When multiple steps in the DAG execute concurrently it can overload very limited dockerized solutions as it becomes difficult to predict how much compute will be needed at different stages of the execution.
By limiting concurrency this can be better controlled.