Despite significant efforts to achieve reliable grid middlewares, grid infrastructures still encounter important difficulties to implement the promise of ubiquitous, seamless and transparent computing. Identified causes are numerous, such as the complexity of middleware stacks, dependence to many distributed resources, heterogeneity of hardware and software operated or incompatibilities between software components declared as interoperable. Based on failures that occurred during a large data challenge run on a grid dedicated to neuroscience, we identify scenarios that can be handled through autonomic management associated to the grid middleware. We also outline a flexible self-adaptive framework that aims at using model-driven development to facilitate the engineering, integration and reuse of MAPE-K loops in large scale distributed systems.