I highly recommend this approach. I run ~all of my digital footprint on FreeBSD jails. After surveying how many jail managers have come and go in the last decade, I decided to just roll my own using a single shell script called jailctl [0].
Nothing fancy, just VNET jails based on ZFS templates (vanilla FreeBSD rootfs) and epair interfaces (which I truck to various VLANs on the host's egress interface).
One pattern that I've found useful is to give each jail a persistently delegated ZFS dataset called "data." This lets me reprovision the OS image for the jail without having to backup and restore its application data (such as a Postgres DB). It also allows each jail to manage its own ZFS snapshots.
The only thing that was a bit hairy was generating unique interface names and MAC addresses for each jail's VNET interface. My first instinct was to derive the interface name from the jail name, but interface names on FreeBSD are limited to 15 characters, and occasionally I'd hit this limit.
In the end I did some dark magic using md5 sums of the jail name / host interface MAC address. Kind of ugly but I really didn't want to introduce any dependencies besides /bin/sh.
I used to do something like this but with ZFS on an OpenSolaris / Illumos storage server, exporting copy-on-write clones of snapshots of iSCSI volumes to boot Xen VMs on neighbouring blades, with tagged VLANs from each host because there were multiple guests to launch with varying roles in the application cluster. We made the VLAN number match 12 bits of the IP addresses and numbered the clones similarly. It merely remained to create new snapshots every release and any dev could launch an bugfix/test/showcase/etc environment for that version, and connect to it via a VPN. I was always worried about scale if we hired more than 4096 developers but fortunately the company was acquired and its product discontinued before that happened.
That was in 2007 so the control plane (scheduler and automation) were built from scratch and we had very few reference points for the overall design. If I was building that today I’d probably still use ZFS clones but at filesystem level instead of block devices, and serve jails over NFS if I can get away with it, the iSCSI part was always a little janky.
I built a reverse proxy server for this kind of usecase specifically. I have a chroot launcher with a simple config file and the reverse proxy is auto configured over RPC and automatically acts as a CA root so internal communication is HTTPS. You can also set the API key as an env var and just run your binary on the command line for testing and it will auto configure and be available on your URL with https.
It's a bad time for me to be mentioning it because I have a major update that's not quite ready to release that changes some client APIs and makes the whole thing much nicer with fully automatic lets encrypt. I haven't had the space to work on it for a while unfortunately.
For those unfamiliar with FreeBSD, this is using base OS tools to manually create this type of immutable jail/container. This can be done with 'less effort' by using a jail manager.
Jail managers come and go. Base OS tools stay and are getting better and better. I would definitely stay away from ezjail as it us quite old, active development or even maintenance has stopped long time ago.
Author of the article seem to know what they are doing so I'm puzzled why they don't use `bsdinstall jail /path/to/jail` to implement basejail instead of manually unpacking archives.
No need for separate custom rc script to start `lo1`, it can be done with `cloned_interfaces` directive in rc.conf.
Updating and upgrading jails by passing `-b /path/to/jail` to `freebsd-update` works, but new recommended way has lately been `-j <jailname>`.
Cool article overall, the beauty of FreeBSD is also in possibility to do things in many different ways.
If you want to do this kind of thing but don’t like the looks of this indepth tutorial, I highly reccomend installing Bastille FreeBSD instead and let the bastille tooling do the heavy lifting of managing zpool cloning and what not. Their getting started tutorial should get you 90% of the way there and has Docker like tooling. [0]
Also found this self host on a Raspberry Pi helpful.[1]
The ezjail[0] port is another option for achieving the article's stated goal:
FreeBSD's native support for ZFS snapshots and jails
provides a powerful foundation for immutable deployments.
I have not used the article's tool(s) and am not comparing the functionality provided by each. I have used ezjail[0] and found it exceptionally useful for similar concerns.
I really don’t like jail managers because I need to learn a non-standard way to operate jails when I already know the manual approach. I just wouldn’t miss the boilerplate. Also, OCI containers > jail managers.
ZFS has been stable in FreeBSD for something like 17 years, and FreeBSD jails have been around for something like 25 years.
By the time Docker hit 1.0 (about 11 years ago), the use of snapshots and jails had already been normal parts of life in the FreeBSD space for over half of a decade.
I never find arguments like this compelling (but I agree with your sentiment). I don’t much care to fire up the time machine to go back 12 or more years to develop software today. If your argument is that ZFS and jails provide the same functionality but are more stable than Docker. But as is it comes off as “get off my lawn you young whipper snappers”.
But at the same time, the reason Docker won was not because it was groundbreaking tech or because it was amazingly well tested or anything. Just as one example, it has a years old bug which actively gets more comments every week having to do with Docker grossly mishandling quotes in env files.
No, the reason it won is because the development experience and the deploy experience is easy, especially when you are on Linux AND on macOS. I can’t run FreeBSD jails or ZFS on macOS, can I? Definitely not with one file and one command.
Jails and ZFS are amazing tech but they are not accessible. Docker made simple things very simple while being cross-platform enough. Do I feel gross using it? Yeah. It’s a kludgy solution to the problem. But it gets the job done and is supported by every provider out there. I am excited that it is being ported to FreeBSD though I know it will be a very long process.
> especially when you are on Linux AND on macOS. I can’t run FreeBSD jails or ZFS on macOS, can I? Definitely not with one file and one command.
On macOS, docker actually launches a Linux VM to run containers. If this counts, then yes, you can run FreeBSD jails or zfs on macOS, by running a FreeBSD VM.
But it works with one command and one (Docker)file. That’s what I mean by Docker being a kludgy solution: this is way less than ideal. But for developer experience this is very nice. And that same Dockerfile runs on everything from AWS, to GCP, to k8s, to Dokku, etc.
I dislike the implementation but I cannot deny that the UX is good enough to be very popular.
Yeah I would love to use FreeBSD jails with ZFS and everything, it’s just that the whole cloud and containerization thing happened based on Linux and FreeBSD just never made it into that ecosystem.
You’ll be sacrificing a lot and have to hand-roll a lot if you want your organization to switch from Linux+docker to FreeBSD+jails
It's all just history now for all I know, but there was work in the past to make Linux containers work on a Solaris fork (SmartOS, specifically) by emulating the Linux syscall table and presenting that to the containers. Joyent did work on this (alas, and there's an excellent and entertaining talk from Bryan Cantrill[1] that goes over it.
I imagine FreeBSD could do something similar if they aren't already. IIRC FreeBSD has a Linux emulation layer (but I don't know how much attention it still gets), and it's had containerization primitives longer than linux, so some amount of filling in the gaps in containerization features and syscall tables (if needed) could possibly yield an OCI compatibility later (for all I know all this already exists).
The problem, and the reason if this doesn't exist why people probably weren't as interested in doing the work, is it would always be "mostly" compatible and working and there would be no guarantee that the underlying software wouldn't exhibit bugs or weird behavior from small behavior differences in the platform when emulating something else. Why open yourself up to the headache when you can just run Linux with containers or build what you want on FreeBSD with jails and their own native containerization primitives.
Some kernels are more similar to others, some are less. Turns out NT is less similar to Linux than required for good performance. I wouldn’t be surprised if Solaris was similar enough given that Linux tries to be Unix-like and Solaris is actually Unix.
In my opinion this is the path forward. I can already imagine Hashicorp Nomad orchestrator, with the podman driver, running fleets of FreeBSD containers.
I think there's FreeBSD images for all the clouds now.
You would need to do more work yourself to fetch and run jails probably, and I don't know if there's a hosted repository of 'jail images', but in return, you'd probably have a nicer system (at least, I'd like such a system more than running containers on google container optimized linux)
You can always upload your own, it's pretty simple doing so in a reproducible manner using something like Packer, but even without it you can just boot a VM into a rescue system, write your OS of choice to the VM disk and reboot.
Docker was the first viable containerization technology on Linux. Despite the 15 year late start vs FreeBSD Jails, it's certainly winning by the numbers.
But that has nothing to do with their respective UXs. It's a Linux vs FreeBSD signal.
> Docker was the first viable containerization technology on Linux.
No it wasn’t. Docker was late to the party even for Linux (and Linux was late compared to every other “UNIX”).
OpenVZ was around for years before docker. Its main issue was that it required out-of-tree kernel code. But there were some distributions that did still ship OpenVZ support. In fact it’s what Proxmox originally used. And it worked very well.
Then LXC came along. Fun fact, Docker was originally used LXC itself. But obviously that was a long time ago too.
I’ve used both OpenVZ and LXC in production systems before Docker came along. But I’ve always preferred FreeBSD Jails + ZFS to anything Linux has offered.
Not really, OpenVZ was/is really quite good, it was just hampered by the fact it requires out of tree modules. LXC has also always been very usable (Docker even used it for several years) but it was IMHO too focused on the VM-like management scheme that Zones and Jails had.
Docker's killer selling point was that it solved a very common and specific developer problem, not that it provided operational improvements over the state of the art on Linux. From an operational perspective, Docker has generally been a downgrade compared to LXC. (I say this as a maintainer of runc, the container runtime that underpins Docker, and as someone who worked a lot on Docker back in the early days and somewhat less today.)
I think it also helps that Docker started with an enormous amount of VC funding because their promise was to bring the “App Store” to Linux and enterprise servers.
Who couldn’t become famous with something like a $200M budget?
Feel like they spent it on marketing instead.
Podman is arguably technically superior yet people stay with Docker out of habit…
Docker engine is one thing, but access to docker hub without rate limits is what people actually pay for if they’re too cheap to host their own proxy registry (which everyone except the smallest companies should regardless).
One of Docker's big advantages is its client/server architecture made it easier to run on all those macOS and Windows boxes by just spinning up a VM for the server side, then making that VM fully managed to where a large percent of users don't even know it's there.
You can't use Docker on Mac or FreeBSD. Now, I am not calling you a liar, you _can_ use Docker on Mac and FreeBSD. But you would probably only want to do so for development, as Docker on Mac and FreeBSD requires running a Linux VM which is the thing which _actually_ runs the containers.
There is work ongoing to try to make this more native on FreeBSD (by using Linux jails) but that work is not complete yet.
So, if you want to get the same kind of experience as Docker on FreeBSD, you are forced to use jails.
The only reason Docker seems accessible is because it's native to the platform people seem to like for running all their services, but if you're dealing with FreeBSD, you most certainly would not just "use Docker" to deploy your stuff. Because you would get worse performance than if you had just used Linux.
So the answer to "Isn't this just Docker with extra steps?" is truly and absolutely "No". Not because of some kind of old man shouting at cloud argument, but because if you are on FreeBSD (for whatever reason that might be) you can't just use Docker as an easier replacement for Jails (at least right now).
I have to imagine systemd’s nspawn with btrfs integration took much inspiration. Combined with systemd’s service configuration it really makes a wonderful way of running distroless, immutable containers.
I highly recommend this approach. I run ~all of my digital footprint on FreeBSD jails. After surveying how many jail managers have come and go in the last decade, I decided to just roll my own using a single shell script called jailctl [0].
Nothing fancy, just VNET jails based on ZFS templates (vanilla FreeBSD rootfs) and epair interfaces (which I truck to various VLANs on the host's egress interface).
One pattern that I've found useful is to give each jail a persistently delegated ZFS dataset called "data." This lets me reprovision the OS image for the jail without having to backup and restore its application data (such as a Postgres DB). It also allows each jail to manage its own ZFS snapshots.
The only thing that was a bit hairy was generating unique interface names and MAC addresses for each jail's VNET interface. My first instinct was to derive the interface name from the jail name, but interface names on FreeBSD are limited to 15 characters, and occasionally I'd hit this limit.
In the end I did some dark magic using md5 sums of the jail name / host interface MAC address. Kind of ugly but I really didn't want to introduce any dependencies besides /bin/sh.
[0] https://github.com/cullumsmith/infrastructure/blob/master/fi...
Very neat!
I used to do something like this but with ZFS on an OpenSolaris / Illumos storage server, exporting copy-on-write clones of snapshots of iSCSI volumes to boot Xen VMs on neighbouring blades, with tagged VLANs from each host because there were multiple guests to launch with varying roles in the application cluster. We made the VLAN number match 12 bits of the IP addresses and numbered the clones similarly. It merely remained to create new snapshots every release and any dev could launch an bugfix/test/showcase/etc environment for that version, and connect to it via a VPN. I was always worried about scale if we hired more than 4096 developers but fortunately the company was acquired and its product discontinued before that happened.
That was in 2007 so the control plane (scheduler and automation) were built from scratch and we had very few reference points for the overall design. If I was building that today I’d probably still use ZFS clones but at filesystem level instead of block devices, and serve jails over NFS if I can get away with it, the iSCSI part was always a little janky.
I built a reverse proxy server for this kind of usecase specifically. I have a chroot launcher with a simple config file and the reverse proxy is auto configured over RPC and automatically acts as a CA root so internal communication is HTTPS. You can also set the API key as an env var and just run your binary on the command line for testing and it will auto configure and be available on your URL with https.
https://github.com/fsmv/daemon/
It's a bad time for me to be mentioning it because I have a major update that's not quite ready to release that changes some client APIs and makes the whole thing much nicer with fully automatic lets encrypt. I haven't had the space to work on it for a while unfortunately.
For those unfamiliar with FreeBSD, this is using base OS tools to manually create this type of immutable jail/container. This can be done with 'less effort' by using a jail manager.
Jail managers come and go. Base OS tools stay and are getting better and better. I would definitely stay away from ezjail as it us quite old, active development or even maintenance has stopped long time ago.
Author of the article seem to know what they are doing so I'm puzzled why they don't use `bsdinstall jail /path/to/jail` to implement basejail instead of manually unpacking archives.
No need for separate custom rc script to start `lo1`, it can be done with `cloned_interfaces` directive in rc.conf.
Updating and upgrading jails by passing `-b /path/to/jail` to `freebsd-update` works, but new recommended way has lately been `-j <jailname>`.
Cool article overall, the beauty of FreeBSD is also in possibility to do things in many different ways.
[dead]
If you want to do this kind of thing but don’t like the looks of this indepth tutorial, I highly reccomend installing Bastille FreeBSD instead and let the bastille tooling do the heavy lifting of managing zpool cloning and what not. Their getting started tutorial should get you 90% of the way there and has Docker like tooling. [0]
Also found this self host on a Raspberry Pi helpful.[1]
[0] https://bastillebsd.org/getting-started/
[1] https://www.sharpwriting.net/project/bastille-jail-managemen...
The ezjail[0] port is another option for achieving the article's stated goal:
I have not used the article's tool(s) and am not comparing the functionality provided by each. I have used ezjail[0] and found it exceptionally useful for similar concerns.0 - https://erdgeist.org/arts/software/ezjail/
Detailed writeup here: https://docs.skunkwerks.at/s/fUiAmi4pE
I hope OCI containers support make this boilerplate obsolete.
FYI, we just merged FreeBSD jail support into the OCI runtime spec v1.3[1]. There is already at least one implementation of it[2] as well.
[1]: https://github.com/opencontainers/runtime-spec/pull/1286 [2]: https://github.com/samuelkarp/runj
It already is, if you use most jail managers, this post is a manual approach.
I really don’t like jail managers because I need to learn a non-standard way to operate jails when I already know the manual approach. I just wouldn’t miss the boilerplate. Also, OCI containers > jail managers.
You’re creating a daemon but it is not restarting when it crashes, is it?
Isn't this just docker with extra steps?
No.
ZFS has been stable in FreeBSD for something like 17 years, and FreeBSD jails have been around for something like 25 years.
By the time Docker hit 1.0 (about 11 years ago), the use of snapshots and jails had already been normal parts of life in the FreeBSD space for over half of a decade.
I never find arguments like this compelling (but I agree with your sentiment). I don’t much care to fire up the time machine to go back 12 or more years to develop software today. If your argument is that ZFS and jails provide the same functionality but are more stable than Docker. But as is it comes off as “get off my lawn you young whipper snappers”.
But at the same time, the reason Docker won was not because it was groundbreaking tech or because it was amazingly well tested or anything. Just as one example, it has a years old bug which actively gets more comments every week having to do with Docker grossly mishandling quotes in env files.
No, the reason it won is because the development experience and the deploy experience is easy, especially when you are on Linux AND on macOS. I can’t run FreeBSD jails or ZFS on macOS, can I? Definitely not with one file and one command.
Jails and ZFS are amazing tech but they are not accessible. Docker made simple things very simple while being cross-platform enough. Do I feel gross using it? Yeah. It’s a kludgy solution to the problem. But it gets the job done and is supported by every provider out there. I am excited that it is being ported to FreeBSD though I know it will be a very long process.
> especially when you are on Linux AND on macOS. I can’t run FreeBSD jails or ZFS on macOS, can I? Definitely not with one file and one command.
On macOS, docker actually launches a Linux VM to run containers. If this counts, then yes, you can run FreeBSD jails or zfs on macOS, by running a FreeBSD VM.
But it works with one command and one (Docker)file. That’s what I mean by Docker being a kludgy solution: this is way less than ideal. But for developer experience this is very nice. And that same Dockerfile runs on everything from AWS, to GCP, to k8s, to Dokku, etc.
I dislike the implementation but I cannot deny that the UX is good enough to be very popular.
Yeah I would love to use FreeBSD jails with ZFS and everything, it’s just that the whole cloud and containerization thing happened based on Linux and FreeBSD just never made it into that ecosystem.
You’ll be sacrificing a lot and have to hand-roll a lot if you want your organization to switch from Linux+docker to FreeBSD+jails
It's all just history now for all I know, but there was work in the past to make Linux containers work on a Solaris fork (SmartOS, specifically) by emulating the Linux syscall table and presenting that to the containers. Joyent did work on this (alas, and there's an excellent and entertaining talk from Bryan Cantrill[1] that goes over it.
I imagine FreeBSD could do something similar if they aren't already. IIRC FreeBSD has a Linux emulation layer (but I don't know how much attention it still gets), and it's had containerization primitives longer than linux, so some amount of filling in the gaps in containerization features and syscall tables (if needed) could possibly yield an OCI compatibility later (for all I know all this already exists).
The problem, and the reason if this doesn't exist why people probably weren't as interested in doing the work, is it would always be "mostly" compatible and working and there would be no guarantee that the underlying software wouldn't exhibit bugs or weird behavior from small behavior differences in the platform when emulating something else. Why open yourself up to the headache when you can just run Linux with containers or build what you want on FreeBSD with jails and their own native containerization primitives.
1: https://www.youtube.com/watch?v=coFIEH3vXPw
Yeah, emulating syscalls is fine until it isn’t. See WSL1.
Some kernels are more similar to others, some are less. Turns out NT is less similar to Linux than required for good performance. I wouldn’t be surprised if Solaris was similar enough given that Linux tries to be Unix-like and Solaris is actually Unix.
What about FreeBSD+OCI containers https://freebsdfoundation.org/blog/oci-containers-on-freebsd...
In my opinion this is the path forward. I can already imagine Hashicorp Nomad orchestrator, with the podman driver, running fleets of FreeBSD containers.
Hacker News serves +4M requests per day using nothing but two FreeBSD servers…
I think there's FreeBSD images for all the clouds now.
You would need to do more work yourself to fetch and run jails probably, and I don't know if there's a hosted repository of 'jail images', but in return, you'd probably have a nicer system (at least, I'd like such a system more than running containers on google container optimized linux)
Bastille has “templates”: https://bastillebsd.org/templates/
Not Hetzner, sadly.
You can always upload your own, it's pretty simple doing so in a reproducible manner using something like Packer, but even without it you can just boot a VM into a rescue system, write your OS of choice to the VM disk and reboot.
Is it doable on a VPS? The documentation made it sound like it was for dedicated only. I went with OVH for this reason (which was cheaper anyway).
It is, yeah, the process is the same: boot into a rescue system running in memory, and do whatever you want with the disk.
Docker was the first viable containerization technology on Linux. Despite the 15 year late start vs FreeBSD Jails, it's certainly winning by the numbers.
But that has nothing to do with their respective UXs. It's a Linux vs FreeBSD signal.
> Docker was the first viable containerization technology on Linux.
No it wasn’t. Docker was late to the party even for Linux (and Linux was late compared to every other “UNIX”).
OpenVZ was around for years before docker. Its main issue was that it required out-of-tree kernel code. But there were some distributions that did still ship OpenVZ support. In fact it’s what Proxmox originally used. And it worked very well.
Then LXC came along. Fun fact, Docker was originally used LXC itself. But obviously that was a long time ago too.
I’ve used both OpenVZ and LXC in production systems before Docker came along. But I’ve always preferred FreeBSD Jails + ZFS to anything Linux has offered.
Not really, OpenVZ was/is really quite good, it was just hampered by the fact it requires out of tree modules. LXC has also always been very usable (Docker even used it for several years) but it was IMHO too focused on the VM-like management scheme that Zones and Jails had.
Docker's killer selling point was that it solved a very common and specific developer problem, not that it provided operational improvements over the state of the art on Linux. From an operational perspective, Docker has generally been a downgrade compared to LXC. (I say this as a maintainer of runc, the container runtime that underpins Docker, and as someone who worked a lot on Docker back in the early days and somewhat less today.)
I think it also helps that Docker started with an enormous amount of VC funding because their promise was to bring the “App Store” to Linux and enterprise servers.
Who couldn’t become famous with something like a $200M budget?
Feel like they spent it on marketing instead.
Podman is arguably technically superior yet people stay with Docker out of habit…
Docker engine is one thing, but access to docker hub without rate limits is what people actually pay for if they’re too cheap to host their own proxy registry (which everyone except the smallest companies should regardless).
One of Docker's big advantages is its client/server architecture made it easier to run on all those macOS and Windows boxes by just spinning up a VM for the server side, then making that VM fully managed to where a large percent of users don't even know it's there.
You can't use Docker on Mac or FreeBSD. Now, I am not calling you a liar, you _can_ use Docker on Mac and FreeBSD. But you would probably only want to do so for development, as Docker on Mac and FreeBSD requires running a Linux VM which is the thing which _actually_ runs the containers.
There is work ongoing to try to make this more native on FreeBSD (by using Linux jails) but that work is not complete yet.
So, if you want to get the same kind of experience as Docker on FreeBSD, you are forced to use jails.
The only reason Docker seems accessible is because it's native to the platform people seem to like for running all their services, but if you're dealing with FreeBSD, you most certainly would not just "use Docker" to deploy your stuff. Because you would get worse performance than if you had just used Linux.
So the answer to "Isn't this just Docker with extra steps?" is truly and absolutely "No". Not because of some kind of old man shouting at cloud argument, but because if you are on FreeBSD (for whatever reason that might be) you can't just use Docker as an easier replacement for Jails (at least right now).
[dead]
The argument was just "No [it is not the same]." The rest was just some supporting facts. :)
You've got a good take on things, and I do not disagree with what you've written.
MacOS runs containers spinning VMs. That’s why it’s slow (processing and filesystem I/O) compared to Linux.
Same thing could be done in FreeBSD. Someone needs to put in the work…
[dead]
[dead]
This is more like making an immutable linux container using only OS base tools. Docker is a whole stack doing the work for you.
I have to imagine systemd’s nspawn with btrfs integration took much inspiration. Combined with systemd’s service configuration it really makes a wonderful way of running distroless, immutable containers.
I second systemd-nspawn being a hidden gem for this usecase. I use git post-recieve hooks that target it for much of my ci/cd pipelines.
I also find myself using nspawn just to isolate apps like firefox, etc.
[dead]
This is like using Lxc with fewer steps and less tape.
FreeBsd has jail managers aka container managers aka “Docker” as well.
Try Bastille FreeBSD!