Loading...
Loading...
Browse 13 real-world technical and behavioral interview questions about Data quality. Review scenarios, edge cases, and architectural best practices.
A steady rejection rate that is cleared daily is not a healthy control - it is a known defect being absorbed by people, and it says nothing about the reports that were accepted while being wrong. Rejections you can see are the smaller half of the problem.
On any protocol that reports by exception, no new value and an unchanged value look identical, so liveness has to be carried deliberately. Status codes, source timestamps, birth and death messages and a source-side counter each establish it at a different layer of the path.
Validate every record against a declared schema at the ingestion boundary, quarantine what fails instead of dropping it, treat additive changes as safe and renames as breaking, and alert on a field that has gone all-null rather than waiting for a dashboard to look wrong.
Agree the business meaning and the grain of each entity before touching field names, profile the real data to find the overloaded and sentinel values, then write a mapping where every target field has a source, a rule, a default and a decision for records that cannot be mapped.
Retrain when there is evidence the learned relationship no longer holds and you have data that reflects the new one. A fixed schedule is a fallback for when you cannot measure that, not a decision, and a drift alert is as likely to mean an upstream pipeline broke as it is to mean the world moved.
A golden customer record in banking is a derived identity view across product systems. Build it with scored entity resolution, source cross-references, attribute survivorship and reversible merge decisions.
Patient matching across healthcare systems uses deterministic rules, probabilistic linkage and human review to connect records without a shared identifier. Blocking, thresholding and reversible links matter because false merges are clinically riskier than false splits. It also connects record linkage to the point an interviewer is testing.
Separate the product a customer shops for from the variant that carries a barcode and stock, decide which attributes are variant-defining rather than merely descriptive, and hold the rest in a typed per-category schema, because a fixed column set cannot grow and an untyped bag cannot be trusted.
Every derived figure that joins on time is affected, and none of them error. Duration and attribution metrics take the skew directly, so a shift with twelve stops mis-measures availability by nearly four points, and only storing both event time and ingestion time makes the offset visible.
Treat ownership as a derived, continuously verified fact rather than a manually entered field: source it from an existing system of record, expire it on a schedule, and fail builds or pages loudly when it goes unverified.
Segment the nulls before drawing any conclusion — by date, source system, record type and creating process — because a null rate that clusters is a documented business path nobody mentioned, while one spread evenly is usually a genuine capture defect.
You cannot know without matching, and matching is the work: suppliers send their own identifiers, spell the same manufacturer three ways, and describe attributes in units you do not use. Treat ingestion as a matching problem with a confidence threshold and a review queue, not as an import.
Reconcile the two numbers row by row before touching the pipelines: the cause is almost always a grain mismatch that fans out a join, a filter difference over refunds or test accounts, or two timestamps meaning different things. The fix is one certified definition that both dashboards are forced to call.