trained models
4. A replacement for Machine Learning evaluation

Why does MLOps require practices beyond traditional DevOps?

1. Machine Learning systems contain no software code
2. ML applications cannot use automated deployments
3. ML behavior depends on data, features, models, thresholds, and changing production conditions in addition to code
4. DevOps can be used only for desktop applications

What is the primary purpose of an ML pipeline?

1. To replace all human review with automation
2. To connect repeatable components with explicit inputs, outputs, and dependencies
3. To store passwords inside a notebook
4. To guarantee that every trained model is suitable for production

What does a directed edge represent in an ML pipeline DAG?

1. A randomly selected hyperparameter
2. A permission assigned to a user
3. An input-output dependency between pipeline steps
4. A production alert

Which sequence represents a sensible training pipeline?

1. Deploy, collect data, define the target, and train
2. Register, monitor, split data, and validate
3. Ingest, validate, split, train, evaluate, and register
4. Evaluate, train, delete data, and deploy

Why should preprocessing normally be fitted only on the training partition?

1. To increase API latency
2. To prevent information from validation or test data leaking into model training
3. To remove the need for an evaluation dataset
4. To ensure that every feature has the same value

Why is application uptime alone insufficient for model monitoring?

1. A healthy API may still serve inaccurate, stale, biased, or unsuitable predictions
2. Uptime cannot be measured for an API
3. Model monitoring applies only before deployment
4. Application health is unrelated to production systems

Which monitoring signal can usually be calculated before ground-truth labels arrive?

1. Observed recall
2. Observed precision
3. Prediction-distribution drift
4. Confirmed concept drift

What is Continuous Integration in an ML project primarily intended to do?

1. Automatically approve every newly trained model
2. Frequently validate changes to code, data contracts, features, pipelines, and model behavior
3. Replace experiment tracking
4. Delete unsuccessful runs immediately

Which statement correctly distinguishes Continuous Delivery from Continuous Deployment?

1. Continuous Delivery prepares an approved release, while Continuous Deployment can promote it automatically after required gates pass
2. Continuous Delivery applies only to databases
3. Continuous Deployment never uses automated checks
4. The two terms always mean manual model training

What should Continuous Training produce?

1. An automatically trusted production model
2. A candidate that still requires evaluation, comparison, registration, and approval
3. An unversioned model file
4. A replacement for production monitoring

Why should the same immutable artifact be promoted through staging and production?

1. To avoid testing the production artifact
2. To ensure that the evaluated release is the release actually deployed
3. To make rollback impossible
4. To eliminate environment configuration

Which item is the most reliable identity for a model artifact?

1. A mutable filename such as latest.pkl
2. A message saying that the model looks correct
3. An immutable model version associated with a recorded checksum and lineage
4. The date on which a developer opened the notebook

What does a cryptographic checksum such as SHA-256 verify?

1. That the model is accurate and fair
2. That a file matches the content represented by a recorded digest
3. That the deployment has sufficient compute capacity
4. That every user has approved the model

What is a model-registry alias?

1. A replacement for an immutable model version
2. A movable human-readable reference to a selected immutable version
3. A secret key used to download a model
4. A duplicate training dataset

Why must a deployment record the model version resolved from an alias?

1. The alias may later point to another model version
2. Aliases cannot be used in model registries
3. Resolved versions cannot be monitored
4. The version is needed only for local development

What is the difference between an experiment and a run?

1. An experiment groups related work, while a run represents one tracked execution
2. A run contains multiple unrelated organizations
3. An experiment is always a deployed model
4. There is no conceptual difference

Which statement correctly distinguishes a parameter from a metric?

1. A parameter is measured after training, while a metric is always selected before training
2. A parameter controls experiment behavior, while a metric measures an observed result
3. Parameters and metrics must always contain text
4. Metrics cannot be compared across runs

Why should compatible evaluation conditions be confirmed before ranking experiment runs?

1. Different datasets, targets, splits, thresholds, or metric definitions can make direct comparison invalid
2. Experiment trackers cannot store multiple runs
3. Every model produces identical predictions
4. Only training duration can be compared

What is data drift?

1. A change in the distribution of production input data relative to a reference
2. A failed API authentication request
3. A change in source-code formatting
4. A guaranteed loss of model accuracy

What is concept drift?

1. A change in the file path used to store the model
2. A change in the relationship between model inputs and the target outcome
3. A missing trace identifier
4. A decrease in endpoint CPU usage

Why should a drift alert not automatically replace the production model?

1. Drift can never be measured
2. Drift may be expected, temporary, unrelated to model quality, or caused by a data defect
3. Production models cannot be replaced
4. Only application logs can justify retraining

What is the primary role of an offline feature store?

1. Provide historical features for training and batch scoring
2. Store only application passwords
3. Replace the model registry
4. Serve only HTTP error pages

What is the primary role of an online feature store?

1. Provide low-latency entity-based feature retrieval for real-time inference
2. Store every historical raw data record indefinitely
3. Replace all model-training workflows
4. Generate source-code revisions

What does point-in-time correctness prevent?

1. Using future feature information in historical training rows
2. Using version control for feature definitions
3. Serving online features at low latency
4. Monitoring feature freshness

What is feature materialization?

1. Deleting feature definitions from source control
2. Synchronizing selected feature values into an online serving store
3. Converting a model into a requirements file
4. Replacing event timestamps with random values

Which set represents the three primary observability signals?

1. Models, labels, and thresholds
2. Logs, metrics, and traces
3. Branches, commits, and pull requests
4. Training, validation, and testing

Why is structured logging preferred for production ML services?

1. It records events through named machine-readable fields that can be searched, filtered, and correlated
2. It guarantees perfect model accuracy
3. It removes the need for metrics and traces
4. It requires complete request payloads to be stored

Which value should normally be avoided as a metric label?

1. Model version
2. Deployment environment
3. Request identifier
4. Response status

A model API is healthy, but delayed evaluation shows declining recall. What is the best next action?

1. Ignore the result because the API is available
2. Delete all monitoring evidence
3. Investigate labels, data, features, thresholds, segments, model versions, and business changes before selecting a controlled response
4. Retrain and deploy a new model immediately without evaluation

Correct Answer: B

MLOps combines Machine Learning, data engineering, software engineering, DevOps, automation, monitoring, security, and governance to manage ML systems throughout their lifecycle.

Why the other options are incorrect: MLOps is not limited to algorithm selection, is not merely a model database, and does not replace independent evaluation.

Correct Answer: C

ML system behavior depends on code and configuration, but also on data, labels, features, preprocessing, model artifacts, thresholds, and changing production populations.

Why the other options are incorrect: ML systems contain software, can use automated deployment, and DevOps is not restricted to desktop applications.

Correct Answer: B

An ML pipeline coordinates focused components through explicit inputs, outputs, dependencies, parameters, and artifacts. This makes the workflow repeatable and testable.

Why the other options are incorrect: Pipelines should preserve necessary review, must not contain credentials, and cannot guarantee production suitability merely because training completed.

Correct Answer: C

In a directed acyclic graph, an edge shows that one component depends on an output produced by another component.

Why the other options are incorrect: Hyperparameters, permissions, and alerts are not represented by pipeline dependency edges.

Correct Answer: C

A sensible workflow ingests data, validates it, creates appropriate partitions, trains candidates, evaluates them, and registers only suitable artifacts.

Why the other options are incorrect: They place deployment, registration, or evaluation before the data and training activities on which those stages depend.

Correct Answer: B

Fitting preprocessing on validation or test observations allows evaluation information to influence the trained pipeline. This can make performance estimates overly optimistic.

Why the other options are incorrect: Training-only fitting is not intended to increase latency, remove evaluation, or make feature values identical.

Correct Answer: A

An endpoint can return technically successful responses while its input data, feature freshness, prediction quality, calibration, fairness, or business usefulness deteriorates.

Why the other options are incorrect: Uptime is measurable, production monitoring happens after deployment, and application health remains important even though it is not sufficient alone.

Correct Answer: C

Score or class distributions can be compared without knowing the eventual true outcome. This makes prediction drift an early proxy signal.

Why the other options are incorrect: Precision, recall, and confirmed concept drift require verified outcomes.

Correct Answer: B

Continuous Integration gives rapid feedback on proposed changes by validating software and ML-specific assets.

Why the other options are incorrect: CI does not approve every model, replace experiment tracking, or require immediate deletion of unsuccessful evidence.

Correct Answer: A

Continuous Delivery produces a validated release ready for controlled deployment. Continuous Deployment can automatically promote that release after applicable gates pass.

Why the other options are incorrect: Delivery is not database-only, deployment should use automated checks, and neither term means manual model training.

Correct Answer: B

Continuous Training automates candidate creation. A newly trained model still requires evaluation, comparison, lineage, registration, approval, and controlled release.

Why the other options are incorrect: Training completion does not create trust, remove versioning, or replace monitoring.

Correct Answer: B

Promoting the same immutable artifact ensures that the content tested in staging is the content released to production.

Why the other options are incorrect: Immutable promotion improves testing and rollback and does not remove the need for external environment configuration.

Correct Answer: C

A reliable model identity is immutable and connected to artifact integrity, code, data, configuration, environment, evaluation, and approval evidence.

Why the other options are incorrect: Mutable filenames, informal messages, and notebook-access dates cannot uniquely establish model content or lineage.

Correct Answer: B

A checksum verifies whether the bytes of a file match the content represented by the recorded digest.

Why the other options are incorrect: Integrity verification does not prove model quality, fairness, infrastructure capacity, or approval.

Correct Answer: B

An alias such as candidate, champion, or rollback provides a readable reference that can be moved between immutable model versions.

Why the other options are incorrect: An alias does not replace version identity, is not a secret, and is unrelated to duplicating training data.

Correct Answer: A

An alias can be reassigned. Recording the resolved immutable version preserves deployment, prediction, monitoring, and incident traceability.

Why the other options are incorrect: Aliases are supported by registry workflows, immutable versions can be monitored, and production traceability is one of their most important uses.

Correct Answer: A

An experiment or project groups related investigations. A run is one execution with its own configuration, metrics, artifacts, status, and conclusion.

Why the other options are incorrect: A run does not represent unrelated organizations, an experiment is not automatically a deployed model, and the concepts have distinct purposes.

Correct Answer: B

A parameter controls execution, while a metric is calculated to measure the result.

Why the other options are incorrect: The first option reverses the definitions, tracked values are not restricted to text, and compatible metrics can be compared across runs.

Correct Answer: A

Run comparison is meaningful only when evaluation conditions are compatible. Differences in targets, data, partitions, thresholds, or metrics can invalidate a ranking.

Why the other options are incorrect: Tracking systems support multiple runs, models are not identical, and quality comparison can include far more than duration.

Correct Answer: A

Data drift means that one or more production input distributions differ from an identified reference population.

Why the other options are incorrect: Authentication, code formatting, and guaranteed accuracy loss do not define data drift.

Correct Answer: B

Concept drift occurs when the relationship between inputs and outcomes changes. Historical patterns may therefore become less predictive.

Why the other options are incorrect: Storage paths, trace identifiers, and CPU usage do not define the predictive relationship.

Correct Answer: B

Drift may be caused by seasonality, an upstream defect, temporary behavior, a new population, or a valid business change. Its effect must be investigated.

Why the other options are incorrect: Drift is measurable, production models can be replaced through controlled processes, and logs alone do not justify retraining.

Correct Answer: A

The offline store provides historical feature values used for point-in-time-correct training, testing, backtesting, and batch inference.

Why the other options are incorrect: It is not a password store, does not replace the model registry, and is unrelated to serving error pages.

Correct Answer: A

An online store provides low-latency retrieval of the latest suitable features for one or more entities during real-time inference.

Why the other options are incorrect: It should not contain every historical raw record, does not replace training, and does not generate code revisions.

Correct Answer: A

Point-in-time correctness ensures that each historical row uses only feature information available at or before that row's prediction timestamp.

Why the other options are incorrect: Point-in-time correctness supports versioned features, online serving, and freshness monitoring rather than preventing them.

Correct Answer: B

Materialization copies or synchronizes selected feature values from an offline source into an online store for inference.

Why the other options are incorrect: It does not delete definitions, create dependency files, or randomize timestamps.

Correct Answer: B

Logs describe events, metrics quantify behavior over time, and traces show the path and timing of distributed requests.

Why the other options are incorrect: The other sets are important ML or engineering concepts, but they are not the primary observability signals.

Correct Answer: A

Structured logs represent events through named fields. This supports searching, filtering, correlation, aggregation, and automated processing.

Why the other options are incorrect: Logging does not guarantee accuracy, replace other telemetry, or justify storing complete payloads.

Correct Answer: C

A request identifier is usually unique and therefore creates unbounded metric cardinality. It belongs in logs and traces rather than metric labels.

Why the other options are incorrect: Model version, environment, and response status are generally bounded dimensions when managed appropriately.

Correct Answer: C

Declining recall is a model-quality signal even when the API is available. The team should verify label maturity and investigate data, features, thresholds, segments, versions, and business changes before choosing a response.

Why the other options are incorrect: Availability does not establish model quality, evidence should be preserved, and immediate unvalidated replacement is unsafe.

Pipeline Design

Draw an ML pipeline that validates data, creates training and evaluation partitions, trains a candidate, applies quality gates, and registers an approved artifact.

Versioning Plan

List every code, data, model, schema, feature, threshold, environment, and monitoring asset required to reproduce one production prediction.

Monitoring Matrix

Define service, data, feature, prediction, delayed-quality, responsible-use, business, and cost metrics for one model.

Incident Investigation

Describe how logs, metrics, traces, model versions, feature metadata, and delayed outcomes would be used to investigate a production-quality alert.

Controlled Retraining

Design a process that converts an approved drift signal into an evaluated candidate without automatically replacing production.