It seems necessary to detect a broken bike rooted at a station in near realtime as the number of bikes within bikeshare systems has reached more than a million in 2015. Indeed, a bike that cannot be moved is not cost effective in terms of number of trips. This brings frustration to users who were expecting to find a bike at that station without knowing that it is actually defective. We thus propose a methodology from feature extraction to anomaly detection on a distributed cloud infrastructure in order to detect bicycles requiring a repair. Through a first step of K-means clustering, and a second step consisting of spotting samples that do not clearly belong to any cluster, we separate anomalies from normal behaviors. The proposal is validated on a publicly available dataset provided by Motivate, the operator of the New-York bikeshare system. The number of distinct bikes that have been classified by this algorithm as broken at least once during a month is close to the number of repairs given in monthly reports of Motivate.