...
- Network congestion or TCP incast in racktop switches.
TODO: REASONS
One possible solution (in case of 2 physical nodes for control) is to host them in different racks. This leaves the ability in case of the problems above in one rack, to recover all control components in the node in the other rack. Hypervisors are wise to keep in separate rack due to the type and nature of network traffic, but also because of the danger of out-of-control virtualized resources, be that then caused by poor configuration or the applications they host. - Low availability markers
Depending on network topology of DC, the reasons in section 1, but also hardware failures may be the cause of unsatisfactory MTBF and MTTR. For example, if both control nodes are placed in one rack, rack-level hardware failure is causing at least short total downtime during diagnostics/repair process.That is not of course the case, when Warren is used with more heavily, than just 3 nodes tryout.
...