Traditional cluster system management tools seldom consider the relevance between managed objects. Such relevance is the reason of related fault and may also lead to human operation errors. Because of this defect, traditional tools do not have the capability of handling of consistency, atomicity and recovery. This article proposes a transaction-based facility, destructive transaction, to solve the problems at some degree. Destructive transaction is a construct to wire down management rules stored in a system administrator's mind. It provides a method to describe managed objects relationships, atomicity facility and recovery for failed operation. And it significantly reduces human caused error possibility.