This document covers the Chef two-phase run system, specifically geared towards how that interacts with the API model we use with Chef at Facebook. If you're not familiar with our model, you should read our Philosophy document and our Cookbooks README.md first. If you are completely new to Chef, you might want to start with some introductory material before delving into this.
Note that this document describes the general cases. Everything said here has exceptions and qualifiers.
The full explanation of what happens during a Chef client run can be found here, but the aim of this document is to elaborate on what happens specifically during the compile-time and run-time phases.
We start with a run_list, which expands into a collection of recipes.
During the compile phase of a Chef run, the recipes are run in the order of the
run_list (or as they are included by include_recipe
statements). As the
recipes run, they result in a list of resources - the Chef resource
collection. When the last recipe has been run, the Chef client run moves onto
the run phase. During the run-time phase, each of the resources is executed in
order of the resource collection. Each resource is "converged" - the point
where Chef actually inspects the system state and enforces the configured state,
if they differ.
Take for example this toy recipe: fb_awesomesoft/recipes/default.rb
package 'awesomesoft' do
action :upgrade
end
template '/etc/awesomesoft.conf' do
source 'awesomesoft.conf.erb'
end
service 'awesomed' do
action [:enable, :start]
end
At compile-time, this recipe evaluates into a list of resources:
package awesomesoft | template awesomesoft.conf | service awesomed |
---|---|---|
action :upgrade |
action :default |
action [:enable, :start] |
If this recipe was the full run_list, then after building this list of resources, the Chef client would proceed from the compile-time into the run-time, running the associated action against each resource in order.
It is sometimes useful to understand what actually happens under the hood here:
compile-time is actually building objects (a Template
object or a Package
object, for example), and storing it along with a list of actions, on a list.
At run-time, it simple calls the run_action
method on each object with the
list of corresponding actions.
In the Facebook model, we build upon the Chef node object to define configuration data, and then implement that configuration by reading from the same node object. This means that the order is critical; if the implementation code reads the configured value prior to the value being defined, the system falls apart.
To make this work, we enforce the following pattern:
- set values at compile-time
- read them at run-time
Therefore, to use a cookbook's API, one sets the corresponding node attribute in
their recipe at compile time. This is denoted with node.default
to specify
that this is a writer:
node.default['fb_swap']['enabled'] = false
This attribute might be set in multiple different recipes, to different values. This is okay, and by design; this follows the last writer wins model, such that whatever recipe set the attribute last will have that value be the one that is actually implemented. Typically in the run_list a later recipe will be more specific to the system, and therefore better suited to define the desired state.
In this example, within the
fb_swap
cookbook, which actually implements this setting, the resources which read this
value are only allowed to do so at run-time. This is denoted with just
node
, to indicate this is a reader:
# e.g. within a template erb (which is rendered at run-time)
enabled = <%= node['fb_swap']['enabled'] %>
These rules - to only write to node attributes during compile time, and only read the same at run time, are only required for node attributes which make up an API.
Many node attributes - like those build by Ohai, including information populated via Ohai plugins - can be safely read from recipes at compile time. They are not part of an API.
Cookbook APIs should define any node attributes which are part of their API
within said cookbook's README. All attributes defined by a cookbook must be
under its namespace (node[<cookbook_name>]
) and treated like an API unless
specifically prefixed with an underscore to denote it is internal. Wherever
possible, non-API data should be put in local variables instead.
You'll need to know whether an attribute will be evaluated at compile time or run time.
You can assume attributes will be evaluated at compile time, except for the following cases:
Case | Details | Example |
---|---|---|
ruby within a template | templates are rendered at runtime |
in some_template.rb:version = <%= node['some']['attribute'] %> |
guards | `only_if`, `not_if`: specify a code block which will be resolved at run time |
package 'some package' do
only_if { node['something']['enabled'] }
end |
lazy resource properties | Specify a code block which will be resolved at run time |
package 'some package' do
version lazy { node['some']['version'] }
end |
provider actions | The vast majority of configuration enforcement happens in providers, which run at run time. |
action :run do
my_temp_var = node['something']['or_other']
...
end |
ruby blocks | An arbitrary block of ruby which will be run at run time. |
ruby_block 'some code' do
block do
my_temp_var = node['something']['someval'] }
end
end |
Consider again fb_awesomesoft
, this time with a new (but broken) feature - an
API to allow the version of the package to be specified. A second recipe sets
this value. The run_list order is fb_awesomesoft::default
,
fb_someapp::default
.
We specify a default value for our attribute
fb_awesomesoft/attributes/default.rb
:
default['fb_awesomesoft'] = {
'version' => 1,
}
We implement the new setting by specifying a version
property on the package
resource in fb_awesomesoft/recipes/default.rb
:
...
# bad example, don't do this!
package 'awesomesoft' do
version node['fb_awesomesoft']['version']
action :install
end
And some app sets it in fb_someapp/recipes/default.rb
:
node.default['fb_awesomesoft']['version'] = 42
This doesn't work, because version
in the package resource is evaluated at
compile time, such that it reads the value prior to all recipes having a
chance to set it. As a result, the assembled resource collection specifies a
version of 1
, based on the default value:
package awesomesoft | template awesomesoft.conf | service awesomed |
---|---|---|
action :upgrade |
action :default |
action [:enable, :start] |
version 1 |
When runtime begins, the version set by chef will not incorporate the value set
by fb_someapp
.
To make the implementation work correctly, we need the implementing recipe to
read the attribute at run time. In this case, we can postpone the resolution
of the version property using lazy
in fb_awesomesoft/recipes/default.rb
:
...
package 'awesomesoft' do
version lazy { node['fb_awesomesoft']['version'] }
action :install
end
With a lazy property, the value saved into the resource collection is a proc, which will be evaluated at run time. The resource collection is thus:
package awesomesoft | template awesomesoft.conf | service awesomed |
---|---|---|
action :upgrade |
action :default |
action [:enable, :start] |
version proc { node['fb_awesomesoft']['version'] } |
When the package resource is converged at run-time, it will evaluate the node
attribute, and correctly set it to 42
.
Take this case:
resource 'some_thing' do
only_if { node['fb_awesomesoft']['enabled'] }
end
resource 'some_other_thing' do
only_if { node['fb_awesomesoft']['enabled'] }
end
This code looks like it could be simplified into the following:
# BAD - DON'T DO THIS
if node['fb_awesomesoft']['enabled'] {
resource 'some_thing'
resource 'some_other_thing'
}
This doesn't work, because the evaluation of
if node['fb_awesomesoft']['enabled']
happens at compile time, and if some
recipe later in the run list were to change it, it would be too late; the
resource would not have been defined. In the original (correct) code:
only_if { node['fb_awesomesoft']['enabled'] }
is a gate, which is evaluated
at run time, and is therefore API safe.
If you have to read the contents of one API in order to write another API, you
can often achieve this by using a whyrun_safe_ruby_block
which is prior to
the implementing resource. For example, to add an entry to the hash for each
fb_timers
job in an API-safe way, you can do:
whyrun_safe_ruby_block 'set fb_timer XAR environments' do
block do
node['fb_timers']['jobs'].each_key do |name|
node.default['fb_timers']['jobs'][name][
'service_options']['Environment=XAR_MOUNT_SEED'] = name
end
end
end
Since this is a ruby_block
, the read will happen at run time, and as long as
this resource is before the implementing resource (in this case the
fb_timers_setup 'fb_timers system setup'
resource in fb_timers
) this will
still be API-safe.
Library calls are often confusing. The important thing to remember is that library code is not inherently compile-time nor runtime - it's simply where you call it. So, for example, in a recipe, this:
# do not do this
package FB::Thingy.determine_packages(node) do
action :upgrade
end
where determine_packages
is selecting packages based on the API attributes
set in the node, is not runtime-safe. The exact same code inside of a
custom resource would be runtime-safe. Or doing the following in a recipe
instead would also be runtime-safe:
package 'thingy packages' do
package_name lazy { FB::Thingy.determine_packages(node) }
action :upgrade
end
And, just to be clear, if a method never references the node, or anything in the node (or only references Ohai data), then it is inherently safe to call any time.