Try to persuade people to backup etcd

[justinsb]

We really need people to back up before attempting this upgrade, and to understand the scenarios where there is no alternative but to restore from backup.

I hope this is enough to persuade users that a backup really is needed this time, but I worry that it's still not enough.

[justinsb]

cc @pwittrock @mml @ethernetdan @smarterclayton

[pwittrock]

I would put This change is irreversible in bold.

[idvoretskyi]

@justinsb will be merged after the review.

[justinsb]

Thanks @pwittrock - emboldened "This change is irreversible".

To me, "HA installations cannot be migrated at the current time using the official kubernetes procedure" is even worse - we are nuking anyone that upgrades their HA kube-up cluster - but we did say that was alpha https://kubernetes.io/docs/admin/ha-master-gce/

[ethernetdan]

To me, "HA installations cannot be migrated at the current time using the official kubernetes procedure" is even worse - we are nuking anyone that upgrades their HA kube-up cluster - but we did say that was alpha https://kubernetes.io/docs/admin/ha-master-gce/

Agreed - it's definitely unfortunate we have to have cluster turndown for upgrade there. The change of media types requires a cluster bringdown because AFAIK there's no support for mixed media types on the API server so disruption of HA is currently inevitable.

[justinsb]

To be clear, I believe that GCE kube-up users that followed https://kubernetes.io/docs/admin/ha-master-gce/ and try to upgrade will have their clusters destroyed. Best case, the upgrade fails and they are stuck on 1.5.

Puts a few minutes downtime into perspective...

[wojtek-t]

To me, "HA installations cannot be migrated at the current time using the official kubernetes procedure" is even worse - we are nuking anyone that upgrades their HA kube-up cluster - but we did say that was alpha https://kubernetes.io/docs/admin/ha-master-gce/

Agreed - it's definitely unfortunate we have to have cluster turndown for upgrade there. The change of media types requires a cluster bringdown because AFAIK there's no support for mixed media types on the API server so disruption of HA is currently inevitable.

I think there may be some misunderstanding here.
The cluster turndown (which I agree is unfortunate) is not about changing the media type from json to protobufs (which we call media type). And we can leave with part of objects be still stored in json and the other part in protobufs.
The cluster turndown is fully because migration to etcd v3.

[justinsb]

Rebased

[calebamiles]

merging this warning as is, @justinsb, @wojtek-t, @ethernetdan please open another PR for further changes.

cc: @idvoretskyi, @kubernetes/kubernetes-release-managers