Leader election process
Automated leader election in Tarantool helps guarantee that
there is at most one leader at any given moment of time in a replica set.
A leader is a writable node, and all other nodes are non-writable –
they accept read-only requests exclusively.
When the election is enabled, the life cycle of
a replica set is divided into so-called
terms. Each term is described by a monotonically growing number.
After the first boot, each node has its term equal to 1. When a node sees that
it is not a leader and there is no leader available for some time in the replica
set, it increases the term and starts a new leader election round.
Leader election happens via votes. The node, which started the election, votes
for itself and sends vote requests to other nodes.
Upon receiving vote requests, a node votes for the first of them, and then cannot
do anything in the same term but wait for a leader being elected.
The node that collected a quorum of votes defined by the replication_synchro_quorum parameter
becomes the leader
and notifies other nodes about that. Also, a split vote can happen
when no nodes received a quorum of votes. In this case,
after a random timeout,
each node increases its term and starts a new election round if no new vote
request with a greater term arrives during this time period.
Eventually, a leader is elected.
If any unfinalized synchronous transactions are left from the previous leader,
the new leader finalises them automatically.
All the non-leader nodes are called followers. The nodes that start a new
election round are called candidates. The elected leader sends heartbeats to
the non-leader nodes to let them know it is alive.
In case there are no heartbeats for the period of replication_timeout * 4,
a non-leader node starts a new election if the following conditions are met:
- The node has a quorum of connections to other cluster members.
- None of these cluster members can see the leader node.
A cluster member considers the leader node to be alive if the member received heartbeats from the leader at least once during the period of
replication_timeout * 4,
and there are no replication errors (the connection is not broken due to timeout or due to an error).
Terms and votes are persisted by each instance to preserve certain Raft guarantees.
During the election, the nodes prefer to vote for those ones that have the
newest data. So as if an old leader managed to send something before its death
to a quorum of replicas, that data wouldn’t be lost.
When election is enabled, there must be connections
between each node pair so as it would be the full mesh topology. This is needed
because election messages for voting and other internal things need direct
connection between the nodes.
In the classic Raft algorithm, a leader doesn’t track its connectivity to the rest of the cluster.
Once the leader is elected, it considers itself in the leader position until receiving a new term from another cluster node.
This can lead to the split situation if the other nodes elect a new leader upon losing the connectivity to the previous one.
The issue is resolved in Tarantool version 2.10.0 by introducing the leader fencing mode.
The mode can be switched on and off by the election_fencing_enabled configuration parameter.
When the fencing is on, the leader resigns its leadership if it has less than the replication_synchro_quorum
of alive connections to the cluster nodes. The resigning leader receives the status of a follower in the current election term and becomes read-only.
Fencing applies to the instances that have the election_mode set to “candidate” or “manual”.
There can still be a situation when a replica set has two leaders working independently (so called split-brain).
It can happen, for example, if a user mistakenly lowered the replication_synchro_quorum below
N / 2 + 1.
In this situation, to preserve the data integrity, if an instance detects the split-brain anomaly in the incoming replication data,
it breaks the connection with the instance sending the data and writes the
ER_SPLIT_BRAIN error in the log.
Eventually, there will be two sets of nodes with the diverged data,
and any node from one set is disconnected from any node from the other set with the
Once noticing the error, a user can choose any representative from each of the sets and inspect the data on them.
To correlate the data, the user should remove it from the nodes of one set,
and reconnect them to the nodes from the other set that have the correct data.
Also, if election is enabled on the node, it won’t replicate from any nodes except
the newest leader. This is done to avoid the issue when a new leader is elected,
but the old leader has somehow survived and tries to send more changes
to the other nodes.
Term numbers also work as a kind of a filter.
For example, you can be sure that if election
is enabled on two nodes and
node1 has the term number less than
node2 won’t accept any transactions from