Open Stack

Gary Kotton came across a doozy of a bug recently:

https://bugs.launchpad.net/nova/+bug/1419785

In short, when you start a Nova compute, it will query the driver for
instances and compare that against the expected host of the the instance
according to the DB. If the driver is reporting an instance the DB
thinks is on a different host, it assumes the instance was evacuated
while Nova compute was down, and deletes it on the hypervisor. However,
Gary found that you trigger this when starting up a backup HA node which
has a different `host` config setting. i.e. You fail over, and the first
thing it does is delete all your instances.

Gary and I both agree on a couple of things:

1. Deleting all your instances is bad
2. HA nova compute is highly desirable for some drivers

We disagree on the approach to fixing it, though. Gary posted this:

https://review.openstack.org/#/c/154029/

I've already outlined my objections to this approach elsewhere, but to
summarise I think this fixes 1 symptom of a design problem, and leaves
the rest untouched. If the value of nova compute's `host` changes, then
the assumption that instances associated with that compute can be
identified by the value of instance.host becomes invalid. This
assumption is pervasive, so it breaks a lot of stuff. The worst one is
_destroy_evacuated_instances(), which Gary found, but if you scan
nova/compute/manager for the string 'self.host' you'll find lots of
them. For example, all the periodic tasks are broken, including image
cache management, and the state of ResourceTracker will be unusual.
Worse, whenever a new instance is created it will have a different value
of instance.host, so instances running on a single hypervisor will
become partitioned based on which nova compute was used to create them.

In short, the system may appear to function superficially, but it's
unsupportable.

I had an alternative idea. The current assumption is that the `host`
managing a single hypervisor never changes. If we break that assumption,
we break Nova, so we could assert it at startup and refuse to start if
it's violated. I posted this VMware-specific POC:

https://review.openstack.org/#/c/154907/

However, I think I've had a better idea. Nova creates ComputeNode
objects for its current configuration at startup which, amongst other
things, are a map of host:hypervisor_hostname. We could assert when
creating a ComputeNode that hypervisor_hostname is not already
associated with a different host, and refuse to start if it is. We would
give an appropriate error message explaining that this is a
misconfiguration. This would prevent the user from hitting any of the
associated problems, including the deletion of all their instances.

We can still do active/passive HA!

If we configure both nodes in the active/passive cluster identically,
including with the same value of `host`, I don't see why this shouldn't
work today. I don't even think the configuration is onerous. All we
would be doing is preventing the user from accidentally running a
misconfigured HA which leads to inconsistent state, and will eventually
require manual cleanup.

We would still have to be careful that we don't bring up both nova
computes simultaneously. The VMware driver, at least, has hardcoded
assumptions that it is the only writer in certain circumstances. That
problem would have to be handled separately, perhaps at the messaging layer.

Matt
-- 
Matthew Booth
Red Hat Engineering, Virtualisation Team

Phone: +442070094448 (UK)
GPG ID:  D33C3490
GPG FPR: 3733 612D 2D05 5458 8A8A 1600 3441 EA19 D33C 3490