On Thu, Jul 30, 2020 at 11:10:35PM +0300, Valter Jansons wrote: > On Thu, Jul 30, 2020 at 10:37 PM Julien Pivotto <[email protected]> > wrote: > > I'm with Lukas on this. 2.1 is a strong release, and we should be > > grateful for everyone which is using that release, as their feedback is > > valuable for the building the next releases of HAProxy. > > My apologies if the message sounded ungrateful, for rolling out new > changes and testing. As the latest 2.2.0 release did show just > recently, there is great benefit in people running upcoming (new) > changes.
No offense, don't worry :-) We're used to say that odd versions being maintained for less time, we're allowed to take more risks with them and we know that most of their users are those autonomous enough to roll back or switch to another one in case of trouble. As such, the stability of an odd version can be a bit more chaotic than the one of an even one, and that's by choice to preserve more users. Also I'm less reluctant to backport small features to odd versions than for even ones (it's a give and take, brave users test & report issues and in exchange they can get a version that better suits their needs). In other areas of the industry, the terms "early availability" and "general deployment" exist to designate these different stability statuses, and I think that they model what we do quite well. Of course when a new version is issued, it needs a little bit of time to dry up, and a few surprises are expected. But the point is that there should be (by design) less risks to upgrade from 2.1.x to 2.2.x than from 2.0.x to 2.1.x two months after the new major release is emitted. Here we're facing something unusual in that 2.1 appeared to be exceptionally stable and 2.2 started with some rough edges, so at this point of the cycle the difference in stability expectations might still be less visible of course. Anyway, the point of maintaining long term supported versions is that anyone is free to use the one that best suits their needs. Anything between the oldest that supports all needed features, to the latest stable enough for the use case is fine. As a rule of thumb, I'd say that it's probably OK to always be late by one or two stable versions on average. This should help one figure what branch to deploy: if the latest stable emits one version every two weeks, it means you need to upgrade your production every two to four weeks. If an older stable one produces one version every 6 months, it may allow you not to care about your prod for 6 months to one year. But in any case there is always the risk of a critical bug requiring an urgent deployment, so you should see this as a rule of thumb only and not a strict rule. Hoping this clarifies the process a bit. Willy

