Immutable Infrastructure and Disposable Components
chadfowler.com
chadfowler.com
The idea of quickly and easily creating these master images in a way that doesn't slow down agility to change infrastructure is crucial, and Packer enables that.
Disclaimer: I wrote Packer.
val x = 1; x = 2;
is an actual error and the compiler/runtime will give a "you can't do that".
Immutable infrastructure, on the other hand, would require read-only enforcement on all aspects of configuration. It can be done, but not easily, and someone always has root. So what you're really talking about is the ability to reconstruct an environment programatically from scratch, with no manual intervention, which is laudable but not exactly immutable.
This is the sort of thing that makes me appreciate the 12 Factor App idea (http://www.12factor.net). Rather than trying to make configuration immutable, make it impossible. Don't rely on the existence/continuity of a filesystem at all. Don't use configuration files.
There's no error message telling you you can't hack an extra vhost into httpd.conf, but once you've got paged at 3am because your hack disappeared, you pretty quickly learn to supply your own error message.
Heroku's an extreme example of this - the dyno filesystem is perfectly writable, but it's definitely going away within 24 hours.
(The large-company version of "all configuration is produced by automation", which produces a surprisingly similar result, is "all configuration is produced by another team".)
In C, it's possible to change a const. But const is still a very valuable construct, even if it is possible to cheat.
same in programming... unsafePerformIO, dropping into native code or assembly, etc. In 2013 the hardware exposes a mutable interface.
In programming, it's a conscious decision to break immutability (assuming a language that supports it). It administration, it's a conscious decision to enforce it.
The first piece of the puzzle is to present configuration to runtime processes in a read only manner and enforce this rule from the remote side (e.g S3 bucket where the machine only has a R/O key or R/O NFS mount). For added points you can make the OS image read only (but that's another post).
The second piece is to make sure that you have easy build automata that can push into that repository easily and reliably. This is really key - you have to have a programatic way of doing things.
Chad claims that they throw away the ssh-keys to the machine. So they don't actually have root. Or maybe that was a goal that hasn't been reached yet, don't remember exactly (Chad gave a talk in Berlin on this and related topics).
But how do we put the idea of immutable infrastructure into practice on bare-metal servers? One option would be to set up one's own virtualization infrastructure, such as Xen or KVM. But that would undermine the I/O performance and resource consolidation advantages of running on bare metal. Docker looks promising; we just need to discover the best practices for running specific kinds of services (e.g. web applications) on top of it.
Nothing beats a bare metal machine with FreeBSD + ZFS pool + ezjail; the really cool thing about ZFS is that you can build a "base jail" then just use ZFS to make snapshots of the base jail when creating new jails. So with that you can create versioned base jails that are running an upgraded ports tree, or more generally, an upgraded world build!
It also makes transition between application versions easy because you can run a separate database jail(s) (maybe different versions of the db?) and have a jail for app v1 running while a jail for app v2 is running at the same time.
What I would love to see is a tool for abstracting this deployment pattern on-top of ezjail. I've been considering doing this myself but my time is stretched too thin :(
Of course, Illumos has zones, ZFS, KVM, and DTrace. But it's much less widely used than even FreeBSD.
btrfs also supports this, and that support is built into some LXC commands (lxc-clone, for instance.) Docker just relies on filesystem overlays and bind-mounts instead, though, so it doesn't really have need for full persisted-to-disk subtrees of the instantiated version (it's more like it holds the base version and the persisted delta on disk, and only brings them together in memory.)
This is basically what we're doing with Docker. We started with lxc but it can and will be ported to other OS virtualization backends, including jails and zones.
Create a new base image from a running container:
CONTAINER_A=$(docker run -d ubuntu apt-get install curl)
docker wait $CONTAINER_A
docker commit $CONTAINER_A shykes/my-ubuntu-with-curl
CONTAINER_B=$(docker run -d shykes/my-ubuntu-with-curl curl --help)
docker push shykes/my-ubuntu-with-curl
V1=$(docker run -d shykes/myapp:v1)
V2=$(docker run -d shykes/myapp:v2)
docker stop $V1
V1=$(docker run -d shykes/mydb:v1)
V2=$(docker run -d -volumes-from=$V1 shykes/mydb:v2)
docker stop $V1I've been doing various things toward the same goal for a long time, but Docker helps on every count (quick, redistributable etc).
On the one hand, the phrase "snowflake server" encapsulates my observations about long-lived servers that are upgraded and extended in place.
On the other hand: "Perfect replication is the enemy of any robust system." -- http://en.wikiquote.org/wiki/Daniel_Suarez (read the other quotes from the same chapter for context)
More information on service arch approach: http://www.youtube.com/watch?v=30v5WVFvlno
A few questions to satisfy my curiosity.
First, I'm assuming that you're running SmartOS on your own dedicated servers, not a public cloud like Joyent. Is that correct? If not, the rest doesn't apply.
Do you use Ansible in the global zone, in individual non-global zones, or both?
Do you build a new zone for each deploy of an application and each new version of a piece of infrastructure? I imagine this is the way to achieve the idea of immutable infrastructure.
And one other little thing: Do you boot SmartOS via a USB key, PXE, or something else?
I personally use USB keys, however it is expected that any one zone and/or server could fail at any moment. Servers are dirt cheap if you do it right. I bet some pay more in RAM costs per month than an entire node costs me with 32GB RAM, E3-1230v3, SSD, etc. There are multiple server types. The E3s are the non-persistent processing nodes... 3RU chassis 8 possible nodes in said chassis... each about 1k with the specs above and they are about 40 ECUs per. Each microcloud is typically around 10RU (1/4 rack). Place at separate datacenters and you get some interesting stuff. I have a whole different approach to the homebase DC. Stuff that REALLY reduce costs.
That being said, he's abstracting the problem a little bit:
> If you absolutely know a system has been created via automation and never changed since the moment of creation, most of the problems I describe above disappear. Need to upgrade? No problem. Build a new, upgraded system and throw the old one away. New app revision? Same thing. Build a server (or image) with a new revision and throw away the old ones.
The key here is created via automation. You get a choice, I think:
* You can spend time/energy/worry/entropy on making sure the server is up to date
* You can spend time/energy/worry/entropy on making sure the automated process is up to date
The second one is a little easier, to be sure, though depending on the setup it could still involve a lot of research, and it's certainly more expensive.
But hey, if you're in a position where trading some money to buy some time is an option and you have the cash to burn, I'd buy the time at every chance.
Instances are ready for use immediately after booting, and deploying new code is as simple as rebuilding the image and booting it. We can also easily scale up most forms of servers by simply booting more and adding them to the proxy as they come up.
The only downside here is that firefighting still kind of remains a two step process. Even after investing a decent bit of time optimizing the process, it still takes 5-10 minutes to build and upload an image. From there it takes another 5-30 minutes to get EC2 to boot it, so there's a strong temptation to tweak things on the server. I've taken to patching, testing, and committing the fix on the tag for that release so I can simply apply a diff to get the server up to date. Not perfect, but it goes a long way to prevent regressions.
After a few years of working with this, I kind of wish we'd bitten that bullet at the start. The AMIs we produce have a root filesystem that's only like 2 gigs so it goes across the network faster. 2 gigs isn't a lot of space when something decides it wants to log or write out temp files in weird places.
Just promotes the author for conference speaking.
Well played good sir you got on the front page of HN, well played.
One of the distinguishing features of a thought experiment is a lack of implementation details in the domain to which the analogy is applied.
Better not to become bogged down trying to build a faster than light locomotive.
Definitely worth a look. I'm not sure if it's fit for production usage, but i think it's the future of operating systems :).
And I find its 'nixos-rebuild build-vm' command quite handy to test changes.
On the other hand immutability of the OS image doesn't mean you can easily roll-back after an upgrade, especially if you've been running the new version for a while, because you might still have mutable state (see below).
Sure if you store all your configuration in Nix, then it is possible to rebuild the old configuration as it was, but your mutable state is not managed by Nix.
One such example is database files. You've upgraded to the new configuration, and perhaps even had a script to auto-upgrade from your old schema. Now you want to go back: what should happen to the database? Should it be entirely rolled back (from backups) to an old snapshot? But then you loose all the new data. Should you roll-back the schema change? Again what happens to the new data that doesn't fit the old schema?
Another example of mutable state is remote nodes that you depend on for a particular service. You can roll back your local machine to an old version, but will it still be able to talk to your new remote nodes?
It seems that the only way to handle this is to make sure you are always at least 1 major version backwards compatible, so you can freely upgrade/roll-back machine that are part of a cluster.
* You make sure everything is automated
* You can always trust the resulting system works; if it were to be built from scratch
Most of the next machine would have the same state as the old one:
; Machine -> Machine
; Makes the next machine
(define (next-machine m)
(make-machine
(machine-name m)
(machine-cpu m)
(machine-storage m)
(update-os CURRENT-OS)
(machine-applications m)))I love the idea of building servers on the fly from an image under some kind of revision control.