On 08/19/2016 04:20 AM, Stef Walter wrote:

On 18.08.2016 22:12, Josh Berkus wrote:

...

On 08/18/2016 02:24 AM, Stef Walter wrote:

...

Yesterday several of us (roughly the top contributors [1] to Cockpit) discussed removing the '0.x' from the Cockpit version number.

The plan is from here on out to just go with integer versions, starting with version '118' following version '0.117'.

Cockpit releases weekly. Some versions of Linux distributions closely track that weekly release. Other choose particular releases to include, based on features and other criteria.

Cockpit now has a stable javascript API and file layout for components [2], as well as backwards compatible transport protocol. I hope to post more documentation about these guarantees soon.

What about making a 1.0 release?

That's a good question. But let me turn it around. What would the benefit of 1.0 be?

Cockpit uses continuous integration and delivery. We deliver every week, and most of them are solid and stable. Certain Linux distributions track the weekly updates, others pick certain releases and avoid churn for those users.

So the numbers go up regularly. What would the benefit of having numbers like "1.110" or "1.202" really be?

I like the systemd or CoreOS model for a project like Cockpit. Cockpit integrates the system, updates steadily as the system changes, and isn't a library or API.

But if you have reasons not to do that, I'd love to hear them.

Well, one of the fundamental problems with going to a whole-integer release number is that it basically doesn't provide you with any opportunity to change your mind. Once you move to version 118, there's no way to revert to saying "Man, I really wish we had a way to say that *this* release was a better one for those distros that pick one and avoid churn". Because if you decided to go back to a major.minor.patch style release, it's going to have to start at 119.0.0 (or else rename the project to go back to 1.0.0, for those distros that don't have something like epoch to deal with it).

There are other, non-technical benefits to a more traditional version scheme. For example, there is a very human response to seeing anything move from 0.x to 1.x. The sense there is that the project has *arrived*. Furthermore, a major version number jump provides a clear notification to tech press that it's a good time to do a re-review and talk about it again. If you make the shift to whole-version numbers, you'll get exactly one more of these short of sending people out to actually talk to journalists (or writing really compelling content for distro release announcements, but those will end up promoting the distro, not the project).

We saw this with NetworkManager... when it finally moved to the 1.x series, all of the major news sites immediately went back and did brand new reviews and perceived it as finally "ready". It didn't matter that the project had been functional and stable for years before that; the move to 1.0 provided a significant opportunity for press and to let people see that they should come back and give it another look. Crazily enough, the number of cargo-cult "never use NetworkMangler(sic)" posts dropped enormously after the 1.0 release because people came back and tried it again.

I'd actually disagree with making Cockpit 1.0 at this particular moment. I think it has come a long way, is very stable and it is a great piece of software. However, I also think that the optimum 1.0 point would have been when the API and transport protocol went stable. That would have been a very clear delineation. So my thoughts right now are that Cockpit should wait a while longer until it's about to land some truly killer feature (maybe the Ansible Galaxy stuff that came up at Flock?) and declare *that* a 1.0 release.

Stef,

Cockpit is continuously developed, integrated, and delivered. We're able to do that with aggressive upstream testing and feedback. On busy days 10,000 of instances are started to test Cockpit changes.

Out of the 50 or so releases in the past year, there's been one or two releases in the last year that were broken in a major way.

So the goal of Cockpit is to be continuously developed, continuously integrated, and stable. Having versions like 1.1, 1.2, 1.3, 2.x doesn't tell that story at all.

So, I'm looking at this from an advocacy perspective.

We've never had a 1.0 release. We haven't really promoted cockpit outside of the Red Hat circle. If Cockpit is all the sudden version 118, people won't know what they're supposed to use ... and we lose an opportunity for promoting the project to external OSS communities and users.

There's also that the other project to use whole numbers for versions, and increase them frequently, has created a bad reputation from that (Firefox). So many users' experience with large version numbers is going to be negative.

I understand where you're coming from on this. However, we will pay a price in advocacy/adoption terms if we do something unexpected with version numbers. It means that we spend time explaining our version numbers to people instead of explaining Cockpit.

On 09/05/2016 12:08 PM, Stef Walter wrote:

On 22.08.2016 18:25, Josh Berkus wrote:

...

Stef,

...

Cockpit is continuously developed, integrated, and delivered. We're able to do that with aggressive upstream testing and feedback. On busy days 10,000 of instances are started to test Cockpit changes.

Out of the 50 or so releases in the past year, there's been one or two releases in the last year that were broken in a major way.

So the goal of Cockpit is to be continuously developed, continuously integrated, and stable. Having versions like 1.1, 1.2, 1.3, 2.x doesn't tell that story at all.

So, I'm looking at this from an advocacy perspective.

We had more discussion about this today in our IRC meeting.

...

We've never had a 1.0 release. We haven't really promoted cockpit outside of the Red Hat circle.

That has indeed been a sore point. We discussed how to help Cockpit into Ubuntu's repositories as a concrete step forward.

...

If Cockpit is all the sudden version 118, people won't know what they're supposed to use ... and we lose an opportunity for promoting the project to external OSS communities and users.

We can promote the fact that it's no longer 0.x and has stable API, plug-ability and so on.

...

There's also that the other project to use whole numbers for versions, and increase them frequently, has created a bad reputation from that (Firefox). So many users' experience with large version numbers is going to be negative.

Firefox is a *great* example of what we're trying to achieve here. Ambitious as it were ... some points:

1. First of all I use "Firefox" and expect it to be updated regularly ... I don't use "Firefox 3.8" ... ditto for Chrome.

   This is the same experience we want. Admins can use port 9090 to access an interface with the specific capabilities of that server. The version number of Cockpit itself is secondary and a diagnostic tool for when things don't function as expected..

That said, I don't disagree that this is what we want. But the version number even today is basically invisible to the *user*. It is interesting primarily to the administrator of the system (and even then, only insofar as it matches up to the feature announcement including the thing they wanted to see added).

I'd claim that this part of the argument is *irrelevant* to the discussion, because the exact nature of the version number doesn't matter as long as it is an index to a set of features (meaning that changing it from 0.x to something else would result in exactly zero change in this experience).

2. Firefox won't get to say: "Oh we now have version 2.x which breaks all the (web) apps you've written so far. Too bad." That's the same goal we'd like to have for Cockpit's functionality and plugins.

Actually, they've done *exactly* this with the change in plugin and extension APIs in Firefox between releases and it has caused *enormous* headaches for many people (for example, it completely broke FreeIPA's autoconfiguration plugin, making it much harder to configure a client). There are plenty of other examples of good extensions that suddenly got broken with basically no advance knowledge of the user. Of course, then the user goes and complains to the extension writer instead...

From the perspective of the user, versions like this are "minor updates" and when things break in a minor update, they get very angry.

As time goes on grow, there will be functionality added, and other parts deprecated or removed. There are independent protocol and javascript API versioning mechanisms in place in Cockpit to account for this.

But a major break to (python) 3.x just isn't part of the plan.

Right, but the best laid plans of mice and men oft go awry...

Consider the possibility that there might come a time when Cockpit would go HTTP 2.0-only, for example. That *would* mean a backwards-incompatible break with any browser that only supported HTTP 1.1 (for example).

Or consider a day in the future where JavaScript is no longer the gold standard in browser scripting and you perform a major rewrite.

I will grant you that these are exceptional circumstances... but not impossible ones (and I'm sure there are others not immediately coming to mind). Version numbers *do* provide a simple, programmatic way of indicating to other software that they may need to change to adapt to the new thing (without the maintainer of that software needing to actively be monitoring release announcements of Cockpit).

3. Various parts of Cockpit can have different version numbers and work together. Just like parts of GNOME, or parts of Fedora (think modularity cough) have different version numbers.

Sure, but the overall project should have a version number that is representative of the whole system. With semantic versioning, you can actually use this as a shorthand to express which portions of the project are stable vs. in-development or experimental, simply by setting the Major version to 0.

So you might have Cockpit 1.2.0 that contains: cockpit-bridge = 1.2.0 cockpit-selinux= 0.56.0

They could be released as the 1.2.0 set but a conservative deployment could easily have tools that would skip installing 0.x portions of it.

4. Cockpit has stable releases weekly. Not all users are forced to cycle through each one ... but this dilutes semantic versioning advantages significantly.

I do understand and respect the traditional "1.0" approach to software, but it's hard to make it work or apply in this case.

I'm not sure in what way it dilutes the semantic versioning at all. I think you skipped to a conclusion without presenting the evidence there.

As I mentioned above, if we version components independently, we can use semver to indicate stability. For the stable components, we increment the Minor release each week until we hit a case where something that was deprecated is removed, at which point we bump the major number for that component and the project as a whole and continue.

The way I see it working is that the individual components would see their own updates following semver and we would just use the highest-versioned of any of them as the overall project version (since that would guarantee that compatibility changes were propagated).

...

I understand where you're coming from on this. However, we will pay a price in advocacy/adoption terms if we do something unexpected with version numbers. It means that we spend time explaining our version numbers to people instead of explaining Cockpit.

The idea is not to explain version numbers at all. But rather "just use Cockpit". Cockpit version numbers are a diagnostic tool for when things don't behave as expected.

Forgive my bluntness, but that's naive. It works *kind of* for Firefox because the majority of users don't have specific feature requests; they just want to browse cat videos. However, for the set of users for which specific features matter (usually either extension writers or integration engineers for major sites like Netflix), this versioning scheme *is* problematic because they don't have any inherent clues when their stuff is going to stop working and their users will be upset. They are forced to either use humans to monitor the project changes (which is costly) or take their chances and try to fix things fast when they break (which has problems with user perception and also means their sustaining engineers are overworked and stressed).

Cockpit is very specifically addressing a different audience than Firefox: it's addressing the administrator of a complex system. This is a class of user that *does* have specific feature needs and *will* notice when things stop working (and will be very noisy about it).

(As an aside, it's also a class of people that are largely in opposition to the idea of constant new features and prefer LTS support for existing features as well, but let's talk about that in a different thread.)

In fact we've added some finishing touches on the diagnostic aspects of version numbers here:

1. https://github.com/cockpit-project/cockpit/pull/4964

2. And we'll probably put a bit of effort into the about dialog to show the various versions of the parts in play when it comes to diagnosing an issue.

3. Lastly we'll better document what's "stable" and what's "internal" before making this move.

This discussion has resulted in some good checks and forethought, but all in all, even after taking the discussion into account, the significant contributors to Cockpit have been in support of simply dropping the '0.x'.

I'm a little surprised that the conversation has focused so heavily on the Firefox and Chrome decision to go to rapid major versions rather than picking systemd which is a much closer analog to Cockpit; it's system software whose feature set is highly relevant to administrators. I don't have any information to share specifically right now, but I'd suggest that it might be useful to do at least a straw poll of real users and get their opinion on whether that versioning scheme works for them, irritates them or is irrelevant to them. That would be a much better guidepost for Cockpit I think than looking at the common-user case of the browsers.

Josh Berkus jberkus@redhat.com writes:

We've never had a 1.0 release. [...]

Yeah, and we couldn't really figure out what the criteria for a 1.0 release would be.

A stable API would be one, but few people should care about the Cockpit plugin API.

Being bug free enough to let people use it for real on real systems would be another, but we have achieved that some time ago already, IMO.

I am very happy with the simple integer versions, precisely because they don't have any meaning, so there is less to discuss... hopefully. :-)

On 06.09.2016 17:15, Josh Berkus wrote:

On 08/22/2016 08:11 AM, Stef Walter wrote:

...

So all that said, I think this is *still* something we want to do: have monotonously increasing version numbers.

OK.

There's one other technical issue: how to handle bugfixes and security updates?

One of the issues with "whole number" releases is that distros are likely to include a specific version. We'll need some way for packagers to pull in things like security updates, and the way packaging systems -- including containers -- work, that expects some kind of major/minor version number.

I realize that's not how Cockpit releases work, but we need an answer which will be acceptable to packagers.

The usual mechanism of package revisions would be in play here.

For example, lets look at system in Fedora:

229 == is an upstream version 229-13 == is a downstream in Fedora 24 (bugfixes, security fixes)

Or systemd in Debian:

231 == is an upstream version 231-4 == is a downstream in Debian unstable (bugfixes, etc.)

Hope that makes sense.

Cheers,

Stef