Does the past say it all? Using history to predict change sets in a CMDB

To avoid unnecessary maintenance costs in large IT systems resulting from poorly planned changes, it is essential to manage and control changes to the system and to verify that all items impacted by each change are updated as needed. This paper presents a method of decision support that helps guarantee that each change set (those items to be updated in the change) contains all the software or hardware components impacted by the proposed change. Today, many IT systems are managed by a Configuration Management Database (CMDB), which can be represented as a large graph in which the nodes are configuration items (CIs), such as software applications or servers, and the edges record dependencies between these items. In this paper we present a new approach to suggesting change sets based on our conjecture that each new change set is likely to be similar to instances of previous change sets. Accordingly, if the analyst determines that CI x is in a new change set, our method essentially searches for previous change sets, stored in the CMDB, that contain x, and suggests that CIs in those sets (appropriately ranked) should be considered for inclusion in the new change set. Our model uses support and confidence measures to estimate how closely nodes x and y are related, based on how often they have appeared together in past change sets. Based on these measures, we implement a prototype that suggests likely items to an analyst who is composing a change set. Based on a history of three years of a particular industrial CMDB, and several filtering techniques, the observed recall and precision values were as high as 69.8% and 88.5% respectively.