why use Recreate strategy in ceph-mgr and rbd-provisioner deployment

Asked by liuzhong chen on 2018-05-26

I wonder why use Recreate strategy in ceph-mgr and rbd-provisioner deployment? if RollingUpdate is better.Thank you!
but I find Alexandre Marangone submit a patch to change the ceph-mgr and rbd-provisioner to recreate,I think he must have some reason,but I can't understand why.

Question information

English Edit question
openstack-helm Edit question
No assignee Edit question
Solved by:
liuzhong chen
Last query:
Last reply:
Deshi Xiao (xiaods) said : #1

ask Alexandre Marangone to give some coin directly.

Alexandre Marangone (alram) said : #2

That was a while ago but if I recall correctly...
The idea for ceph-mgr was that we only wanted to run one at a time in cluster deployed with k8s since there could only be one active ceph-mgr daemon in a Ceph cluster. Figuring out which one is active was up to the operator who had to manually exec to a ceph-mon Pod to do so.
Since the ceph-mgr is not critical to I/O, having only 1 was a good enough solution at the time. Only 1 ceph-mgr at time -> Recreate.
I haven't followed this very carefully but I do remember changes in ceph-mgr where a standby mgr can now redirect requests to the active one. That needs to be checked but if that's the case, it could be updated to run 3 and a RollingUpdate strategy.

Regarding rbd-provisioner, I'm not sure. I haven't look at this bit but a RollingUpdate strategy sounds reasonable.

liuzhong chen (huigui65) said : #3

@Alexandre Marangone
Thank you!
As mgr share some responsibility for ceph-mon.If it is better keep three replications for mgr?

Alexandre Marangone (alram) said : #4

They scale independently of each other but I do think the common stance for now is to have the same amount, just easier that way

liuzhong chen (huigui65) said : #5

@Alexandre Marangone
okļ¼ŒThank you!