Data engineering·Orchestration and running it in anger
you cap the number of tasks that may touch the production replica at four, so a backfill cannot take the source down.
Concurrency limit
Also calledpool, parallelism cap, task slots
A cap on how many tasks may run at once, globally or against a named resource. It is the orchestrator's version of a bulkhead, and it is the mechanism that stops one greedy pipeline from starving everything else or overwhelming a shared dependency. Setting them per resource rather than globally is what makes them useful, because the thing you are protecting is the source database, not the scheduler.