This paper describes an experimental methodology used to evaluate the effectiveness of partial fault tolerance (PFT) techniques in data stream processing applications. Without a clear understanding of the impact of faults on the quality of the application output, applying PFT techniques in practice is not viable. We assess the impact of PFT by injecting faults into a synthetic financial engineering application running on top of IBM's stream processing middleware, System S. The application output quality degradation is evaluated via an application-specific output score function. In addition, we propose four metrics that are aimed at assessing the impact of faults in different stream operators of the application flow graph with respect to predictability and availability. These metrics help the developer to decide where in the application he should place redundant resources. We show that PFT is indeed viable, which opens the way for considerably reducing the resource consumption when compared to fully consistent replicas.