Resources
Explainer · 4 articles
Duplicates from fan-out joins: the bug that inflates every number
A join that used to be one-to-one quietly becomes one-to-many, and every sum downstream doubles while each row still looks correct. How to detect fan-out, and how to fix the grain rather than the symptom.
Freshness: 'the pipeline ran' is not the same as 'the data arrived'
A job can exit 0 and still load nothing. The ways that happens, why freshness has to be measured on the table rather than the orchestrator, and how to write a check that respects a table's own cadence.
Why fixed thresholds fail, and what a learned baseline does instead
A rule like "alert if row count < 1000" is wrong on weekends, wrong after growth, and wrong after a backfill. Here is what a baseline learned from a table's own history captures, and where it breaks too.
The six ways data breaks silently
Broken data rarely throws an error. Dashboards keep rendering and the numbers are just wrong. Here are the six failure classes, what each looks like in the warehouse, and the query that catches it.
Guide · 4 articles
Build vs buy: what a homegrown data-quality check actually costs
The first data-quality check takes an afternoon. The tenth one is a product you now maintain. An honest accounting of when to write the test yourself, and where the hidden costs land in month three.
Data quality monitoring for a team of one
You build the pipelines, answer the questions, and find out data broke when the CFO does. A realistic operating model for one person: what to automate, what to keep manual, and what to simply not do.
How to choose which tables to monitor, and which to leave alone
Monitoring every table is how you get noise and a bill. Monitoring none is how you get surprised. A method for picking the handful that matter, assigning priority tiers, and reviewing the list quarterly.
Alert fatigue in data monitoring: a noise budget for small teams
A monitoring tool that pages you forty times a day gets muted within a fortnight — and a muted tool is worse than none, because you still believe you are covered. How to set a noise budget and hold it.
Playbook · 2 articles
When a dashboard number looks wrong: a debugging checklist
Someone just asked whether revenue is really that low and you have twenty minutes. An ordered checklist from the cheapest check to the most expensive, with the query to run and what to tell them at each step.
A fifteen-minute triage playbook for the morning digest
A repeatable routine for one person clearing a morning's data-quality findings before standup: what to read first, how to verify a claim, the three outcomes every finding gets, and how to tell stakeholders.