This repo contains attribute-driven-API cookbooks maintained by Facebook. It's a large chunk of what we refer to as our "core cookbooks."

It's worth reading our Philosophy.md doc before reading this. It covers how we use Chef and is important context before reading this.

It is important to note that these cookbooks are built using a very specific model that's very different from most other cookbooks in the community. Specifically:

• It assumes an environment in which you want to delegate. The cookbooks are designed to be organized in a "least specific to most specific" order in the run-list. The run-list starts with the "core cookbooks" that setup APIs and enforce a base standard, which can be adjusted by the service owners using cookbooks later in the run-list.
• It assumes a "run from main" type environment. At Facebook we use Grocery Delivery to sync the main branch of our source control repo with all of our Chef servers. Grocery Delivery is not necessary to use these cookbooks, but since they were built with this model in mind, the versions never change (relatedly: we do not use environments).
• It assumes you have a testing toolset that allows anyone modifying later cookbooks to ensure that their use of the API worked as expected on a live node before committing. For this, we use Taste Tester.

Cookbooks in this repo all being with fb_ to denote that not only do they use the Facebook Cookbook Model, but that they are maintained in this repo.

Local cookbooks or cookbooks in other repositories that implement this model should not use this prefix, but should reference this document in their docs.

Unlike other cookbook models, we do not use resources as APIs, we use the node object. Configuration is modeled in arrays and hashes as closely and thinly as possible to the service we are configuring. Ideally, you should only have to read the docs to the service to configure it, not the docs to the cookbook.

For example, if the service we are configuring has a key-value pair configuration file, we will provide a simple hash where keys and values will be directly put into the necessary configuration file.

There are two reasons we use attribute-driven APIs:

1. Cascading configuration Since our cookbooks are ordered least specific (core team that owns Chef) to most specific (the team that owns this machine or service) it means that the team who cares about this specific instance can always override anything. This enables stacking that is not possible in many other models. For example, you can have a run-list that looks like:

   • Core cookbooks (the ones in this repo)
   • Site/Company cookbooks (site-specific settings)
   • Region cookbooks (overrides for a given region/cluster)
   • Application Category cookbooks (webserver, mail server, etc.)
   • Specific Application cookbook ("internal app1 server")

   So let's say that you want a specific amount of shared memory by default, but in some region you know you have different size machines, so you shrink it, but web servers need a further different setting, and then finally some specific internal webserver needs an even more specific setting... this all just works.

   Further, a cookbook can see the value that was set before it modifies things, so the 'webserver' cookbook could look to see what the value was (small or large) before modifying it and adjust it accordingly (so it could be relative to the size of memory that the 'region' cookbook set).

   Using resources for this does not allow this "cascading", it instead creates "fighting". If you use the cron resource to setup an hourly job, and then someone else creates a cron for that same job but only twice a day, then during each Chef run the cron job gets modified to hourly, then re-modified to twice a day.

2. Allows for what we refer to as "idempotent systems" instead of "idempotent settings." In other words, if you only manage a specific item in a larger config, and then you stop managing it, it should either revert to a less-specific setting (see #1) or be removed, as necessary.

   For example let's say you want to set a cron job. If you use the internal cron resource, and then delete the recipe code that adds that cronjob, that cron isn't removed from your production environment - it's on all existing nodes, but not on any new nodes.

   For this reason we use templates to take over a whole configuration wherever possible. All cron jobs in our fb_cron API are written to /etc/cron.d/fb_crontab. If you delete the lines adding a cronjob, since they are just entries in a hash, when the template is generated on the next Chef run, those crons go away.

   Alternatively, consider a sysctl set by the "site" cookbook, then overwritten by a later cookbook. When that later code is removed, the entry in the hash falls back to being set again by the next-most-specific value (i.e. the "site" cookbook in this case).

How you formulate your run-lists is up to your site, as long as you follow the basic rule that core cookbooks come first and you order least-specific to most-specific. At Facebook, all of our run-lists are:

recipe[fb_init], recipe[$SERVICE]

Where fb_init is similar to the sample provided in this repo, but with extra "core cookbooks."

We generally think of this way: fb_init should make you a "Facebook server" and the rest should make you a whatever-kind-of-server-you-are.

Grab a copy of the repo, rename fb_init_sample to fb_init, and follow the instructions in its README.md (coordinating guidance is in comments in the default recipe).

It is often useful to factor out logic into a library - especially logic that doesn't create resources. Doing so makes this logic easier to unit test and makes the recipe or resource cleaner.

Our standard is that all cookbooks use the top-level container of module FB, and then create a class for their cookbook under that. For example, fb_fstab creates a class Fstab inside of the module FB. We will refer to this as the cookbook class from here.

We require all cookbooks use this model for consistency.

Since we don't put anything other than other classes inside the top-level object, it's clear that a module is the right choice.

While there is no reason that a cookbook class can't be one designed to be instantiated, more often than not it is simply a collection of class methods and constants (i.e. static data and methods that can then be called both from this cookbook and others).

Below the cookbook class, the author is free to make whatever class or methods they desire.

When building a complicated Custom Resource, the recommended pattern is to factor out the majority of the logic into a module, inside of the cookbook class, that can be included in the action_class. This allows the logic to be easily unit tested using standard rspec. It is preferred for this module to be in its own library file, and for its name to end in Provider, ala FB::Thing::ThingProvider.

When more than 1 or 2 methods from this module are called from the custom resource itself, it is highly recommended you include it in a Helper class for clarity, ala:

action_class do
  class ThingHelper
    include FB::Thing::ThingProvider
  end
end

In this way, it is clear where methods come from.

You may have noticed that some of our cookbooks will extend the node object, while others have self-contained classes that sometimes require the node be passed as a parameter to some methods.

In general, the only time when extending the node is acceptable is when you are simply making a convenience function around using the node object. So, for example, instead of making people do node['platform_family'] = 'debian', there's a node.debian?. This is simply syntactic sugar on top of data entirely in the node.

In all other cases, one should simply have the node be an argument passed on, so as to not pollute the node namespace. For example, a method that looks at the node attributes, but also does a variety of other logic, should be in a cookbook class and take the node as an argument (per standard programming paradigms about clear dependencies).

Sometimes it is convenient to put a method directly in a recipe. It is strongly preferred to put these methods in the cookbook class, however there are some cases where methods directly in recipes make sense. The primary example is a small method which creates a resource based on some input to make a set of loops more readable.

Methods should not be put into templates. In general, as little logic as possible should be in templates. In general the easiest way to do this is to put the complex logic into methods in your cookbook class and call them from the templates.

Chef is an ordered system and thus is designed to fail a run if a resource cannot be converged. The reason for this is that if one step in an ordered list cannot be completed, it's likely not safe to do at least some of the following steps. For example, if you were not able to write the correct configuration for a service, then starting it may open up a security vulnerability.

Likewise, the Facebook cookbooks will err on the side of failing if something seems wrong. This is both in line with the Chef philosophy we just outlined, but also because this model assumes that code is being tested on real systems before being released using something like taste-tester and that monitoring is in place to know if your machines are successfully running Chef.

Here are some examples of this philosophy in practice:

• If a cookbook is included on a platform it does not support, we fail. It might seem like returning in this case is reasonable but there is a good indication the run-list isn't as-expected, so it's a great idea to bail out before this machine is mis-configured.
• If a configuration was passed in that we don't support, rather than ignore it we fail.

Many cookbooks rely on the service underneath and the testing of the user to be the primary validator of inputs. Is the software we just configured, behaving as expected?

However, sometimes it's useful to do our own validation because there are certain configurations we don't want to support, because the software may accepted dangerous configurations we want to catch, or because the user could pass us a combination of configurations that is conflicting or impossible to implement.

In this model, however, this must be done at runtime. If your implementation is done primarily inside of an internally-called resource, then this validation can also be done there. However, if your implementation is primarily a recipe and templates, doing the validation in templates is obviously not desirable. This is where whyrun_safe_ruby_blocks come in.

Using an ordinary ruby_block would suffice to have ruby code run at runtime to validate the attributes, however that means that the error would not be caught in whyrun mode. Since this validation does not change the system, it is safe to execute in whyrun mode, and that's why we use whyrun_safe_ruby_blocks: they are run in whyrun mode.

It is worth noting that this is also where you can take input that perhaps was in a structure convenient for users and build out a different data structure that's more convenient to use in your template.

This model intentionally draws the complexity of Chef into the "core cookbooks" (those implementing APIs) so that the user experience of maintaining systems is simple and (usually) requires little more than writing to the node object. However, the trade-off for that simplicity is that implementing the API properly can be quite tricky.

How to do this is a large enough topic that it gets its own document. However, some style guidance is also useful. This section assumes you have read the aforementioned document.

The three main ways that runtime-safety is achieved are lazy, templates, and custom resources. When should you use which?

The template case is fairly straight forward - if you have a template, read the node object from the within the template source instead of using variables on the template resource, and all data read is inherently runtime safe since templates run at runtime.

But what about lazy vs custom resources? For example, in a recipe you might do:

package 'thingy packages' do
  package_name lazy {
    pkgs = 'thingy'
    if node['fb_thingy']['want_devel']
      pkgs << 'thingy-devel'
    end
    pkgs
  }
  action :upgrade
end

Where as inside of a custom resource you could instead do:

pkgs = 'thingy'
if node['fb_thingy']['want_devel']
  pkgs << 'thingy-devel'
end

package pkgs do
  action :upgrade
end

Which one is better? There's not an exact answer, both work, so it's a style consideration. In general, there are two times when we suggest a custom resource:

The first is when you need to loop over the node in order to even know what resources to create. Since this isn't possible to (well, technically it's possible with some ugliness, but by and large not using the standard DSL), this must go into a custom resource. Example might be:

# This MUST be inside of a custom resource!
node['fb_thingy']['instances'].each do |name, config|
  template "/etc/thingy/#{instance}.conf" do
    owner 'root'
    group 'root'
    mode '644'
    variables({:config => config})
  end
end

The second is when you're using lazy on the majority of the resources in your recipe. If your recipe has 15 resources and you've had to pepper all of them with lazy, it's a bit cleaner to make a custom resource that you call in your recipe.

It's important here to reiterate: we're not referring to using a Custom Resource as an API, but simply making an internal custom resource, called only by your own recipe, as a way to simplify runtime safety.

Outside of these two cases, you should default to implementations inside of recipes. This is for a few reasons.

The first reason is that dropping entire implementations in custom resources leads to confusion and sets a bad precedent for how runtime-safety works. For example, consider the custom resource code we saw earlier where you assemble the package list in "naked" ruby:

pkgs = 'thingy'
if node['fb_thingy']['want_devel']
  pkgs << 'thingy-devel'
end

This code works fine in a resource, but serves as a bad reference for others - since this absolutely won't work in a recipe (even though it'll run).

The second reason is that quite often implementations need both compile-time and runtime code, and by blindly dropping the implementation into a custom resource, you can often miss this and create bugs like this:

# only safe because we're in a custom resource
packages = FB::Thingy.determine_packages(node)

package packages do
  action :upgrade
end

if node['fb_thingy']['want_cron']
  node.default['fb_cron']['jobs']['thingy_runner'] = {
    'time' => '* * * * *',
    'command' => '/usr/bin/thingy --quiet',
  }
end

service 'thingy' do
  action [:enable, :start]
end

Note here that while this code all seems reasonable in a custom resource (if statements are runtime safe when inside of a custom resource), that cronjob will never get picked up, because you're using an API at runtime, but APIs must be called at compile time and consumed at runtime. In reality, this needs to be in the recipe in order to work, and should look like this, in a recipe:

package 'thingy packages' do
  package_name lazy { FB::Thingy.determine_packages(node) }
  action :upgrade
end

node.default['fb_cron']['jobs']['thingy_runner'] = {
  'only_if' => proc { node['fb_thingy']['want_cron'] },
  'time' => '* * * * *',
  'command' => '/usr/bin/thingy --quiet',
}

service 'thingy' do
  action [:enable, :start]
end

In general, always start your implementation as a recipe and then escalate to Custom Resources where necessary.

You can set up kitchen using the same commands as in .github/workflows/ci.yml, but once Chef runs you won't have access to connect, so modify fb_sudo/attributes/default.rb and uncomment the kitchen block.

Then you can do bundle exec kitchen login <INSTANCE> after a failed run, and sudo will be passwordless so you can debug.

See the LICENSE file in this directory