Why we decided for and against Ubuntu Core
nitrokey.com
nitrokey.com
Bare metal management really feels like an unsolved problem. Whilst everybody working with cloud environments is whisked away by the latest shiny tools like Docker and Ansible, those of us working with bare metal are still trying to find a way to keep machines up and running with an OS that doesn't get corrupted from unexpected poweroffs or permanently cut itself off from the network because of a bad config.
The only existing candidate I've seen is Balena, but it only supports specific hardware and the cost is probably so high that we wouldn't be making a profit if we went with it.
At my current employer we are building a custom flavor of Ubuntu and provision it with Puppet, but we still get failures, and it's far from the immutable haven that DevOps guys would be used to.
I think it would be a battle to convince my colleagues and managers to try NixOS precisely because of the learning curve and lack of experts in the hiring pool out there.
It manages a fleet of physical servers and provides an API to provision servers, configure networks, and install operating systems. Basically an easier alternative to OpenStack for bare metal servers.
It supports ipmi and a bunch of other BMC's, uses pxe to deploy the operating systems, optionally runs its own DNS and DHCP servers and integrated with stuff like SMART to verify server health before deploying operating systems.
I use it in my research group to manage a fleet of about two dozen servers.
- Network topology management, IP address management (IPAM) and host associations (DNS, DHCP services or Static IP assignment)
- IPMI to power on/off machine and change boot orders, serial consoles
- Disk configuration (MD raid sets, LVM configuration, partitioning)
- Firmware of all the components
- Burn-in testing
- Passing configuration to the host OS for application/OS configuration
- Detecting hardware failures and tracking response.
- Install the OS
MaaS does most of these things, but it's all the life of bare metal. So "just a tool to install a chosen OS" undersells the complexity in this space.
There's even a terraform provider in development.
It does require a lot of planning though.
But the company has been great to work with and super helpful in slack.
Ubuntu OS is now stinky doodoo, which is a shame as it used to be the cat's pajamas for ease of use. Snapd is a debacle.
I think there's an OpenStack project for provisioning bare metal servers via an API from images; I wonder how that's doing nowadays.
Puppet is pretty good for configuration management when your systems do actually require the occasional change instead of just being continuously redeployed. Maybe there are people who rebuild their database servers every hour via CI, but I'd rather not.
"Immutable infra" is definitely not the default state of things in DevOps land either. Often people talking about how their infra is immutable just conveniently ignore the parts that aren't. The data has to live somewhere.
I'm guessing there's more to this story than you've summarised because those points are pretty easily solved with:
- UPS (if the power outs are that much of a problem then you might need to invest in a generator as well).
- iLo / IPMI (remote management). Though even just running a serial cable out the back of the server is good enough for a remote console in the event of a network failure.
As for managing the config of them, the usual tools like Ansible and Puppet work just as well (in some cases actually better since they were initially designed for on-prem hardware). Likewise for Docker. So don't think you can't run those tools on bare metal Linux. But if you don't want the containerisation-like aspects of Docker but still wanted the deployment tools then you can go a long way with git and shell scripts.
While DevOps really came into popularity with cloud hosting, there's nothing fundamentally new about a lot of the tooling that wasn't possible in the old days of bare metal UNIX and Linux. Us older sysadmins were still doing a lot of the same stuff back then too, we just didn't given it trendy names back then.
10+ years ago I was installing Linux remotely on bare metal hardware via a console server. Then once I had one machine installed I'd take one HDD out of the mirrored RAID array and plug it in the 2nd server so the RAID controller would clone. I could repeat the disk swapping as many times as I wanted as long as I remembered to change the IPs (again, via remote serial console). And this was the lazy way of deploying a fleet of servers quickly. More professional places would push the install onto the server via netboot cloning solutions.
Puppet works okayish but when you have a large fleet of what are basically IoT devices, you start getting an unpleasant failure rate. Push one bad network config and you've 'bricked' thousands of machines.
I am being a bit vague intentionally, hope you can understand. But to help imagine, the scale of our problem is pretty big, you've probably even had some interaction with one of our machines.
Hopefully it makes some sense!
With regards to bricking thousands of machines, with great power comes great responsibility. Is there no way to test the config before mass deployment? In batches or a canary deployment?
I got _super_ lucky at the startup where I was responsible for OS/updates/security for our IoT devices. Between the design stage and the production run, the price of 4GB SD cards dropped below the price of the 2GB cards on our BOM, so I had an entire spare partition to play with where I could keep a "spare" copy of the entire device image. And we had a "watchdog" microprocessor that could switch the main processor's boot config if it failed to boot. (We were basically running a RaspberryPi and an Arduino connected together. The prototype were exactly that, the final hardware was an iMX233 and an Atmel328 on our own custom board.)
We used Arch Linux with our own pacman repo, so the devices all pulled their own updates automatically. (Also it was super low risk, these were xmas tree lights, so our problem was "we don't want to ruin anyone's xmas!" instead of "If we fuck this up people might go broke and/or die...")
For a household-name IoT device, we did about 10000+ hours worth of testing on an array of devices for every update candidate. Think a walk-in closet chock full of devices covering every surface. This included thousands of hours of pure power-interruption scenarios, all automated.
We had a "alpha" and "beta" branch for internal users for ~a month before updates hit customers. If we bricked a device we could replace it.
For all update channels, we set a percentage of devices to deterministically receive updates. We start with 1% (probably 0.0001% now...) and double the rollout every day or so if things look good.
We had the ability to "roll back" to the last good firmware, which is stored on a partition on the device. This is almost never used.
Devices update in the background, unobtrusively. Once the update is complete, the device waits for a quiet window in which to reboot and try out the new firmware. Boot-time tests and a custom watchdog monitor the device to make sure everything works, including networking, filesystem, all services start up normally. If there is an anomaly, the device reboots back to the previous firmware, and this is reported home.
After some period of stability (minutes to hours usually) the device marks the update as good and will keep using it. If the device crash loops three times in a row within some window, it reverts to the previous version.
Yes, losing power sucks, and yes devices will get bricked and the filesystem will be corrupt, but you can do lots to ensure this is minimized. Like redundant partitions, leaving ample room in flash storage for bad blocks to prolong the device life, tuning log storage to avoid wasting write cycles and losing data on power loss, lots and lots of metrics and lots and lots and lots of automated integration tests.
If you're really struggling, I can help. Lots of fun problem solving in this space.
That's actually 90% of the solution these days: Use Mender. It wasn't around when I did this, but we likely would've used it, as we built essentially the same thing.
The other 90% is quality control. Distros can't really solve that for you.
> To be fair, to qualify as a solved problem in the industry one would expect either a standard, possibly community-supported, software implementation of the important parts, or at least documentation such that others could read up, learn about your wins, and apply them consistently to their projects, or both.
These are solved problems. Mender exists and documents almost everything I mentioned. Google has published incredibly thorough technical documents describing how Chromecasts and Chromebooks update, and at least the latter solution is Open Source.
If you look, you'd see these problems are not novel, which is why I declared it "not a hard problem space." The prior art is tremendous.
10, 15 years ago? Not so much.
I hear this a lot, but do they really? As best I can tell they require writing everything from scratch. I went down this path just an exercise to see what end-users were having to deal with. Everything in ansible appears to be re-inventing the wheel.
Sure it's powerful, and it provides a robust framework to do everything, but out of the box it does NOTHING on its own. Want to update ilo(m)? Someone probably has a playbook somewhere you can build on, but you have to go find the playbook, and modify it to meet your needs. Same with updating an OS, or software package or X.
There's no "point at this server, scan the server, figure out the hardware model, os, application packages, and build a default list of things you might want to manage".
Cisco UCS handles a chunk of this for the physical hardware, but it has plenty of issues itself. Same with HPe and Synergy, and Dell/EMC with Openmanage (this may have been replaced, I honestly haven't had to deal with Dell recently).
Honestly it gets a bit frustrating when the response is always "there's this tool that can do it if you just spend 6 months customizing it to your environment!". I think what OP is asking for is some intelligence to automatically discover.
If Ansible can do that, I'm all ears, but I haven't figured out how.
If you want something that just works automatically and is free, well, I don't think you will find. Learning Ansible (or another configuration management tool) well will pay dividends through a lot of your career tho.
Ansible is not even an option for IoT because it doesn't scale well and requires a stable connection. Puppet's agent model works better for large fleets especially if you're crossing multiple gateways.
So if I've got 40 different models of servers, with 10 different models of switches, 8 different operating systems, you're telling me right now there's a playbook for every possible hardware configuration I could have that will update all of that to the latest firmware without significant work on my part? Where is this magical repo?
You will also be able to share as much of the deployment as you want/can between the different models/OSes. Templating configuration variables based on their characteristics.
What you're asking to is akin to "Why cannot I download an app that does what $PRODUCT does but for my business for free? Why do I need to write my own software?"
Having to hire someone to customize ansible is the exact opposite of what op asked for and just reaffirms my point: ansible does nothing without significant customization.
Yes. Ansible is a framework with plugins. Just like Django or Ruby on Rails.
So again: Ansible is not an answer for someone looking for a ready-made solution to managing bare metal, which was my entire premise. Telling people to just use Ansible or Puppet when they're asking for a solution to the problem is just barely more helpful then telling them you can do it with a bash script and SSH for a consulting fee. It's pedantic and misses the point entirely.
You can use Ansible to manage bare metal - I've done it at two different companies.
Trust me, it works great.
I'm not as big of a fan of Puppet but actually puppet also works great for on-prem systems given that's what it was originally designed for. If anything, Puppet makes less sense in the cloud than it does on bare metal.
I'm not disputing you have a complex problem but that just means you need to spend a little more time tuning your solution (not less time like you seem to assume).
And if you want my advice about how to approach a daunting build: break your problem down. First start with delivering easy systems which will have the least impact if it goes wrong. This is to get your confidence up in working with the tool. Then start picking the harder targets that will give you the most reward, so even if your project ends up unfinished you've still fixed the biggest problems in your org. Then work your way backwards until everything is fixed. After a while, some of the easier deliveries will become background jobs you can fit in between support queries or half day sprint tickets (depending on whether you Kanban or Sprint). Before you know it, you'll have everything automated and realise it was far less painful than it appeared before you'd started the project.
Disclaimer: I'm DevOps Manager who has transitioned several orgs to through this process :)
When you decide that your cattle is going to be pets and pets get individual care and feeding, you have to hire a lot of people to do the individual care and feeding.
I've done done management of ~2k Linux systems across 17 generations with Ansible. It is not a big deal. You enforce conventions so you no longer have 2k different servers but rather a feet of Betsys, a fleet of Franks, a dozen of Marshas and a couple of Jacks. And you do not touch the boot configuration because even in 2020 you do not need to touch the boot/network configuration.
Other people's Ansible playbooks are often nice, because people mostly only publish things that stick to "best practice". I'll often use these if the exist and seem well thought out (and can do what I need done).
For things that there isn't an obvious "good choice of existing playbook", I'll sometimes write a "proper" one (we've got a quite good one here that got used a lot when we were deploying a lot of fundamentalist similar Grails/Tomcat/Apache apps with a bunch of common dependancies), or I'll just resort to using Ansible to run remote commands over ssh. If I know how to do it on the server from the command line, it's trivial to do exactly the same thing using Ansible from my laptop (or our bastion or config host). The trick here is making sure you don't write Ansible/ssh that screws up if you run it twice - which mostly isn't too hard to avoid so long as you remember to do so.
> I think what OP is asking for is some intelligence to automatically discover. > If Ansible can do that, I'm all ears, but I haven't figured out how.
I suspect this is a difference in approaches. It's easier in the cloud/vm world, bit I still treat bare metal servers more like cattle than pets. If I get a problem like "OS that doesn't get corrupted from unexpected poweroffs or permanently cut itself off from the network because of a bad config." I just get the box reimaged to a known state, then run the Ansible "update and deploy/configure from scratch" on it.
I don't think trying to build something to "intellegently auto discover" and repair a corrupted OS from hard powerdowns or a totally botched configuration change is a good use of my time... I'll just stand up another server from scratch using tried/tested automation.
There is no really great deployment story. In the very best case, you have to connect a keyboard, a PxE LAN cable, change some things in the bios, and select the OS you want to clone. But creating the image in the first place takes the better part of a day. Another option is to maybe clone the SSD beforehand. It gets more complicated if you have to flash a certain bios, change settings, and so on. Ideally, we'd just want to connect one cable bundle and "pressure tank" the new system with OS and configuration in a couple of minutes.
This is for the PCs that we sell, the story for the office laptops we use is even worse. There are good tools in the Windows world for deployment, but they all seem geared for installations of 1000s of computers. What if the office you manage (on the side of your normal work) just has 20? There is little point setting up SCCM, WSUS, or newer stuff like Autopilot (which seems pretty cool, but I couldn't figure out how to install MS office with the user's license, or how to install an ERP from Microsoft themselves... that should be 1-click or 1 line of code if you offer such a solution).
What I'd really like is a mixture of Ansible or Puppet with a stupid simple monitoring GUI. Then I'd be able to say
choco install firefox,7zip,vlc
install office,erp
join-domain mycorp,$credentials
For what it's worth, I've had some success with Powershell for package management and domain management, and tools like Clonezilla / Norton Ghost for managing images on small to medium sized fleets of machines (again, both desktop and server). There are also a plethora tools that can interrogate what machines are on a given network, the software installed and their patch levels -- but most of them are not going to be free. However there definitely are alternative options to SCCM and WSUS if they're too "enterprisey" for your needs (I've used a few different ones but I'm afraid I can't recall the names of the more effective solutions in terms of ease of us and features vs license fee).
Somebody please correct me if this thing exists and is available, I'd love to be wrong!
However, the Docker-based strategy is far ahead.
CoreOS has indeed since been deprecated, however when Redhat acquired CoreOS in Jan 2018 they merged CoreOS and Project Atomic to create Redhat CoreOS. They also released Fedora CoreOS which shares some technologies with vanilla Fedora and Silverblue.
Shortly after the acquisition, Flatcar Container Linux was released which is an updated derivative of the original CoreOS.
Thanks again!
I do recommend Flatcar linux though, and strongly encourage everyone to check it out. Not as jazzed about the name, but what’s in a name?
However, you might find QBee relevant for the other stuff.
Edit: as said in another comment in this thread, there are now alternatives to CoreOS!
- Fedora CoreOS: https://docs.fedoraproject.org/en-US/fedora-coreos/
- Flatcar Container Linux: https://kinvolk.io/flatcar-container-linux/
"Silverblue is a variant of Fedora Workstation. It looks, feels and behaves like a regular desktop operating system, and the experience is similar to what you find with using a standard Fedora Workstation.
However, unlike other operating systems, Silverblue is immutable. This means that every installation is identical to every other installation of the same version. The operating system that is on disk is exactly the same from one machine to the next, and it never changes as it is used.
Silverblue’s immutable design is intended to make it more stable, less prone to bugs, and easier to test and develop. Finally, Silverblue’s immutable design also makes it an excellent platform for containerized applications as well as container-based software development. In each case, applications (apps) and containers are kept separate from the host system, improving stability and reliability.
Silverblue’s core technologies have some other helpful features. OS updates are fast and there’s no waiting around for them to install: just reboot as normal to start using the next version. With Silverblue, it is also possible to roll back to the previous version of the operating system, if something goes wrong."
Silverblue user here. Haven't tried CoreOS myself.
Also one thing to note is that CoreOS as well as IoT is bootstrappable with a configuration file via ignition(a tool for automation).
It is and isn't solved. It usually takes a lot of work or custom scripts. One of the best is the Nerves Project, and is what I use for IoT deploymens [0] or even simple cloud deployments.
Nerves is setup to run Elixir/Erlang, but it's really just a wrapper around buildroot and Elixir can start programs in any language desired with some work. One of the core authors wrote a tool called `fwup` for doing immutable updates on Linux [1]. The ability to do an A/B update and have the device do an automatic rollback if an update fails is crucial.
A year or so ago they changed the default boot process to still allow networking and remote connections to work even if the main application crashes. Surprisingly it's all done using Erlang tooling, AFAICT. There are still rough edges, like limited ipv6 support. You can still get devices failing from dead SD cards -- even if your system boots from a read-only partition as power outages during a write on any partition can effectively destroy the SD card, so skip the SD cards.
0: https://www.nerves-project.org/ 1: https://github.com/fwup-home/fwup
We ended up with a mostly hand rolled system based on kexec, grub chainloading, and deploying rootfs images as big tarballs. It works, but I really wish there was something we could have just taken off the shelf.
> matchbox is a service that matches bare-metal machines to profiles that PXE boot and provision clusters. Machines are matched by labels like MAC or UUID during PXE and profiles specify a kernel/initrd, iPXE config, and Ignition config.
Years ago I did this to "rapidly" provision a couple of thousand machines we bought for the stock exchange. You can do most of your testing locally in vagrants, even simulating the networks you need to provision.
You can go a step further and trigger api updates at the end of your ansible runs so that cobbler updates collins (https://tumblr.github.io/collins/) so that you can track inventorying, cm changes and the like. At one point in time we reimaged entire subnets when they lost power - and ran tests against the hardware to ensure the machines functioned post-provisioning.
As far as I know the boursa still uses this system.
Though, that's more of the traditional server management approach, if properly set up, you can have it worry about the network configuration and pushing it to clients.
To me, it beats the entire IoT pull setup that Canonical is pushing with Ubuntu Core and snaps.
Do they want all the "features" for free, I went to https://ubuntu.com/core and it says 10 year security update commitment, it doesn't say it's going to be free, how Canonical will make money?
Source: spending my teenage years on installing various distros on my dad's computer, simultaneously pissing him off and ensuring I'd never have a girlfriend
In my mind, I'm here wondering -- if you really need the features, spending your own dev time for self-maintenance has to cost more than $30k/year... what are they thinking?
> In addition, it became apparent that we had not selected sufficiently strict according to open source criteria. Assuming Canonical would eventually cease to exist or discontinue Ubuntu Core, it would be nearly impossible with Ubuntu Core for the open-source community to ensure that NextBox would continue to be usable in a meaningful way.
It's about more than the monetary costs over the coming couple of years. Canonical pulling a RedHat here would be much worse for Ubuntu Core users than it is/was for CentOS users.
We'll have a more clear form for that kind of application soon, so that we can streamline such requests, community or otherwise.
It once was a Debian that just worked, nowadays is some kind of a trap. As you start getting deeper you start to notice non-standard things that get in the way more than they should, besides being utterly non-standard.
Things i noticed so far:
- auto-update enabled by default. if I boot a vm it's going to be nearly unusable (can't install packages) because it's going to spend the first 30 minutes doing a full upgrade
- netplan -- not sure why that's there
- snaps. for everything. the last straw for me was realizing that gnome-calculator packaged as a snap. it took almost 20 seconds to show the f-ing calculator. every time an app is slow i suspect that's because it's packaged as a snap.
- doing weird stuff with motd. why?
At this point my next reinstall will be a good old Debian.
I initially disliked Flatpak, but I've come around to it recently in light of where Snaps seem to be heading.
The same goes for flatpak too unfortunately.
Currently I feel like AppImage, while kind of a hack, might just be the best solution.
The one drawback is that you need to use an ancient compiler, but for our purposes that ancient compiler supports C++17 so that is okay for the time being.
Appimage is super cool for being able to quickly test builds of stuff but for software you actually use its not great.
I have been using Fedora Silverblue for a few months now and using flatpak to install every gui tool and its been excellent. I use a tool called flatseal which lets me tighten or loosen permissions for apps based on what I need which has been awesome. I can just flat out disable networking on apps if I don't need the networked parts.
Not 100% true. See e.g.
https://github.com/probonopd/go-appimage/blob/master/src/app...
Since the AppImage format is always the same and not just a random .exe it is possible to build tools around them.
It's the repository that's locked; Snap only supports a single repository; the "Snap Store".
I can't see a single thing Snap has going for it when Flatpak is the same idea but better in every single way.
https://linuxmint-user-guide.readthedocs.io/en/latest/snap.h...
In reality, it's a pain in the backside which may not yet support a config option you want to use for the chosen backend.
These days my preference is to bypass the unnecessary abstraction and just use raw systemd-networkd instead.
When it came out in an LTS, I was impressed by being able to declaratively describe networking. And there's even a way to test a configuration with auto rollback.
These are features that I find great in thing like juniper routers.
But when I went online to see what people thought there just was annoyance.
It made no sense to me.
I don't see the advantage over networkmanager, honestly. And now there's another thing I have to learn and support, with no real benefit.
And for all the swearing i did when i add to change the packer conf, then the ansible conf, i do think netplan is in fact easier to understand, read and change than brctl/bridge-utils.
I hate snaps though.
Yeah you have to look the details for the specific kind of configuration, but reading the fine manual is the norm, not the exception.
And regarding me personally: yes I would and I do. It's not that bad.
I can't be bothered making a new packer image each time a new ubuntu image is up. And people to use USB keys to make their friend and family try Ubuntu (or you know, as a backup).
Also, because even if your point is wrong, you're kinda right: you can disable autoupdate in the seed file or during install. And also, minimal image are a good idea/good practice and won't ever take 30 minute even if your image is really old.
Snaps killed it for me. I moved my daily driver(s) on Fedora a few months after 20.04. It turns out, RPMfusion does have all the packages I want.
...but this Kubuntu system (my primary workstation) has been rock steady for the past ~5 years with no signs of problems. So as annoyed as I am, I gotta hand it to them for keeping things working smoothly.
I haven't tried Debian on desktop, but it is great for servers.
https://cdimage.debian.org/images/unofficial/non-free/images...
- auto-update enabled by default. if I boot a vm it's going to be nearly unusable (can't install packages) because it's going to spend the first 30 minutes doing a full upgrade
I just disabled it and moved on with my life as bugs like these in my experience get ignored forever when you report them.
Somehow nobody at canonical thinks that snap performance is an issue at all, he was genuinely surprised. Snap is just unusable, full stop. If everything takes like 10x to open i can not work with my system anymore. I'd rather work with windows or a chalkboard instead.
At the same time, snaps are a new packaging format, and when you change the layout of applications to include things such as restrictions or making things read-only, suddenly all kinds of things can go wrong, and some of these can cause major performance impact.
Two easy and real examples from the snap world: early on there was a bug where .pyc files would be out of date, and the filesystem was read-only. This meant every single time the application was opened Python would recompile the entire application and fail to write its cache files in every case. Major performance impact. That was fixed.
Another one: fontconfig cache changed its format, and as a side effect applications running could not make use of the one in the system and had to rebuild their own copy every time. Extreme performance impact. That was fixed.
And the list goes on. So the point is: snaps are not slow, because there's nothing fundamental happening there to make them slow. But snap applications can be slow, of course, potentially by orders of magnitude. These are bugs, and we fix them when we see them.
In my experience everything i opened had significantly longer startup times - especially for VS Code this was a problem for me. I can not explain why, i just experienced the symptoms as did a few other people i talked to.
So if it is just an app problem great, hopefully there will be a day when not every app with UI i try has that problem.
"In other words openSUSE MicroOS is an operating system you don't have to worry about. It's designed for but not limited to container hosts and edge devices.
...
- Read-only root filesystem to avoid accidental modifications of the OS
- Transactional Updates leveraging btrfs snapshots to apply updates without interfering with the running system
- health-checker to verify the OS is operational after updates. Automatically rolls back in case of trouble.
..."
This line made me somewhat sad. But it is possible that I am just being old and grumpy.
One number, Debian Buster for example is LTS until June 2024 + 2 yrs. Just to mention one strength, I don't tell you about the well-known ones such as large community, fully openness, among others.
I see the point about long term support being deceptive, and also the mixed messaging being confusing.
Its rock solid and you can depend on it being engineered well.
Snaps , docker, and various other solutions are symptoms of a problem not a solution. The problem is userland is in a complete mess in linux. Fix the kernel and userland together, and everything flows.
95% of those reasons are that it is not GPL licensed.