We present a protocol for the distributed detection of garbage in a distributed system subject to common failures such as lost and duplicated messages, network partition, dismounted disks, and process, site and disk crashes. The protocol uses only information local to each site, or exchanged between pairs of sites; no global mechanism is necessary. Since overhead is low, the protocol can be integrated into the operating system. We show than a logical separation of responsibility allows to integrate in the operating system the detection role leaving up to the application the collection role