Diffstat (limited to 'doc')
1 files changed, 44 insertions, 0 deletions
diff --git a/doc/features/server-quorum.md b/doc/features/server-quorum.md
new file mode 100644
@@ -0,0 +1,44 @@
+# Server Quorum
+Server quorum is a feature intended to reduce the occurrence of "split brain"
+after a brick failure or network partition. Split brain happens when different
+sets of servers are allowed to process different sets of writes, leaving data
+in a state that can not be reconciled automatically. The key to avoiding split
+brain is to ensure that there can be only one set of servers - a quorum - that
+can continue handling writes. Server quorum does this by the brutal but
+effective means of forcing down all brick daemons on cluster nodes that can no
+longer reach enough of their peers to form a majority. Because there can only
+be one majority, there can be only one set of bricks remaining, and thus split
+brain can not occur.
+Server quorum is controlled by two parameters:
+ * **cluster.server-quorum-type**
+ This value may be "server" to indicate that server quorum is enabled, or
+ "none" to mean it's disabled.
+ * **cluster.server-quorum-ratio**
+ This is the percentage of cluster nodes that must be up to maintain quorum.
+ More precisely, this percentage of nodes *plus one* must be up.
+Note that these are cluster-wide flags. All volumes served by the cluster will
+be affected. Once these values are set, quorum actions - starting or stopping
+brick daemons in response to node or network events - will be automatic.
+## Best Practices
+If a cluster with an even number of nodes is split exactly down the middle,
+neither half can have quorum (which requires **more than** half of the total).
+This is particularly important when N=2, in which case the loss of either node
+leads to loss of quorum. Therefore, it is highly advisable to ensure that the
+cluster size is three or greater. The "extra" node in this case need not have
+any bricks or serve any data. It need only be present to preserve the notion
+of a quorum majority less than the entire cluster membership, allowing the
+cluster to survive the loss of a single node without losing quorum.