Yes. The testing should be done on users that opt in (likely less than 0.1% but that should be OK if it's done properly) and it should be rolled out by releasing an update to everyone.

These so-called "best practices" often aren't.

By relying on opt-in you won't test whether a feature actually works for the average user, so if you don't test it, you might roll something useless/confusing/actively harmful out to all of them instead of just a few.

Exactly. The recent Windows 10 update fiascos demonstrate this perfectly: they are opt-in (Insider program) and there were too few users experiencing the bugs to overcome the noise of those who weren't. You need a truly random sample.

So instead of having enough voluntary or paid testers to detect catastrophic, file deleting bugs in a new version, your solution is to just roll out the new changes to random users expecting a stable system?

Something went terrible wrong somewhere, that we got to this.

I think they updated too many parts of Windows at once, not too many computers at once. And I'm not certain that a delayed rollout would have fixed it. Is it fair that a random sampling of people are given the problem?

And that's OK. Being honest to your users is much more important that having the most efficient deploy system.

I, for one, as a matter of principle, opt-in to all user tracking systems that are disabled by default, and try to opt-out whenever they are enabled by default. In the case of firefox, it was never clear to me that the tracking/remote install capability was enabled by default.

I don't see the purpose. There seems to be an assumption that A/B testing is inherently bad, therefor should be opt-in.

A/B testing is necessary for stable rollouts, that much seems self evident.

So what is the evidence that A/B testing is harmful to users?

It isn't A/B testing that is bad, it is running and/or modifying stuff in the user's computer without their knowledge that is bad. If a site modifies some part of their layout or whatever isn't bad, but a browser installing addons or changing settings to see if it will crash is bad.

Well, it's not like they're hoping it will crash. Presumably they're hoping it won't.

Wouldn't it be more irresponsible to deploy to every user without first testing it out?

that's what the stable/experimental channels are for - a user who don't want to be tested on (such as a business user) can easily be sure that they are getting a stable release and their settings won't update under them.

Experimental/beta channels can have these changes pushed to it, since that's implied by the name, and those people who like living on the edge of new features do so knowingly.

Let the users control their destiny, let the users decide when. That's the hallmark of a trustworthy company.

Users who intentionally run beta / nightly are not normal users. How many of them run on cheap "desktop" computers they bought on sale at Target, do you think?

Wait, so running and modifying stuff on all users' computers is good, but a random sample isn't? That's what A/B testing deals with.

The discussion isn't "modify stuff" vs. "don't modify stuff". The discussion is "some users" vs. "all users".

And obviously if you have an auto-update mechanism you're modifying stuff on the user's computer. Chrome updates are near invisible on modern machines. That's a good thing. It keeps the browser evergreen and protects users. If everything were opt-in because "Google shouldn't modify stuff on my computer" that would be an absolute disaster for everyone. No way.

This is a technicality, users do not care about technicalities, they have their own expectations and if you are on -say- a stable branch with autoupdates enabled your expectation is that you'll get the latest version that everyone else is using.

A/B testing through that branch breaks that expectation (and no, mentioning it somewhere hidden in some EULA or equally obfuscated place does not mean people will not have the expectation - hell, even mentioning it right below the download wont make a difference).

Isn't that the entire point of the nightly and beta branches?

Those populations are not representative of the release population.