Quality, tests and knowing it is wrong

nothing errored, every job is green, and the number on the board is wrong by eleven percent.

Data quality

you split 'the data is bad' into six specific claims, and it turns out only one of them is actually true.

Data quality dimensions

the transform ships with a check that this column is never null, and the check runs against real output every night rather than against a fixture.

Data test

you write down that this column should be between zero and one hundred and that ninety-nine percent of rows should have an email, and it is checked automatically.

Data expectation

the failed check stops the pipeline so the bad numbers never reach the dashboard, and someone gets paged instead.

Blocking data test

the load succeeded and inserted eleven rows, where every previous day inserted about four hundred thousand.

Volume anomaly

the column has been two percent null for a year, this morning it is sixty percent, and nothing else changed.

Null-rate check

the key you have been treating as unique turns out to have four hundred duplicates, which is why the join inflated the total.

Uniqueness test

the same event is in the table three times because the connector retried, and you keep the latest by version per key.

Deduplication

eight thousand fact rows point at a customer id that does not exist in the customer dimension.

Referential integrity check

the average order value did not change but the shape did, because a whole segment of small orders stopped arriving.

Distribution check

you count the rows in the source and in the warehouse for the same period and prove that the numbers agree.

Row-count reconciliation

before merging, you run both versions of the model and see that exactly nine hundred rows differ, all in one region.

Data diff

every task is green, the run took its usual eleven minutes, and it processed zero rows.

Silent failure

the numbers were wrong from Tuesday 02:00 until someone noticed on Thursday afternoon, and that gap is the thing worth measuring.

Data downtime

last quarter's published figures change after a backfill, and somebody has to tell the people who already presented them.

Restatement

you get told that a table's volume, freshness and schema look unusual without anyone having written a test for that table.

Data observability

the table is wrong and the question 'who owns this' takes two days and three Slack channels to answer.

Data ownership

some tables are marked as certified with tests and an owner, and the rest are clearly labelled as somebody's experiment.

Trust threshold