On 06/25/13 at 09:42am, Joe Gordon wrote:
On Tue, Jun 25, 2013 at 7:22 AM, Andrew Laski <andrew.la...@rackspace.com>wrote:
I have a couple of reviews up to introduce the concept of shelving an
instance into Nova. The question has been raised as to whether or not this
belongs in Nova, or more rightly belongs in Heat. The blueprint for this
feature can be found at https://blueprints.launchpad.**
net/nova/+spec/shelve-instance<https://blueprints.launchpad.net/nova/+spec/shelve-instance>
**, but to make things easy I'll outline some of the goals here.
The main use case that's being targeted is a user who wishes to stop an
instance at the end of a workday and then restart it again at the start of
their next workday, either the next day or after a weekend. From a service
provider standpoint the difference between shelving and stopping an
instance is that the contract allows removing that instance from the
hypervisor at any point so unshelving may move it to another host.
the part that caught my eye as something that *may* be in heat's domain and
is at least worth a discussion is the snapshotting and periodic task part.
from what I can tell, it sounds like the use case is for this is: I want
to 'shutdown' my VM overnight and save money since I am not using it, but I
want to keep everything looking the same.
But in this use case I would want to automatically 'shelve' my instance off
the compute-server every night (not leave it on the server) and every
morning I would want it to autostart before I get to work (and re-attach my
volume and re-associate my floating-ip). All of this sounds much closer to
using heat and snapshotting then using 'shelving.'
The periodic task for removing a shelved instance from the hypervisor is
a first pass attempt at a mechanism for reclaiming resources, and is
under discussion and will probably evolve over time. But the motivation
for reclaiming resources will be driven by deployment capacity or the
desire to reshuffle instances or maybe something else that's important
to a deployer. Not the user. Since I see Heat as an advocate for user
requests, not deployer concerns, I still think this falls outside of its
concerns.
There's no concept of autostart included in shelving. I agree that that
gets beyond what should be performed in Nova.
Additionally, storing the shelved instance locally on the compute-node
until a simple periodic task to migrates 'shelved' instances off into deep
storage seems like it has undesired side-effects. For example, as long as
the shelved instance is on a compute-node, you have to reserve CPU
resources for it, otherwise the instance may not be able to resume on the
same compute-node invalidating the benefits (as far as I can tell) of
keeping the instance locally snapshotted.
You're correct that there's not a large benefit to a deployer unless
resources are reclaimed. Perhaps some small power savings, and the
freedom to migrate the instance transparently if desired. I would
prefer to remove the instance when it's shelved rather than waiting for
something, like a periodic task or admin api call, to trigger it. But
booting disk based images can take a fairly long time so I've optimized
for the case of an instance being shelved for a day or a weekend. That
way users get acceptable unshelve times for the expected case, and
deployers benefit when an instance is shelved longer term. I don't
think this needs to be set in stone and the internal working can be
modified as we find ways to improve it.
From a user standpoint what they're looking for is:
The ability to retain the endpoint for API calls on that instance. So
v2/<tenant_id>/servers/<**server_id> continues to work after the instance
is unshelved.
All networking, attached volumes, admin pass, metadata, and other user
configurable properties remain unchanged when shelved/unshelved. Other
properties like task/vm/power state, host, *_at, may change.
The ability to see that instance in their list of servers when shelved.
This sounds like a good reason to keep this in nova.
Again, the objection that has been raised is that it seems like
orchestration and therefore would belong in Heat. While this is somewhat
similar to a snapshot/destroy/rebuild workflow there are certain properties
of shelving in Nova that I can't see how to reproduce by handling this
externally. At least not without exposing Nova internals beyond a
comfortable level.
What properties are those, and more importantly why I need them?
Mainly uuid, but also the server listing. If Heat
snapshots and removes an instance it has no way to recreate it with the
same uuid. As much as I wish it wasn't the case, this is important to users.
So I'd like to understand what the thinking is around why this belongs in
Heat, and how that could be accomplished.
______________________________**_________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.**org <OpenStack-dev@lists.openstack.org>
http://lists.openstack.org/**cgi-bin/mailman/listinfo/**openstack-dev<http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev>
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
