Data Quality & Validation quiz

Question 1

What is the difference between data quality dimensions: completeness, accuracy, consistency, and timeliness?

Accepted Answer

Completeness=presence, Accuracy=correctness, Consistency=no contradictions, Timeliness=currency — four core data quality dimensions

Answer

Accuracy and completeness are the only internationally recognized data quality dimensions

Answer

Consistency measures the speed of database queries

Answer

Timeliness is only relevant for real-time streaming data, not batch pipelines

Question 2

What is data drift and how is it detected and monitored in production ML systems?

Accepted Answer

Data drift is detected using KS tests, PSI, or MMD to compare training vs. production distributions; PSI>0.2 indicates significant drift

Answer

Data drift can only be detected after model accuracy drops significantly

Answer

Concept drift and covariate shift are identical phenomena

Answer

Data drift only affects classification models, not regression

Question 3

What is the Population Stability Index (PSI) and how is it calculated?

Accepted Answer

PSI = Σ (P_current - P_expected) × ln(P_current/P_expected) per bin; PSI>0.2 indicates significant distribution shift

Answer

PSI measures the percentage of null values in a feature distribution

Answer

PSI is equivalent to KL divergence but only applicable to categorical features

Answer

PSI < 0.2 always indicates the model needs immediate retraining

Question 4

What is data observability and how does it differ from traditional data quality monitoring?

Accepted Answer

Data observability uses ML-based anomaly detection across freshness, volume, distribution, schema, and lineage vs. rule-based monitoring

Answer

Data observability is identical to data quality monitoring but uses a different marketing term

Answer

Data observability only monitors SQL databases, not streaming systems

Answer

Traditional data quality monitoring automatically detects schema changes without configuration

Question 5

What is the data validation pyramid and how should validation tests be prioritized?

Accepted Answer

The validation pyramid prioritizes schema checks first (fast/cheap), then constraints, statistics, business rules, and ML-specific checks at top

Answer

All validation checks should run at equal frequency regardless of computational cost

Answer

ML-specific drift checks should run first as they catch the most critical issues

Answer

Statistical checks are never necessary if schema validation passes

Question 6

What is the difference between fail-fast and warn-and-continue validation strategies in data pipelines?

Accepted Answer

Fail-fast halts on critical violations; warn-and-continue logs and proceeds—appropriate severity depends on downstream impact

Answer

Fail-fast is always preferred because data quality must be perfect before proceeding

Answer

Warn-and-continue should be used for all checks to maximize pipeline availability

Answer

Both strategies produce identical pipeline behavior with different log formats

Question 7

What is feature store validation and how does it prevent training-serving skew?

Accepted Answer

Feature stores prevent training-serving skew by centralizing transformation code run identically at training and serving time with point-in-time correctness

Answer

Training-serving skew only occurs with streaming features, not batch features

Answer

Feature stores eliminate the need for data validation by enforcing schema automatically

Answer

Point-in-time correctness means features must be computed in real-time, not pre-computed

Question 8

How does Apache Iceberg's table format improve data quality management in lakehouses?

Accepted Answer

Iceberg provides ACID transactions, time travel for rollback, schema evolution, and row-level deletes for data quality management in lakehouses

Answer

Iceberg is only for data warehouse environments, not data lakes

Answer

Iceberg requires Spark as the execution engine and cannot be used with other engines

Answer

Iceberg's time travel feature only supports rolling back one snapshot

Question 9

What is data masking and when is it required for ML training data compliance?

Accepted Answer

Data masking replaces PII with realistic non-identifiable substitutes, required by GDPR, HIPAA, and PCI-DSS for ML training data compliance

Answer

Data masking is only required for production user-facing systems, not ML training data

Answer

Format-preserving encryption decrypts data at query time, making it equivalent to no masking

Answer

GDPR only requires data masking for European users' data accessed from Europe

Question 10

What is the Write-Audit-Publish (WAP) pattern and how does it enforce data quality before exposure?

Accepted Answer

WAP writes to staging, validates quality, then atomically publishes to production—consumers never see data that failed validation

Answer

WAP writes and publishes simultaneously, then runs audits for monitoring purposes

Answer

WAP is only applicable to streaming pipelines, not batch ETL

Answer

The Publish step in WAP duplicates data to multiple regions for availability

Question 11

How do you perform statistical hypothesis testing to validate that two dataset samples come from the same distribution?

Accepted Answer

KS test compares empirical CDFs non-parametrically; chi-squared tests categorical distributions; Wasserstein measures shift magnitude—all used for distribution validation

Answer

Only the KS test is statistically valid for comparing ML dataset distributions

Answer

Chi-squared test is appropriate for continuous numerical feature comparisons

Answer

Statistical significance alone is sufficient to determine practical data quality impact

Test yourself