Multi Version Concurrency Control

Multi Version Concurrency Control or MVCC is an efficient method to let multiple processes access the same data in parallel without corrupting the data and the possibility of deadlocks. It is an alternative to the Lock based approaches, where every process first has to request an exclusive lock on a data item, before it can be read or updated. MVCC is used in some relational databases as well as in most distributed databases.

Instead of letting each process access the data exclusively for a certain amount of time, MVCC allows processes to read the data in parallel, even if a process is updating the data. To maintain consistency, each data item has some kind of time stamp or revision. If a process reads a data item, it does not only get the value of it, the process also retrieves the revision of the data item. So if this process attempts to update this data item, then it writes the new value with the previously read revision number to the database. If the actual revision in the store is the same, then the new value is written and the revision of the data item is incremented.

But if the revision in the store is not the same as the revision read by the writing process, then there must have been another process which has updated the data item in the meantime. In a relational database, the writing process would be a transaction, which would in this case be aborted or restarted.

In distributed databases, there are at least two cases for such a conflict: The first one is that two processes are attempting to write the same data item on the same node. In this case, the database could detect the conflict during the write operation and abort it, so the client would need to re-read the data item and retry its desired update, like the behavior of a RDBMS in this situation.

Another case is that multiple clients update the same data item on different nodes. If the distributed database uses asynchronous replication, then this conflict can not be detected and handled during the write operations. The nodes first have to be synchronized before they can handle the conflict. The conflict resolution can happen during the replication or during the first read operation on the conflicting data item.

Some databases, which implement this, store all conflicting revisions of the data items and let the client decide how the conflict should be handled. In such systems a read request returns all conflicting versions of the value and the client has to choose one or has to merge the versions and to write the corrected revision back to the database.