Uploaded image for project: 'Couchbase Server'
  1. Couchbase Server
  2. MB-6511

We should try to prioritize move of active vbuckets off most loaded node and into least loaded node

    Details

    • Type: Bug
    • Status: Resolved
    • Priority: Critical
    • Resolution: Fixed
    • Affects Version/s: None
    • Fix Version/s: 2.0.1
    • Component/s: ns_server
    • Security Level: Public
    • Labels:

      Description

      SUBJ.

      It was found that since 1.8.1 which does reliable replica building we tend to 'move replicas' faster than we move active vbuckets. Sometimes this is just misreading of stats, but in some cases indeed we can try to order movements better

      No reviews matched the request. Check your Options in the drop-down menu of this sections header.

        Activity

        Hide
        alkondratenko Aleksey Kondratenko (Inactive) added a comment -

        See MB-6512 as good example what we're doing 'wrong'.

        Show
        alkondratenko Aleksey Kondratenko (Inactive) added a comment - See MB-6512 as good example what we're doing 'wrong'.
        Hide
        alkondratenko Aleksey Kondratenko (Inactive) added a comment -

        Particularly, because change of replicas without change of active node for any vbucket is still counted as move, those moves are still limited on basis of 'outgoing' node (which in this case is same as incoming).

        So we often see a confusing case where at the beginning of rebalance we 'move' only replicas. That happens exactly because for all movements 'out' from each node we sometimes first do moves 'into' same node, thus active vbuckets are indeed not moved.

        For sake of equalizing load on cluster as soon as possible we indeed should avoid this. We need to move 'replica only' vbuckets in the end. Or at least, clearly, not at the beginning.

        Show
        alkondratenko Aleksey Kondratenko (Inactive) added a comment - Particularly, because change of replicas without change of active node for any vbucket is still counted as move, those moves are still limited on basis of 'outgoing' node (which in this case is same as incoming). So we often see a confusing case where at the beginning of rebalance we 'move' only replicas. That happens exactly because for all movements 'out' from each node we sometimes first do moves 'into' same node, thus active vbuckets are indeed not moved. For sake of equalizing load on cluster as soon as possible we indeed should avoid this. We need to move 'replica only' vbuckets in the end. Or at least, clearly, not at the beginning.
        Hide
        alkondratenko Aleksey Kondratenko (Inactive) added a comment -

        This is actually done and shipped in 2.0.1

        Show
        alkondratenko Aleksey Kondratenko (Inactive) added a comment - This is actually done and shipped in 2.0.1

          People

          • Assignee:
            alkondratenko Aleksey Kondratenko (Inactive)
            Reporter:
            alkondratenko Aleksey Kondratenko (Inactive)
          • Votes:
            0 Vote for this issue
            Watchers:
            2 Start watching this issue

            Dates

            • Created:
              Updated:
              Resolved:

              Gerrit Reviews

              There are no open Gerrit changes