I85PREP
⌘ K
IA MLO study materials
IA MLO study resources

IA MLO practice questions

One question at a time. Build your IA MLO understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1MLOps engineering practice for machine learning lifecycle and deployment

In an ML deployment pipeline, what is the primary purpose of a feature store?

  1. To schedule retraining jobs at fixed calendar intervals
  2. To store raw training datasets in their original file formats
  3. To serve consistent feature values for both training and online inference Correct answer.
  4. To host the model artifact registry used during rollout

Explanation

A feature store centralizes feature definitions and computed values so that offline training pipelines and low-latency online serving read the same data, which prevents training-serving skew. It is not primarily a raw data lake, a scheduler, or a model registry.

Question 2MLOps engineering practice for machine learning lifecycle and deployment

A team wants reproducible training runs. Which practice most directly supports this goal?

  1. Compressing model checkpoints before upload
  2. Increasing the number of parallel training workers
  3. Pinning data snapshots, code commits, and environment dependencies per run Correct answer.
  4. Running all experiments on the largest available GPU type

Explanation

Reproducibility requires capturing the exact inputs to a run: the data version, the code revision, and the dependency environment. Hardware scaling, compression, and GPU size affect performance or storage, not the ability to recreate a run.

Question 3MLOps engineering practice for machine learning lifecycle and deployment

What distinguishes a model registry from a general-purpose artifact repository?

  1. It replaces the need for any experiment tracking tool
  2. It provides the compute cluster used for distributed training
  3. It tracks model lineage, versions, and stage transitions such as staging to production Correct answer.
  4. It automatically retrains models whenever new data arrives

Explanation

A model registry is purpose-built to record model versions along with metadata, lineage, and lifecycle stages so promotion and rollback are auditable. It does not itself train models, supply compute, or fully substitute for experiment tracking.

Question 4MLOps engineering practice for machine learning lifecycle and deployment

A monitoring system detects that the distribution of input features has shifted while model accuracy is still acceptable. What is this condition called?

  1. Overfitting
  2. Gradient explosion
  3. Data drift Correct answer.
  4. Label leakage

Explanation

Data drift (covariate shift) means the input feature distribution has changed relative to training data, even if current predictions still look fine. Label leakage, gradient explosion, and overfitting are training-time problems, not input-distribution monitoring signals.

Question 5MLOps engineering practice for machine learning lifecycle and deployment

Which metric best captures whether a deployed classifier's predicted positive rate matches the observed positive rate in production?

  1. Training loss on the original dataset
  2. Number of features in the model
  3. Prediction drift measured as a change in output distribution Correct answer.
  4. Wall-clock latency of the inference endpoint

Explanation

Prediction drift compares the distribution of model outputs over time, so a shift in the predicted positive rate is captured here. Training loss, feature count, and latency describe model quality, structure, or performance, not output distribution change.

Question 6MLOps engineering practice for machine learning lifecycle and deployment

What role does a container image play in deploying an ML model to a serving environment?

  1. It packages the model, runtime, and dependencies into a portable, reproducible unit Correct answer.
  2. It defines the feature transformation logic only
  3. It schedules retraining jobs across the cluster
  4. It stores the labeled training data used to fit the model

Explanation

A container image bundles the model artifact with its runtime and library dependencies so the same environment runs consistently across machines. Training data storage, job scheduling, and feature logic are handled by other components.

Question 7MLOps engineering practice for machine learning lifecycle and deployment

In a CI/CD pipeline for ML, what is the purpose of an automated model quality gate?

  1. To replace human review of the model architecture
  2. To reduce inference latency on the serving cluster
  3. To block promotion when evaluation metrics fall below defined thresholds Correct answer.
  4. To increase the training dataset size automatically

Explanation

A quality gate evaluates the candidate model against agreed metric thresholds and halts promotion if it underperforms. It does not alter dataset size, remove architecture review, or optimize serving latency.

Question 8MLOps engineering practice for machine learning lifecycle and deployment

What does a canary deployment of a new model version accomplish?

  1. It retrains the model on only a small sample of data
  2. It routes a small fraction of live traffic to the new version to observe behavior before full rollout Correct answer.
  3. It deletes the previous model version to save storage
  4. It runs the new model only in offline batch mode

Explanation

Canary deployment sends a limited share of production traffic to the new version so real-world metrics and errors can be observed before broad release. It does not change training data, remove prior versions, or restrict the model to batch mode.

Question 9MLOps engineering practice for machine learning lifecycle and deployment

A model serving endpoint must return predictions within 50 ms at the 99th percentile. Which deployment pattern is most appropriate?

  1. Online real-time serving with the model loaded in memory behind a low-latency API Correct answer.
  2. A weekly offline report generated by analysts
  3. Nightly batch scoring written to a data warehouse
  4. Manual scoring requested through a ticketing system

Explanation

A strict per-request latency budget requires an online serving pattern where the model is resident and answers API calls directly. Batch, weekly report, and manual workflows cannot meet a 50 ms interactive deadline.

Question 10MLOps engineering practice for machine learning lifecycle and deployment

A team needs to detect training-serving skew. Which comparison is most diagnostic?

  1. Number of GPUs used in training versus serving
  2. Model file size before and after quantization
  3. Feature value distributions computed in the training pipeline versus those logged at inference time Correct answer.
  4. Container image tag used in staging versus production

Explanation

Training-serving skew arises when the features seen in production differ from those used in training, so comparing the two distributions reveals it. File size, GPU counts, and image tags are infrastructure details unrelated to feature consistency.

Question 11MLOps engineering practice for machine learning lifecycle and deployment

A pipeline must retrain a model automatically when new labeled data lands in storage. Which component orchestrates this?

  1. The load balancer in front of the inference API
  2. The model registry's version counter
  3. A workflow orchestrator with a storage-event trigger Correct answer.
  4. The feature store's online serving cache

Explanation

Workflow orchestration tools watch for events such as new data arrival and then launch the training pipeline steps. Registry counters, serving caches, and load balancers do not trigger training workflows.

Question 12MLOps engineering practice for machine learning lifecycle and deployment

After promoting a new model, error rates spike sharply. What is the fastest safe remediation given a registry with versioned artifacts?

  1. Increase the batch size of the inference server
  2. Roll back to the previously promoted model version and investigate Correct answer.
  3. Retrain the model from scratch on all available data
  4. Disable monitoring to reduce alert noise

Explanation

Because prior versions are preserved in the registry, rolling back restores known-good behavior quickly and buys time for root-cause analysis. Retraining is slow, silencing monitoring hides the problem, and batch size tuning does not fix a faulty model.

Question 13MLOps engineering practice for machine learning lifecycle and deployment

A fraud model's positive rate is stable, but its precision drops because the relationship between features and labels has changed. Which concept describes this?

  1. Data partitioning
  2. Schema evolution
  3. Cold start
  4. Concept drift Correct answer.

Explanation

Concept drift means the mapping from inputs to target has changed, so the same feature values now imply different outcomes, degrading precision. Schema evolution, cold start, and partitioning describe data structure, new users, or storage layout, not a changed input-target relationship.

Question 14MLOps engineering practice for machine learning lifecycle and deployment

Why is shadow deployment useful when introducing a new model version?

  1. It runs the new version on mirrored live traffic without affecting user-facing responses Correct answer.
  2. It trains the model on production traffic labels automatically
  3. It replaces the production model immediately to reduce cost
  4. It splits user traffic evenly to measure statistical significance

Explanation

Shadow deployment duplicates live requests to the candidate model and logs its outputs while the incumbent still serves users, so behavior can be compared safely. Even splitting is A/B testing, immediate replacement is risky, and shadow mode does not auto-label or train.

Question 15MLOps engineering practice for machine learning lifecycle and deployment

An inference service must handle sudden traffic spikes without over-provisioning for peak load at all times. Which capability addresses this?

  1. Storing model weights on a slower network filesystem
  2. Horizontal autoscaling of stateless model replicas Correct answer.
  3. Running a single large replica with no redundancy
  4. Increasing the training batch size

Explanation

Autoscaling adds or removes stateless replicas in response to load, matching capacity to demand and avoiding permanent peak provisioning. Slow storage, a single non-redundant replica, and training batch size do not provide elastic serving capacity.

Question 16MLOps engineering practice for machine learning lifecycle and deployment

What is the main benefit of logging model inputs and outputs in production?

  1. It guarantees the model will not degrade over time
  2. It automatically improves model accuracy
  3. It eliminates the need for a feature store
  4. It enables drift analysis, debugging, and audit trails for predictions Correct answer.

Explanation

Logging inputs and outputs creates the evidence needed to detect drift, reproduce issues, and satisfy audit requirements. Logging itself cannot prevent degradation, replace a feature store, or improve accuracy.

Question 17MLOps engineering practice for machine learning lifecycle and deployment

A data scientist's notebook produces a model that cannot be reproduced by the deployment team. Which missing practice is the most likely cause?

  1. Environment and dependency versions were not captured with the code Correct answer.
  2. The notebook used pandas for data loading
  3. The model was evaluated on a holdout set
  4. The model used gradient boosting

Explanation

Without pinned dependencies and a recorded environment, the deployment team cannot recreate the same runtime, so results diverge. Using a holdout set, pandas, or gradient boosting are normal choices that do not by themselves break reproducibility.

Question 18MLOps engineering practice for machine learning lifecycle and deployment

Which practice best protects a model registry from accidental overwriting of a production model artifact?

  1. Allowing any engineer to delete artifacts after testing
  2. Immutable versioned artifacts with promotion controlled by role-based access Correct answer.
  3. Naming artifacts by date only with no version metadata
  4. Storing all models in a single shared folder with write access for everyone

Explanation

Immutable versions combined with access control ensure artifacts cannot be silently replaced and that promotion requires authorization. Shared writable folders, date-only naming, and open deletion all risk losing or corrupting production artifacts.

Question 19MLOps engineering practice for machine learning lifecycle and deployment

A pipeline uses the same transformation code for training and serving, yet skew still appears. Which cause is most plausible?

  1. The serving path computes a rolling statistic over a different time window than training Correct answer.
  2. The training job used a GPU
  3. The serving API returns JSON responses
  4. The model was trained with a fixed random seed

Explanation

Even with shared code, differing window definitions for time-dependent features produce different values at training and serving, causing skew. Fixed seeds, GPU training, and JSON responses do not alter feature computation semantics.

Question 20MLOps engineering practice for machine learning lifecycle and deployment

A team wants to compare a new ranking model against the incumbent using live traffic while limiting user exposure and measuring a business metric. Which approach fits best?

  1. A full immediate rollout to all users
  2. A shadow deployment that logs outputs without serving them
  3. An offline evaluation on the historical test set only
  4. An A/B test with randomized user assignment and a predefined success metric Correct answer.

Explanation

An A/B test randomizes users between variants and measures a predefined business metric, giving a controlled estimate of real-world impact while capping exposure. Shadow mode never serves the new model, full rollout removes the comparison, and offline evaluation cannot measure live business outcomes.

Question 21MLOps engineering practice for machine learning lifecycle and deployment

A feature pipeline backfills historical data and recomputes features with a corrected join, but the deployed model was trained on the old values. What risk arises?

  1. Training-serving skew because the feature semantics changed after training Correct answer.
  2. Gradient vanishing because the join changed
  3. Overfitting because the model saw too many examples
  4. Label leakage because the backfill used labels

Explanation

Changing feature semantics after training means production features no longer match what the model learned, which is a form of training-serving skew. Overfitting, gradient vanishing, and label leakage are unrelated to a corrected join altering feature values.

Question 22MLOps engineering practice for machine learning lifecycle and deployment

A monitoring system fires frequent drift alerts because a feature is highly seasonal, causing alert fatigue. Which adjustment is most principled?

  1. Raise the model's learning rate to adapt faster
  2. Compare current values against a seasonally adjusted or same-period baseline Correct answer.
  3. Retrain the model on every alert regardless of cause
  4. Disable drift monitoring for that feature permanently

Explanation

Seasonal features naturally vary, so comparing against a seasonally adjusted baseline distinguishes expected cycles from genuine drift and reduces false alerts. Disabling monitoring loses coverage, and learning rate or blanket retraining do not address alert design.

Question 23MLOps engineering practice for machine learning lifecycle and deployment

An organization wants guarantees that a model artifact in production is byte-identical to the one that passed evaluation. Which mechanism provides this?

  1. A spreadsheet listing model filenames
  2. A shared network drive with manual copy steps
  3. Email approval from the data science lead
  4. Content-addressed artifact storage with recorded digests verified at deploy time Correct answer.

Explanation

Content-addressed storage identifies artifacts by cryptographic digest, and verifying that digest at deploy time proves the deployed bytes match the evaluated artifact. Manual copies, spreadsheets, and email approvals do not provide integrity guarantees.

Question 24MLOps engineering practice for machine learning lifecycle and deployment

A team deploys a model that must comply with audit requirements to explain why each prediction was made. Which practice supports this most directly?

  1. Persisting model version, input snapshot, and prediction with request identifiers Correct answer.
  2. Storing only the final model weights
  3. Discarding per-request inputs after scoring
  4. Aggregating only daily average predictions

Explanation

Recording the model version, the exact inputs, and the resulting prediction under a request identifier allows each decision to be reconstructed and explained later. Aggregated averages, discarded inputs, and weights alone cannot support per-prediction audit.

Question 25MLOps engineering practice for machine learning lifecycle and deployment

Which scenario most clearly requires a batch rather than online serving pattern?

  1. Autocompleting a search query as the user types
  2. Answering a single user's recommendation request in under 100 ms
  3. Scoring an entire customer base each night to refresh segment assignments Correct answer.
  4. Rejecting a fraudulent transaction during checkout

Explanation

Nightly scoring of the full customer base is a high-volume, non-interactive job well suited to batch processing. The other scenarios demand immediate per-request responses and therefore require online serving.

Question 26MLOps engineering practice for machine learning lifecycle and deployment

In an ML lifecycle, what distinguishes a model registry from a general-purpose artifact repository?

  1. It stores model files together with version lineage, stage transitions, and metadata used for promotion decisions Correct answer.
  2. It automatically retrains models whenever new data arrives
  3. It compiles model code into native binaries for each target operating system
  4. It replaces the feature store by storing raw training datasets

Explanation

A model registry is a metadata-aware catalog: it tracks versions, lineage, evaluation metrics, and lifecycle stages such as staging or production, which is what promotion decisions depend on. A plain artifact repository only stores bytes and offers little lifecycle semantics.

Question 27MLOps engineering practice for machine learning lifecycle and deployment

A team wants every production model to be traceable back to the exact commit, dataset snapshot, and hyperparameters that produced it. Which practice most directly delivers this?

  1. Keeping model files in a shared network drive named by date
  2. Manual spreadsheet tracking of experiments by the data science lead
  3. Relying on the model's accuracy score as the unique identifier
  4. Recording experiment runs with logged parameters, metrics, and data versions in a tracking system Correct answer.

Explanation

Experiment tracking systems log parameters, metrics, code version, and dataset references per run, creating reproducible lineage. Manual or filesystem-based tracking loses the link between artifact and inputs.

Question 28MLOps engineering practice for machine learning lifecycle and deployment

What is the primary purpose of a feature store in an MLOps platform?

  1. To serve as the system of record for labeling annotations
  2. To replace the model registry for storing serialized models
  3. To train gradient boosting models faster on GPU hardware
  4. To provide consistent feature definitions for both offline training and online inference Correct answer.

Explanation

A feature store centralizes feature definitions and computes them for both offline training and low-latency online serving from the same logic, which reduces training-serving skew. It is not a training accelerator or a labeling or model storage system.

Question 29MLOps engineering practice for machine learning lifecycle and deployment

Which statement best describes infrastructure as code in an ML deployment context?

  1. It provisions serving infrastructure through version-controlled declarative definitions rather than manual console changes Correct answer.
  2. It means writing model training code as declarative configuration files
  3. It guarantees that model predictions remain stable across environments
  4. It converts notebook code into production Python modules automatically

Explanation

Infrastructure as code means environments such as clusters, endpoints, and networking are defined in version-controlled files and applied reproducibly. It concerns infrastructure provisioning, not model code or prediction stability.

Question 30MLOps engineering practice for machine learning lifecycle and deployment

A model is deployed to a serving endpoint and predictions become slow under load. Which metric set is most appropriate for ongoing operational monitoring of that endpoint?

  1. Training accuracy and F1 score on the original validation split
  2. Size of the training dataset in gigabytes
  3. Number of git commits merged per week
  4. Request latency percentiles, error rates, and throughput Correct answer.

Explanation

Operational monitoring of a serving endpoint centers on latency percentiles, error rates, and throughput, which reflect the health of live inference. Training metrics and repository statistics do not describe runtime behavior.

Question 31MLOps engineering practice for machine learning lifecycle and deployment

What does versioning a training dataset snapshot accomplish for an ML pipeline?

  1. It removes the need for a model registry
  2. It lets a past training run be reproduced against the same input data Correct answer.
  3. It reduces the storage cost of the feature store
  4. It guarantees the model will generalize to future data

Explanation

Pinning a dataset snapshot lets a run be re-executed against identical inputs, which is essential for reproducibility and audits. Generalization and storage savings are separate concerns.

Question 32MLOps engineering practice for machine learning lifecycle and deployment

Which of the following best describes the role of a batch inference job in an ML system?

  1. It routes live traffic between two model versions
  2. It scores a large stored dataset on a schedule and writes results for downstream consumption Correct answer.
  3. It answers individual user requests within a strict millisecond budget
  4. It continuously trains the model on incoming labeled data

Explanation

Batch inference processes a bounded, stored dataset on a schedule and persists the scores, which suits use cases like nightly churn scoring. It is distinct from online request-response serving, continuous training, and traffic routing.

Question 33MLOps engineering practice for machine learning lifecycle and deployment

A pipeline step must fail the build when a newly trained model's validation metric falls below an agreed threshold. Which mechanism implements this?

  1. A code review requirement on the training script
  2. A dashboard alert sent to the on-call channel after deployment
  3. A scheduled batch job that rescoring historical data
  4. A quality gate in the pipeline that blocks promotion when the metric check fails Correct answer.

Explanation

A quality gate is an automated check embedded in the pipeline that compares the metric against a threshold and halts promotion on failure. Dashboards and reviews are not automated blocking checks on the artifact.

Question 34MLOps engineering practice for machine learning lifecycle and deployment

A team must guarantee that the container running a model in production is built from the exact code and dependencies that passed testing. Which practice provides this?

  1. Copying model files manually onto the production server
  2. Rebuilding the image on the production host at deploy time from the latest branch
  3. Tagging the image as latest and pulling it during deployment
  4. Pinning the image by immutable digest produced by the CI build that passed tests Correct answer.

Explanation

An immutable digest identifies a specific image build, so the deployed container is byte-identical to the tested one. Rebuilding or using a mutable latest tag can silently pull different code or dependencies.

Question 35MLOps engineering practice for machine learning lifecycle and deployment

A feature transformation is implemented in SQL for training and in Python for online serving. Which risk does this create?

  1. Training-serving skew from divergent transformation semantics Correct answer.
  2. Slower model registry lookups
  3. Reduced model accuracy on the training set
  4. Higher storage cost in the feature store

Explanation

Two independent implementations of the same transformation can differ in edge cases such as nulls, time zones, and rounding, producing skew between training and serving inputs. Centralizing the transformation in a feature store or shared library avoids this.

Question 36MLOps engineering practice for machine learning lifecycle and deployment

An inference service needs to scale to zero during idle periods to control cost while still handling occasional bursts. Which deployment approach fits best?

  1. A permanently provisioned GPU cluster sized for peak traffic
  2. A shadow deployment with a fixed replica count
  3. A serverless or autoscaling endpoint that spins up on demand Correct answer.
  4. A nightly batch inference job

Explanation

Serverless or autoscaling endpoints can scale down when idle and expand on demand, matching intermittent traffic to cost. Fixed clusters sized for peak waste resources during idle periods.

Question 37MLOps engineering practice for machine learning lifecycle and deployment

What is the main purpose of an automated rollback policy tied to serving health metrics?

  1. To delete the new model artifact from the registry permanently
  2. To restore the previous model version quickly when the new one degrades key metrics Correct answer.
  3. To increase the number of replicas during traffic spikes
  4. To retrain the model automatically when accuracy drops

Explanation

A health-linked rollback policy detects degradation such as rising errors or latency and reverts to the last known-good version, limiting user impact. It does not retrain, delete artifacts, or scale replicas.

Question 38MLOps engineering practice for machine learning lifecycle and deployment

A pipeline must run the same preprocessing and evaluation steps across development, staging, and production without drift. Which design choice best supports this?

  1. Maintaining separate hand-written scripts per environment
  2. Running preprocessing only in production to save time
  3. Allowing each environment to install the newest library versions
  4. Packaging preprocessing and evaluation as versioned, reusable pipeline components Correct answer.

Explanation

Versioned reusable components ensure the same code and dependencies execute in every environment, which reduces environment-specific drift. Hand-written per-environment scripts and unpinned libraries reintroduce divergence.

Question 39MLOps engineering practice for machine learning lifecycle and deployment

A team wants to detect data quality problems before a model consumes a new batch of features. Which control is most appropriate?

  1. An increase in the model's serving replica count
  2. A larger training batch size
  3. A manual spot check performed weekly by an analyst
  4. A schema and distribution validation step that fails the pipeline on violation Correct answer.

Explanation

Automated validation of schema, ranges, and distributions at ingestion catches bad data before it reaches training or serving. Manual spot checks are slow and non-blocking, and scaling or batch size does not address data quality.

Question 40MLOps engineering practice for machine learning lifecycle and deployment

A model is retrained nightly, but its evaluation metrics vary widely between nights with no code changes. Which factor most plausibly explains this?

  1. The number of replicas in the registry
  2. Instability in the labeled data or sampling used for each training run Correct answer.
  3. The choice of container base image
  4. The serving endpoint's autoscaling policy

Explanation

When code is unchanged, metric volatility usually traces to data: shifting label distributions, small or noisy samples, or inconsistent data snapshots. Serving and infrastructure settings do not alter offline training metrics.

Question 41MLOps engineering practice for machine learning lifecycle and deployment

Which practice most directly enables auditing who approved a model for production and when?

  1. Compressing the training logs to save space
  2. Running inference on GPU instead of CPU
  3. Storing approval records with the model version in the registry Correct answer.
  4. Encrypting the model artifact at rest

Explanation

Recording approvals and timestamps alongside the model version creates an auditable promotion trail. Encryption, compression, and hardware choice are unrelated to approval governance.

Question 42MLOps engineering practice for machine learning lifecycle and deployment

A team wants to catch silent model degradation where labels arrive only weeks later. Which monitoring approach is most suitable?

  1. Tracking only request latency and HTTP error codes
  2. Waiting for the next quarterly business review
  3. Disabling alerts until labels become available
  4. Monitoring input feature distributions and prediction distributions for drift in the interim Correct answer.

Explanation

When ground-truth labels are delayed, drift in input and prediction distributions serves as an early proxy for degradation. Latency and error codes reflect infrastructure health, not model quality.

Question 43MLOps engineering practice for machine learning lifecycle and deployment

What is the purpose of a smoke test executed immediately after deploying a new model version?

  1. To verify the endpoint responds correctly to a few representative requests Correct answer.
  2. To permanently compare two models across all users
  3. To measure the model's long-term business impact
  4. To retrain the model on production traffic

Explanation

A smoke test sends a small set of representative requests to confirm the endpoint loads the model and returns valid responses. It is a fast sanity check, not an impact measurement or retraining step.

Question 44MLOps engineering practice for machine learning lifecycle and deployment

A team wants reproducibility across machines for a training job. Which combination is most essential?

  1. A marketing dashboard for stakeholders
  2. More frequent model deployments
  3. Pinned dependency versions and a fixed random seed with recorded environment details Correct answer.
  4. A larger GPU and a faster network

Explanation

Reproducibility requires controlling sources of variation: pinned dependencies, fixed seeds, and recorded environment and hardware details. Hardware speed and deployment frequency do not affect determinism.

Question 45MLOps engineering practice for machine learning lifecycle and deployment

A pipeline artifact is promoted to production, but later investigation shows the evaluation used a stale cached feature table. Which control would most likely have prevented this?

  1. Increasing the number of serving replicas
  2. Pinning the feature snapshot version used for evaluation and validating its freshness Correct answer.
  3. Adding more logging to the inference endpoint
  4. Shortening the model's input timeout

Explanation

Binding evaluation to an explicit, freshness-checked feature snapshot prevents stale cached data from silently entering the assessment. Replica counts, timeouts, and endpoint logging do not validate training inputs.

Question 46MLOps engineering practice for machine learning lifecycle and deployment

An organization requires that every deployed model be reproducible from source months later. Which practice is most critical?

  1. Capturing code commit, dependency lockfile, data snapshot, and configuration alongside the artifact Correct answer.
  2. Storing only the final model weights and discarding pipeline configuration
  3. Relying on the data scientist's memory of the training setup
  4. Rebuilding the environment ad hoc when the need arises

Explanation

Reproducibility months later requires the full set of inputs: code version, locked dependencies, the exact data snapshot, and configuration. Weights alone or human memory cannot reconstruct the run.

Question 47MLOps engineering practice for machine learning lifecycle and deployment

A feature is computed from a join across two tables, and the join key occasionally has duplicates in production but not in the historical snapshot. Which failure mode is most likely?

  1. Reduced container image size
  2. Faster inference due to duplicate keys
  3. Automatic improvement in model calibration
  4. Training-serving skew caused by differing join cardinality between offline and online data Correct answer.

Explanation

Duplicate keys change the cardinality and aggregation of features at serving time, so the model receives inputs that differ from those seen in training. This is a classic source of training-serving skew.

Question 48MLOps engineering practice for machine learning lifecycle and deployment

A team runs an A/B test of a new model and observes a statistically significant lift, but the new model also increases p99 latency beyond the service level objective. What is the most sound decision?

  1. Treat latency as a release criterion and either optimize the model or reject the release Correct answer.
  2. Ship the new model immediately because accuracy is the only criterion
  3. Disable latency monitoring to avoid noisy alerts
  4. Ignore latency because it only affects a small fraction of requests

Explanation

A release decision must satisfy all service level objectives, including latency, not just the primary quality metric. Shipping a model that violates the latency SLO degrades user experience regardless of accuracy gains.

Question 49MLOps engineering practice for machine learning lifecycle and deployment

A model's predictions are used to make automated decisions that regulators require to be explainable per case. Which design choice best satisfies this while keeping the model in production?

  1. Storing only the final prediction value
  2. Retraining the model weekly to improve transparency
  3. Removing all input logging to protect privacy
  4. Logging feature attributions or explanation outputs alongside each prediction Correct answer.

Explanation

Per-case explainability requires capturing explanation artifacts such as feature attributions together with the prediction and its inputs. Storing only the output or removing logs leaves no basis for explaining a decision.

Question 50MLOps engineering practice for machine learning lifecycle and deployment

An ML platform team wants to reduce mean time to recovery after a bad model release. Which combination of practices contributes most?

  1. Removing monitoring to simplify operations
  2. Deploying directly to all users without staged rollout
  3. Immutable versioned artifacts, health-based automated rollback, and preserved prior deployments Correct answer.
  4. Longer manual approval chains for every release

Explanation

Fast recovery depends on being able to revert quickly to a known-good artifact, which requires immutable versions, automated health-based rollback, and retained prior deployments. Manual chains and absent monitoring slow recovery.