As probably pretty much everyone at this point, I agree the tick-tock experiment isn't working as well as it should and that it's probably worth course correcting. I happen to have been thinking about this quite a bit already as it turns out so I'm going to share my reasoning and suggestion below, even though it's going to be pretty long, in the hope it can be useful (and if it isn't, so be it).
My current thinking is that a good cycle should accommodate 2 main constraints: 1) be useful for users 2) be realistic/limit friction on the development side and let me develop what I mean by both points slightly first. I think users mostly want 2 things out of the release schedule: they want a clearly labeled stable branch to know what they should run into production, and they want new features and improvements. Let me clarify that different users will want those 2 in different degrees and with variation over time, but I believe it's mainly some combination of those. On the development side, I don't think it's realistic to expect more than 2/3 branches/series to be supported at any one time (not going to argue that, let's call it a professional opinion). I also think accumulating new work for any meaningful length of time before releasing, as we used to do, is bad as it pushes devs to rush things to meet a given release deadline as they don't want to wait for the next one. This in turn impacts quality and creates unnecessary drama. It's also good imo to have a clear policy regarding where a given work can go (having to debate on each ticket on which branch it should go is a waste of dev time). With those "goals" in mind, I'll note that: - the fixed _and_ short cadence of tick-tock is imo very good, in particular in (but not limited to) avoiding the 'accumulate unreleased stuffs' problem. - we have ample evidence that stuffs don't get truly stable until they get only bug fixes for a few months. Which doesn't mean at all that we shouldn't continue to make progress on increasing the quality of new code btw. - Simple is also a great quality of a release cycle. I think we should try to define what's truly important to achieve (my opinion on that is above) and do the simplest thing that achieve that. This does imply the release cycle won't make the coffee, but that's alright, it probably shouldn't anyway. In light of all this, my suggesting for a release cycle woud be: - To have 3 branches: 'features', 'testing' and 'stable', with an X month rotation: 'features' becomes 'testing' after X months and then 'stable' after X more, before getting EOL X months later. - The feature branch gets everything. The testing branch only gets bug fixes. The stable branch only gets critical bug fixes. And imo, we should be very strict on this (I acknowledge that there is sometimes a bit of subjectivity on whether something is a bug or an improvement, and if it's critical or not, but I think it's not that hard to get consensus if we're all reasonable (though it might worth agreeing on some rough but written guideline upfront)). - We release on a short and fixed cadence of Y month(s) for both the feature and testing branch. For the stable branch, given that it already had X months of only bug fixes during the testing phase, one can hope critical fixes will be fairly rare, less than 1 per Y period on average). Further, it's supposed to be stable and fixes are supposed to be critical, so doing hot-fix releases probably makes the most sense (though it probably only work if we're indeed strict on what is considered critical). And that's about it. I think it would believably achieve stability (with a clear label on which releases are stable), but also provide new features and improvements quickly for those that wants that. The testing phase is imo a necessary intermediate step to get the stable one. On thing to define is X and Y. For Y (the cadence of feature/testing), I think 1 or 2 months are the only options that make sense (less than 1 month is too fast, and more than 2 months is imo starting to get too long). For X, that's more debatable but it's open-source and we should recognize volunteers generally don't want to maintain things for too long either. My 2 is that 6 or 8 months are probably the best options here. We'd also have to put some numbering scheme on top of that, but that's not really the important part (the meaning is in the branch labels, not the numbers). To give just one possible option (and assuming X=6, Y=1), in January 2017 we could cut 4.0 as the start of both 'feature' and 'testing'. We'd then have 4.1, 4.2, ... on the 'feature' branch, and 4.0.1, 4.0.2, ... on the testing branch for the next 6 months. In July, we'd switch from 4.5 to 5.0, with that becoming the new 'feature' and 'testing' base. At the same time, we'd cut 4.0.6 from 4.0.5 as the new 'stable' branch. Hot-fix on that stable branch would be versioned 4.0.6.1, 4.0.6.2 and so on. Of course, there can be variations on all this. We could have a different X and/or Y for the different branches to name just one. I'm just trying to abide by the "simpler is better" rule though. Lastly, I'll note that the versioning wouldn't use semver. I like semver alright but I think it's more suited to libraries than something like Cassandra. Amongst other things, it strongly rely on a fine grained notion of backward compatibility which I think is ill-defined for Cassandra in general. Sylvain On Fri, Sep 16, 2016 at 4:33 AM, Mick Semb Wever <m...@thelastpickle.com> wrote: > Totally agree with all the frustrations felt by Jon here. > > > TL;DR > Here's a proposal for 4.0 and beyond: that is puts together the comments > from Benedict, Jon, Tyler, Jeremy, and Ed; > > - keep bimonthly feature releases, > - revert from tick-tock to SemVer numbering scheme, > - during the release vote also vote on the quality label (feature branches > start with a 'Alpha' and the first patch release as 'Beta'), > - accept that every feature release isn't by default initially supported, > and its branch might never be, > - maintain 3 'GA' branches at any one time, > - accept that it's not going to be the oldest GA branches that necessarily > reach EOL first. > > > Background and rationale… > > IMO the problem with Tick-Tock is that it introduces two separate concepts: > - incremental development, and > - limiting patch releases. > > The first concept: having bimonthly tocks; made C* development more > incremental. A needed improvement. > No coincidence, at the same time as tick-tock was introduced, there was > also a lot of effort being put into testing and a QA framework. > From this we've seen a lot of fantastic features incrementally added to C*! > > The second concept: having bimonthly ticks; limited C* to having only one > patch release per tock release. > The only real benefit to this was to reduce the effort involved in > maintenance, required because of the more frequent tock releases. > The consequence is instability has gone bananas, as Jon clearly > demonstrates. Someone went and let the monkey out. > > A quick comparison of before to tick-tock: > > * Before tick-tock: against 6-12 months of development it took a > time-frame of 3-6 months and 6+ patch releases to stabilise C*. > > * After tick-tock: against 2 months of development we could have > expected the same time-frame of 3-6 months (because adoption is dictated by > users, not developers) and *over* this period 1-2 patch releases to > stabilise. It seemed to have been a fools errand to force this to 1 patch > release after only one month. It seems that the notion of incremental > development was applied for the developers where-as the waterfall model was > applied to QA in production for the users. (note: all this is not taking > into account advantages of incremental development, an improved QA > framework, and a move towards a stable-master.) > > The question remains to how many of these releases can the community afford > to support. And being realistic much of this effort relies upon the > commercial entities around the community. For example having 1 year of > support means having to support 6 feature releases, and there's probably > not the people power to do that. It also means that in effect any release > is actually only supported for 6-9 months, since it took 3-6 for it to get > to production-ready. > > A typical Apache release process is that each new major release gets voted > on as only 'Alpha' or 'Beta'. As patch releases are made it is ascertained > whether enough people are using it (eg in production) and the quality label > appropriately raised to either 'Beta' or 'GA'. The quality label can be > proposed in the vote or left to be voted upon by everyone. The quality > label is itself not part of the version number, so that the version number > can follow strict SemVer. > > Then the community can say, for example, it supports 3 'GA' branches. This > permits some major releases to never make it to GA, and others to hang over > for a bit longer. It's something that the community gets a feel for by > appreciating the users and actors around it. The number of branches > supported depends on what the community can sustain (including the new > non-GA branches). The community also becomes a bit more honest about the > quality of x.y.0 releases. > > The proposal is an example that embraces incremental development and the > release-often mentality, while keeping a realistic and flexible approach to > how many branches can be supported. The cost of supporting branches is > still very real, and pushing for a stable master means no feature branch is > cut without passing everything in the QA framework and 100% belief that it > can be put into a user's production. That is there's not a return to > thinking about feature branches as a place for ongoing stabilisation > efforts, just because they have a 'Alpha/Beta' label. The onus of work is > put upon the developer having to maintain branches for features targeted > for master, and not on the community having to stabilise and support > feature branches. > > BTW has anyone figured out whether it's the tick or the tock that > represents the feature release?? I probably got it wrong here :-) > > > ~mck >