Abstract: A distributed data processing system includes a distributed resource manager which detects dependencies between transactions caused by conflicting lock requests. A distributed transaction manager stores a wait-for graph with nodes representing transactions and edges connecting the nodes and representing dependencies between the transactions. Each edge is labelled with the identities of the lock requests that caused the dependency. The distributed transaction manager propagates probes through the wait-for graph, to detect cyclic dependencies, indicating deadlock. A deadlock message is then sent to the resource manager identifying a particular lock request as a victim for deletion to resolve the deadlock. Resilience to failure is achieved by duplicating between agents and servers, rather than by duplicating the servers. As a result, the number of messages between agents and servers in normal operation is not increased.