Performance for dbts
dbt performance is about making builds faster and cheaper. Incremental models process only new or changed rows using the is_incremental macro and a unique_key, while state-based selection with the state method and defer runs only modified models and their dependents, which keeps CI and production builds short.
Stage: Scale · Part of the dbt roadmap · Last reviewed 19 November 2026
What to learn in Performance
- Write an incremental model using is_incremental and a unique_key
- Choose an incremental strategy such as merge or insert_overwrite
- Select only modified models and children with state:modified+
- Use defer to read unchanged upstream models from production
- Force a full refresh when logic changes
Performance interview questions
How does an incremental model work in dbt?
On the first run dbt builds the full table. On later runs the is_incremental block filters to new rows, such as updated_at greater than the table's max, and dbt merges or appends them. A full refresh rebuilds from scratch.
Why can an incremental model silently drift from the source?
Late-arriving data, updates to old rows or logic changes are not picked up if the filter only reads new rows. Use a lookback window, a reliable unique_key, periodic full refreshes and tests comparing totals to the source.
Where Performance fits in the dbt roadmap
Performance belongs to the Scale stage (intermediate level). Efficient builds.