7 ms·

Immutable Infrastructure is the future

125 points by geerlingguy 12y ago

geerlingguy 12y ago
A couple relevant snippets: Caveat - I think images must come from a build system. Using anybody’s canned images as a baseline is reckless. Config tools still have a place - in building the automated systems that allow folks to focus on business logic, but I think there are going to move down-stack to doing things like automating OpenStack. The longer term future is about cloud/virt systems like this provisioning themselves, but they are still yet a bit too complicated in the way they are going about it, but I feel it’s coming. (The post's author, Michael DeHaan, wrote Ansible in 2012, but left the project earlier this year to pursue other interests.) My thoughts on the subject are much the same—but I think it'll take a few years for the tooling to reach the level of polish we currently have with best-of-breed CM and general provisioning tools. It's not as easy to go from 0-to-infrastructure with a Docker-based workflow (especially if you have your own registry, etc.) as opposed to a VM-plus-CM-based workflow.
- mpdehaan2 12y ago
  author here. "but left the project earlier this year to pursue other interests" - Hi Jeff! I haven't really shared reasons for leaving the company to date and can't go into that. I also don't think it will take years and think immutable is there today. I think we're there today with a bit of skill and minor bit of work on top, but a little bit of work from cloud/virt solutions will push us over the hump - to be able to codify the upgrade-flip idiomatically within cloud/virt systems, so you don't need something else to do it for you. I do look forward to that being surfaced. Things like Amazon's ECS in particular strike me as particularly interesting.
  - geerlingguy 12y ago
    > I think we're there today with a bit of skill and minor bit of work on top, but a little bit of work from cloud/virt solutions will push us over the hump. I think that's true, but my timeline is based on how long I see it taking 'mainstream' enterprise companies to adopt immutable infra—for many of the orgs I've worked with, they've only recently completed a sea change from dedicated servers to virtualized 'cloud' deployments. So it really depends on who you're working with; many businesses are making the move today, but cautiously. Once the early adopters complete rolling things out to production and working out the kinks, we'll finally get to the late majority/stability, and I won't have to commiserate with the thousands of sysadmins stuck with dozens, hundreds, or thousands of VMs :)
runlevel1 12y ago
I'd like to see a polished common configuration API. Currently, we have Chef, Puppet, Ansible, and Salt (amongst others) all working to create a sane common abstraction on top of a rainbow of configuration files and command-line arguments. Each of them have their strengths, and frankly they're all impressive in their own right. But there's only so much you can do to reign in the chaos of the bazaar. The Docker workflow certainly has its merits. It's great for narrowing dev-prod parity and repeatability. I've embraced it, but I'm a bit anxious that we, as a community, might end up creating bigger black boxes -- Dockerfiles upon Dockerfiles.
- zdw 12y ago
  > I'd like to see a polished common configuration API. https://xkcd.com/927/ https://xkcd.com/927/
- atsaloli 12y ago
  What effort is this? URL please?
- mpdehaan2 12y ago
  Hmm, I think your boxes can be still pretty transparent in that case - if anything more so, as you've got single-bash files versus perhaps directories and directories of content? In most organizations, all those Dockerfiles should be yours, and in source control. Relying on the blobs from others shouldn't be neccessary for most applications. Yes - in large organizations, a management problem probably does arise, but I think if you find copy/paste rising up you have to push that down into the base image, basically? I don't think a configuration API helps to be honest - what you would do is turn them all into a "greatest common divisor" type scenario, where as they feel different because they were trying to be different. I think you're best solved by organizationally mandating one of them within your company, as painful as that may be.
kolev 12y ago
Wrong! It's the present!
akurilin 12y ago
Very interested in this subject. Is there a comprehensive guide on how I can pull this off TODAY with e.g. AWS?
- eropple 12y ago
  It's only one piece of the puzzle, but at my last company I wrote deploy_thing[1] for exactly this. An immutable config bundle and an immutable artifact tag combine to create a deploy that can be launched and put into production with an AWS auto-scaling group. Need to change a configuration, you burn down the auto-scaling group. This was a transitional step for us before moving to Mesos, but it's a good example, I think, of how to approach the problem from a minimum-viable level. (I'm currently playing with moving it to OpenStack.) [1] https://github.com/eropple/deploy_thing https://github.com/eropple/deploy_thing
  - bender80 12y ago
    The URL is throwing 404, probably due to it being private. Would love to see it. I am working on something similar.
    - zerocrates 12y ago
      It looks like it was supposed to a hyphen, not an underscore: https://github.com/eropple/deploy-thing https://github.com/eropple/deploy-thing
      - eropple 12y ago
        Whoops, yeah. My bad, I switched from an underscore to a hyphen halfway through the project. It's still inconsistent, because we had a case of company-going-unders and I haven't needed it since.
- yo-code-sucks 12y ago
  Make a dockerfile, make sure it works and you understand it. Then go into Elastic Beanstalk and click new application and there is an option for loading a dockerfile, boom. For you this may work, for me it's more complex.
andrewchambers 12y ago
Stateless servers with kubernetes seem like a great thing to do this sort of thing with.
- nickbauman 12y ago
  Why go that low, though? Why not go with Managed VMs first?
  - andrewchambers 12y ago
    Kubernetes gives you freedom about where to run your stuff since it works on multiple providers and your own hardware. Once you learn how to set it up, it isn't too bad. Google managed kubernetes cluster gives you the best of both worlds anyway.
    - KaiserPro 12y ago
      A decent VM system will give you HA and load balancing (ie migrating VMs from overloaded hosts) for free. Something docker/kubernetes still can't do. Its also worth bearing in mind that the "vm penalty" is almost trivially small compared to the cost of dev time.
      - deathanatos 12y ago
        > […] (ie migrating VMs from overloaded hosts) for free. > Something docker/kubernetes still can't do. This doesn't solve issues arising from the VM outright crashing though, does it? (I don't see how it could.) At the end of the day, I still need to solve that myself, and if my application can gracefully withstand being crashed, then docker/kubernetes can migrate from overloaded hosts: just crash container. I could probably also throw in something to gracefully die on a SIGTERM, but I feel like once you can withstand a crash, throwing in a SIGTERM is fairly straightforward. The only difference is that one trains traffic, the other just outright 500s it. (Though, perhaps there's more work involved there than I realize.) Ultimately, I want crash resistance. I want to run Chaos Monkey. I can't, right now, because I know all too well what would happen.
falcolas 12y ago
> Ultimately in the end it means more time for everyone to focus on business logic versus thinking about deployments, so I’m a fan of that. I find this amusing: our company is migrating to Docker right now, and over in operations I'm spending more and more time thinking about deployments than I've had to since I started doing operations work. I have to think about "what major functionality has changed in Docker in the last three months, and how does that impact any/all of our images?" I have to think about creating frequently updated base images, and how to deploy these updated base images across our regions. I have to think about making the image layers as small as possible, to limit the bandwidth impact of updating images. I have to think about making as few image layers as possible to avoid DevMapper IO overhead. I have to think about externally mapped volumes to keep necessary IO as fast as possible. I have to think about asymmetric TCP, docker vs host network interfaces, natting, and service discovery. I have to think about private registries, token based authentication systems, and redundancy. My developers have been able to focus on business logic since they hired an ops team. Will the amount of thinking go down over time? Some of it: the bits which can be automated fade to the back (at least until the core Docker functionality changes yet again), but there are bits and pieces will necessarily change with every deploy.
- parasubvert 12y ago
  Or you could adopt a platform that has made a lot of those opinionated decisions in a pre-integrated config. There's a chance it might not be a fit but then it's a matter of whether it's easier to start somewhere and tweak it vs. starting ground up.
- dasil003 12y ago
  The difference is the pain you are experiencing with Docker is potentially transient, whereas the problems with configuration management on the wild internet are intractable, and even unquantifiable. With Docker the scope of changes you have to deal with is at least knowable. I say "potentially" because I think the jury is still out whether Docker can live up to its promise. Clearly it has a real purpose and value add above traditional configuration management, but the churn is an indication of how hard a problem it is. Plus, even in an ideal case, you still have to deal with security updates and other inevitable version updates and things that invalidate your current images.
nickbauman 12y ago
> While PaaS has not become a reality for everyone... A telling aside. Much of what he wants already exists in Google App Engine (and probably in Elastic Beanstalk, too, but I haven't tried it) — sans the need to manage the underlying OS. Interesting Google is going the other way with Managed VMs running on top of Kubernetes. Perhaps they will meet in the middle somewhere?
- w4tson 12y ago
  I thought Google was running containers in kubernetes? AFAIK you can install Kubs on VMs
  - nickbauman 12y ago
    We're both correct.
  - kevinschumacher 12y ago
    "Managed VMs" is a service offering within Google App Engine that lets you use any runtime (e.g., Python, Node.js, Go, C) but get the benefits of the GAE PaaS auto-scaling, versioning, etc. It actually works by shipping a docker container to the Google Container Registry (privately scoped to your project), spinning up a "VM" that shows up in your Compute Engine instances list that you can then SSH into if you want (this is why you're both right; they call it Managed VMs, it shows up in your GCE instance list, but they literally are, I think, running your docker container, probably with Kubernetes under the hood).
    - nickbauman 12y ago
      This is exactly what's happening as I've seen it.
dorfsmay 12y ago
Yes, I agree, but the issue is still how do you build that infrastructure (be it an AMI, a docker image etc..)? Docker file sare simplistic, so you end up using the same tools that were available before (ansible, salt, bash etc...). What we're doing right now is move what we used to do at "configuration time" to "build time".
- mpdehaan2 12y ago
  My point was you don't need to use the tools you use today, except to do the version flip (and probably to deploy the subtrate - unless you are doing a cloud-image solution and not trying to do something on-premise). And this will get better when cloud providers embrace higher level cluster operations more natively. And build time is the way to go for everything you can push down into build time. There still may be some need for service-discovery type applications.
  - dorfsmay 12y ago
    Agreed, moving more and more stuff to the build, lighting up an image should be nearly instant, so we can do auto-scala etc... so no more "pip install -r" when bringing up an image, that's now part of the build. But the landscape for tools to build that image, be it a docker image or an AMI, right now, is the same. I find myself "going back" to ansible but now using it to build images rather than change online VMs (could be any tool, I know ansible and it suits my way of thinking).
tomc1985 12y ago
So by immutable he means... systems you don't patch? That's it?
KaiserPro 12y ago
I quite like immutable infrastructure. It means you have a rollout/roll back with a flip of a dns/connection. However I see a lot of things about docker and think two things: 1) most people really wany a mainframe 2) shared filesystems really are brlilliant. Let me unpack that: a mainframe provides a massive amount of hardware, a hypervisor and scheduler. This means that you can control a bunch of isolated processes through one one system easily. Something that you can't really do in docker yet (fleet/kubernetes arn't there yet. in a mainframe you can use a single script to control an entire system) Quite a lot of web config management is about shipping packages to different machines. Almost all of this config, and therefore complication can be removed by using a shared filesystem. For example we have many version of java. For some reason they are in seperate repos. This means that if you want to change version (openJDK to oracle etc) you have to change config. With a shared file system, in the $PATH, you'd just say oracalejava-1.8 $programme or java-1.7 $programme If you make your deploys to a similar shared filesystem, then to change config is basically a case of ssh machine, kill old process, launche new process. Job done. This makes each machine a dumb execution system, minimal state is contained in each node. With NFS coming to AWS, that'll be the model for my new infrastructure. the reason why? because it works. In HPC thats how we've been managing high scale workloads for years.
ABS 12y ago
For those looking into this topic only now we interviewed 6 people who have been doing this for some time, collected their answers and published them + did a hangout with them here: https://highops.com/insights/immutable-infrastructure-6-questions-6-experts/ https://highops.com/insights/immutable-infrastructure-6-ques... (pure content, no sales pitch anywhere)

125 points by geerlingguy 12y ago

A couple relevant snippets: Caveat - I think images must come from a build system. Using anybody’s canned images as a baseline is reckless. Config tools still have a place - in building the automated systems that allow folks to focus on business logic, but I think there are going to move down-stack to doing things like automating OpenStack. The longer term future is about cloud/virt systems like this provisioning themselves, but they are still yet a bit too complicated in the way they are going about it, but I feel it’s coming. (The post's author, Michael DeHaan, wrote Ansible in 2012, but left the project earlier this year to pursue other interests.) My thoughts on the subject are much the same—but I think it'll take a few years for the tooling to reach the level of polish we currently have with best-of-breed CM and general provisioning tools. It's not as easy to go from 0-to-infrastructure with a Docker-based workflow (especially if you have your own registry, etc.) as opposed to a VM-plus-CM-based workflow.

mpdehaan2 12y ago
author here. "but left the project earlier this year to pursue other interests" - Hi Jeff! I haven't really shared reasons for leaving the company to date and can't go into that. I also don't think it will take years and think immutable is there today. I think we're there today with a bit of skill and minor bit of work on top, but a little bit of work from cloud/virt solutions will push us over the hump - to be able to codify the upgrade-flip idiomatically within cloud/virt systems, so you don't need something else to do it for you. I do look forward to that being surfaced. Things like Amazon's ECS in particular strike me as particularly interesting.
- geerlingguy 12y ago
  > I think we're there today with a bit of skill and minor bit of work on top, but a little bit of work from cloud/virt solutions will push us over the hump. I think that's true, but my timeline is based on how long I see it taking 'mainstream' enterprise companies to adopt immutable infra—for many of the orgs I've worked with, they've only recently completed a sea change from dedicated servers to virtualized 'cloud' deployments. So it really depends on who you're working with; many businesses are making the move today, but cautiously. Once the early adopters complete rolling things out to production and working out the kinks, we'll finally get to the late majority/stability, and I won't have to commiserate with the thousands of sysadmins stuck with dozens, hundreds, or thousands of VMs :)

runlevel1 12y ago

I'd like to see a polished common configuration API. Currently, we have Chef, Puppet, Ansible, and Salt (amongst others) all working to create a sane common abstraction on top of a rainbow of configuration files and command-line arguments. Each of them have their strengths, and frankly they're all impressive in their own right. But there's only so much you can do to reign in the chaos of the bazaar. The Docker workflow certainly has its merits. It's great for narrowing dev-prod parity and repeatability. I've embraced it, but I'm a bit anxious that we, as a community, might end up creating bigger black boxes -- Dockerfiles upon Dockerfiles.

zdw 12y ago
> I'd like to see a polished common configuration API. https://xkcd.com/927/ https://xkcd.com/927/
atsaloli 12y ago
What effort is this? URL please?
mpdehaan2 12y ago
Hmm, I think your boxes can be still pretty transparent in that case - if anything more so, as you've got single-bash files versus perhaps directories and directories of content? In most organizations, all those Dockerfiles should be yours, and in source control. Relying on the blobs from others shouldn't be neccessary for most applications. Yes - in large organizations, a management problem probably does arise, but I think if you find copy/paste rising up you have to push that down into the base image, basically? I don't think a configuration API helps to be honest - what you would do is turn them all into a "greatest common divisor" type scenario, where as they feel different because they were trying to be different. I think you're best solved by organizationally mandating one of them within your company, as painful as that may be.

kolev 12y ago

Wrong! It's the present!

akurilin 12y ago

Very interested in this subject. Is there a comprehensive guide on how I can pull this off TODAY with e.g. AWS?

eropple 12y ago
It's only one piece of the puzzle, but at my last company I wrote deploy_thing[1] for exactly this. An immutable config bundle and an immutable artifact tag combine to create a deploy that can be launched and put into production with an AWS auto-scaling group. Need to change a configuration, you burn down the auto-scaling group. This was a transitional step for us before moving to Mesos, but it's a good example, I think, of how to approach the problem from a minimum-viable level. (I'm currently playing with moving it to OpenStack.) [1] https://github.com/eropple/deploy_thing https://github.com/eropple/deploy_thing
- bender80 12y ago
  The URL is throwing 404, probably due to it being private. Would love to see it. I am working on something similar.
  - zerocrates 12y ago
    It looks like it was supposed to a hyphen, not an underscore: https://github.com/eropple/deploy-thing https://github.com/eropple/deploy-thing
    - eropple 12y ago
      Whoops, yeah. My bad, I switched from an underscore to a hyphen halfway through the project. It's still inconsistent, because we had a case of company-going-unders and I haven't needed it since.
yo-code-sucks 12y ago
Make a dockerfile, make sure it works and you understand it. Then go into Elastic Beanstalk and click new application and there is an option for loading a dockerfile, boom. For you this may work, for me it's more complex.

andrewchambers 12y ago

Stateless servers with kubernetes seem like a great thing to do this sort of thing with.

nickbauman 12y ago
Why go that low, though? Why not go with Managed VMs first?
- andrewchambers 12y ago
  Kubernetes gives you freedom about where to run your stuff since it works on multiple providers and your own hardware. Once you learn how to set it up, it isn't too bad. Google managed kubernetes cluster gives you the best of both worlds anyway.
  - KaiserPro 12y ago
    A decent VM system will give you HA and load balancing (ie migrating VMs from overloaded hosts) for free. Something docker/kubernetes still can't do. Its also worth bearing in mind that the "vm penalty" is almost trivially small compared to the cost of dev time.
    - deathanatos 12y ago
      > […] (ie migrating VMs from overloaded hosts) for free. > Something docker/kubernetes still can't do. This doesn't solve issues arising from the VM outright crashing though, does it? (I don't see how it could.) At the end of the day, I still need to solve that myself, and if my application can gracefully withstand being crashed, then docker/kubernetes can migrate from overloaded hosts: just crash container. I could probably also throw in something to gracefully die on a SIGTERM, but I feel like once you can withstand a crash, throwing in a SIGTERM is fairly straightforward. The only difference is that one trains traffic, the other just outright 500s it. (Though, perhaps there's more work involved there than I realize.) Ultimately, I want crash resistance. I want to run Chaos Monkey. I can't, right now, because I know all too well what would happen.

falcolas 12y ago

> Ultimately in the end it means more time for everyone to focus on business logic versus thinking about deployments, so I’m a fan of that. I find this amusing: our company is migrating to Docker right now, and over in operations I'm spending more and more time thinking about deployments than I've had to since I started doing operations work. I have to think about "what major functionality has changed in Docker in the last three months, and how does that impact any/all of our images?" I have to think about creating frequently updated base images, and how to deploy these updated base images across our regions. I have to think about making the image layers as small as possible, to limit the bandwidth impact of updating images. I have to think about making as few image layers as possible to avoid DevMapper IO overhead. I have to think about externally mapped volumes to keep necessary IO as fast as possible. I have to think about asymmetric TCP, docker vs host network interfaces, natting, and service discovery. I have to think about private registries, token based authentication systems, and redundancy. My developers have been able to focus on business logic since they hired an ops team. Will the amount of thinking go down over time? Some of it: the bits which can be automated fade to the back (at least until the core Docker functionality changes yet again), but there are bits and pieces will necessarily change with every deploy.

parasubvert 12y ago
Or you could adopt a platform that has made a lot of those opinionated decisions in a pre-integrated config. There's a chance it might not be a fit but then it's a matter of whether it's easier to start somewhere and tweak it vs. starting ground up.
dasil003 12y ago
The difference is the pain you are experiencing with Docker is potentially transient, whereas the problems with configuration management on the wild internet are intractable, and even unquantifiable. With Docker the scope of changes you have to deal with is at least knowable. I say "potentially" because I think the jury is still out whether Docker can live up to its promise. Clearly it has a real purpose and value add above traditional configuration management, but the churn is an indication of how hard a problem it is. Plus, even in an ideal case, you still have to deal with security updates and other inevitable version updates and things that invalidate your current images.

nickbauman 12y ago

> While PaaS has not become a reality for everyone... A telling aside. Much of what he wants already exists in Google App Engine (and probably in Elastic Beanstalk, too, but I haven't tried it) — sans the need to manage the underlying OS. Interesting Google is going the other way with Managed VMs running on top of Kubernetes. Perhaps they will meet in the middle somewhere?

w4tson 12y ago
I thought Google was running containers in kubernetes? AFAIK you can install Kubs on VMs
- nickbauman 12y ago
  We're both correct.
- kevinschumacher 12y ago
  "Managed VMs" is a service offering within Google App Engine that lets you use any runtime (e.g., Python, Node.js, Go, C) but get the benefits of the GAE PaaS auto-scaling, versioning, etc. It actually works by shipping a docker container to the Google Container Registry (privately scoped to your project), spinning up a "VM" that shows up in your Compute Engine instances list that you can then SSH into if you want (this is why you're both right; they call it Managed VMs, it shows up in your GCE instance list, but they literally are, I think, running your docker container, probably with Kubernetes under the hood).
  - nickbauman 12y ago
    This is exactly what's happening as I've seen it.

dorfsmay 12y ago

Yes, I agree, but the issue is still how do you build that infrastructure (be it an AMI, a docker image etc..)? Docker file sare simplistic, so you end up using the same tools that were available before (ansible, salt, bash etc...). What we're doing right now is move what we used to do at "configuration time" to "build time".

mpdehaan2 12y ago
My point was you don't need to use the tools you use today, except to do the version flip (and probably to deploy the subtrate - unless you are doing a cloud-image solution and not trying to do something on-premise). And this will get better when cloud providers embrace higher level cluster operations more natively. And build time is the way to go for everything you can push down into build time. There still may be some need for service-discovery type applications.
- dorfsmay 12y ago
  Agreed, moving more and more stuff to the build, lighting up an image should be nearly instant, so we can do auto-scala etc... so no more "pip install -r" when bringing up an image, that's now part of the build. But the landscape for tools to build that image, be it a docker image or an AMI, right now, is the same. I find myself "going back" to ansible but now using it to build images rather than change online VMs (could be any tool, I know ansible and it suits my way of thinking).

tomc1985 12y ago

So by immutable he means... systems you don't patch? That's it?

KaiserPro 12y ago

I quite like immutable infrastructure. It means you have a rollout/roll back with a flip of a dns/connection. However I see a lot of things about docker and think two things: 1) most people really wany a mainframe 2) shared filesystems really are brlilliant. Let me unpack that: a mainframe provides a massive amount of hardware, a hypervisor and scheduler. This means that you can control a bunch of isolated processes through one one system easily. Something that you can't really do in docker yet (fleet/kubernetes arn't there yet. in a mainframe you can use a single script to control an entire system) Quite a lot of web config management is about shipping packages to different machines. Almost all of this config, and therefore complication can be removed by using a shared filesystem. For example we have many version of java. For some reason they are in seperate repos. This means that if you want to change version (openJDK to oracle etc) you have to change config. With a shared file system, in the $PATH, you'd just say oracalejava-1.8 $programme or java-1.7 $programme If you make your deploys to a similar shared filesystem, then to change config is basically a case of ssh machine, kill old process, launche new process. Job done. This makes each machine a dumb execution system, minimal state is contained in each node. With NFS coming to AWS, that'll be the model for my new infrastructure. the reason why? because it works. In HPC thats how we've been managing high scale workloads for years.

ABS 12y ago

For those looking into this topic only now we interviewed 6 people who have been doing this for some time, collected their answers and published them + did a hangout with them here: https://highops.com/insights/immutable-infrastructure-6-questions-6-experts/ https://highops.com/insights/immutable-infrastructure-6-ques... (pure content, no sales pitch anywhere)