That's for those who can't pay for RHEL subscriptions and they have such critical experiments going on that getting updates earlier than RedHat screws them up?

They can surely opt out of updating automatically and only apply updates after reading the changelogs.

In many cases systems run out-of-tree drivers like Lustre, MOFED, NVIDIA etc. that tend to break due to kernel internal API changes. Yes, also for changes incorporated into minor RHEL releases like X.Y -> X.(Y+1), though not for kernel security updates within a minor X.Y release. At the same time they might need to keep up to date wrt security updates. So if CentOS stream implies that the kernel will randomly get an API-breaking update, with no security support for the previous version, that is not workable.

I always found that to be a little weird. Why is scientific projects that dependent on even minor version changes?

Isn’t that normally an indication that there’s something horribly wrong in your code?

The reason they don't want to do anything other than security updates, and sometimes not even that, is because they need consistency across calculations over time.

This might be a poor example, but lets say an error was found in a computation in a library, for example a rounding error or something, in that case, more often than not, the community will rather live with the known error and have consistency instead.

Ideally, yes, but in my experience it's more often due to software management processes being, well, non-existent, and whatever exact version that is 'required' is the version that the original developer used and no attention whatsoever has been given to portability or even testing on other versions.

Luckily containers have become quite popular in the HPC world, to isolate such 'tricky' applications from the host OS.

Won’t that make reproducing results difficult? I mean it seems like a result won’t be worth much if it’s only reproducable with a specific version of say glibc from five years ago.

Over the cause of one experiment sure, I can see why you’d like to know the calculation are done identically. I just question the validity/usefulness of a calculation that can only be repeated on one version of some library. Surely the overall result should remain the same.

There are always inconsistencies in the scientific world:

* https://en.wikipedia.org/wiki/Fudge_factor

A good portion of experimentalists' time is probably spent in trying to figure just how badly they've messed their experiments as compared to reality. Major discoveries are sometimes based on teasing apart those differences:

* https://en.wikipedia.org/wiki/Muon_g-2

In my EE studies we had a specific course just in examining the limitations of (digital) numbers:

* https://en.wikipedia.org/wiki/Numerical_analysis

* https://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.h...