Hi Justin,
Speaking in terms of Consul, not sure where you found that
information, but it's not (generally) correct. That's referred to as a
split-brain problem, and consensus algorithms (like Raft, the one in
Consul) are engineered to avoid that situation. That's why you should
always have an odd number of servers in your Consul cluster; that way,
a network partition can be handled appropriately -- if there is still
a majority, business can continue as usual, but if you no longer have
a majority of the known cluster members, actions that require
consensus refuse to be performed.
This is true even if you start with an odd number of peers and lose a
majority in all situations due to multiple network partitions (e.g. A,
B, and C are all isolated nodes) -- they will know that they each had
two other peers, and no node will declare itself the leader. If you
have an even number, a network partition can leave you with 2/2, in
which case no node can declare itself the leader -- so you can have
four nodes still online but no leader, whereas a 2/1 split with three
nodes will still allow a (single!) leader to be elected.
Network blips and outages can occur within a datacenter as well, so
there is nothing inherently different about going across datacenters.
You may get more frequent network partitions, so you may get more
frequent leader elections, but again -- network outages can happen
even within a datacenter, so Consul's consensus/leader election
handling is designed to handle this case without having a split brain.
--Jeff
>
https://groups.google.com/d/msgid/vault-tool/9dcfc193-746e-46d3-ae9b-6ceea1dd6339%40googlegroups.com.