helm install should rollback on failure

[adamreese]

helm should cleanup any successful manifest uploads if one fails

[mboersma]

:+1: In testing the "deis" chart I ran into errors, and I expected to be able to helm install again, but I couldn't because some objects were still installed. helm uninstall <chart> && helm install <chart> works, but I did expect helm install essentially to be atomic.

[technosophos]

As we're starting to do the automatic dependency resolution work, I'm thinking that...

• On failure, a chart should uninstall any of the components it already installed, but...
• A chart should not uninstall any of the charts it depends upon, even if those were installed expressly to satisfy this chart's dependency requirements.

That will keep us from potentially harmful bugs where we remove charts that were actually being used.

Does that make sense?

[mboersma]

@technosophos that behavior makes sense for me when dependency resolution is in place, unless running helm install again would error on the dependent charts already being installed. But it sounds like dependency management would be smart enough not to.

[arschles]

:+1: also makes sense to me. tiny thing - it'd be helpful to log when helm install runs into already-installed dependencies.