Proposal for new release schedule / users are not interested in releases that will not become LTS

[mcollina]

I took some time to create an small web app that render the Node.js Downloads data from the CDN, aggregating them by months:

https://nodedownloads.nodeland.dev/

(Code at https://github.com/mcollina/nodejs-download-stats. I will put this under a proper domain as soon as I get back).

The download numbers show that very few people are downloading the releases that are not going to become LTS. This is matched by anecdotal feedback from collaborators that bugs are not reported for releases that are not going to become LTS.

What people expect from LTS is our "maintenance" phase, and they also expect our "active" LTS to land new features regularly.

Therefore, I propose a simplification to our LTS process:

1. Stop doing all releases that are not going to become LTS
2. eliminate the "current" phase
3. increase the "active" LTS phase to 1 year.

For example, we would cut v22 in April 2024, which would stay "active" LTS until April 2025, when we will ship v23. v22 will go into "maintenance" LTS at that point.

This has the benefit of simplifying our messaging to:

• if you want new features, stay on active LTS
• if you want to avoid surprises, stay on maintenance LTS
• as a maintainer, you should add support to all releases (right now quite a few people skip non-LTS releases)

We should also consider an increase of the length of the maintenance phase, if possible by our dependencies.

(Thanks to @Qard for the review)

[aduh95]

That would mean that we can ship semver-major changes every year, instead of every 6 months. For example, today a full deprecation cycle can take a minimum of 1 year. With this proposal, the minimal would be 2 years.

It’d also mean more errors will slip through to end-users.

I’m not sure our current messaging is confusing, I’d argue that the graph you shared on the contrary shows that everyone understands actually quite well our release cycle.

[lifeisfoo]

The download numbers show that very few people are downloading the releases that are not going to become LTS

If you look at the first six months after each release you can see that the download rate is similar between LTS and non-LTS releases.
When a new LTS is released, the previous non-LTS release downloads decrease quickly - as expected.

The same pattern is visible here.

As a user, I use non-LTS versions to test my existing applications: in this way, the migration to the next LTS will be easier.

[BethGriggs]

Agree with @aduh95.

This is matched by anecdotal feedback from collaborators that bugs are not reported for releases that are not going to become LTS.

Wouldn't this proposal mean we'd potentially have semver major changes sitting on the main branch unreleased for up to almost 1 year? I feel that could become difficult, and would result in even less feedback (as we get very few users of nightly and rc builds) and therefore more surprises. I do think we still get value from the alternating LTS/Current releases as we have a chance to land riskier breaking changes in an odd release giving us a window before the project is committed to supporting it long-term.

Generally, I feel the benefits of adjusting the release policy and schedule need to be quite high for it to outweigh the disruption and unlearning required by consumers and end users. It would be good to frame what problem we're trying to solve for end users before jumping ahead to restructuring the policy and schedule.

[benjamingr]

For the change to make sense to me we need to address deprecation cycles and semver-major changes. If the release plan can be adjusted while keeping those at an as-good or better cadence -I'm for it.

[mcollina]

My understanding from the session we had at NodeConf.eu was that very little feedback is provided by the non-LTS releases, with bugs introduced there got reported way in the following release. From what we have seen recently, quite a few bugs are reported only until a release
is flagged as LTS (which then include a long cycle for updates).

This is the starting point of this discussion, and the download data seems to back it up, as odd releases have minimal downloads.

[Qard]

Yeah, our own data shows users basically never install odd releases, so they're kind of superfluous. We're just adding extra complication to the release cycle by doing them.

[BethGriggs]

I think we should take into account that download data will be made up of a significant number of automated builds and deployments which, based on recommendations, target LTS over current. There were 155,314 downloads of v21.1.0 on 8th November- I wouldn't consider that an insignificant number for one day even though there is a huge relative difference to LTS.

[benjamingr]

My understanding from the session we had at NodeConf.eu was that very little feedback is provided by the non-LTS releases,

We have a concrete recent counter-example with navigator

[targos]

We have data from production deployments

That's highly biased. We don't expect many people to use odd versions in production. It doesn't mean they won't use it in CI.
Also, open source libraries can run their tests on odd versions.

[gurgunday]

That's highly biased. We don't expect many people to use odd versions in production. It doesn't mean they won't use it in CI.

Ok, but what would be the difference between running CI on a more recent version of node that practically never gets deployed to production and just running it on main (or an active LTS in the proposed plan)?

Apart from maybe having more CI breakages that don't affect any real user, I can't see it

Even if it was granted that testing on a proper interim release is better, after a non-LTS release, the next one that would become the LTS is again a major bump, which has its own breaking changes and bugs (i.e. node 20)

So for instance I'm not sure if anything concrete got accomplished by testing on node 19

No opinions, just providing perspective

[Qard]

As far as I can tell odd versions are only ever used in CI because we've trained people to think people might actually be using those versions in production. But no one actually does so it's really just a slightly more stable equivalent of nightly for the use case--catch when new things in Node.js break the code--which the proposed model would equally cover.

[targos]

It's not the same as using a nightly. My default version currently is 21. I would not bother always updating to the latest nightly.