A system and method is disclosed for recording checkpoints in a distributed system. A distributed system comprises one or more computers implementing a plurality of nodes coordinating with one another to maintain a shared state of the distributed system. The system chooses a given one of the plurality of nodes to record a checkpoint of the shared state. In response, the given node records the checkpoint by isolating itself from communication with the other nodes, storing the checkpoint, restarting, and attempting to reinitialize its state from the stored checkpoint. Restarting may include deliberately causing a runtime error in the node. If the reinitialization is successful, the node restores communication with the other nodes and indicates to them that the newly stored checkpoint is valid.