@martinesmann passed me logs from your cluster. Problems that I’m seeing on your server:
node has just 2 cores, minimum recommended is 4.
Transparent HugePages is enabled, recommendation is to disable it.
Bucket resident ratio is close to 0%, that can leads to loads of other problems.
background fetches from disk at peak is around 17K ops/sec per node, which is pretty high! and this is expected because of low residency. And if your disk IO capacity isn’t sufficient, you will see timeout errors.
For info, we have add a new node : solve the issue but only during few days
I know about background fetch and it’s the issue. I want to solve them. I don’t understand why the server store more replica in memory and not active.
Also it’s same servers but we have more doc in RAM on the last server…
It does appear to be. I’d recommend filing an issue and attaching a cbcollect_info. Also, if you have an enterprise subscription, you may want to raise awareness with Couchbase support since that’ll escalate getting the issue looked at.