Currently, if the task is deemed unhealthy, i.e. it failed a health check a certain number of times, it is killed by both default executors: command and docker. This is what can be called "local" kill policy.
While local kill policy can save some network traffic and unload the scheduler, there are cases, when a scheduler may want to decide what—and when—to do. This is what can be called "global" policy, i.e. the health check library reports whether a health check failed or succeeded, while the executor forwards this update to the scheduler without taking any action.