Michael Smith

Could part of getting stuck be that the service doesn't have enough time to perform all shutdowns? Services have a 90s timeout for ExecStop on shutdown.

We use https://gist.github.com/MikaelSmith/2603b9c09079f02fc914f7860603f652 for a single node, which is specific to containerd but shouldn't be too hard to adapt for Docker if needed. Not sure how well it would work...

Host preflights seem to be in Troubleshoot and kURL, but I haven't found any documentation of them yet.

Also the command documented at https://kurl.sh/docs/install-with-kurl/upgrading#airgap isn't valid, should probably be something like `cat tasks.sh | sudo bash -s load-images`.

Well, I still think it's poor default behavior, but I believe the example does work.

One thing I'd really like is that there's a way to define the collector and analyzer that works for both before and after your changes. Otherwise it's going to be...

I haven't thought through your proposed changes enough to have a specific problem. I was stating a possible requirement that would help smooth upgrades.

Maybe this requires a change in support bundle and troubleshoot api version? I guess that would handle the migration issue.

That seems like it could work. To some extent resources are still needed to perform the upgrade: need to spin up a new replica before terminating the previous one. It...

Yeah, that seems fine.