Stratosphere: AWS CloudFormation in Haskell

AWS CloudFormation is a system that provisions and updates Amazon Web Services (AWS) resources based on declarative templates. Common criticisms of CloudFormation include the use of JSON as the template language and limited error-checking, often only available in the form of run-time errors and stack rollbacks. By wrapping templates in Haskell, we are able to easily construct them and help ensure correctness.

The goals of stratosphere are to:

• Build a Haskell EDSL to specify CloudFormation templates. Since it is embedded in Haskell, it is type-checked and generally much easier to work with than raw JSON.
• Have a simple checking/linting system outside of the types that can find common errors in templates.
• Be able to also read valid CloudFormation JSON templates so they can be type-checked. This also gives us free integration tests by using the huge amount of example templates available in the AWS docs.

Example

Here is an example of a Template that creates an EC2 instance, along with the JSON output:

{-# LANGUAGE OverloadedLists #-}
{-# LANGUAGE OverloadedStrings #-}

module Main where

import qualified Data.ByteString.Lazy.Char8 as B
import Stratosphere

main :: IO ()
main = B.putStrLn $ encodeTemplate instanceTemplate

instanceTemplate :: Template
instanceTemplate =
  template
  [ resource "EC2Instance" (
    EC2InstanceProperties $
    ec2Instance
    "ami-22111148"
    & eciKeyName ?~ (Ref "KeyName")
    )
    & deletionPolicy ?~ Retain
  ]
  & description ?~ "Sample template"
  & parameters ?~
  [ parameter "KeyName" "AWS::EC2::KeyPair::KeyName"
    & description ?~ "Name of an existing EC2 KeyPair to enable SSH access to the instance"
    & constraintDescription ?~ "Must be the name of an existing EC2 KeyPair."
  ]

{
  "Description": "Sample template",
  "Parameters": {
    "KeyName": {
      "Description": "Name of an existing EC2 KeyPair to enable SSH access to the instance",
      "ConstraintDescription": "Must be the name of an existing EC2 KeyPair.",
      "Type": "AWS::EC2::KeyPair::KeyName"
    }
  },
  "Resources": {
    "EC2Instance": {
      "DeletionPolicy": "Retain",
      "Type": "AWS::EC2::Instance",
      "Properties": {
        "KeyName": {
          "Ref": "KeyName"
        },
        "ImageId": "ami-22111148"
      }
    }
  }
}

Please see the examples directory for more in-depth examples.

Value Types

CloudFormation resource parameters can be literals (strings, integers, etc), references to another resource or a Parameter, or the result of some function call. We encapsulate all of these possibilities in the Val a type.

We recommend using the OverloadedStrings extension to reduce the number of Literals you have to use.

Note that CloudFormation represents numbers and bools in JSON as strings, so we had to some types called Integer' and Bool' to override the aeson instances. In a future version we plan on using our own JSON encoder/decoder to get around this.

Lenses

Almost every CloudFormation resource has a handful of required arguments, and many more optional arguments. Each resource is represented as a record type with optional arguments wrapped in Maybe. Each resource also comes with a constructor that accepts required resource parameters as arguments. This allows the user to succinctly specify the resource parameters they actually use without adding too much noise to their code.

To specify optional arguments, we recommend using the lens operators & and ?~. In the example above, the ec2Instance function takes the AMI as an argument, since it is required by the EC2Instance resource type. Then, the optional EC2 key name is specified using the & and ?~ lens operators.

This approach is very similar to the approach taken by the amazonka library. See this blog post for an explanation.

Auto-generation

All of the resources and resource properties are auto-generated from JSON files and are placed in library-gen/. The gen/ directory contains the auto-generator code and the JSON model files. We include the library-gen/ directory in git so the build process is simplified. To build library-gen from scratch and then build all of stratosphere, just run the very short build.sh script. You can pass stack args to the script too, so run ./build.sh --fast to build the library without optimization. This is useful for development.

In the future, it would be great to not have to include the auto-generated code in git.

Also, there is a file called scraper.py that scrapes a given CloudFormation resource documentation page to produce the JSON model. It isn't perfect, but it helps a lot.

Contributing

Feel free to raise any issues, or even just make suggestions, by filing a Github issue.

Future Work

The library is usable in its current state and it is already much more enjoyable to work with than writing JSON templates by hand, but there are of course a few possible future improvements:

• Not all resources implemented. Adding resources is very easy though. Just request them and I will implement them :)
• Implement basic checker for things like undefined Refs and duplicate field names. This stuff would be too unwieldy to do in types, and performing a checking pass over a template should be pretty straightforward.
• Use a custom JSON encoder so the templates look a little more idiomatic. We also create a lot of empty whitespace and newlines using aeson-pretty. There are limits on the size of CloudFormation templates, and we want readable output without hitting the limits. Also, we have some newtypes that just exist to override aeson instances, and we could get rid of those.
• Use a custom JSON decoder with useful error messages. Although we don't use them, we have implemented FromJSON instances for everything. Theoretically, stratosphere could be used as a checker/linter for existing JSON CloudFormation templates.