Details
-
Improvement
-
Resolution: Done
-
Major
-
6.6.0
Description
Although DCP V2 is faster and less resource intensive for the backup client, the same can't be said for the cluster. When performing a backup, there is a high likelihood that a single node in the cluster will be targeted before moving onto the next. This is the nature of sequential vBucket backup; we can alleviate this by distributing the vBuckets evenly among our DCP agents. This will mean instead of all the agents targeting a single node, they will each target a different node in the cluster; obviously this won't be perfect because there is no guarantee that the number of node in the cluster matches the number of agents cbbackupmgr is using.
See the cbmonitor output of some performance testing on Leto for a graphical explanation of the problem we are facing. Looking at the cbmonitor output you can see a spike/drop in memcached CPU usage as we move from node to node.
Attachments
Issue Links
- duplicates
-
MB-27550 cbbackupmgr: Should limit the number of vbuckets streamed from one node at a time.
- Closed