packer-cn/website/source/docs/extend/builder.html.markdown

7.6 KiB

layout
docs

Custom Builder Development

Builders are the components of Packer responsible for creating a machine, bringing it to a point where it can be provisioned, and then turning that provisioned machine into some sort of machine image. Several builders are officially distributed with Packer itself, such as the AMI builder, the VMware builder, etc. However, it is possible to write custom builders using the Packer plugin interface, and this page documents how to do that.

Prior to reading this page, it is assumed you have read the page on plugin development basics.

Warning! This is an advanced topic. If you're new to Packer, we recommend getting a bit more comfortable before you dive into writing plugins.

The Interface

The interface that must be implemented for a builder is the packer.Builder interface. It is reproduced below for easy reference. The actual interface in the source code contains some basic documentation as well explaining what each method should do.

type Builder interface {
	Prepare(...interface{}) error
	Run(ui Ui, hook Hook, cache Cache) (Artifact, error)
	Cancel()
}

The "Prepare" Method

The Prepare method for each builder is called prior to any runs with the configuration that was given in the template. This is passed in as an array of interface{} types, but is generally map[string]interface{}. The prepare method is responsible for translating this configuration into an internal structure, validating it, and returning any errors.

For multiple parameters, they should be merged together into the final configuration, with later parameters overwriting any previous configuration. The exact semantics of the merge are left to the builder author.

For decoding the interface{} into a meaningful structure, the mapstructure library is recommended. Mapstructure will take an interface{} and decode it into an arbitrarily complex struct. If there are any errors, it generates very human friendly errors that can be returned directly from the prepare method.

While it is not actively enforced, no side effects should occur from running the Prepare method. Specifically, don't create files, don't launch virtual machines, etc. Prepare's purpose is solely to configure the builder and validate the configuration.

In addition to normal configuration, Packer will inject a map[string]interface{} with a key of packer.DebugConfigKey set to boolean true if debug mode is enabled for the build. If this is set to true, then the builder should enable a debug mode which assists builder developers and advanced users to introspect what is going on during a build. During debug builds, parallelism is strictly disabled, so it is safe to request input from stdin and so on.

The "Run" Method

Run is where all the interesting stuff happens. Run is executed, often in parallel for multiple builders, to actually build the machine, provision it, and create the resulting machine image, which is returned as an implementation of the packer.Artifact interface.

The Run method takes three parameters. These are all very useful. The packer.Ui object is used to send output to the console. packer.Hook is used to execute hooks, which are covered in more detail in the hook section below. And packer.Cache is used to store files between multiple Packer runs, and is covered in more detail in the cache section below.

Because builder runs are typically a complex set of many steps, the multistep library is recommended to bring order to the complexity. Multistep is a library which allows you to separate your logic into multiple distinct "steps" and string them together. It fully supports cancellation mid-step and so on. Please check it out, it is how the built-in builders are all implemented.

Finally, as a result of Run, an implementation of packer.Artifact should be returned. More details on creating a packer.Artifact are covered in the artifact section below. If something goes wrong during the build, an error can be returned, as well. Note that it is perfectly fine to produce no artifact and no error, although this is rare.

The "Cancel" Method

The Run method is often run in parallel. The Cancel method can be called at any time and requests cancellation of any builder run in progress. This method should block until the run actually stops.

Cancels are most commonly triggered by external interrupts, such as the user pressing Ctrl-C. Packer will only exit once all the builders clean up, so it is important that you architect your builder in a way that it is quick to respond to these cancellations and clean up after itself.

Creating an Artifact

The Run method is expected to return an implementation of the packer.Artifact interface. Each builder must create their own implementation. The interface is very simple and the documentation on the interface is quite clear.

The only part of an artifact that may be confusing is the BuilderId method. This method must return an absolutely unique ID for the builder. In general, I follow the practice of making the ID contain my GitHub username and then the platform it is building for. For example, the builder ID of the VMware builder is "mitchellh.vmware" or something similar.

Post-processors use the builder ID value in order to make some assumptions about the artifact results, so it is important it never changes.

Other than the builder ID, the rest should be self-explanatory by reading the packer.Artifact interface documentation.

Provisioning

Packer has built-in support for provisioning, but the moment when provisioning runs must be invoked by the builder itself, since only the builder knows when the machine is running and ready for communication.

When the machine is ready to be provisioned, run the packer.HookProvision hook, making sure the communicator is not nil, since this is required for provisioners. An example of calling the hook is shown below:

hook.Run(packer.HookProvision, ui, comm, nil)

At this point, Packer will run the provisioners and no additional work is necessary.

Note: Hooks are still undergoing thought around their general design and will likely change in a future version. They aren't fully "baked" yet, so they aren't documented here other than to tell you how to hook in provisioners.

Caching Files

It is common for some builders to deal with very large files, or files that take a long time to generate. For example, the VMware builder has the capability to download the operating system ISO from the internet. This is timely process, so it would be convenient to cache the file. This sort of caching is a core part of Packer that is exposed to builders.

The cache interface is packer.Cache. It behaves much like a Go RWMutex. The builder requests a "lock" on certain cache keys, and is given exclusive access to that key for the duration of the lock. This locking mechanism allows multiple builders to share cache data even though they're running in parallel.

For example, both the VMware and VirtualBox support downloading an operating system ISO from the internet. Most of the time, this ISO is identical. The locking mechanisms of the cache allow one of the builders to download it only once, but allow both builders to share the downloaded file.

The documentation for packer.Cache is very detailed in how it works.