#1907
  • Created on
    2012-03-14 12:47
  • Reported by
    Jonas Bonér
  • Status
    New
  • Priority
    Normal (3)
  • Assigned to
    -
  • Milestone
    Rollins
  • Plan Level
    Subtask
  • Due Date
    -

Attachments

Related Tickets

Implement support for DOWNING a node that is reachable

for maintenance etc.
tricky part is that he will still be able to gossip, but not receive gossip so his failure detector will detect all nodes as 'unreachable'
how to force him to go through JOINING phase again?

Now we have 'ClientDowningNodeThatIsUpMultiJvmSpec' test. But it doesn't test:
  • rejoining after being DOWNed
  • managing the gossip while being DOWN
User picture
  on 2012-06-01 09:08 *
By Patrik Nordwall
Priority changed from Normal (3) to High (2)
User picture
  on 2012-06-04 21:21 *
By Jonas Bonér
Assigned to set to Jonas Bonér
Status changed from New to Accepted
User picture
  on 2012-06-04 21:22 *
By Jonas Bonér
Description changed from for maintenance etc. tricky... to for maintenance etc. tricky...
User picture
  on 2012-07-05 15:16 *
By Jonas Bonér
Assigned to changed from Jonas Bonér to -none-
Status changed from Accepted to New
User picture
  on 2012-07-05 20:59 *
By Jonas Bonér
Estimate changed from None to Small
Sum of child estimates changed from 0.0 to 1.0
User picture
  on 2012-08-30 14:10 *
By Patrik Nordwall
Priority changed from High (2) to Normal (3)
User picture
  on 2012-12-05 16:53 *
By rkuhn
Is this ticket not invalid? Maintenance should be done via the LEAVING state as far as I thought.
User picture
  on 2012-12-06 11:17 *
By Patrik Nordwall
Well, I guess the use case would be to immediately bring it down without doing all the graceful handoff stuff.
I don't know if we need that. I think we should do this when we know how the partitioning and handoff is designed.
What is implemented now is a way to take a node out of the member ring (it's added to the unreachable set) without involving the failure detector, but that's all.
User picture
  on 2012-12-07 10:03 *
By rkuhn
okay, I am still not convinced that we actually need it, but deferring the decision to Rollins
User picture
  on 2012-12-07 10:03 *
By rkuhn
Milestone changed from Coltrane to Rollins
User picture
  on 2012-12-07 10:17 *
By Jonas Bonér
I think we can mark it as invalid. We can always open it up later if needed. I agree that LEAVING should be the default way of leaving a cluster.