This is part 3 of my Cloud programming series where I explore Chef, a very cool infrastructure automation tool.

When I started this blog series, I had a project in mind which required me to learn Ruby and how to use the Fog cloud abstraction library so that I could start and stop groups of Ninefold virtual servers at will. That project is still on the drawing board :), for now I have moved on to something far more important to our customers.

On my rubyist reinvention journey, I kept coming across something called Chef, an open source configuration management tool managed by OpsCode out of Seattle. And as I listened to people’s enthusiasm about the power of Chef, I realised that this was the answer to two of our biggest challenges:

  • managing Ninefold’s expanding infrastructure without expanding our ops team at the same rate;
  • providing simple to use app deployment for our customers that also allows them to leverage the full power of our production grade multi-zone infrastructure.

Here’s the quick explanation of Chef: your infrastructure as code.

OK, a bit cryptic, so try this longer definition:

  1. Chef is a systems integration framework, designed to bring configuration management to your entire infrastructure. It is an open source project with a vibrant community contributing to the base software and sharing cookbooks (abstracted definitions of resource configuration).

  2. Servers are known as Nodes. Nodes have attributes that describe the current or desired state of its configuration and provide one of several mechanisms for communicating config changes.

  3. Cookbooks are collections of files which configure the nodes to a desired end-state. Recipes are written using ruby and a recipe DSL which is readily extensible.

  4. The various elements of the configuration are known as resources e.g. apache web server, iptables firewall, mysql database etc. Recipes are (should be) idempotent i.e. re-running the recipe on a node multiple times always returns the system to an identical state.

In terms of software, chef-client runs on the node and talks to a server which holds the cookbooks and attribute status of all registered nodes. There is a special chef-solo version of the server which runs on the node, an open source chef-server, an OpsCode Hosted Chef server as a service and a licensed, multi-tenant, HA edition called Private Chef.

A chef-client run operates in two stages. During compilation, the various code files – libraries, attributes, definitions, recipes etc – are loaded and evaluated. A resource list is built from the DSL in each recipe in the order they appear. During convergence, each resource is configured according to the relevant DSL configuration. Rather than go into a lot more detail about Chef – and believe me, there is a lot of detail to get your head around – I am going to share my Top Four Tips for Chef.

Tip 1 - Understand the difference in Compilation vs Convergence

During compilation, any normal Ruby code is evaluated when the relevant file – attribute, library, recipe etc – is first loaded. If you want to calculate some value after a resource has been converged and then save it in an attribute, you will need to delay the evaluation by wrapping the relevant code in a ruby_block resource:

ruby_block save node state do
  block do
    node.set['some_host']['configured_time'] = Time.now.to_s
    node.save
  end
  action :create
end

For a more detailed and highly readable explanation, check out the Anatomy of a Chef Run.

Tip 2 - The cause of chef run terminating with “undefined method ‘[]’ for nil:NilClass”

This initially perplexing message is almost always due to trying to reference a node attribute that doesn’t exist. For example if I am expecting to use node['my_cookbook']['app']['version'] somewhere in my recipe, I will get the value ‘nil’ if node['my_cookbook']['app'] exists but there is no ['version'] attribute. But if node['my_cookbook']['app'] also doesn’t exist then I get the dreaded undefined method ‘[]’ since I am in effect trying to reference nil['version'] and nil doesn’t have a ‘[]’ method.

To avoid this common occurrence in your early recipe writing, ensure attributes exist before you use them. Two methods are:

if node.attribute?('version')
  # there is some attribute somewhere called 'version'
  # perhaps not the best method
end

if node['my_cookbook']['app'] && node['my_cookbook']['app']['version']
  # there is an 'app' and a 'version' for 'app'
end

Tip 3 - Essential knife plug-in: knife-block

Knife is the command line tool for managing your chef server and nodes from your workstation. If you use Chef Server it is likely that you will have more than one of these. At Ninefold, we use Private Chef and as well as multiple instances of the server, we also have multiple organisations within each server – you can think of each organisation as a logical Chef Server with separate cookbooks etc but sharing the one client-validation key.

Managing multiple knife.rb configuration files is a bit of a nightmare, and requires you to place the relevant knife.rb, chef-validator.pem and client.pem files into each project. But then if you want to push a cookbook that you are working on from your development chef to your testing chef, there is a spot of juggling required. That is until you install the knife-block plugin by Green and Secure IT Limited.

Place all your knife.rb and .pem files into your /home/.chef/ directory and rename each knife.rb file as knife-{something}.rb e.g. knife-dev.rb, knife-test.rb, knife-wazza-is-awesome.rb. knife block dev will create a knife.rb symlink to the knife-dev.rb.

knife walks up the directory structure looking in ../.chef/ to find the knife.rb configuration so placing all those files in /home/.chef/ means knife will always find the configuration required for your current chef context. At any time you can find out your choices by

$ knife block list
  The available chef servers are:
    * dev
    * test
    * wazza-is-awesome [ Currently Selected ]

And switch context simply using

$ knife block test
  The knife configuration has been updated to use test

$ knife block list
  The available chef servers are:
    * dev
    * test [ Currently Selected ]
    * wazza-is-awesome

Tip 4 - Manage cookbook dependencies using Berkshelf

As Berkshelf states “If you’re familiar with Bundler, then Berkshelf is a breeze”.

A cookbook’s metadata.rb file uses ‘depends’ clauses to specify cookbook dependencies – at Ninefold, we almost always specify exact versions to isolate us from potential breaking changes. Chef ensures that these versions are loaded onto the node at the start of a run provided they are present on the Chef Server. How do they get on to the server in the first place? You upload them using knife. But if you have a number of cookbook development projects and if you are reliant on specific branch versions of community cookbooks, then managing this process is very difficult. Until Berkshelf.

And ermagherd, Berkshelf ers ersum!

Cookbooks can be easily managed in single repositories. Dependency data can be drawn from the cookbook’s metadata.rb and the source of dependent cookbooks can be defined in various ways. Cookbooks are installed into /home/.berkshelf and this is where they are sourced by default but if missing from there they can be sourced from the Opscode Community Cookbooks site, a specific branch of a git repository, a path on the workstation or from a chef server. Once the cookbooks have been installed or updated they can be bulk uploaded into the Chef Server.

Ninefold’s cookbook development and customer provisioning process makes extensive use of Berkshelf and I highly recommend reading an introduction to authoring cookbooks by Jamie Winsor, the creator of Berkshelf.

Originally published at Ninefold (2010-2015), a cloud services provider I helped found.