Sorry for asking an exceptionally end-user question, but what am I doing wrong?
I find hard time accepting this is the way it works; we have theoretical fix in place reasonably fast for security issues, but in reality noone access them unless we do something manual e.g. changing the input to $release-small.
Not like it’s important or anything for this specific case, but as a general question.
Or is the -small version is already suggested instead of full on server-ish/production things?
-small is much more about less beeing cached by hydra before the channel advances. So you potentially get less substitution from cache.nixos.org and therefore potentially have to build more locally by yourself.
So it depends on your package set whether you can use that without relevant problems.
Yes, there are more channel blockers for the regular channels, but that is more of a side effect.
And just to be clear, not every test or package failing is a channel blocker.
So users just have to test wheter -small works for them or not.
Building everything just takes time. The current limitation here is build capacity. But extending that aint that cheap unfortunately
Hopefully it will still happen at some point in the somewhat near future. But even then it still will take some time given that some packages depend on one another and can only be build once others have finished building and the like.
Either there is something broken somewhere in the pipeline, or we are caching combinations of all optional values like withGeoIP, withDebug etc (x all supported architectures?). Because I just changed the flake input to github:NixOS/nixpkgs/release-26.05 and (afaiu) told my system “I don’t care, build locally if you need it”. It didn’t take long to finish.
Anyway I am asking questions above my understanding level now. I think I’ll keep the new line to apply everything merged as fast as possible. Let’s see how it will turn out.
The docs explicitly disagree with you, see the small channels item in the upgrade docs. They are specifically for getting security updates through the pipeline.
The impact of less testing is quite minimal. AIUI it avoids testing a full boot, and it doesn’t do any of the GUI testing. For incremental headless server updates, it will almost certainly be fine.
I don’t think it’s just build capacity. No matter how much money you throw at the problem, there is a critical chain of dependencies that cannot be further parallelized.
Compiling that chain will likely still be long enough even with the fastestest infrastructure that it cannot be pushed through faster than commits come in.
NixOS struggles with this more than other distros because nix demands full rebuilds for the whole reverse tree; things like guix grafts could be a solution here, but this is a trade-off the nix world doesn’t want to accept for diamond dependency problem reasons.
NixOS’ problem is a mix of that, and the level of transparency of the build infrastructure. Nobody wants to involve us in embargoes, because we can’t push fixes before they’re announced anyway (and we have 99% drive-by contributors so there’s nobody to contact or collaborate with).
For most real users just following the -small channel is the correct solution, though.
More I think about this, more it becomes clear, thank you.
The PR only touched nginx, which causes nginx and anything depends on it to be rebuild. We cannot gradually switch either by saying “nginx itself is built, cache it and update on nixos-26.05 channel” all other upstream packages will say “wait nginx’ hash does not match, let me get the new one from cache and build everything else locally” (afaiu).
The docs don’t disagree with me. It are two parts of the same coin. Faster bumps means less builds can be done. So less caching. If you look at the quote i was replying to there it was stated (until it was edited) that the less testing is the important thing too consider.
So thanks for providing the link and more or less agreeing with me.
Please just read what i wrote after that. I specifically stated the dependency problem afterwards.
And while that is an important aspect, the fact remains that we are (according to the infra lead whose assessment of the situation i would trust) currently constrained in regards to build resources.
Worse; because nix embeds the full file path in the resulting binaries. Nothing would be rebuilt, you’d just continue to use the old version for packages that haven’t been updated.
Nix is explicitly designed to work this way, it’s the entire point of eelco’s original thesis. It’s so not every package needs to update when just one wants to use a newer version of a library. This is great when you need to run one package that still depends on openssl 1.1, but has the downside that you need these reverse-tree rebuilds if a low-level dependency changes.
You can think of every NixOS package as a hyper-efficient docker container. Just like containers, all NixOS packages need to be updated to use a new version of their dependency; it isn’t enough to update the alpine container most of them include as a layer (because you have to change the image config to do that, which means rebuilding the container).
In theory you could build a NixOS-style distro that allows partial upgrades, but it would be quite a bit more complex to use. It would also only help with security flaws that cannot be exploited through a dependant, so benefiting from such an OS would require deep analysis of all security issues, which doesn’t mesh well with our CVE-brained security ecosystem.
Right, sorry, I misread. I thought you were saying that the small channels weren’t there to get security patches out the door quicker, didn’t realize you were extending on the testing thing.
I still think the speed of the “critical chain” is improtant enough of a constraint that “currently” is overly optimistic regarding the build pipeline.
More build capacity is always nice, but I sincerely doubt it will ever get to a point where updates can be instant enough to avoid having this conversation every time someone new tries to patch a CVE.
Now I fully get it. Everything needs to be one solid snapshot. All reverse deps will be damned..
Also I’ll switch to -small, because what I was thinking “small doesn’t have all packages” but it’s not the case, it’s just a smaller subset of packages hydra requires/builds/tests before advancing the channel.