Paco Internals

Discussions on what Paco does under the hood.

Paco Architecture

What happens when you run “paco provision” to create cloud resources?

Paco will go through the following steps:

  1. PacoContext: Command line arguments are read and parsed. An object of the class paco.config.paco_context.PacoContext is created which holds the command-line arguments. Most notably, this object will have a .home attribute, which is the path to the Paco project and a .project attribute which will contain that project loaded as a model.
  2. Load the project config: After the home directory is set on PacoContext, then paco_ctx.load_project() will call paco.models.load_project_from_yaml with that directory. The paco.models loader will read all of the YAML files and construct Python objects. The Paco model is a tree of objects, the root node is a Project object. Every object in the tree has a name and __parent__ attribute. This allows any object to know it’s paco.ref by walking up the parents to the root node and concatenating the names.
  3. AccountContext: Next and object of the class paco.config.paco_context.AccountContext is created for the master account. This will ask the user for their MFA credentials. AccountContext objects manage connections to the AWS accounts.
  4. Global Initialization: Paco has Controllers which are objects which initialize and orchestrate the CloudFormation templates and stacks. They are also responsible for connecting the model to the stacks they create, so that resources can find the outputs that they create. Global controllers that are widely depended upon are initialized (Route53, S3 and SNS Topics). Finally once everything is almost ready, Service controllers are loaded - these are Paco Add-Ons. These are last in the process to give them a chance to react/modify the final set-up without limit.
  5. Scope Initialization: Depending on the scope that is being provisioned (e.g. netenv.mynet.dev or resource.s3) a controller of the appropriate type will be looked up and initialized.
  6. Perform Cloud Action: The cloud action (validate, provision, delete) is called on the controller for the scope. It is up to the controller to determine how it goes about doing that action, but most controllers follow the common pattern of iterating through their StackGroups and calling the cloud action on each StackGroup.

Stacks and Templates

AWS CloudFormation

Paco uses the AWS CloudFormation service to provision AWS resources and maintain resource state. CloudFormation has two core concepts: a template and a stack. A template is a CloudFormation document the declares resources. A stack is when a template is uploaded to AWS to create those resources. A stack will always belong to a specific account and region.

Paco has several Classes which it uses to model stacks and templates and control how they interact with the AWS CloudFormation service.

Controller

Controller objects initialize and set-up other objects. They create StackGroups and add Stacks to them. The can also interact with commands from the CLI.

Controllers also inject a resolve_ref_obj into model objects, to allow model objects to use Paco References to refer to Stack outputs.

PacoContext

The paco.config.paco_context.PacoContext class contains the arguments and options parsed from the CLI. PacoContext also makes a call to load a Paco project into a model and make the project root node available as a .project attribute.

The .get_controller(<controller-name>) method on PacoContext is used to fetch a controller. This ensures that controllers are initialized once and only once.

StackGroup

The paco.stack.stack_group.StackGroup class implements a StackGroup. A StackGroup is a logical collection of stacks that support a single concept. StackGroups apply the same operation against all Stacks that it contains and ensure that they are executed in the correct order and if necessary wait for stacks to be created if one stack depends upon the output of another stack.

StackGroups are often subclassed and the subclass adds logic to related to that subclasses purpose. For example, a BackupVault needs an IAM Role to assume. If you have a BackupVault Stack, you also need an IAM Role Stack with a Role. The BackupVaultsStackGroup adds the ability to create a Stack for that IAM Role.

Stack

The paco.stack.Stack class defines a Paco Stack. A Stack is connected to an account and region, and can fetch the state of the Stack as well as create, update and delete a stack in AWS.

Every Stack expects a StackTemplate object to be set on the .template attribute. This happens by calling add_new_stack() on a StackGroup. This method ensures that the Stack is created first, then the StackTemplate is created with the stack object passed in the constructor, after the new StackTemplate object is set on the .template attribute any commands that need to happen after are applied and the stack is given orders to the StackGroup.

Every Stack is created with a .resource attribute. This is the model object which contains the configuration for that Stacks template. The IResource interface in the models provides an is_enabled() method, and a .order and .change_protected attributes. This helps inform the stack if it should be modified, and in which order, or if it shouldn’t be touched at all.

Every Stack as a stack_ref property. This is normally the paco.ref for the .resource but it can also be extended with a support_resource_ref_ext when the Stack is created. For example, an ASG resource needs a LogGroup stack where it will log to. This is a supporting resource that isn’t explicitly declared in the configuration. The same happens for Alarms, which add a ‘.alarms’ extension to the ref.

StackTemplate

The paco.cftemplates.StackTemplate class defines a Paco Template. A StackTemplate has a .body attribute which is a string of CloudFormation in YAML format.

A StackTemplate requires a Stack object to be passed to the constructor. In Paco, a StackTemplate can provision a CloudFormation template in several different locations and potentially look different in each of those locations. The StackTemplate has access to the Stack. The StackTemplate typically sets Parameters on the Stack. It can also change the behaviour of Stack updates, for example, certain Parameters can be set to use the previously existing value of the Stack.

A troposphere.Template class defines a StackTemplate’s .template attribute. Troposphere is an external Python dependency of Paco. It’s a great library with a complete and updated representation of CloudFormation objects. However a StackTemplatecan provide any kind of return string, so simple Python strings can also be constructed and set as the template body.

When Paco uses a StackTemplate it never instantiates it directly. It’s a base class that resource specific templates inherit from. These subclasses are responsible for creating the template.

The .paco-work directory

Paco creates a directory in every Paco project named .paco-work. This directory contains several sub-directories that Paco will read/write to while it’s working.

.paco-work/applied

Paco starts a provision command by showing you a diff of the configuration from the last provision. It does this by keeping a cache of YAML configuration files after it applies them here.

Paco will also show you changes between previously CloudFormation stacks and Parameters and the new ones it wants to apply. Paco creates a cache of stacks here when after they have been applied.

If this directory gets out-of-sync then Paco can skip updates to Resrouces believing that they haven’t changed. You can remedy this by using the -n --nocahce flag with the Paco CLI.

Alternatively, you could run rm -rf .paco-work/applied/cloudformation to remove this cache and Paco will simply run slower on it’s next run as it fetches state from CloudFormation.

.paco-work/build
This is a scratch space that Paco can use. For example, the EC2LaunchManager creates a zip file bundles of files used to configure EC2 instances. These zip files are created in here.
.paco-work/outputs
Stack outputs are cached here. These outputs are organized according to the structure of the Paco model as opposed to the structure of the CloudFormation stacks.