Sampling a model run before a full build

The --sample flag puts a dbt invocation into 'sample mode,' filtering input data to a limited, relative time window so developers can build or test models faster during local development and CI without processing full production volumes. It uses each model's configured event_time (the same field used for microbatch incremental models) to determine which rows to include.

1 · Learn the must-know

• 1Sample mode filters rows based on the model's (and its sources'/seeds') configured event_time column; models without event_time set are not sampled and run against their full dataset.
• 2You invoke it as part of a normal invocation, e.g. dbt run --sample=<relative_time_range> or dbt build --sample=<relative_time_range>, restricting data to that recent window relative to the current run time.
• 3Sample mode is meant for development and CI workflows to speed up iteration, not for production runs, since it intentionally works on incomplete data.
• 4Because upstream data is truncated, downstream tests (row counts, uniqueness, referential integrity, etc.) can behave differently or fail compared to running against full data.
• 5Sample mode combines with normal node selection (--select/--exclude), so you can scope both which models run and how much data each processes for fast local cycles.
• 6It only affects what data is read/processed during that invocation—it does not modify or truncate your actual source tables or persist any sampling logic into the model's compiled SQL.

2 · Check your understanding