During the execution of a Composite Web Service (CWS), different faults may occur and cause a Web Service (WS) to fail. To repair failures some strategies can be applied, such as WS retry or substitution, compensation of the performed execution, roll-back, replication, or take checkpoints to later restart the execution. Each strategy has advantages and disadvantages on different execution scenarios and can produce different impact on the CWS QoS, depending on the execution environment and execution state at the moment of the failure. In this paper we propose a model and show experimental results to dynamically decide which recovery strategy is the best choice in terms of the impact on the CWS QoS. © 2013 ACM.