In the automatic scaling scenario, when a scale-in rule is triggered, if an ApplicationMaster (AM) is running on the task node to be terminated, the running AM will also be terminated, which will cause the current job to fail.
No AMs will be assigned to task nodes automatically added through the transformation of YARN source code and the addition of configuration items to the automatic scaling process. Instead, all AMs will be assigned to those that are not automatically added. Automatically added task nodes are responsible for compute tasks only. Therefore, an automatic scale-in action will only terminate the nodes processing compute tasks, while the AMs remain active and will try to take over the compute tasks on other nodes to keep the current job running.