Hadoop stacks when competing with mapreduce and spark

Question

I manage a cluster with several machines that are shared by other colleagues. Some use sparks, and some use Map Reduce.

Typically, Spark users open a context and open it for several days or weeks, while in MR tasks begin and end.

The problem is that the MR task is stacked because:

• After the X% phase of the card, it starts working with gearboxes.
• In the end, you have a lot of gear starts and only 5-15 cards are expected.
• At the moment, there is not enough memory to launch a new card, and gearboxes cannot move by 33%, because the cards have not yet completed work, creating a dead end.

The only way to solve this problem is to kill one of the spark contexts and let the cards finish.

Is there any way to adjust the yarn to avoid this problem?

Thanks.