Une approche pragmatique pour mesurer la qualité des applications à base de composants logiciels

摘要

Over the past decade, many companies proceeded with the introduction of component-oriented software technology in their development environments. The component paradigm that promotes the assembly of autonomous and reusable software bricks is indeed an interesting proposal to reduce development costs and maintenance while improving application quality. In this paradigm, as in all others, architects and developers need to evaluate as soon as possible the quality of what they produce, especially along the process of designing and coding. The code metrics are indispensable tools to do this. They provide, to a certain extent, the prediction of the quality of « external » component or architecture being encoded. Several proposals for metrics have been made in the literature especially for the component world. Unfortunately, none of the proposed metrics have been a serious study regarding their completeness, cohesion and especially for their ability to predict the external quality of developed artifacts. Even worse, the lack of support for these metrics with the code analysis tools in the market makes it impossible to be used in the industry. In this state, the prediction in a quantitative way and « a priori » the quality of their developments is impossible. The risk is therefore high for obtaining higher costs as a consequence of the late discovery of defects. In the context of this thesis, I propose a pragmatic solution to the problem. Based on the premise that much of the industrial frameworks are based on object-oriented technology, I have studied the possibility of using some « conventional » code metrics unpopular to component world, to evaluate component-based applications. Indeed, these metrics have the advantage of being well defined, known, equipped and especially to have been the subject of numerous empirical validations analyzing the predictive power for imperatives or objects codes. Among the existing metrics, I identified a subset of them which, by interpreting and applying to specific levels of granularity, can potentially provide guidance on the compliance of developers and architects of large principles of software engineering, particularly on the coupling and cohesion. These two principles are in fact the very source of the component paradigm. This subset has the ability to represent all aspects of a component-oriented application : internal view of a component, its interface and compositional view through architecture. This suite of metrics, identified by hand, was then applied to 10 open-source OSGi applications, in order to ensure, by studying of their distribution, that it effectively conveyed relevant information to the component world. I then built predictive models of external quality properties based on these internal metrics : reusability, failure, etc. The development of such models and the analysis of their power are only able to empirically validate the interest of the proposed metrics. It is also possible to compare the « power » of these models with other models from the literature specific to imperative and/or object world. I decided to build models that predict the existence and frequency of defects and bugs. To do this, I relied on external data from the history of changes and fixes a panel of 6 large mature OSGi projects (with a maintenance period of several years). Several statistical tools were used to build models, including principal component analysis and multivariate logistic regression. This study showed that it is possible to predict with these models 80% to 92% of frequently buggy components with reminders ranging from 89% to 98%, according to the evaluated projects. Models for predicting the existence of a defect are less reliable than the first type of model. This thesis confirms thus the interesting « practice » of using common and well equipped metrics to measure at the earliest application quality in the component world.