[Maas-devel] RFC: "Serialising" power actions

[Julian Edwards]

On Wednesday 17 Sep 2014 09:39:05 Raphaël Badin wrote:
> [...]
> 
> >> I think there is a middle ground here: it's true that MAAS should
> >> consider it is in control of the nodes; controlling (and monitoring) the
> >> power state is certainly part of it.  But as Gavin pointed out, once a
> >> node is deployed, it's out of the question to consider it an *error* (as
> >> in, something so bad MAAS will transition a node to an error state) if
> >> the actual power state is different from the expected power state.
> > 
> > No, I disagree.  It's an error because MAAS is in control and expects it
> > to
> > be in a particular state.
> 
> That's true for all the MAAS-controlled steps: namely commissioning,
> deploying and releasing.  Not so much for 'deployed': a node at this
> stage is potentially powering up a user-managed service.  If something
> looks wrong with this node (for instance if the power state seems wrong)
> we want to issue a warning but we don't want to take the initiative to
> mark this node as failed.
> 
> > You cannot have a half way house, either MAAS is managing it or it is not.
> > MAAS owns the network.
> 
> We want MAAS to be the tool through which you control the power state of
> the nodes; that's why we should *flag* deployed nodes that seem to be in
> the wrong power state.  That being said, I don't think we can rule out
> the possibility that a deployed server's power state will be controlled
> externally; this is mostly a matter of blending in with existing tooling
> and practices.

I believe flagging to be wrong, MAAS should be taking corrective action.  If 
someone wants to turn off a machine, do it through MAAS, not elsewhere.

For example, the power could have failed, and the BIOS might not be set to 
restart the machine upon power restoration.  MAAS could be useful and power up 
the node if it knows it should be on.

Just because a node is deployed does not mean we can abdicate responsibility 
for this.