Design a data quality framework from source to consumption

What's really being asked

The interviewer wants to see if you treat data quality as a socio-technical system rather than a checklist of tools. They are looking for architectural thinking that connects business governance to pipeline implementation, with clear ownership, preventive controls, observable metrics, and closed-loop remediation. Senior candidates should demonstrate that they understand quality must be designed into every stage from ingestion to consumption, not inspected in at the end.

The full answer

First, establish the governance foundation before any code is written. This means identifying critical data elements, assigning data stewards who own business definitions, and publishing a RACI matrix. Second, implement preventive controls at ingestion using schema registries and data contracts to reject malformed records before they enter the pipeline. Third, embed continuous validation inside CI/CD through automated profiling that tracks null rates, cardinality, and distributions, plus anomaly detection that flags statistical deviations or schema drift in near real time. Fourth, build column-level lineage so you can trace a quality issue back to the specific source system and transformation logic. Fifth, create business-aligned scorecards that roll technical metrics into KPIs the executive team cares about, such as revenue impact or compliance posture. Finally, define remediation workflows including automated quarantine, alerting thresholds, and root cause analysis.

The mistakes people make

A red flag is jumping immediately to vendor names like Great Expectations, Monte Carlo, or dbt tests without first describing ownership and policy. Another mistake is treating quality as a nightly batch report rather than a continuous, pipeline-integrated function. Candidates often describe lineage as a coarse DAG of table dependencies while missing column-level granularity, which is what data consumers actually need for impact analysis. Proposing metrics that are purely technical vanity scores with no tie to business outcomes is also a weak signal. Finally, forgetting error handling entirely, such as what happens when a check fails, shows immature design.

What usually comes next

The interviewer may ask how you would handle schema evolution without breaking downstream consumers, or how you prioritize which critical data elements get quality rules first when faced with hundreds of sources. They might probe the cost of implementing full lineage versus the value, or ask how you would design anomaly detection for low-volume event streams where statistical significance is hard to achieve. Another common follow-up is how you prove ROI on this framework to a CFO.

A concrete example

Imagine a retail company ingesting order events from a POS system into a real-time analytics platform. A data steward from finance owns the order total and tax fields as critical data elements. A schema registry at ingestion enforces that order totals are numeric and non-null. CI/CD profiling runs every fifteen minutes, flagging if the average order value drops three standard deviations below the rolling seven-day mean. Column-level lineage maps the order total from the POS API through a Flink job to the Snowflake fact table and finally to the executive revenue dashboard. When the anomaly fires, the scorecard shows a projected two percent revenue impact, triggering automatic quarantine of events from that store ID and a Slack alert to the steward and on-call engineer. The steward confirms whether it is a true bug or a legitimate promotion, and the incident is logged for root cause review.