I am using unattended-upgrades across multiple servers. I would like package updates to be rolled out gradually, either randomly or to a subset of test/staging machines first. Is there a way to do that for APT on Ubuntu?
An obvious option is to set some machines to update on Monday and the others to update on Wednesday, but that only gives me only weekly updates...
The goal of course is to avoid a Crowdstrike-like situation on my Ubuntu machines.
edit: For example. An updated openssh-server comes out. One fifth of the machines updates that day, another fifth updates the next day, and the rest updates 3 days later.
Ubuntu only does security updates, no? So that seems like a bad idea.
If you still want to do that, I guess you'd probably need to run your own package mirror, update that on Monday, and then point all the machines to use that in the sources.list and run unattended-upgrades on different days of the week.
Ubuntu only does security updates, no?
No, why do you think that?
run your own package mirror
I think you might be on to something here. I could probably do this with a package mirror, updating it daily and rotating the
staging,production, etc URLs to serve content as old as I want. This would require a bit of scripting but seems very configurable.Thanks for the idea! Can't believe I didn't think of that. It seems so obvious now, I wonder if someone already made it.
Using scheduling is not a good option IMO, it's both too slow (some machines will wait a week to upgrade) and too fast (significant part of machines will upgrade right away).
It seems that making APT mirrors at the cadence I want is the best solution, but thanks for the answer.
A cron job that runs when you want it to instead of the unattended updates metapackage.
unattended-upgrades can already do that actually, i e. you can configure the systemd timers. But that's insufficient for my needs. Using a mirror seems like the best option so far.
In an ideal world, there should be 3 separated environments of the same app/service:
devel → staging → production.Devel = playground, stagging = near identical to the production.
So you can test the updates before fixing production.
So you can test the updates before fixing production.
My question is how to do that with APT.
I think there is no a out-of-the-box solution.
You can run security updates manually, but it's too much to do.Try to host apt mirrors in different stages, with
unattended-updatestuned on.
Devel will have the latest.
Staging the latest positively tested on the devel.
Production the latest positively tested on the staging.Making multiple mirrors seems like the best solution. I will explore that route.
I was hoping there was something built into APT or unattended-upgrades, I vaguely remembered such a feature... what I was remembering was probably Phased Updates, but those are controlled by Ubuntu not by me, and roll out too fast.
Maybe you could switch to an image based distro which is easy to roll back and won't boot into a broken image.
Which distro is image based and have the staggered rollout feature I'm after?
You don't need the staggered rollout since it won't boot into a broken image and you can boot easily into an old one if you don't like the new one. E.g. fedora atomic.
I'm not up to date with vanilla os for the debian world if it is on par with fedora.
I am not worried about upgrades so bad that they literally don't boot. I am worried about all the possible problems that might break my service.
You also roll back package versions. I'm not sure what problems could arise.
I can roll back with APT too, my question is how to do the staggered rollout.
You have to reboot for an image update. Hence, you can update the computers at different times and days.
This doesn't seem to enhance my workflow at all. Seems I now would have to reboot, and I still need to find a separate tool to coordinate/stagger updates, like I do now. Or did I miss something?
If the os works always (atomic image based distro), and the docker container work, and both can roll back easily. What else could go wrong?
Don't overthink it :)
No, OP absolutely still need staggered rollout. Immutable distros are a blue-green deployment self-contained. Yet, all the instance can upgrade and switch all at once and break all of them. OP still need some rollout strategy externally to prevent the whole service being brought down.
Small number of machines?
Disable unattended-upgrades and use crontab to schedule this on the days of the week you want.
Eg, Monday each week at 4 am - every combination of dates and days is possible with crontab. 2nd Tuesdays in a month? No problem.
0 4 * * MON apt-get update && apt-get upgrade && reboot(You can also be more subtle by calling a script that does the above, and also does things like check whether a reboot is needed first)
Dozens, hundreds or thousands of machines? Use a scheduling automation system like Uyuni. That way you can put machines into System Groups and set patching schedule like that. And you can also define groups of machines, either ad-hoc or with System Groups, to do emergency patching like that day's openssh critical vuln by sending a remote command like the above to a batch at a time.
All of that is pretty normal SME/Enterprise sysadminning, so there's some good tools. I like Uyuni, but others have their preference.
However - Crowdstrike on Linux operates much like CS on Windows - they will push out updates, and you have little or no control over when or what. They aren't unique in this - pretty much every AV needs to be able to push updates to clients when new malware is detected. But! In the example of Crowdstrike breaking EL 9.4 a few months ago when it took exception to a new kernel and refused to boot, then yes, scheduled group patching would have minimised the damage. It did so for us, but we only have CS installed on a handful of Linux machines.
