N+1 redundancy for shared storage




Partially Implemented



This document describes how N+1 redundancy is achieved for instances using shared storage.

Current state and shortcomings

For instances with DRBD as disk template, in case of failures of their primary node, there is only one node where the instance can be restarted immediately. Therefore, htools reserve enough memory on that node to cope with failure of a single node. For instances using shared storage, however, they can be restarted on any node—implying that on no particular node memory has to be reserved. This, however, motivated the current state where no memory is reserved at all. And even a large cluster can run out of capacity.

Proposed changes

Definition on N+1 redundancy in the presence of shared storage

A cluster is considered N+1 redundant, if, for every node, all DRBD instances can be migrated out and then all shared-storage instances can be relocated to a different node without moving instances on other nodes. This is precisely the operation done after a node breaking. Obviously, simulating failure and evacuation for every single node is an expensive operation.

Basic Considerations

For DRBD, keeping N+1 redundancy is affected by moving instances and balancing the cluster. Moreover, taking is into account for balancing can help Improving allocation efficiency by considering the total reserved memory. Hence, N+1 redundancy for DRBD is to be taken into account for all choices affecting instance location, including instance allocation and balancing.

For shared-storage instances, they can move everywhere within the node group. So, in practice, this is mainly a question of capacity planing, especially is most instances have the same size. Nevertheless, offcuts if instances don’t fill a node entirely may not be ignored.

Modifications to existing tools

  • hail will compute and rank possible allocations as usual. However, before returning a choice it will filter out allocations that are not N+1 redundant.

  • Normal gnt-cluster verify will not be changed; in particular, it will still check for DRBD N+1 redundancy, but not for shared storage N+1 redundancy. However, hcheck will verify shared storage N+1 redundancy and report it that fails.

  • hbal will consider and rank moves as usual. However, before deciding on the next move, it will filter out those moves that lead from a shared storage N+1 redundant configuration into one that isn’t.

  • hspace computing the capacity for DRBD instances will be unchanged; In particular, the options --accept-exisiting and --independent-groups will continue to work. For shared storage instances, however, will strictly iterate over the same allocation step as hail does.