A system for managing distributed storage includes a collection of one or more compute sites that are connected and that perform a joint computation. A witness node is connected to the one or more compute sites. A membership service allows a change in collection membership, in response to a failure, to continue the joint computation without causing data inconsistency or corruption. If the membership service is unable to reach a decision on the change in collection membership in response to the failure, the witness node controls the decision on the change in collection membership to continue the joint computation. A distributed storage system provided by the system described herein may advantageously meet consistency guarantees and maximize data access even when individual compute sites and/or their interconnects fail.