The present application discloses an instance-based distributed data recovery method. A specific implementation of the method includes: detecting a non-master down node; allocating multiple secondary storage units corresponding to the down node to at least one online node; performing hash grouping on instances stored on logs and allocating the instances to multiple threads; and recovering data of multiple primary storage units in parallel inside the online node. Embodiments of the present invention recover data of a down node in a distributed database in parallel in nodes.