Skip navigation.

My job description

http://dilbert.com/strips/comic/2010-04-07

I bet a lot of sysadmins are putting this comic up on their office wall today. It neatly encapsulates that management information problem that goes with the job.

What's wrong with this image?

The problem

Is this a familiar nightmare? “The server you built yesterday is great,” your boss enthuses. “Now we need a hundred more just like it. And when you’ve done that, could you take a look at my laptop before lunch? I can’t seem to get into my email.”

It’s common nowadays for sysadmins to manage very large numbers of servers - often virtual servers created on-demand in cloud platforms like Amazon EC2. How do you build so many servers and keep them updated - automatically, reliably, and quickly?

One common strategy is the golden image pattern: set up the server the way you want it, and then copy the disk image and use that to boot as many clone machines as you need. Another is the automation pattern: start with a generic server, and then apply automatic configuration tools such as Puppet to bring the machine into a specified state where it is ready for production.

The golden image

The golden image is attractive to many because it’s easy and cheap. It’s quick to create images, and you just need a big disk to store them.

Imaged machines also have low latency: they’re ready for production in just a few minutes from startup. Finally, it’s platform agnostic: you can image virtually any kind of machine, even Windows (which most config automation tools can’t handle very well).


So what’s the problem with the golden image? Luke Kanies, creator of Puppet, prefers to call it the “foil ball” approach: your images can quickly become an unmanageable, crumpled mess. Firstly, how many images do you need? Well, as many as you have different types of machines: web servers, databases, load balancers, DNS servers, and so on.

The number of images multiplies again when you take history into account. What happens when you need to recreate the state of the database server last month to troubleshoot a problem? You need to keep old images around, which means a new image every time you make a change.

And here’s a thing: if you do need to make a change (for example a security upgrade for a critical package), you’ll have to rebuild the golden image and then re-image and reboot all affected machines - which means lots of bandwidth and possibly downtime. And, unfortunately, even imaged machines usually need some final tweaking (eg IP addresses) before they’re ready for use.

Automation tools

By comparison, building your machines with an automatic configuration tool such as Puppet, Chef or even cfengine gives you many advantages. The configuration is self-documenting; it’s easy to see what should be on each machine, and more importantly, why it’s there. It will also stay in sync with reality.

Although your server types will be different, they will also likely have a lot in common. With images, these supposedly common settings could easily drift apart over time. With automation tools, you can put these settings into modules or classes, just as a software developer re-uses a module or library in another project.

In fact, all the advantages of software modules are available: you can share code with others, download public recipes and manifests, and maintain a complete version history of your network.

Automated configuration is intrinsically multi-platform. Puppet and Chef can abstract away some of the differences between operating systems, and you can add minor tweaks of your own. For example, you might need a slightly different MySQL config setting for MySQL 4.x as opposed to 5.x. A one-line switch in the configuration can address this problem, instead of creating two disk images.

Of course, everything comes with a downside. It is more difficult and usually more time-consuming to build an automated configuration, especially from scratch and when you’re still learning to use the tools. And some things are hard to automate. Some software just wasn’t written with automated installation and management in mind.

Also, not all software will be packaged for your platform (or you need special compilation options). In these cases you will need to package the software yourself.

Finally, configuration management doesn’t always mix well with human intervention. In particular, someone may make a config change on the machine which is then automatically reversed by Puppet. At best this can be a nasty surprise, at worst it can result in a site outage. Managing desktops automatically can create similar problems. This needs to be addressed at a business level with a proper change control process and making it clear to everyone where the boundaries of automatic management lie.

Combined strategies

So both imaging and automation have their place; neither is a magic bullet. The best strategy for scaling your deployments is often a judicious combination of these two approaches. You can, for example, start with a ‘stem cell’ image containing an approved base build, security fixes and settings which don’t often change, which is then configured by Puppet or Chef. Or you could use provisioning tools such as Kickstart and Cobbler and then hand over to Puppet. Or even build the machine with Puppet and then create a golden image from this.

One thing is certain: ignoring the problem is not an option. Moving to an automated network might seem like a big project, but the key is to start small. Set aside a block of time every week to spend on improving your infrastructure, and automate one small thing at a time as part of a gradual continuous process. The more you invest in this, the more benefits you’ll see - and you can bet your competitors will be doing the same.

Further reading

An earlier version of this article appeared on the Agile Sysadmin blog.

Chicken Techa Masala

A good time was had by all at the 4th monthly Devops Curry Night in London - a chance for technically-minded folk to get together over a curry and a few beers, and talk about servers and software and science fiction. This month, sysadmin entrepreneur David Mytton from Boxed Ice dropped in to talk about Server Density, what makes a good server monitoring product, blogging, SEO, and other fascinating topics. Also in attendance were sysadmin/devop Will Pink from jobsgopublic, Lisp hacker Donald Fisk, Python hacker Andy Nelis and special guest, information designer Matt Gourd.

If you couldn’t make it this time, don’t worry - it’s happening again next month. Watch this space for details!

Can you master Puppet?

Pay attention at the back. Reductive Labs just announced Puppet training dates for London:

Puppet Training London: March 29 - April 2

There are two courses available: Becoming a Puppet Master (3 days) covers Puppet and Puppetmaster configuration, classes, modules, definitions, resource types, the resource abstraction layer, virtual resources, exported resources, stored configs, metaparameters, dependencies, events, tags, environments, Puppet language patterns and best practices. There’ll be a mix of teaching and interactive class exercises using your own laptop and virtual machines. It costs £1595 for the 3 day course.

If you’re a seasoned old Puppet hand and you know all that stuff by heart, then the Puppet Developer course (2 days) is probably for you. You’ll get an introduction to Ruby for Puppet, as well as learning how to craft your own custom resource types and providers, develop custom functions and Facter facts. In addition there’s a segment on testing practices and RSpec for Puppet. The cost is £995 for two days.

Register for the training (get £100 off each course if you register before March 15th) or contact Scott Campbell to find out more. And don’t forget to bring an apple for the teacher.

Puppet dry run

Is the web site down?” asked my boss. Probably the five words I least want to hear, along with “I’ve run over your cat” and “I’ll pay you next month”.

Yes, the site was down - I’d made a minor Puppet change, and assumed this would roll out without any problems. Of course, the problem with assumptions is that they make an ASS out of U in front of your BOSS. Whenever I use Puppet to make changes to production servers, I like to do it by running Puppet manually, rather than waiting for the half-hourly auto-run, and I like to do a dry run first to see what’s going to happen.

Puppet’s dry-run feature is a powerful tool that’s often overlooked by busy sysadmins. Even if you test your Puppet manifests on a virtualised replica of your production site, which many people don’t have the time or the budget to do, pushing changes out live can have unforeseen side effects which are best avoided.

To dry-run Puppet, use the --noop flag:

# puppetd --test --noop
info: Caching catalog at /var/puppet/state/localconfig.yaml
notice: Starting catalog run
762a763
> This change will screw up your Apache server.
notice: //Node[soupnazi]/nagios::server/apache/File[/etc/httpd/conf/httpd.conf]/source: 
  is {md5}327dc6b2d2337e4467840c007a0edd61, should be puppet:///apache/httpd.conf (noop)
info: //Node[soupnazi]/nagios::server/apache/File[/etc/httpd/conf/httpd.conf]: Scheduling refresh 
  of Service[httpd]
notice: //Node[soupnazi]/nagios::server/apache/Service[httpd]: Would have triggered refresh from 1 
  dependencies
notice: Finished catalog run in 8.93 seconds

Puppet’s ‘noop’ (no-operation) mode shows you what would happen, but doesn’t actually do it. As you can see, Puppet reports that it would have updated the httpd.conf file and restarted Apache, and it helpfully shows us the line it would have inserted into the file:

> This change will screw up your Apache server.

Aren’t you glad you added the --noop flag? Of course, whenever you’re contemplating rolling out changes to a live server like this, you’ll want to stop the Puppet daemon running on the box temporarily (or disable the cron job if you run Puppet out of cron). The last thing you want is Puppet running automatically in “really do it” mode while you’re scrambling to fix the broken manifest.

Puppet’s dry run mode is not perfect - resource Y may fail because it requires resource X, for example, and resource X has not really been applied. Puppet isn’t so clever as to model the state of the box if each resource had been really applied - which is not surprising, because that’s a hard problem. Luckily, it doesn’t need to be perfect: dry-running Puppet in this way will catch a lot of small but potentially disastrous errors before they happen.

Opinions differ as to how meaningful or correct a dry-run mode can be, since changing the state of the box makes things unpredictable (Chef doesn’t have a dry run option at all), but I’ll take avoiding a disaster over academic arguments any day.

Thanks for getting the site back up,” said my boss. “Now, something else I meant to tell you. I’ll pay you next month.” Read more »

Syndicate content