fleetctl: fleetctl load stops already scheduled unit

[kayrus]

/cc @jonboulle @tixxdz

Steps to reproduce:

• create unit file app.service:

[Service]
ExecStart=/usr/bin/bash -c 'while true; do echo hello1; sleep 1; done'

• start it: fleetctl start app
• modify local app.service: sed -i 's#hello1#hello2#g' app.service
• load it: fleetctl load app ->

$ fleetctl load app.service 
WARNING: Unit app.service in registry differs from local unit file app.service
Unit app.service loaded on 659caee1.../coreos1

And this unit will be stopped.

Have to review lazyLoadUnits https://github.com/coreos/fleet/blob/master/fleetctl/fleetctl.go#L705

And it should be much smarter, i.e. like lazyCreateUnits one https://github.com/coreos/fleet/blob/master/fleetctl/fleetctl.go#L597

In addition we have to create unit tests for this behavior.

[guilhem]

This behavior is also a problem when using Global=true. If load doesn't stop, it can be used for a manual rolling upgrade:

• start globalUnit (Filev1)
• update Filev1 -> Filev2
• load globalUnit (Filev2)
• rolling upgrade:
  • ssh machine1 "systemctl restart globalUnit"
  • ssh machine2 "systemctl restart globalUnit"
  • ...

So I'm 👍 about changing current behavior

[zeisss]

FYI load/start/stop describe the desired state, not the action to perform, thus fleetctl load will move the unit from whichever state it is in to the loaded state.