Can Ubuntu Server Roll Too?
Wow…I just realized how long it’s been since I did a blog post, so apologies for that first off. FWIW, it’s not that I haven’t had any good things to say or write about, it’s just that I haven’t made the time to sit down and type them out….I need a blog thought transfer device or something :-). Anyway, with all the talk about Ubuntu doing a rolling release, I’ve been thinking about how that would affect Ubuntu Server releases, and more importantly….could Ubuntu Server roll as well? In answering this question, I think it comes down to two main points of consideration (beyond what the client flavors would already have to consider).
How Would This Affect Ubuntu Server Users?
We have a lot of anecdotal data and some survey evidence that most Ubuntu Server users mainly deploy the LTS. I doubt this surprises people, given the support life for an LTS Ubuntu Server release is 5 years, versus only 18 months for a non-LTS Ubuntu Server release. Your average sysadmin is extremely risk adverse (for good reason), and thus wants to minimize any risk to unwanted change in his/her infrastructure. In fact, most production deployments also don’t even pull packages from the main archives, instead they mirror them internally to allow for control of exactly what and when updates and fixes roll out to internal client and/or server machines. Using a server operating system that requires you to upgrade every 18 months, to continue getting fixes and security updates, just doesn’t work in environments where the systems are expected to support 100s to 1000s of users for multiple years, often without significant downtime. With that said, I think there are valid uses of non-LTS releases of Ubuntu Server, with most falling into two main categories: Pre-Production Test/Dev or Start-Ups, with the reasons actually being the same. The non-LTS version is perfect for those looking to roll out products or solutions intended to be production ready in the future. These releases provide users a mechanism to continually test out what their product/solution will eventually look like in the LTS as the versions of the software they depend upon are updated along the way. That is, they’re not stuck having to develop against the old LTS and hope things don’t change too much in two years, or use some “feeder” OS, where there’s no guarantee the forked and backported enterprise version will behave the same or contain the same versions of the software they depend on. In both of these scenarios, the non-LTS is used because it’s fluid, and going to a rolling release only makes this easier…and a little better, I dare say. For one, if the release is rolling, there’s no huge release-to-release jump during your test/dev cycle, you just continue to accept updates when ready. In my opinion, this is actually easier in terms of rolling back as well, in that you have less parts moving all at once to roll back if needed. The second thing is that the process for getting a fix from upstream or a new feature is much less involved because there’s no SRU patch backporting, just the new release with the new stuff. Now admittedly, this also means the possibility for new bugs and/or regressions, however given these versions (or ones built subsequently) are destined to be in the next LTS anyway, the faster the bugs are found out and sorted, the better for the user in the long term. If your solution can’t handle the churn, you either don’t upgrade and accept the security risk, or you smoke test your solution with the new package versions in a duplicate environment. In either case, you’re not running in production, so in theory…a bug or regression shouldn’t be the end of the world. It’s also worth calling out that from a quality and support perspective, a rolling Ubuntu Server means Ubuntu developers and Canonical engineering staff who normally spend a lot of time doing SRUs on non-LTS Ubuntu Server releases, can now focus efforts on the Ubuntu Server LTS release….where we have a majority of users and deployments.
How Would This Affect Juju Users?
In terms of Juju, a move to a rolling release tremendously simplifies some things and mildly complicates others. From the point of view of a charm author, this makes life much easier. Instead of writing a charm to use a package in one release, then continuously duplicating and updating it to work with subsequent releases that have newer packages, you only maintain two charms…maximum of three if you want to include options for running code from upstream. The idea is that every charm in the collection would default to using packages from the latest Ubuntu Server LTS, with options to use the packages in the rolling release, and possibly an extra option to pull and deploy direct from upstream. We already do some of this now, but it varies from charm to charm…a rolling server policy would demand we make this mandatory for all accepted charms. The only place where the rules would be slighlty different, are in the Ubuntu Cloud Archives, where the packages don’t roll, instead new archive pockets are created for each OpenStack release. From a users perspective, a rolling release is good, yet is also complicated unless we help…and we will. In terms of the good, users will know every charmed service works and only have to decide between LTS and rolling as the deployment OS, where as now, they have to choose a release, then hope the charm has been updated to support that release. The reduction in charm-to-release complexity also allows us to do better testing of charms because we don’t have to test every charm against oneiric, precise, raring, “s”, etc, just precise and the rolling release….giving us more time to improve and deepen our test suites.
With all that said, a move to a rolling Ubuntu Server release for non-LTS also adds the danger of inconsistent package versions for a single service in a deployment. For example, you could deploy a solution with 5 instances of wordpress 3.5.1 running, we update the archive to wordpress 3.6, then you decide to add 3 more units, thus giving you a wordpress service of mixed versions….this is bad. So how do we solve this? It’s actually not that hard. First, we would need to ensure that Juju never automatically adds units to an existing service if there’s a mismatch in the version of binaries between the currently deployed instances and the new ones about to be deployed. If Juju detected the binary inconsistency, it would need to return an error, optionally asking the user if he/she wanted it to upgrade the currently running instances to match the new binary versions. We could also add some sort of –I-know-what-I-am-doing option to give the freedom to those users who don’t care about having version mismatches. Secondly, we should ensure an existing deployment can always grow itself without requiring a service upgrade. My current thinking around this is that we’d create a package caching charm, that can be deployed against any existing Juju deployment. The idea is much like squid-deb-proxy (accept the cache never expires or renews), where the caching instance acts as the archive mirror for the other instances in the deployment, providing the same cached packages deployed in that given solution. The package cache should be ran in a separate instance with persistent storage, so that even if the service completely goes down, it can be restored with the same packages in the cache.
So…Can Ubuntu Server Roll?
I honestly think we can and should consider it, but I’d also like to hear the concerns of folks who think we shouldn’t.