Each mapped copy is its own task instance: it retries on its own, logs on its own and can be cleared on its own. If one copy still fails after its retries, a task that collects the results is upstream_failed under the default rule.
Mapping can start many copies at once. To protect a partner system, limit how many run together:
max_active_tis_per_dag=2 on the task: at most two copies of it running at a time, across all runs of the Dag.
max_active_tis_per_dagrun=2: at most two per run.
- A pool with two slots, shared with other tasks that call the same system.
A mapped output can feed another mapped task: message.expand(result=track.expand(parcel=open_parcels())) makes one message per tracking result. Keep what you pass small: each result is an XCom.
At most two tracking calls at once, each retried on its own.