Synchronous replication | Tarantool
Tarantool
Check out the new release policy
Concepts Replication Synchronous replication

Synchronous replication

By default, replication in Tarantool is asynchronous: if a transaction is committed locally on a master node, it does not mean it is replicated onto any replicas. If a master responds success to a client and then dies, after failover to a replica, from the client’s point of view the transaction will disappear.

Synchronous replication exists to solve this problem. Synchronous transactions are not considered committed and are not responded to a client until they are replicated onto some number of replicas.

To learn how to enable and use synchronous replication, check the guide.

A killer feature of Tarantool’s synchronous replication is its being per-space. So, if you need it only rarely for some critical data changes, you won’t pay for it in performance terms.

When there is more than one synchronous transaction, they all wait for being replicated. Moreover, if an asynchronous transaction appears, it will also be blocked by the existing synchronous transactions. This behavior is very similar to a regular queue of asynchronous transactions because all the transactions are committed in the same order as they make the box.commit() call. So, here comes the commit rule: transactions are committed in the same order as they make the box.commit() call—regardless of being synchronous or asynchronous.

If one of the waiting synchronous transactions times out and is rolled back, it will first roll back all the newer pending transactions. Again, just like how asynchronous transactions are rolled back when WAL write fails. So, here comes the rollback rule: transactions are always rolled back in the order reversed from the one they make the box.commit() call—regardless of being synchronous or asynchronous.

One more important thing is that if an asynchronous transaction is blocked by a synchronous transaction, it does not become synchronous as well. This just means it will wait for the synchronous transaction to be committed. But once it is done, the asynchronous transaction will be committed immediately—it won’t wait for being replicated itself.

Until version 2.5.2, there was no way to enable synchronous replication for existing spaces, but since 2.5.2 it can be enabled by space_object:alter({is_sync = true}).

Synchronous transactions work only for master-slave topology. You can have multiple replicas, anonymous replicas, but only one node can make synchronous transactions.

Since Tarantool 2.10.0, anonymous replicas do not participate in the quorum.

Starting from version 2.6.1, Tarantool has the built-in functionality managing automated leader election in a replica set. For more information, refer to the corresponding chapter.

Found what you were looking for?
Feedback