High Availability

What do you mean by High Availability?

With the help of multiple chunk servers and good goals, files can be stored multiple times. Therefore, a certain level of high availability on a file level can be achieved easily.

In addition, it is important to know that per default, the master service only can be active in a master role on one node at the time. If this node fails, e.g. because of broken hardware or out-of-memory situations, the current master has to be demoted (if still possible) and an existing shadow has to be promoted manually.

If the fail over happens automatically, a good state of high availability is achieved on a service level. Thus the term “High Availability” refers to keeping the master role alive when everything goes down under.

How can I achieve High Availability of the master?

There are multiple ways of keeping the master highly available.

One would be to demote and promote manually if you need to. The better way would be to delegate that task to a mechanism which knows the current state of all (possible) master nodes and can perform the fail over procedure automatically.

Known methods, when only using open-source software, are building Pacemaker/ Corosync clusters with self-written OCF agents. Another way could be using keepalived.

This is too complicated! I need a better solution for High Availability.

An officially supported way to achieve high availability of the master is to obtain the uRaft component from Skytechnology Sp. z o.o., the company behind LizardFS. Based on the raft algorithm, the uRaft service makes sure that all master nodes talk to each other and exchange information regarding their health states.

In order to ensure that a master exists, the nodes participate in votes. If the current master fails, uRaft moves a floating IP from the formerly active node to the new designated master. All uRaft nodes have to be part of one network and must be able to talk to each other.

The uRaft component can be obtained by signing a support contract with SkyTechnology Sp. z o.o..