when issuing pcs cluster stop --all the stop sequence will:
stop all services on node1 -> stop cluster on node1
stop all services on node2 -> stop cluster on node2
There are 2 major problems with approach:
1) at somepoint there will be loss of quorum that can potentially block some resources that need quorum to stop (so when you get to the N node the resource will fail to stop and node will kill itself)
2) some applications (for example galera) have their own internal quorum and needs to be stopped all together to avoid corner case data corruption. Generally we can assume (for now at least) that external application quorum follows our quorum rules.
What we see in OpenStack is that stop --all on 3 node, will result in:
- moving tons of services around unnecessarily (since they will all die sooner or later)
- node fencing when stopping the last node because services can´t write to the db anymore (lack of internal quorum)
My recommendation here is:
on pcs cluster stop --all -> stop all resources -> stop cluster (at this point it´s irrelevant if it´s done one at a time or all together)
on pcs cluster stop -> this operation should verify if removing a node from the cluster will cause loss of quorum and abort. Of course we want to allow a manual override --pretty-please-i-know-i-will-lose-quorum-go-on
Created attachment 980519 [details]
proposed fix for stop --all
related patch is already in upstream: https://github.com/feist/pcs/commit/7221995f47fa33eae271cce01d7b361d3578ef15
Created attachment 981508 [details]
proposed fix - loss of quorum warning
Created attachment 981896 [details]
proposed fix for destroy --all
*** Bug 1165412 has been marked as a duplicate of this bug. ***
Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.
For information on the advisory, and where to find the updated
files, follow the link below.
If the solution does not work for you, open a new bug report.