Avoiding a data center failure

Use this information to understand how to avoid a data center failure.

Configuring the data center infrastructure

Each data center within a stretch cluster can have a different storage cluster configuration to reflect local capabilities and dependencies. Set up replication between the data centers to help preserve the data. If one data center fails, the other data centers in the storage cluster contain copies of the data.

Setting up failure domains within the CRUSH map hierarchy

Failure, or failover, domains are redundant copies of domains within the storage cluster. If an active domain fails, the failure domain becomes the active domain.

By default, the CRUSH map lists all nodes in a storage cluster within a flat hierarchy. However, for best results, create a logical hierarchical structure within the CRUSH map. The hierarchy designates the domains to which each node belongs and the relationships among those domains within the storage cluster, including the failure domains. Defining the failure domains for each domain within the hierarchy improves the reliability of the storage cluster.

When planning a storage cluster that contains multiple data centers, place the nodes within the CRUSH map hierarchy so that if one data center goes down, the rest of the storage cluster stays up and running.

Designating failure nodes within the domains

If you plan to use three-way replication for data within the storage cluster, consider the location of the nodes within the failure domain. If an outage occurs within a data center, it is possible that some data might reside in only one copy. When this scenario happens, there are two options:

  • Leave the data in read-only status with the standard settings.

  • Live with only one copy for the duration of the outage.

With the standard settings, and because of the randomness of data placement across the nodes, not all the data will be affected, but some data can have only one copy and the storage cluster would revert to read-only mode. However, if some data exist in only one copy, the storage cluster reverts to read-only mode.