none
Question about dynamic quorum

    Question

  • Hello,

    We have three nodes Exchange 2013 DAG with DAC enabled. All three servers have identical Exchange roles installed. There are two datacenters: primary and DR - each configured as separate AD site. In primary DC there are two nodes, in DR - one. OS is Windows Server 2012 R2. By default Dynamic quorum feature is enabled. When all three nodes are up and running and there is connectivity between sites every node has DynamicWeight of 1 which fits my understanding. Quorum model is NodeMajority.

     Now, during test we found out that when one server (any of three) goes offline the one from two which are alive and which have databases mounted has DynamicWeight of 1 and the second one (with passive DBs) changes its DynamicWeight to 0. Again it does not matter which server goes offline of three behavior is always the same. It looks like that it is by design. Can you please give some comments on this? Does it mean that if the server with mounted DBs goes down the last one will not be able to become 'last man standing' - in other words to maintain a quorum and to mount DBs?

    Many thanks in advance!

    Nikolay



    • Edited by Nikolay__ Thursday, May 31, 2018 7:04 AM
    Thursday, May 31, 2018 7:03 AM

All replies

  • Hello,

    We have three nodes Exchange 2013 DAG with DAC enabled. All three servers have identical Exchange roles installed. There are two datacenters: primary and DR - each configured as separate AD site. In primary DC there are two nodes, in DR - one. OS is Windows Server 2012 R2. By default Dynamic quorum feature is enabled. When all three nodes are up and running and there is connectivity between sites every node has DynamicWeight of 1 which fits my understanding. Quorum model is NodeMajority.

     Now, during test we found out that when one server (any of three) goes offline the one from two which are alive and which have databases mounted has DynamicWeight of 1 and the second one (with passive DBs) changes its DynamicWeight to 0. Again it does not matter which server goes offline of three behavior is always the same. It looks like that it is by design. Can you please give some comments on this? Does it mean that if the server with mounted DBs goes down the last one will not be able to become 'last man standing' - in other words to maintain a quorum and to mount DBs?

    Many thanks in advance!

    Nikolay



    You have a witness defined yes? Assuming the last server and the witness are available, the last server will remain up and the DBs will stay mounted. 

    https://blogs.technet.microsoft.com/askcore/2016/03/21/behavior-of-dynamic-witness-on-windows-server-2012-r2-failover-clustering-3/

    If you lose the witness as well, then it will go down

    Thursday, May 31, 2018 12:00 PM
    Moderator
  • Hi David,

    Thank you for your answer and link - very helpful! However my post was mainly about odd node number clusters (three in the mentioned environment). We have no witness server defined at the moment in this environment. And in this situations as I understood if servers will go offline one by one database will remain mounted even if only one server left (last man standing). But it case only one node is offline, dynamic weight of remaining servers will not become equal 1 for all the nodes. Only one server with active databases will have dynamic weight of 1. Second will get 0. And this is by design as I understood: the majority of two node cluster is two! So if one will goes offline databases have to dismount.

    Regards,

    Nikolay


    • Edited by Nikolay__ Monday, June 4, 2018 12:40 PM
    Monday, June 4, 2018 12:19 PM
  • Hi David,

    Thank you for your answer and link - very helpful! However my post was mainly about odd node number clusters (three in the mentioned environment). We have no witness server defined at the moment in this environment. And in this situations as I understood if servers will go offline one by one database will remain mounted even if only one server left (last man standing). But it case only one node is offline, dynamic weight of remaining servers will not become equal 1 for all the nodes. Only one server with active databases will have dynamic weight of 1. Second will get 0. And this is by design as I understood: the majority of two node cluster is two! So if one will go offline databases have to dismount.

    Regards,

    Nikolay

    You need to define a witness.

    Monday, June 4, 2018 12:26 PM
    Moderator
  • David, as far as I understand - no, with odd number of nodes in the cluster and NodeMajority quorum model (which is default for odd number of nodes clusters), I do not need to define a witness. It will make no sense. Even when cluster looses one node and the number of survived nodes becomes even it still does not automatically change quorum model to Node and FileShare majority and cluster still does not need a witness as far as I understand. Here is an article precisely describing our case:

    http://techgenix.com/exchange-2013-dag-dynamic-quorum-part2/

    There is no witness defined and it is not needed.

    Please correct me if I'm wrong.

    Regards,

    Nikolay


    • Edited by Nikolay__ Monday, June 4, 2018 3:06 PM
    Monday, June 4, 2018 3:05 PM
  • David, as far as I understand - no, with odd number of nodes in the cluster and NodeMajority quorum model (which is default for odd number of nodes clusters), I do not need to define a witness. It will make no sense. Even when cluster looses one node and the number of survived nodes becomes even it still does not automatically change quorum model to Node and FileShare majority and cluster still does not need a witness as far as I understand. Here is an article precisely describing our case:

    http://techgenix.com/exchange-2013-dag-dynamic-quorum-part2/

    There is no witness defined and it is not needed.

    Please correct me if I'm wrong.

    Regards,

    Nikolay


    Sorry for the confusion, I was referring specifically to the dynamic witness that the cluster will maintain, hence the link.

    When you added the second node to the DAG, a witness would be created automatically. There is always a witness *defined*,- even if its not used with odd members. Do you see it defined in your cluster?

    The only real way to know is to test unfortunately. When you get to the last 2 servers, assuming that its a planned downing, the last one should remain standing. 

    BTW, My first name is Andy  :)

    Monday, June 4, 2018 5:28 PM
    Moderator
  • Andy,

    Looks like we are getting closer! Thank you for the patience! Actually I've read many articles regarding Exchange cluster behavior with Dynamic quorum and witness but still have some questions which I'm sure you will be able to answer :)

    Yes, technically speaking we have witness defined (it is mentioned in the DAG's properties) and we added third node not so far ago - before that it was two nodes cross-site cluster with the witness in the third site. So I think when we added third node the quorum model automatically switched to NodeMajority but witness server have not gone from the DAG configuration and if for some reasons quorum model will again change in the future to the Node + FileShare majority it will start play its role again automatically. Is this understanding correct so far?

    If yes then I should tell you that according to our tests when we started to power off mailbox nodes one by one in three-node configuration as I said quorum model did not change and the witness did not come into play. I still think this is by design.

    So just to summarize:

    1. Is it recommended practice to define witness server in all DAG configurations - no matter how many nodes in the cluster - just to make witness selection much easier and guaranteed in case quorum model will change to Node + FileShare majority?

    2. In which cases does quorum model actually change? As far as I understand only in two cases: when administrator adds/removes node in the cluster or executes Start/Stop-DatabaseAvailabilituGroup which in turn also makes changes to underlying cluster config. When nodes go offline consequently one by one or several at the same time - quorum model does not change (which is proved by our tests). Is this correct understanding?

    3. Is there any difference from the node's dynamic weight perspective when the node is gracefully shut down or unexpectedly powered off/failed? I heard that if you just power off the node the databases wont mount on neither of the remaining nodes. I think this is not true.

    Many thanks in advance!

    Nikolay

    Tuesday, June 5, 2018 7:19 AM