Goals:

• Track each node's STONITH_WATCHDOG_TIMEOUT
  • Most likely in the controller
  • A node_state attribute would be good, so it persists across reboots
  • Cluster nodes could advertise it as part of the join process
  • Pacemaker Remote nodes already have a timeout verification when it could be done
  • The local STONITH_WATCHDOG_TIMEOUT would be the default before a target node's value is known (watchdog fencing is unreliable before a target is first seen anyway, and a never-seen node shouldn't be running resources)
• The stonith-watchdog-timeout cluster option should be deprecated and replaced with one or more new options
  • The new options should use standard types/validators hen practical
  • The new option names should use some variant of "fencing" rather than "stonith"
  • Something like watchdog-fencing-duration as a duration (nonnegative interval spec) to set a specific wait time
  • Something like watchdog-fencing-auto=true/false/increase/decrease where true = use twice the target-specific value, false = use specified duration exactly, increase = use the higher of the specified duration or twice the target-specific value, decrease = use the lower of the specified duration or twice the target-specific value
  • The new syntax should be designed so it is straightforward to XSL-transform the old syntax to it when we eventually drop it