Computer services are normally assumed to work well all the time. In this work we examined the operation and the errors of metadata harvesting services and tried to find clues that will help predicting the consistency of the behavior and the quality of the harvesting. The large number of such services, the huge amount of harvested information and the possibility of meeting transient conditions makes this work hard. We studied 395530 harvesting tasks from 2138 harvesting services in 185 harvesting rounds during a period of 9 months, of which 214163 ended with error messages and the remaining tasks occasionally returning fewer records. A significant part of the OAI services never worked or have ceased working while many other serves occasionally fail to respond. It is not trivial to decide when a tasks is successful, as tasks that return without an error message do sometimes return records and also tasks that declare that complete normally sometimes return less or no records. This issue is fundamental for further analysis of the harvesting outcome and any assessment that may follow. Therefore, on this work we studied the error messages and the task outcome patterns in which they appear and also the tasks that returned no records, to decide on which is the most essential condition to decide when a task is successful. Our conclusion is that a task should be considered successful when it returns some records.