Assuming a healthy cluster where all members are online, all IO to an object is completely synchronous. When a WRITE operation is executed, the acknowledgement is not sent back to the writer until the write has been acknowledged as written into the SSD write buffer associated with the disk groups where the object's various components reside. Reads will be serviced from whichever mirror/read cache is most convenient.

In this way, data consistency is absolutely and aggressively maintained across the entire distributed object. As a result of this logic, any given mirror can fail and production will resume seamlessly with no loss of data or on-disk consistency.