A pool is a named number of slots. Tasks that name it can only run while a slot is free; the rest wait in the scheduled state. Use one per shared system that cannot take unlimited load: a warehouse with two spare connections, a partner API.
airflow pools set warehouse 2 "Marigold warehouse, two connections"
@task(pool="warehouse")
def load(...): ...
Tasks without a pool use default_pool (128 slots). A heavy task can take several slots with pool_slots=2. A task that names a pool that does not exist never starts, so create pools before deploying the Dags that use them.
When slots are scarce, priority weights decide who goes first. Each task has priority_weight (default 1), and weight_rule decides how it counts: downstream (the default) adds the weights of everything downstream, so tasks early in long chains go first; absolute uses the task's own weight as written. Dag-level limits work alongside: max_active_tasks per Dag and max_active_runs.
Finance gets the first free warehouse slot.