I85PREP
⌘ K
Intel MLOps Professional study materials
Intel MLOps Professional study resources

Intel MLOps Professional practice questions

One question at a time. Build your Intel MLOps Professional 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 Foundations and Solution Scoping

A team wants to move a notebook-based fraud model into production. Which artifact should be treated as the deployable unit in an MLOps pipeline?

  1. The exploratory notebook with the highest validation accuracy
  2. The raw feature table used during training
  3. The hyperparameter search log from the tuning run
  4. The serialized model plus its preprocessing pipeline and pinned dependency set Correct answer.

Explanation

The deployable unit must reproduce both transformation and prediction behavior, so the model artifact is packaged together with its preprocessing steps and pinned dependencies. Notebooks, raw tables, and search logs are development inputs rather than versioned serving artifacts.

Question 2MLOps Foundations and Solution Scoping

During scoping, a stakeholder asks for a model that predicts customer churn probability next quarter. Which framing best fits a supervised MLOps solution?

  1. Reinforcement learning with churn as a reward signal
  2. Time-series forecasting of total revenue only
  3. Binary classification with a calibrated probability output Correct answer.
  4. Unsupervised clustering of churn reasons

Explanation

Predicting whether a labeled churn outcome occurs within a defined horizon is a binary classification problem, and calibration makes the probability usable for targeting decisions. Clustering, reinforcement learning, and revenue forecasting do not directly estimate per-customer churn probability.

Question 3MLOps Foundations and Solution Scoping

What is the primary purpose of establishing a baseline model early in an MLOps project?

  1. To satisfy regulatory audit requirements for model documentation
  2. To eliminate the need for a validation split
  3. To provide a reference performance level that later models must beat before promotion Correct answer.
  4. To reduce the size of the training dataset needed

Explanation

A baseline, such as a simple heuristic or majority-class predictor, sets the minimum bar a candidate model must exceed to justify added complexity. It does not replace validation design or reduce data needs, and documentation obligations are separate.

Question 4MLOps Foundations and Solution Scoping

A retail team defines success as reducing manual review workload by 30 percent while keeping recall above 0.9. Which scoping element does this statement represent?

  1. Model explainability method
  2. Infrastructure capacity plan
  3. Data retention policy
  4. Business and model success criteria Correct answer.

Explanation

The statement pairs a business outcome with a model performance threshold, which is exactly what success criteria capture during scoping. Retention, capacity, and explainability are downstream design choices that follow from agreed criteria.

Question 5MLOps Foundations and Solution Scoping

Which practice most directly supports reproducibility of a training run?

  1. Recording data snapshot identifiers, code commit, hyperparameters, and environment versions Correct answer.
  2. Re-running training manually whenever results are questioned
  3. Storing only the final model weights in a shared drive
  4. Relying on the analyst's local machine configuration

Explanation

Reproducibility requires capturing every input that influences the result: data version, code revision, hyperparameters, and environment. Storing only weights or depending on local setups loses the context needed to regenerate the same model.

Question 6MLOps Foundations and Solution Scoping

In an MLOps architecture, what distinguishes a feature store from a plain data warehouse?

  1. It replaces the need for a model registry
  2. It serves consistent feature definitions for both training and low-latency online inference Correct answer.
  3. It stores only raw event logs for archival purposes
  4. It automatically retrains models on a fixed schedule

Explanation

A feature store centralizes feature logic and provides both offline access for training and online access for real-time serving, reducing training-serving skew. Warehouses store data but do not inherently guarantee consistent online feature serving.

Question 7MLOps Foundations and Solution Scoping

A candidate model improves offline AUC by 0.03 but raises median inference latency from 40 ms to 400 ms against a 100 ms service-level objective. What is the most appropriate scoping decision?

  1. Lower the service-level objective to 400 ms
  2. Deploy and monitor latency after launch
  3. Reject or optimize the model because it violates the latency objective Correct answer.
  4. Deploy immediately because offline AUC is the primary metric

Explanation

Serving constraints are part of the agreed solution scope, so a model that breaches the latency SLO should be optimized or rejected regardless of offline gains. Deploying first and adjusting targets later shifts risk onto users and undermines the scoping agreement.

Question 8MLOps Foundations and Solution Scoping

Which scenario best justifies investing in an automated retraining pipeline rather than periodic manual retraining?

  1. Stakeholders request a single offline report
  2. Input feature distributions drift frequently and labels arrive with short delay Correct answer.
  3. The model is a one-time analysis with no production consumer
  4. The training dataset is static and will never change

Explanation

Frequent drift combined with timely label availability means model staleness accrues quickly, making automation worthwhile. Static data or one-off reporting does not create recurring retraining need.

Question 9MLOps Foundations and Solution Scoping

A team must choose between batch scoring nightly and real-time inference for a credit limit adjustment. Which factor most strongly drives the decision?

  1. The choice of gradient boosting versus neural network
  2. How quickly the downstream decision must react to new input events Correct answer.
  3. The size of the training dataset
  4. The number of features in the model

Explanation

Serving mode is driven by decision latency requirements: if the business can wait until the next day, batch is simpler and cheaper. Feature count, algorithm family, and training size affect cost and complexity but not the fundamental latency need.

Question 10MLOps Foundations and Solution Scoping

Which element belongs in a model card produced as part of MLOps governance?

  1. The internal IP addresses of training nodes
  2. The salary bands of the data science team
  3. Intended use, training data description, and known limitations Correct answer.
  4. The private API keys used by the serving cluster

Explanation

Model cards document intended use, data provenance, evaluation results, and limitations so consumers can judge fitness. Credentials, personnel data, and infrastructure addresses are operational secrets unrelated to model documentation.

Question 11MLOps Foundations and Solution Scoping

A project sponsor asks why data versioning is needed when the code repository is already under version control. What is the strongest justification?

  1. It allows analysts to skip writing unit tests
  2. It guarantees the model will generalize to new data
  3. Model behavior depends on the exact training data, which code versioning does not capture Correct answer.
  4. Data versioning reduces cloud storage costs significantly

Explanation

Code versioning tracks logic but not the datasets that determine learned parameters, so data versioning is required to reproduce or audit a model. Cost savings, testing shortcuts, and generalization claims are not consequences of data versioning.

Question 12MLOps Foundations and Solution Scoping

Which metric pair is most appropriate for scoping an imbalanced fraud detection solution where missed fraud is costly?

  1. R-squared and mean absolute error
  2. Silhouette score and cluster count
  3. Overall accuracy and training loss
  4. Recall at a fixed precision threshold and expected cost per transaction Correct answer.

Explanation

For rare positive events with asymmetric costs, recall at an acceptable precision level plus a cost-based measure reflects business impact. Accuracy is misleading under imbalance, and regression or clustering metrics do not apply to classification.

Question 13MLOps Foundations and Solution Scoping

What is the main purpose of a model registry in an MLOps platform?

  1. To execute hyperparameter search jobs
  2. To host the feature computation code for online serving
  3. To track model versions, lineage, and promotion stage across environments Correct answer.
  4. To store raw training data in Parquet format

Explanation

A registry records model versions together with metadata such as lineage, metrics, and lifecycle stage, enabling controlled promotion and rollback. Data storage, tuning execution, and feature computation are handled by other components.

Question 14MLOps Foundations and Solution Scoping

A stakeholder requests a fully autonomous decision system, but regulations require human sign-off for adverse actions. How should the scoping be adjusted?

  1. Replace the model with a random selection process
  2. Abandon the project because automation is impossible
  3. Design human-in-the-loop review while automating scoring and prioritization Correct answer.
  4. Proceed with full automation and address compliance after launch

Explanation

Scoping should align automation with regulatory constraints by automating the parts that are permitted, such as scoring and case prioritization, while retaining required human approval. Ignoring the constraint or scrapping the project both fail to deliver value within the rules.

Question 15MLOps Foundations and Solution Scoping

Which factor most affects the choice between managed ML services and self-hosted open-source tooling during solution scoping?

  1. Team skills, compliance constraints, and total cost of ownership over the project lifetime Correct answer.
  2. The number of questions in the vendor's certification exam
  3. The color scheme of the monitoring dashboard
  4. The popularity of the tool on social media

Explanation

Platform choice should weigh operational skills, data residency and compliance requirements, and lifetime cost including maintenance. Superficial popularity or cosmetic factors are not valid scoping inputs.

Question 16MLOps Foundations and Solution Scoping

A team plans to collect user behavior data for a recommendation model. Which scoping artifact should define what is collected and why?

  1. A hyperparameter search space
  2. A data collection and usage plan with purpose limitation Correct answer.
  3. A container base image specification
  4. A model serving autoscaling policy

Explanation

Purpose limitation and collection scope belong in a data plan agreed before collection begins, supporting privacy compliance and clear justification. Autoscaling, base images, and search spaces are technical details defined later.

Question 17MLOps Foundations and Solution Scoping

During scoping, a team discovers that labels for a defect detection task arrive two weeks after the image is captured. What is the most significant implication?

  1. The problem must be reformulated as reinforcement learning
  2. The model cannot use image data at all
  3. Labels become unnecessary for evaluation
  4. Delayed feedback constrains monitoring and retraining cadence and must be designed for Correct answer.

Explanation

Label delay means performance cannot be measured immediately and retraining loops are slower, so monitoring and pipeline design must account for the lag. It does not invalidate image inputs, force a reinforcement learning framing, or remove the need for labels.

Question 18MLOps Foundations and Solution Scoping

A solution architect proposes using the same feature computation code for offline training and online serving. Which risk does this design primarily mitigate?

  1. Overfitting caused by too many features
  2. Label leakage across train and test splits
  3. Training-serving skew caused by divergent feature logic Correct answer.
  4. Class imbalance in the target variable

Explanation

Sharing one feature definition between batch and online paths prevents the two implementations from drifting apart, which is the root of training-serving skew. Overfitting, leakage, and imbalance arise from data and modeling choices rather than duplicated feature logic.

Question 19MLOps Foundations and Solution Scoping

An organization must decide whether a new model can be promoted to production. Which combination best represents a complete promotion gate?

  1. Offline accuracy alone
  2. Successful training job completion
  3. Approval by the model's original author
  4. Offline metric thresholds, fairness and robustness checks, and operational readiness such as latency and rollback Correct answer.

Explanation

A promotion gate should verify predictive quality, responsible-AI checks, and operational readiness including rollback capability. Single metrics, author approval, or job completion each cover only a fragment of the evidence needed.

Question 20MLOps Foundations and Solution Scoping

A team scoping an MLOps solution finds that the model's predictions influence loan approvals. Which requirement should shape the architecture from the outset?

  1. Using the largest available deep learning model
  2. Removing all human review to reduce processing time
  3. Maximizing the number of features in the model
  4. Auditable decision logging and explainability sufficient for adverse action notices Correct answer.

Explanation

High-stakes decisions require traceable records of inputs, model version, and rationale so decisions can be explained and audited. Feature count, model size, and eliminating review do not address accountability and may worsen it.

Question 21MLOps Foundations and Solution Scoping

Which statement best describes the relationship between MLOps maturity and the level of automation in a solution scope?

  1. Automation should be maximized immediately regardless of safeguards
  2. Automation should increase as monitoring, testing, and rollback capabilities mature Correct answer.
  3. Maturity only concerns the size of the data science team
  4. Automation and maturity are unrelated concepts

Explanation

Higher automation is safe only when detection, testing, and recovery mechanisms are in place, so maturity and automation advance together. Maximizing automation without safeguards increases risk, and maturity is about process capability rather than headcount.

Question 22MLOps Foundations and Solution Scoping

What does a data contract between an upstream source system and an ML pipeline primarily specify?

  1. The visualization style for monitoring charts
  2. The model's hyperparameter values
  3. Schema, semantics, freshness, and quality expectations for the delivered data Correct answer.
  4. The deployment region for the serving cluster

Explanation

Data contracts formalize the structure, meaning, timeliness, and quality guarantees that producers commit to, protecting downstream training and serving. Hyperparameters, deployment regions, and chart styling are unrelated concerns.

Question 23MLOps Foundations and Solution Scoping

A team estimates that a proposed model will require 10x more labeled data but only promises a 1 percent accuracy gain. What scoping conclusion is most reasonable?

  1. Reassess the value proposition and consider a simpler model with comparable business impact Correct answer.
  2. Approve the plan because accuracy is the only relevant measure
  3. Approve the plan because more data always improves outcomes
  4. Cancel all ML work for the use case

Explanation

Marginal gains should be weighed against labeling cost and risk, so a simpler alternative with similar business impact is often preferable. Assuming more data is always better ignores diminishing returns, and abandoning the use case entirely is an overreaction.

Question 24MLOps Foundations and Solution Scoping

Which activity is part of the discovery phase rather than the deployment phase of an ML project?

  1. Interviewing stakeholders to define the decision the model will support Correct answer.
  2. Setting up alerting on prediction drift
  3. Packaging the model into a container image
  4. Configuring canary traffic splitting in the serving layer

Explanation

Discovery focuses on understanding the business decision, users, and constraints before technical design begins. Canary releases, drift alerting, and containerization are deployment-phase engineering tasks.

Question 25Data Engineering and Feature Pipelines

In a feature pipeline, which component is primarily responsible for computing aggregate statistics such as mean and standard deviation used later during model serving?

  1. The online feature store lookup layer
  2. The inference request router
  3. The model registry metadata service
  4. The training-time transformation stage Correct answer.

Explanation

Training-time transformations compute statistics over the training dataset and persist them so that serving-time transforms apply identical logic. The online store only serves precomputed values, the registry stores model artifacts, and the router dispatches requests without computing statistics.

Question 26Data Engineering and Feature Pipelines

A team wants point-in-time correctness when building training datasets from an event table. Which practice achieves this?

  1. Joining each label to features whose event timestamp is at or before the label timestamp Correct answer.
  2. Joining each label to the most recent feature row regardless of timestamp
  3. Joining labels to features after deduplicating on the label identifier
  4. Joining labels to features using the ingestion date of the feature table

Explanation

Point-in-time correctness requires that only feature values known before the label event are used, preventing label leakage from future data. Joining to the latest row or by ingestion date can include information unavailable at prediction time. Deduplicating on label ID does not address temporal ordering.

Question 27Data Engineering and Feature Pipelines

Which storage pattern is designed to serve low-latency single-entity feature lookups during online inference?

  1. A columnar data lake with Parquet files
  2. A key-value store indexed by entity ID Correct answer.
  3. A message queue partitioned by timestamp
  4. A distributed batch processing cluster

Explanation

Online serving requires millisecond lookups by entity key, which key-value stores such as Redis or Cassandra provide. Columnar lakes and batch clusters target throughput-oriented analytics, and message queues are for streaming transport rather than random-access retrieval.

Question 28Data Engineering and Feature Pipelines

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

  1. To catalog feature definitions with metadata, ownership, and lineage Correct answer.
  2. To schedule periodic retraining jobs across clusters
  3. To execute distributed SQL queries over raw event logs
  4. To host trained model binaries for A/B testing

Explanation

A feature registry provides a discoverable catalog of feature definitions, including ownership, descriptions, and lineage, enabling reuse and governance. Query execution, model hosting, and job scheduling are handled by other platform components.

Question 29Data Engineering and Feature Pipelines

A pipeline must guarantee that a downstream training job only starts after all upstream feature partitions succeed. Which orchestration capability enforces this?

  1. Schema evolution of the feature table
  2. Role-based access control on datasets
  3. Horizontal autoscaling of worker nodes
  4. Task dependency management with failure propagation Correct answer.

Explanation

Dependency management ensures downstream tasks wait for upstream success and fail when prerequisites fail, providing the required ordering guarantee. Autoscaling, schema evolution, and access control do not enforce execution-order dependencies.

Question 30Data Engineering and Feature Pipelines

Which technique best mitigates training-serving skew caused by divergent feature transformation code?

  1. Increasing the online store replication factor
  2. Reimplementing transforms separately in each environment
  3. Sharing a single transformation library between training and serving Correct answer.
  4. Freezing the training dataset after the first model release

Explanation

Using one shared transformation implementation eliminates divergence between batch and online code paths. Separate reimplementations invite drift, freezing data does not fix code differences, and replication affects availability rather than transform consistency.

Question 31Data Engineering and Feature Pipelines

Which data format is most appropriate for storing large feature tables that are queried column-wise by a batch analytics engine?

  1. CSV
  2. Parquet Correct answer.
  3. JSON lines
  4. XML

Explanation

Parquet is a columnar format with compression and predicate pushdown, making it efficient for column-wise analytical scans. JSON lines, CSV, and XML are row-oriented text formats with higher parse and I/O overhead for wide tables.

Question 32Data Engineering and Feature Pipelines

What does a feature freshness SLA define in a streaming feature pipeline?

  1. The number of replicas maintained for the online feature store
  2. The maximum number of columns allowed in a feature group
  3. The retention period for raw event logs in cold storage
  4. The maximum acceptable delay between an event occurring and its feature being queryable Correct answer.

Explanation

Freshness SLAs bound the latency from event occurrence to feature availability, guiding streaming architecture and alerting. Replica counts, column limits, and log retention are capacity or governance settings unrelated to freshness.

Question 33Data Engineering and Feature Pipelines

A training dataset join accidentally includes feature values recorded after the prediction timestamp. What is the most direct consequence?

  1. The model registry will refuse to version the artifact
  2. The model may learn from future information and overstate offline performance Correct answer.
  3. The feature store will reject the dataset during materialization
  4. The online serving latency will increase proportionally

Explanation

Including post-prediction values is temporal leakage, which inflates offline metrics because the model exploits information unavailable at inference. Feature stores, serving latency, and registry versioning are not directly affected by this leakage.

Question 34Data Engineering and Feature Pipelines

A feature pipeline computes a rolling 7-day average per user. Which window specification avoids leaking the current day's label into the feature?

  1. A full-history window ending at dataset creation time
  2. A trailing window that excludes the current event Correct answer.
  3. A forward-looking window starting at the current event
  4. A centered window spanning three days on each side

Explanation

A trailing window ending before the current event uses only prior observations, preserving temporal causality. Centered and forward-looking windows incorporate future data, and a full-history window ending at dataset creation includes events after the label.

Question 35Data Engineering and Feature Pipelines

An online feature store shows stale values for a subset of entities while the batch job reports success. Which cause is most plausible?

  1. The offline table uses Parquet compression
  2. The model registry was updated recently
  3. The training job used a larger batch size
  4. Materialization to the online store skipped partitions for those entities Correct answer.

Explanation

Stale online values with a successful batch run typically indicate that materialization did not publish the affected partitions to the online store. Format choice, registry updates, and training batch size do not govern online freshness.

Question 36Data Engineering and Feature Pipelines

Which approach best supports backfilling a newly defined feature across two years of historical events?

  1. Replay the transformation logic over partitioned historical data with an idempotent job Correct answer.
  2. Delete the online store and rebuild it from the model registry
  3. Increase the TTL of cached inference responses
  4. Retrain the model on the current snapshot only

Explanation

Backfilling requires re-running deterministic, idempotent transformation code over historical partitions so reruns do not duplicate or corrupt data. The registry, current snapshots, and cache TTLs do not reconstruct historical feature values.

Question 37Data Engineering and Feature Pipelines

A streaming job computes features with exactly-once semantics into the online store. Which mechanism is essential to achieve this?

  1. Multiple consumer groups reading the same topic
  2. Larger consumer batch sizes
  3. Compression of the event payload
  4. Idempotent writes keyed by event or offset Correct answer.

Explanation

Exactly-once delivery into a store depends on idempotent or transactional writes so retries and replays do not duplicate updates. Batch size, consumer group count, and payload compression affect throughput or isolation but not duplicate suppression.

Question 38Data Engineering and Feature Pipelines

What is the main benefit of defining features as versioned transformation code rather than manually curated tables?

  1. Guaranteed lower storage cost for all feature groups
  2. Automatic improvement of model accuracy
  3. Elimination of the need for data quality checks
  4. Reproducibility of feature values across environments and time Correct answer.

Explanation

Versioned transformation code lets teams recompute identical feature values in any environment and audit changes over time. It does not remove the need for quality checks, guarantee storage savings, or by itself improve model accuracy.

Question 39Data Engineering and Feature Pipelines

A team observes that offline AUC is high but online metrics are poor, and investigation shows the online transform uses a different default for missing categorical values. Which fix addresses the root cause?

  1. Retrain the model with a lower learning rate
  2. Add more replicas to the online feature store
  3. Unify missing-value handling in the shared transformation library Correct answer.
  4. Shorten the feature freshness SLA

Explanation

Divergent missing-value defaults between batch and online transforms create training-serving skew, so consolidating the logic in one shared library resolves it. Learning rate, replication, and SLA changes do not reconcile the transform mismatch.

Question 40Data Engineering and Feature Pipelines

Which data quality check is most appropriate for detecting a sudden drop in the proportion of non-null values for a critical feature?

  1. A schema compatibility check for column types
  2. A referential integrity check against a dimension table
  3. A uniqueness constraint on the primary key
  4. A completeness threshold evaluated per pipeline run Correct answer.

Explanation

Completeness checks measure the share of populated values and can alert when null rates spike. Uniqueness, referential integrity, and type compatibility validate different properties and would not catch a null-rate shift.

Question 41Data Engineering and Feature Pipelines

In a feature pipeline, what role does a schema contract play between an upstream producer and a downstream consumer?

  1. It defines expected fields and types so breaking changes are caught before consumption Correct answer.
  2. It encrypts feature values in transit between services
  3. It specifies the hardware profile for training clusters
  4. It determines the model's hyperparameter search space

Explanation

Schema contracts declare field names, types, and compatibility rules, allowing validation to fail fast when producers introduce breaking changes. Encryption, hardware profiles, and hyperparameter search are unrelated concerns.

Question 42Data Engineering and Feature Pipelines

A feature group must support both batch training over years of history and low-latency online reads. Which architecture satisfies both requirements?

  1. A message queue retained for the full history
  2. An offline store for history plus a materialized online store kept in sync Correct answer.
  3. A single online key-value store holding only the latest values
  4. A data lake queried directly at inference time

Explanation

Dual-store architectures keep full history offline for training and materialize recent values into a low-latency online store for serving. Online-only storage loses history, direct lake queries are too slow, and queues are not designed for long-term analytical reads.

Question 43Data Engineering and Feature Pipelines

Which metric best monitors the operational health of a streaming feature pipeline's event-to-feature latency?

  1. The average size of training batches
  2. The count of model versions in the registry
  3. The number of registered feature definitions
  4. The 99th percentile of processing delay per event Correct answer.

Explanation

Tail latency such as p99 processing delay directly measures how long events take to become queryable features. Feature definition counts, model versions, and batch sizes do not reflect event-to-feature latency.

Question 44Data Engineering and Feature Pipelines

A pipeline deduplicates events using a non-deterministic timestamp such as processing time. Why is this problematic for feature correctness?

  1. It increases the compression ratio of the output files
  2. It forces the online store to use a different key type
  3. It reduces the number of partitions in the dataset
  4. Reruns may select different records, producing non-reproducible feature values Correct answer.

Explanation

Processing time varies across runs, so deduplication can keep different records and yield different feature values on recomputation. Compression, partitioning, and key types are unaffected by the choice of deduplication timestamp.

Question 45Data Engineering and Feature Pipelines

A team needs to add a new feature to an existing feature group without breaking consumers that read the previous schema. Which strategy is safest?

  1. Add the field as optional with a backward-compatible default Correct answer.
  2. Rename an existing required field to the new name
  3. Change the type of a required field in place
  4. Remove the least-used existing field first

Explanation

Optional fields with defaults preserve backward compatibility so existing consumers continue to work. Renaming or retyping required fields and removing fields are breaking changes that can fail downstream jobs.

Question 46Data Engineering and Feature Pipelines

Two pipelines compute the same churn feature: one derives it from raw clickstream events, the other from a pre-aggregated daily table. What risk should be evaluated before treating them as interchangeable?

  1. Increased model registry version churn
  2. Lower availability of the daily table during weekends
  3. Semantic drift from differing aggregation boundaries and late-arriving events Correct answer.
  4. Higher storage cost of the clickstream pipeline only

Explanation

Deriving from raw events versus pre-aggregates can differ in window boundaries, handling of late events, and time zones, producing semantically different values. Storage cost, weekend availability, and registry churn do not make the two definitions equivalent.

Question 47Data Engineering and Feature Pipelines

A backfill job reruns over historical partitions but writes duplicates because it uses append mode. Which change most directly restores idempotency?

  1. Increase the number of shuffle partitions
  2. Enable compression on the output files
  3. Run the job with a larger executor memory
  4. Write with partition overwrite or merge keyed by entity and timestamp Correct answer.

Explanation

Idempotent backfills replace or merge target partitions deterministically rather than appending, so reruns converge to the same result. Shuffle partitioning, compression, and executor memory affect performance but not duplicate writes.

Question 48Data Engineering and Feature Pipelines

A feature store enforces a TTL on online values. What failure mode appears if the TTL is shorter than the interval between materialization runs?

  1. The model registry deletes older model versions
  2. The schema contract rejects new optional fields
  3. The offline store loses historical partitions permanently
  4. Entities whose features were not recently refreshed return missing values at serving time Correct answer.

Explanation

If values expire before the next materialization, requests for entities not refreshed in time find no online value and fall back or fail. TTL settings do not delete offline history, affect schema contracts, or remove registry versions.

Question 49Data Engineering and Feature Pipelines

Which design most effectively prevents a single faulty upstream event schema change from corrupting a production feature table?

  1. Write directly to production and rely on dashboards to detect issues
  2. Validate incoming data against a contract in a staging layer before promotion Correct answer.
  3. Disable schema validation to avoid pipeline failures
  4. Allow producers to push changes without notifying consumers

Explanation

Validating against a contract in staging isolates bad data before it reaches production tables. Direct writes, disabled validation, and silent producer changes let corruption propagate to consumers.

Question 50Data Engineering and Feature Pipelines

In a feature pipeline, what distinguishes an idempotent write to a partitioned offline feature table from a merely retryable one?

  1. Re-running the same partition with the same inputs produces the same final table state regardless of how many times it executes Correct answer.
  2. Each retry writes to a new timestamped partition so no data is ever overwritten
  3. Retries always run on a different worker node to avoid correlated hardware failures
  4. Retries are attempted a fixed number of times before the job is marked failed and paged to on-call

Explanation

Idempotency is a property of the write semantics: replaying the same logical unit of work must converge on the same table contents, typically via partition overwrite or deterministic upsert keys. Retryability only concerns whether a failed attempt can be re-attempted; a job can be retried many times and still duplicate rows if it appends.