|Ryan Lane||Aug 28, 2012 2:25 pm|
|Gabriel Hurley||Aug 28, 2012 2:38 pm|
|Troy Toman||Aug 28, 2012 2:58 pm|
|Michael Still||Aug 28, 2012 3:28 pm|
|Ryan Lane||Aug 28, 2012 3:50 pm|
|Ryan Lane||Aug 28, 2012 3:52 pm|
|Robert Collins||Aug 28, 2012 4:21 pm|
|Jay Pipes||Aug 28, 2012 11:31 pm|
|Ryan Lane||Aug 29, 2012 12:10 am|
|Tim Bell||Aug 29, 2012 1:57 am|
|stua...@hp.com||Aug 29, 2012 6:21 am|
|James E. Blair||Aug 29, 2012 8:24 am|
|Eoghan Glynn||Sep 4, 2012 1:53 am|
|Thierry Carrez||Sep 4, 2012 3:06 am|
|Subject:||Re: [Openstack] A plea from an OpenStack user|
|Date:||Aug 29, 2012 6:21:04 am|
I think its great that we're having this discussion.
In the hope that its informative, I'd like to give some info on issues we're looking at when moving our Glance deployment to Folsom. A lot of this is in common with Ryan, but there are a couple of twists due to our goals of maximization of uptime (ie we are hoping to do a rolling rather than lock-step upgrade) and decoupling upgrades. Also, I mention the case where you may have existing Glance clients which you don't control.
In our case the upgrade of the various components (swift/glance/nova etc) will be staggered rather than performed simultaneously. (For large organisations I imagine this may often be the case.) If we are to avoid downtime for Glance we must simultaneously support ids and uuids. These must be presented appropriately to Nova nodes which we must assume are moving targets -- ie will initially be on older code, but will upgrade to Folsom.
We have some ideas on how this may be possible but haven't worked through all the details yet (in particular the Nova database entries)... but there could be some coding for Nova/Glance and some deploys of the interoperable code before eventually switching to standard Folsom.
(Jay -- I don't think scripts are sufficient here?)
If Glance were publically available its not clear how the id/uuid change could be worked through gracefully where we didn't have control over the glance client. Ie the upgrade would break existing customers' clients which expected an id rather than a uuid.
I agree with everyone about testing upgrade paths between releases -- if we could include paths for interoperability/rolling upgrades and perhaps define an official order in which components (swift/nova etc) should be upgraded that would be great too.
I should mention that in the same way as Nova has a dependency on Glance Glance has a dependency on Swift ... so far that upgrade path seems absolutely painless -- kudos to the Swift folks.
On Wed, 29 Aug 2012, Tim Bell wrote:
We should not also forget the documentation for the migration process itself. We need to have stable documentation (such as perform software installation, run upgrade script and restart service) while allowing post-release migration bugs to be resolved (when migration problems are found in the field). Automated testing can find some problems but the flexibility of OpenStack configuration would make it likely that there are other scenarios not covered by the test suite.
Another item are to clearly identify the order in which services can be upgraded and what are the possibilities for pushing a small part of the cloud to a more recent version ohile maintaining the majority on the previous stable release.
This sort of feedback loop is exactly what I hope for from the interaction between the user and technical communities.
Since we're Essex based, we've not had to face this yet but Essex to Folsom will be a good chance for improvements to be included in the core code.
-----Original Message----- From: openstack-bounces+tim.bell=cern...@lists.launchpad.net [mailto:openstack-bounces+tim.bell=cern...@lists.launchpad.net] On Behalf Of Jay Pipes Sent: 29 August 2012 08:32 To: open...@lists.launchpad.net Subject: Re: [Openstack] A plea from an OpenStack user
Ryan, thank you for your excellent and detailed comments about problems you encountered during the upgrade process. This is precisely the kind of constructive feedback that is needed and desired.
Someone mentioned automated testing of upgrade paths. This is exactly what needs to happen. Hopefully the Tempest folks can work with the CI team in the G timeframe to incorporate upgrade path testing for the OpenStack components. It likely won't solve ALL the issues -- such as the poor LDAP port in Keystone Light -- but it will at least serve to highlight where the major issues are BEFORE folks run into them. It will also help identify those tricky things like the Glance issue below: Glance itself upgraded its data effectively, but failed to produce scripts to modify the Nova image database IDs at the same time.
Thanks again, -jay
On 08/28/2012 05:26 PM, Ryan Lane wrote:
Yesterday I spent the day finally upgrading my nova infrastructure from diablo to essex. I've upgraded from bexar to cactus, and cactus to diablo, and now diablo to essex. Every single upgrade is becoming more and more difficult. It's not getting easier, at all. Here's some of the issues I ran into:
1. Glance changed from using image numbers to uuids for images. Nova's reference to these weren't updated. There was no automated way to do so. I had to map the old values to the new values from glance's database then update them in nova. The mention of testing upgrade paths go 2. Instance hostnames are changed every single release. In bexar and cactus it was the ec2 style id. In diablo it was changed and hardcoded to instance-<ec2-style-id>. In essex it is hardcoded to the instance name; the instance's ID is configurable (with a default of instance-<ec2-style-id>, but it only affects the name used in virsh/the filesystem. I put a hack into diablo (thanks to Vish for that hack) to fix the naming convention as to not break our production deployment, but it only affected the hostnames in the database, instances in virsh and on the filesystem were still named instance-<ec2-style-id>, so I had to fix all libvirt definitions and rename a ton of files to fix this during this upgrade, since our naming convention is the ec2-style format. The hostname change still affected our deployment, though. It's hardcoded. I decided to simply switch hostnames to the instance name in production, since our hostnames are required to be unique globally; however, that changes how our puppet infrastructure works too, since the certname is by default based on fqdn (I changed this to use the ec2-style id). Small changes like this have giant rippling effects in infrastructures.
3. There used to be global groups in nova. In keystone there are no global groups. This makes performing actions on sets of instances across tenants incredibly difficult; for instance, I did an in-place ubuntu upgrade from lucid to precise on a compute node, and needed to reboot all instances on that host. There's no way to do that without database queries fed into a custom script. Also, I have to have a management user added to every single tenant and every single tenant-role.
4. Keystone's LDAP implementation in stable was broken. It returned no roles, many values were hardcoded, etc. The LDAP implementation in nova worked, and it looks like its code was simply ignored when auth was moved into keystone.
My plea is for the developers to think about how their changes are going to affect production deployments when upgrade time comes.
It's fine that glance changed its id structure, but the upgrade should have handled that. If a user needs to go into the database in their deployment to fix your change, it's broken.
The constant hardcoded hostname changes are totally unacceptable; if you change something like this it *must* be configurable, and there should be a warning that the default is changing.
The removal of global groups was a major usability killer for users. The removal of the global groups wasn't necessarily the problem, though. The problem is that there were no alternative management methods added. There's currently no reasonable way to manage the infrastructure.
I understand that bugs will crop up when a stable branch is released, but the LDAP implementation in keystone was missing basic functionality. Keystone simply doesn't work without roles. I believe this was likely due to the fact that the LDAP backend has basically no tests and that Keystone light was rushed in for this release. It's imperative that new required services at least handle the functionality they are replacing, when released.
That said, excluding the above issues, my upgrade went fairly smoothly and this release is *way* more stable and performs *way* better, so kudos to the community for that. Keep up the good work!