[openstack-dev] [nova] Austin summit unconference sessions recap
mriedem at linux.vnet.ibm.com
Mon May 2 21:09:42 UTC 2016
The Nova team held unconference sessions on Wednesday and Thursday after
the lunch break. An unconference session is a 40 minute session broken
into 10 minute talks/proposals for a blueprint/spec/feature idea. It's
for things that need some discussion but don't warrant a full session,
generally because they are very new.
Unfortunately I wasn't able to attend either session due to conflicts,
but Dan Smith graciously took meticulous notes.
1. Rob Crittenden discussed getting and injecting credentials when
booting an instance . This is proposing a new driver that gets and
injects a credential into an instance when it's created. It's meant to
replace some existing code that's using the hooks mechanism, which was
deprecated in Mitaka.
There are some other specs related to this, like instance users  and
calling an external REST API for getting vendor metadata .
Ultimately we ran out of time to discuss all of this so it has to move
to the specs.
2. Randall Nortman talked about scoped policies for server groups .
The basic idea is that server groups let you define compute host
affinity/anti-affinity and Randall's spec proposes to expand that to
more than just compute hosts.
There was interest in the idea but it gets complicated because you can't
currently add/remove members from server groups, which would be a
pre-requisite of this work. Randall has a spec up for solving that 
and there has been work and specs proposed in the past from people at
IBM to add support for adding/removing server group members.
This also came up in the Friday meetup session. Andrew Laski pointed out
that server groups in their current form are somewhat useless since they
can't be updated, so we should fix that to make them useful (which is
complicated for the scheduler), or remove them. Jay Pipes has an idea
for potentially replacing server groups with server tags, but we didn't
really get into details.
3. Discussion was had on a spec proposed to attach/detach filesystem
shares created by Manila to Nova instances . The idea is to have the
hypervisor connect to share-level resources to pass the filesystem
through to the guest using something like 9p or virtio-vsock. There are
concerns from the Nova team on complexity, stability and security of the
proposed solutions. Further discussion was moved to the spec.
4. The EMC ScaleIO team discussed a spec for supporting ephemeral
storage backed by ScaleIO . For the most part this looks like
ceph-backed ephemeral storage to Nova. However, concerns were raised
about the 8gb block size limitation which will be discussed in the spec.
Also, it needs to come after Matthew Booth refactors the imagebackend
code in the Nova libvirt driver, which is part of a larger effort to
support libvirt storage pools for migration.
1. Tom Patzig discussed a spec for disabling local/ephemeral disk in
flavors . The Nova API would fail fast if you used one of these
flavors and didn't provide a bootable volume. There is general agreement
in adding this feature, the details just need to be worked out in the spec.
2. Mix and match resource federation was discussed . There are two
alternative specs proposed for this. This has come up briefly over the
last few summits, in hallway talk between sessions or during
unconference. The Nova team is generally uncomfortable with this given
the existing level of complexity between Nova and a single Cinder
endpoint, adding more Cinders that Nova has to share state with will
compound that problem. But the team is open to discussing this, so that
will happen in the specs.
3. Specs were discussed for direct download of images from the RBD
glance backend  and direct upload of images to the RBD glance
backend . The consensus was that the direct download spec makes
sense so we should move forward on that. As for the direct upload,
that's less clear since it bypasses glance, so it requires more
discussion. That will move to the spec.
4. Chris Friesen talked about a spec to improve debugging scheduler
failures . There is interest in this from the operator community
since it's related to the dreaded NoValidHost problem. It's mostly a
two-part change. The first is an audit of the existing log messages when
a scheduler filter skips a host to make sure they are useful enough to
know why the host was skipped. The second part is batching the messages
(in case debug logging is not enabled) so they can be dumped on failure
(NoValidHost). There are implementation details about how operators want
to use this, and that discussion has moved to the spec.
More information about the OpenStack-dev