It is implemented using the ACL permission information of the zookeeper node in the resourcemanager of the yarn framework.
Suppose there are two resourcemanagers in the yarn cluster: different resourcemanagers create lock nodes through competition to achieve the determination of the ready state. Suppose there are currently two resourcemanagers: m1, m2. The current m1 is active, if it is suspended at this time, zookeeper will delete the lock node of m1. m2 becomes active and creates its own lock node. You must carry your own ACL information when creating a lock node. When m1 recovers, it is found that the current lock node is not the lock node created by itself, and it will be automatically converted to the standby state.