Design Rules for Producing Controllable Computer Services

Automatic management of computer services such as databases, e-commerce systems and other Web services, is a highly desirable but challenging problem. We argue that the main reason for this is that the services themselves are usually not designed to be managed by an automatic management algorithm, such as one based on machine learning or control theory. Many times, the services themselves expose hundreds of tunable parameters and measurements with obscure and highly unpredictable relationships between them. This have given rise to many complicated, error prone and specialized algorithms for each single management goal and system, mainly because these services were not designed with control in mind. This paper presents eight design rules for the implementation of computer services that are easy to control. These rules have been derived from practical experience with several computer services and from theoretical results. If these rules are followed in the design process of the service, it is easy to apply standard solutions from fields such as control theory and machine learning to control the service and reach the desired management goals. We show how to apply these rules by using examples from existing computer services