Data Exploration and Quality Assessment Questions

Investigate a dataset thoroughly before analysis or reporting by profiling its structure, contents, and reliability. Typical steps include examining row counts and data volume, inspecting column data types and sample values, validating date formats and ranges, and identifying missing values, duplicates, outliers, and impossible values. Understand schema and relationships between tables or files, check data freshness and latency, and characterize data completeness and coverage with simple metrics and queries. Document discovered issues, their likely causes and impacts on conclusions, and pragmatic workarounds or transformation strategies to mitigate risk. Use exploratory queries and summary statistics to quantify data quality, note limitations and assumptions, and allocate an appropriate portion of case study time to data assessment before proceeding to modeling or visualization.

MediumTechnical

22 practiced

Compare and contrast the following imputation techniques for numeric features: mean/median imputation, k-NN imputation, regression-based imputation, and multiple imputation. For each method describe assumptions about missingness, computational cost, and expected impact on downstream predictive performance.

Sample Answer

Mean/median imputation- What: Replace missing numeric values with the column mean (sensitive to outliers) or median (robust).- Missingness assumption: Effectively assumes MCAR (missing completely at random) or that missingness is unrelated to the value; otherwise it biases distributions.- Computational cost: O(n) per column — trivial, very fast and memory-light.- Impact on modeling: Shrinks variance and can distort correlations, often underestimates uncertainty. Works as a quick baseline or for models that are robust to biased imputations (e.g., tree models), but can degrade linear-model coefficients and uncertainty estimates.

k-NN imputation- What: Impute using weighted average of k nearest neighbors in feature space.- Missingness assumption: Assumes MAR (missing at random) conditional on observed features — similar instances predict the missing value.- Computational cost: O(n^2) naive; with KD-trees/approx neighbors ~ O(n log n) for query; expensive for high-dimensional data.- Impact on modeling: Preserves local structure and multivariate relationships better than mean/median; can improve predictive performance but risks propagating noise, biased neighbors if many features missing, and leakage if not done inside cross-validation.

Regression-based imputation- What: Predict missing feature using a regression model (linear, tree-based) trained on rows where it’s observed.- Missingness assumption: Best under MAR — missingness depends on other observed variables. If MNAR, biased unless model accounts for missingness mechanism.- Computational cost: Moderate — training one model per imputed feature; cost depends on model complexity.- Impact on modeling: Preserves relationships between features, often improves downstream model performance vs simple imputation. Risks overfitting / optimistic variance reduction because predicted values are treated as real; can inflate correlation and understate uncertainty.

Multiple imputation (e.g., MICE)- What: Generate multiple plausible complete datasets by sampling from the posterior predictive distribution for missing values, analyze each, then pool results.- Missingness assumption: Designed for MAR (can incorporate auxiliary variables); provides more defensible inference under this assumption. MNAR requires explicit modeling of missingness.- Computational cost: Highest — iterative model fitting and generating m datasets (m typically 5–20); can be parallelized.- Impact on modeling: Best for valid uncertainty quantification and unbiased parameter estimates; generally yields better predictive performance and calibrated standard errors than single imputation. More complex to implement and requires careful model specification.

Practical guidance- Use mean/median for quick EDA or when missingness is tiny and model choice is robust. Use k-NN or regression when you need to preserve multivariate structure and missingness is MAR, remembering to impute within CV folds. Use multiple imputation when inference and uncertainty quantification matter or missingness is nontrivial. Always analyze missingness patterns (MCAR/MAR/MNAR), include missingness indicators as features when informative, and validate imputation impact by comparing model performance and residual diagnostics.

EasyTechnical

21 practiced

Explain what data exploration and data quality assessment mean in the context of a data science project. List the typical steps you would perform when you first receive a new dataset (including automatic checks and manual inspections) and explain why each step matters. Provide three concrete examples of issues you might discover during profiling and how each would impact downstream modelling or reporting.

Sample Answer

Data exploration is the initial investigation of a dataset to understand distributions, relationships, and patterns. Data quality assessment evaluates completeness, accuracy, consistency, and provenance to determine fitness for analysis or modelling.

Typical first-step checklist (automatic checks + manual inspection) and why each matters:- Ingest and schema check (automatic): verify column names, types, row counts, primary keys, and readability. Why: prevents downstream type errors and ensures joins work.- Missingness profiling (automatic): per-column null rates, missing-by-group. Why: guides imputation strategy or exclusion.- Value-range and distribution checks (automatic + visual): histograms, boxplots, quantiles. Why: detects outliers, skewness, need for transforms.- Uniqueness and duplicate detection (automatic): duplicate rows, duplicate keys. Why: duplicates bias aggregates and leak information into models.- Referential integrity and foreign-key checks (automatic): confirm joins across tables. Why: broken relations cause data loss or wrong joins.- Type/coercion and parsing checks (automatic): dates, numerics parsed correctly. Why: mis-parsed types break time-series logic and aggregations.- Spot-checks and contextual validation (manual): sample rows, compare to business expectations, check units. Why: human domain knowledge catches semantic errors.- Provenance and metadata review (manual): source, update frequency, collection method. Why: affects modelling assumptions and drift risk.- Quick feature correlation matrix and missingness patterns (automatic + manual): identify collinearity and informative missingness. Why: informs feature selection.

Three concrete profiling issues and impact:1) High null rate in a predictor (e.g., 60% missing credit_income): Impact — imputation may introduce bias; models may underperform or require feature engineering (missingness indicator).2) Outliers / impossible values (e.g., negative ages, timestamps in future): Impact — skew metrics, break time-series models, produce misleading aggregates; requires cleaning or capping.3) Duplicate customer IDs with conflicting labels: Impact — label leakage or incorrect target frequency leading to overoptimistic performance; must deduplicate or reconcile labels before modelling.

These steps balance automated checks for scale with manual domain validation to ensure models and reports are trustworthy.

HardTechnical

25 practiced

You have dozens of data quality issues across datasets but only a small team. Propose a prioritization framework that balances business impact, severity, effort to fix, and downstream risk to models and reports. Provide a scoring rubric, an example scoring of three hypothetical issues, and explain how you would engage stakeholders to get fixes approved.

Sample Answer

Situation: Our team inherited dozens of data-quality (DQ) issues but only two engineers and myself to remediate them; stakeholders expect minimal disruption and model/report reliability.

Approach / Framework (what I’d implement):- Create a transparent, numeric prioritization score combining four axes: Business Impact (BI), Severity (SEV), Effort (EFF), Downstream Risk (DR).- Use weighted sum: Priority = 0.35*BI + 0.25*DR + 0.20*SEV + 0.20*(1/EFF) — higher = higher priority. I invert Effort so low-effort-high-priority surfaces.

Scoring rubric (0–5 each):- Business Impact (0 none → 5 strategic revenue/cost-critical)- Downstream Risk (0 no consumers → 5 breaks models/regulated reports)- Severity (0 cosmetic → 5 corrupts primary key or values)- Effort (0 trivial fix → 5 multi-team infrastructure change)

Example scoring:1) Missing customer IDs in transaction feed:- BI=5 (affects billing), DR=5 (breaks churn model + invoices), SEV=4, EFF=3Priority = 0.35*5 + 0.25*5 + 0.20*4 + 0.20*(1/3≈0.33) = 1.75+1.25+0.8+0.066 = 3.866 → Top priority

2) Sporadic nulls in experimental feature table:- BI=3, DR=4 (A/B analysis risk), SEV=3, EFF=2Priority ≈ 0.35*3 +0.25*4 +0.20*3 +0.20*0.5 =1.05+1.0+0.6+0.1 =2.75 → Medium

3) Typos in non-critical product descriptions:- BI=1, DR=1, SEV=1, EFF=1Priority ≈ 0.35+0.25+0.20+0.20 =1.0 → Low

How I’d engage stakeholders:- Present the rubric and a ranked backlog in a short SLA-aligned report; show metrics: affected models/reports, estimated user/financial impact, and shipping risk.- Run a 30–45 minute prioritization workshop with product, analytics, and engineering to validate weights and capture risk tolerances; get formal sign-off on top 5 items.- For top items: propose quick mitigations (monitoring/feature flags) + a 2-week fix plan; request resource/time commitments or temporary re-prioritization.- Use lightweight governance: publish weekly status, incident postmortems, and evolve the rubric with stakeholder feedback.

Result (expected): Fast reduction of highest-risk failures, fewer model regressions, and stakeholder buy-in because decisions are data-driven, auditable, and time-boxed.

HardTechnical

38 practiced

Describe techniques to optimize SQL queries used for profiling very large tables (hundreds of GB or more). Cover strategies such as incremental processing, pre-aggregations, approximate functions, columnar storage advantages, partition pruning, and use of system-level statistics. Provide examples and guidelines for choosing techniques based on data size and freshness needs.

Sample Answer

To profile very large tables efficiently, combine incremental processing, pre-aggregation, approximate functions, columnar storage benefits, partition pruning, and system-level statistics. Start by clarifying freshness vs cost: if near-real-time is required, favor incremental + approximate; if daily/weekly is fine, use full pre-aggregations.

Approach summary:- Incremental processing: compute deltas using watermark columns (ingest_time, updated_at) and store running aggregates.- Pre-aggregations/materialized views: maintain summarized tables for heavy GROUP BYs.- Approximate functions: use HyperLogLog, t-digest, or APPROX_COUNT_DISTINCT for cardinalities and quantiles.- Columnar storage: parquet/ORC on cloud data warehouses reduces IO for wide tables; pushdown predicates speed scans.- Partition pruning: partition by date or logical shards and ensure queries include partition keys.- System stats: use table statistics, histograms, and query plans to detect skew and choose indexes.

Examples:Partition pruning + incremental aggregate

sql

-- incremental aggregate by day
INSERT INTO profile_user_day (day, country, cnt)
SELECT dt AS day, country, COUNT(*) FROM events
WHERE dt = CURRENT_DATE - INTERVAL '1' DAY
GROUP BY dt, country;

Approx distinct with HLL (Postgres + extension):

sql

SELECT hll_cardinality(hll_add_agg(hll_hash_text(user_id))) FROM events WHERE dt >= '2025-11-01';

Columnar read reduction (BigQuery example):

sql

SELECT country, COUNT(*) FROM `project.dataset.events`
WHERE DATE(event_time) BETWEEN '2025-11-01' AND '2025-11-30'
GROUP BY country;
-- relies on partitioned table by date and column pruning

Guidelines for technique selection:- Small-medium (<100GB): full scans + indexes acceptable; materialize only heavy queries.- Large (100GB–5TB): partitioning + pre-aggregations; use approximate functions for expensive stats.- Very large (>5TB): rely on columnar storage, distributed compute engines, sample-based profiling, and HLL/t-digest. Keep fresh needs in mind: for hourly freshness use streaming increments; for daily, schedule batch pre-aggregations.

Best practices:- Always collect/refresh table statistics and histograms so optimizers pick good plans.- Monitor skew and heavy keys; consider bloom filters or secondary partitioning.- Validate approximate results against exact on sample data periodically.- Automate materialized view refresh and maintain lineage for reproducibility.

Performance considerations:- Incremental reduces work from O(N) to O(delta)- Approx functions trade exactness for O(1) memory and much faster runtime- Partitioning reduces scanned data proportional to partition selectivity

Edge cases:- Late-arriving data: design reprocessing windows or backfills- Highly skewed joins: broadcast smaller table or use salting- Low-cardinality columns: avoid unnecessary indexes

This combination yields fast, cost-effective profiling while balancing accuracy and freshness.

HardTechnical

19 practiced

You are hired as lead data scientist to create a company-wide data quality program. Outline a 12-month roadmap including policies, tooling, ownership model, SLAs, monitoring and alerts, training, and integration with existing data governance. Propose KPIs to measure program success and describe change-management considerations to onboard teams.

Sample Answer

Month 0–1 — Kickoff & assessment- Conduct inventory of data assets, pipelines, owners, existing governance, and major quality pain points (accuracy, completeness, timeliness, consistency).- Form cross-functional steering committee (Analytics, Data Engineering, Prod, Legal, BI, PM).

Month 2–3 — Policy & ownership model- Publish Data Quality (DQ) policy: definitions, severity levels, SLA targets, escalation paths, permitted remediation windows.- Define ownership: Data Owner (business accountable), Data Steward (operational custodian), Data Engineer (pipeline owner), DQ Lead (program owner).

Month 4–6 — Tooling & SLAs- Select tooling stack: data catalog + lineage (e.g., Collibra/Amundsen), DQ framework (Great Expectations/Deequ), monitoring (Prometheus/Datadog/Monte Carlo) and alerting integrated to ticketing (Jira/ServiceNow).- Define SLAs by dataset criticality (e.g., P0: <1% error, 1-hour detection & 4-hour workaround; P1: <3% error, 24-hour remediation).

Month 7–9 — Monitoring, alerts, remediation- Implement automated tests in CI for schemas, null/uniqueness checks, distribution drift, referential integrity, freshness.- Build dashboards and alerting rules; route incidents to stewards with runbooks and automated rollback or quarantine for critical flows.

Month 10–11 — Training & process integration- Run role-based training (hands-on Great Expectations, reading lineage, triaging alerts).- Embed DQ checks into model development and deployment pipelines; require DQ sign-off in PRs and data contracts.

Month 12 — Governance integration & continuous improvement- Formalize integration with data governance: metadata exchange, compliance reporting, audit trails.- Conduct program review, update policies, roadmap next 12 months.

KPIs to measure success- % of critical datasets meeting SLAs- Mean time to detect (MTTD) and mean time to remediate (MTTR)- Number of production incidents attributed to data quality per quarter- % of pipelines with automated tests- User trust score (survey) and reduction in ad-hoc data fixes

Change-management considerations- Start with high-impact pilot teams to show quick wins and metrics.- Communicate value (business impacts, saved engineering hours, model stability) using case studies.- Provide low-friction onboarding: templates, starter test suites, prebuilt runbooks.- Incentivize adoption: include DQ metrics in team OKRs, recognize stewards.- Maintain a feedback loop via steering committee and monthly DQ reviews to iteratively refine policy and tooling.

This plan balances governance, engineering automation, and cultural change to make data quality measurable, owned, and sustainable.

Unlock Full Question Bank

Get access to hundreds of Data Exploration and Quality Assessment interview questions and detailed answers.

Join thousands of developers preparing for their dream job.