February 23, 2014

What are Bags?

Chef-server offers an API for “data bags”. These are mostly unstructured JSON data with a strict two-level hierarchy:

bag1/
  item1
  item2
bag2/
  item1
  item2

Both bags and items offer basic CRUD operations, and items can be used with Chef’s search API so you can do simple key/value queries.

This is commonly used, and indeed considered by most to be a best-practice, for storing data that doesn’t map one-to-one with nodes. For example, having a bag calls users with one item for each of your system users with keys like name and ssh_keys. You can then write a single Chef recipe to loop over each item in the bag and create their user account, populate authorized_keys, and so on.

So what is the problem?

The data bags (hereafter: bags) system has two main use cases in my experience. The first is the one mentioned above, storing structured data as files in git and then using it in a recipe. The thing is, we already have that, they are just called cookbooks. Storing the data separately as JSON files which you manage in Git and then upload them to the bags API, and then do the exact same dance for the cookbook that consumes them is just more work for little gain.

A simpler solution is to use resources and recipes like with anything else in Chef. Each item maps to a single resource, and each bag to a recipe. The code used to process each bag can either go in an LWRP or a class-based resource. This means we get all the benefits of cookbook versioning as well as fewer moving pieces and a unified workflow. What do we lose? We can’t use JSON-based tools to process our user data and we can’t use Chef search on it. Most users don’t use either of these features, so this is not generally a significant downside.

The second use case is using the bags API as a simple database. An example would be storing which version of your application to deploy and updating that from Jenkins or similar. Unfortunately the Chef API is very susceptible to race conditions when you have multiple writers/updaters. At the first Chef community summit we discussed some additions to the API to improve this, but nothing has materialized so far. If you need a simple database like this, you are probably better off building a small API service with Flask/Sinatra/Express that provides the semantics you need. ZooKeeper and Etcd are also options, though that rabbit hole goes very deep and may be more complex than you need.

What about encrypted data bags?

Added a few years ago, encrypted data bags offer a limited way to store secret information in a data bag without the server being able to read it. Unfortunately secrets management is a very hard problem and encrypted bags only address a very specific attack vector. If you are using Hosted Enterprise Chef and you want to store something like database connection information in a bag (though I hope after reading this that you won’t) then you don’t want a compromise of Hosted Chef to leak your database passwords and other secrets. For this case, encrypted bags protect you from Chef Inc and vice versa.

For this to work you need to distribute a decryption key to every machine that has to read encrypted data. This means you must have a system in place to distribute secrets to all your servers, so why not just use that to distribute the original secret data and skip the whole bag thing?

Secrets management is easily worthy of its own post but a tl;dr is that we at Balanced are using S3 via the citadel cookbook with a close eye on Barbican and Red October. You could also do worse than chef-vault, though it is based around bags at heart so shares many of the same issues.

So I should just use resources for everything?

In most cases, yes. There are some more advanced uses for bags like using the same bag in two difference recipes which can solved by inverting things (build a new resource that handles both of the things you want to do with the data) but this doesn’t always work out cleanly. Another option is to make a recipe that simply stores the data in node.run_state and then process it exactly as you would with a bag in the other recipes.

Overall though, yes, take the code you used to have inside the loop which processed your bag and build a resource (LWRP or class-based), and then use that in a recipe.

How do I rewrite my recipe?

Let’s start with a users bag item like:

{
  "id": "asmithee",
  "comment": "Alan Smithee",
  "ssh_keys": ["ssh-rsa ..."]
}

Our current recipe looks like:

search(:users, '*:*').each do |u|
  user u['id'] do
    comment u['comment']
    supports manage_home: true
  end

  directory "/home/#{u['id']}/.ssh" do
    owner u['id']
  end

  file "/home/#{u['id']}/.ssh/authorized_keys" do
    owner u['id']
    mode '644'
    content u['ssh_keys'].join("\n")
  end
end

So in short, for each user, we create the user account, create a .ssh folder, and create an authorized_keys file.

To convert this to being resource-based, let’s create a mycompany_user LWRP:

# resources/default.rb
default_action(:create)

attribute(:comment)
attribute(:ssh_keys)


# providers/default.rb
action :create do
  user new_resource.name do
    comment new_resource.comment
    supports manage_home: true
  end

  directory "/home/#{new_resource.name}/.ssh" do
    owner new_resource.name
  end

  file "/home/#{new_resource.name}/.ssh/authorized_keys" do
    owner new_resource.name
    mode '644'
    content new_resource.ssh_keys.join("\n")
  end
end

As you can see, the provider code very closely mirrors our old recipe code, just using the new_resource helper instead of u. We can then convert our old bag item into a recipe:

mycompany_user 'asmithee' do
  comment 'Alan Smithee'
  ssh_keys ['ssh-rsa ...']
end

And again you can see a very tight correlation between the recipe code and the bag item. If you would like to see a more full-featured example of this, check out the balanced-user cookbook.