Hacker News Reader

Comment by eropple | Hacker News Reader

Chef Server has a central server. Chef Zero doesn't, especially when using tools like cfn-init/cfn-hup and S3-backed minimart to self-bootstrap. This is the approach I take, and it's the approach I see becoming more and more common. Letting nodes figure out how to deal with their own problems (better able to auto-scale, more fault tolerant) is, to me, much better than having Jenkins or whatever have to SSH into them in the first place.

You can do it with Ansible, if you're going home-roll it, but I haven't seen too many people do so.

View on HN

Does anyone actually use Ansible server-less? Can it even be done?

(Edit: my bad, ignore this comment, I misphrased my question; the real question should have been: can you run Ansible locally on the target server, just like chef-solo/chef-zero?)

Ansible was designed to be server-less from the beginning. Are thinking of Chef or Puppet? Ansible Tower (which is the subject of this article) is a web application frontend for running Ansible.

Ansible is serverless by design - all you need to control a remote system is SSH access and python installed on the target.

Calling that "serverless" is something of an abuse of the term. You can call it "push-based" rather than "pull-based" (a Chef/Puppet model), but there is definitely a "server" to be had--it's the machine running SSH and with the canonical datastore. It is--and this is one of many reasons I don't like Ansible very much--just a pretty poor server and often the developer's workstation.

"Serverless" would be more like what I described with regards to Chef Zero, where a machine, as it bootstraps, is able to fetch its playbooks from somewhere and self-execute with some sort of sourceable configuration data. The standard Ansible workflow is not only not "serverless", but it is antithetical to cloud-friendly scaling and fault-tolerance practices. (Think about how you're going to manage auto-scaling groups with it. It hurts.)

There is no need to have a server at all. Any configuration can be run locally without ssh.

Sure! If you read the thread, you will notice that I explicitly describe how to do that. And that this post is not referring to doing that--it's explicitly referring to not doing that, instead doing remote provisioning via SSH. And, as a bonus, I went into some of the contributing factors that lead to people not doing it. So I'm not super sure why you felt compelled to reply?

You were right in that your question was misphrased. Nobody seem to understand what you were asking.

we [1] do most of our AMI provisioning by booting a machine, checking out our ansible codebase and running `ansible-playbook -c local -i "localhost," ...` as part of cloud-init. Build progress is piped over SQS and a management process [2] waits for completion and triggers an AMI snapshot. It works well. I'm not sure if that's what you mean by "server-less", but in our use case there is no controller.

1. https://github.com/edx/configuration/ 2. https://github.com/edx/configuration/blob/master/util/vpc-to...

Hey Fred - I do believe we met when I was chatting with edX a little while back, funny seeing you around the internet. =) And yeah, this is much more of what I would describe as "serverless". It kind of sounds like what you're doing is pretty Packer-compatible; any particular reason you guys went the way you did?

For my purposes, Packer isn't always available, which is a bummer. The Chef Zero approach I described above is nice for my purposes because it works with either an AMI or a live instance; when I write cookbooks I break them into "bake" and "configure" recipes and sub-recipes, and the "bake" steps are effectively memoization of steps I also run (idempotently) when a machine comes up.

Yes, it can absolutely be done, but you'll need to build your own infrastructure. You just run the playbook locally using "-c local -i localhost,". Obviously, you'll have to figure out how to set variables and get the execution results back to the master.

It also means if any of your servers is compromised, the attacker will be in possession of your whole playbook.

> It also means if any of your servers is compromised, the attacker will be in possession of your whole playbook.

Which should, if you are designing your systems wisely, not matter at all, in any way, because configuration management systems should not contain secret or sensitive data (which should be provided from a more secure option--Credstash, Vault, whatever).

I could open-source all of my Chef cookbooks or Ansible playbooks and not have a security or correctness concern; this is pretty straightforward design.

You can run Ansible locally with `-c local`.

But, as far as serverless/self-bootstrapping deploys go, it's less common. Ansible has less of a "culture of dependencies"; the simpler, more approachable-looking nature of the Ansible playbook format seems to lend itself to people one-offing whatever they need rather than looking for best-practices solutions that already exist. Because of this, there's no real Berkshelf equivalent for Ansible. The tooling doesn't exist, outside of Tower (sorta), because nobody wants it, and nobody wants it because the tooling doesn't exist. So the people who are doing with Ansible something similar to the Chef Zero stuff I mentioned above are mostly home-rolling it. (I just use a S3 bucket as a Minimart berkshelf endpoint and move on with my day.)

Last-mile configuration is also tricky. In my Chef Zero stuff, I use CloudFormation metadata to provide Chef attributes. You can do something similar with Ansible...but it's duct-tapey. There are times when simple is better; IMO, Ansible's core tooling errs too far on that side and the ecosystem has not caught up to make more rigorous approaches really viable.

Apologies if I'm wrong, but it sounds like you're not that familiar with Ansible's roles and the public Ansible Galaxy repo for them?

https://galaxy.ansible.com/intro

There are tons of roles available, for just about everything, and the quality isn't always great, but still higher on average than what I've found for most Chef recipes and Puppet modules.

And that's not to mention the very high number of high quality modules that are builtin to Ansible.

I am familiar with them, yeah. In practice, across a pretty wide spread of clients, I have never seen them used or written by anyone who isn't me. This is why I referred to it as a culture problem; the tooling problem is the lack of a Berkshelf equivalent.

I would strongly, strongly disagree as to the quality of most Ansible modules that I have dealt with, but it's probably more based on exactly what you need than anything else.

On AWS, we bake our AMIs with packer and include the Ansible roles and playbooks.

We use CloudFormation to deploy, so in the instance metadata we have it run Ansible locally to bootstrap and return the exit status to cfn-signal.

We retrieve secrets via Parameter Store. For environment specific configs that are not secrets (ie passing in vars from CloudFormation), we have cloud-init write a json file that we include with our ansible-playbook command.

The command ends up looking something like:

ansible-playbook -v -i 'localhost,' -c local /some/path/playbook.yml --extra-vars '@/some/path/vars.json' && /opt/aws/bin/cfn-signal -e $? --stack ${AWS::StackName} --resource ${AsgName} --region ${AWS::Region}

Yes, my friend uses ansible to store his work-pc configuration. After upgrade/replacement he just installs git, ansible and his private key and then rest of the the setup is done by ansible from playbook in his private config repo.

In my mind, though, the biggest reason I used ansible was it's simple push-through-ssh nature. I knew that if there are some maintanance tasks I would have to do by sshing from my machine to some server, I could as well create and run an ansible playbook against it. This was espacially useful for configuring ephemeral boxes spawned on i.e. openstack where I know it will have ssh running and that is it.

Yes, I reprovisioned our dying Jenkins machine like this, because we usually run ansible from Jenkins. I was tweaking and updating the playbooks on the new Jenkins machine then running ansible there, then adding things that worked back into git.

Ansible also works well over ssh too. It’s pretty flexible.

Yes - you can either use your own tooling to distribute your playbooks to the hosts and then run ansible-playbook using the 'local' connection method, or you can use ansible-pull, which retrieves the playbooks from an scm (git) repo and runs them locally. You don't need AWX/Tower for any of this.

Do you happen to know of any fuller description, maybe a blog post, of the Chef Zero + minimart (etc.) setup? (Or, really, any similar setup.)

I'm having a bit of a hard time 'getting' the handling of secrets, etc. in such a setup.

(We're currently using Ansible, and frankly it's turned into a huge pain -- even on just a few hosts it's incredibly slow and horrible to debug. I'd really like to eventually transition to a more "build-a-pristine base system image" + "self-by-applying-more-recent-playbooks" type approach.)

No, but I can probably write one. We just use Credstash (rcredstash) for secrets and are done with it. (One of the nice things about not-using-Ansible is the lowered bar to entry of just pulling a library and writing the code you need.)

PLEASE write a decently detailed blog post about this! I and several other people would be eternally grateful.

I find that this is one of the most frustrating things about the Brave New World of immutable/container/VM/cloud/etc. It's actually REALLY hard to separate the hype from actual working things because all you seem to find is the hype and... 50 page "tutorials" on how to set Kubernetes[1] up.

[1] Random example, but the mere fact that they use a "simplified" admin program (minikube) for the tutorial tells me volumes about how fun it's going to be and how little administration it's going to require in production. /s

I'll see what I can do. But--would an online course do? Because I have written one, I just haven't recorded the voiceovers yet...

The fun part is, I think most of the sexy tools are lousy. Ansible demos well but works poorly; Terraform has bitten me so many times I don't trust it; Kubernetes doesn't make sense to me in a universe where I am buying already-provisioned-and-segmented resource bunches (we can call them "virtual machines", maybe) where I have to incur extra deadweight loss because fault-tolerance implies requiring extra space in case any existing node goes away.

I have spent enough time thinking about this that I am reasonably confident in my approach, and I would like to share it. It's a little more than a blog, though.

Assuming you can motivate it decently, I think a course might be good... However, that's quite a different value proposition, so it might take a bit to convince my employees to do it -- whereas a 'I read this series of articles by this very knowledgeable guy' doesn't require much convincin'.

To quote the infinite wisdom of The Simpsons: "Do what you feel" [is right]. :)

ansible-pull gives a similar workflow to chef-zero in an autoscaling environment.

There's also a balance to be struck between pre-baking AMIs and running config management at instance launch time.

Sure. But that's pretty awful. ansible-pull relies on a git repository, which relies on key provisioning, which means that you need to configuration-manage your configuration management and you don't have a stump-dumb, easy solution for it in any major cloud. And you have no dependency management (submodules, at best, are not dependency management), so I hope that you've vendored (which is gross) all of your dependencies.

This really is what I do for a living. I'm speaking from a position of entirely too extensive experience when I say that Ansible has no good solution here in common use. If I thought Ansible was good enough for me to be spending a lot of time with (it's not, and I advocate that clients not use it if they have a choice), I'd have probably already had to write it.

As far as machine images go, they are an optimization, not a core system. Your configuration management systems need to be able to bootstrap from either an AMI, to lay on last-mile (configuration, as opposed to setup, stuff) and converge any updates since the last AMI build, or to start from scratch. And that is another weakness of Ansible; writing idempotent Ansible scripts is significantly harder than it needs to be.

Reply on news.ycombinator.com