The short answer
Test severity controls whether a failing test fails the dbt invocation or only warns.
error (default)
If the test finds failing rows, dbt marks the test as failed and the CLI exits non-zero (depending on flags / build behavior). Use for hard contracts: , critical FKs, amounts that must never be null.
warn
Failures show up as warnings but do not necessarily fail the whole job. Useful for:
- Newly added tests while cleaning historical bad data
- Soft (e.g. "prefer unique emails, but legacy dupes exist")
- Monitoring drift before enforcing a hard gate
columns:
- name: email
tests:
- unique:
config:
severity: warnMental model
severity: error → "block the pipeline / fail CI" severity: warn → "raise visibility, keep going (for now)"
Related knobs
You may also see store_failures, where filters on tests, and warn_if / error_if style thresholds on some tests/packages to tolerate small failure counts.
Interview tip: "Start new quality rules as warn while remediating data, then flip to error once the warehouse is clean so CI protects the contract."
How severity behaves in dbt build
An error test that fails stops the models after it from running. A warn test that fails is reported but does not skip anything, so a rule you set to warn never blocks the pipeline by itself.
You can also set thresholds, so one bad row does not break a large table:
columns:
- name: order_id
tests:
- not_null:
config:
severity: error
warn_if: ">0"
error_if: ">100"Here any failing row is a warning, and more than 100 failing rows is an error. This is a good way to introduce a rule on messy history: start with a wide error_if, watch the count fall, then tighten it to zero.