Data quality · Extra High-Value

Root-cause analysis for data incidents

HardRCAincidentlineagebackfillpostmortem

Interview question

How do you perform root-cause analysis (RCA) when a data quality incident happens?

How to answer

The short answer

Root-cause analysis finds the underlying reason a quality incident happened, not only the symptom ("dashboard revenue is wrong"). Good RCA prevents repeat failures.

A practical DE workflow

  1. 1Detect: alert on freshness, volume, test failure, or stakeholder report.
  2. 2Contain: pause publish / mark asset unhealthy / communicate stale-but-safe status.
  3. 3Scope blast radius: use lineage: which marts, dashboards, and jobs are affected?
  4. 4Localize the first bad hop: source empty? ? bad join? late ? wrong filter in yesterday's deploy?
  5. 5Confirm with evidence: reconciliations, diffs, git/deploy timeline, raw vs staging vs mart samples.
  6. 6Fix + : correct logic or upstream, reprocess impacted partitions idempotently.
  7. 7Prevent: add a test/monitor/contract that would have caught this earlier; write a short postmortem.
Output
Symptom: revenue_dashboard -30%
Lineage: dashboard <- fct_orders <- stg_orders <- raw.orders
Finding: raw.orders partition missing for Asia region after producer deploy
Fix: restore producer feed, backfill partition, add source freshness + row-count anomaly gate

Interview tip: Emphasize contain → lineage → first bad hop → evidence → backfill → preventive test. RCA is a process, not a guess.