This is a replacement for travis-yaml.

Travis CI build config processing

This library is used for parsing, normalizing, and validating Travis CI build configuration (.travis.yml).

It has three main functions:

  • Defining and producing a spec for build configuration.
  • Applying the spec to an actual configuration in order to normalize and validate it.
  • Expanding a build config to job matrix configs.

The library produces a specification that is used internally, but also can be served on the API in order to be consumed by, for example:

  • The UI build process (stop duplicating "expand keys").
  • A travis.yml wizard that fully conforms the specification.

Applying the specification to a build configuration produces messages on the levels info, warn, and error. Messages include a message code (e.g. unknown_key, unsupported_value etc.) and metadata (e.g. the given key, value) so they can be translated to human readable messages. These can be used in the UI, and links to the documentation, suggesting fixes.

yaml = 'rvm: 2.3'
config = Travis::Yaml.load(yaml)

# {
#   language: 'ruby',
#   os: ['linux'],
#   dist: 'trusty',
#   sudo: false
# }

# [
#   [:info, :language, :default, key: :language, default: 'ruby'],
#   [:info, :root, :alias, { alias: :rvm, name: :ruby }]
# ] { |msg| Travis::Yaml.msg(msg) }
# [
#   '[info] on root: missing :language, defaulting to: default',
#   '[info] on language: :rvm is an alias for :ruby, using :ruby'
# ]


Diverging from the YAML specification this library does not:

  • convert numbers to integer or float types
  • convert the strings "yes" or "on" to the boolean true
  • convert the strings "no" or "off" to the boolean false

For these reasons:

  • Version numbers like 1.10 used to be truncated to 1.1 by the YAML parser, causing confusion, and requiring users to specify these as strings ("1.10") instead.
  • Deployment config uses the key on. This being converted to true as a hash key does not make sense for us.
  • Integers add an unnecessary type to the system. Build config values that could be given as integers are going to be passed through the Bash build script. Bash does not distinguish integers from strings.


Borrowing from YAML's terminology the library uses:

  • Map for hashes (key value mappings)
  • Seq for arrays (sequence)
  • Scalar for all other types (bool, string, secure string)

Also, a Fixed type is a Scalar that has a number of known and allowed values. Examples for keys that hold fixed types are: language, os, dist etc.

A Map can be strict or not strict. Maps are strict by default. A strict map disallows keys that are not known.

Some scalars are marked as wanting a Secure, so we can alert the user if they pass a string.

Build config specification

The spec is included to the repository as spec.json.

The spec is defined in Spec::Def, and can be produced by:


Classes in Spec::Def use classes in Spec::Type to build a tree of nodes that define allowed keys, types, and various options in a readable and succinct way (using a DSL for describing the spec).

This tree is then serialized to a lengthy Hash which serves as a specification, and both an internal and, optionally, external API.

A good starting point for exploring the definition is the root node.

Examples for various nodes on this specification can be found in the tests, e.g. for the git, heroku, or os nodes.

Most nodes can be sufficiently specified by mapping known keys (e.g. language) to types (fixed, scalar, map, seq) with certain options, such as:

  • values: known values on a scalar type
  • default: default value
  • required: the node has to have a value
  • cast: try casting the node value to another type
  • edge: the node represents an experimental feature
  • flagged: the node represents a feature that requires a feature flag
  • only: the node is valid only if the given value matches the current language or os
  • except: the node is valid only if the given value does not match the current language or os
  • expand: the node represents a matrix key for matrix expansion

The fully expanded spec would be huge due to the repition of large sections that need to be included on N+1 child sections. For example, the deploy.on (which is allowed on currently 42 deployment providers) section allows to include all language specific sections, such as rvm, python etc., and so do the matrix.include, matrix.exclude, and matrix.allow_failures sections.

In order to keep the JSON payload reasonably small nodes can specify an option include in order to include shared sections. Shared sections are kept on the key includes on the root node and will be expanded into the respecitive target nodes at load time, and kept in memory from there on. We currently have 2 sections that can be included:

  • job (shared by root, matrix.*, keys like addons, branches, caches etc.)
  • support (language specific keys, such as rvm, python etc.)

Loading the spec

Before the spec can be applied to an actual build configuration it will be expanded (i.e. shared sections will be included to nodes that require them), and turned into an object oriented representation, so non-parametrized methods can be memoized for better performance.

The method Travis::Yaml.expanded returns the fully expanded, object oriented tree.

Applying the spec to a build config

This representation of the spec can be applied to a build configuration by:

# given a YAML string

# given a Ruby hash

Both methods also accept an optional options hash. There is one known option:

Travis::Yaml.apply(yaml, alert: true)

The option alert rejects plain strings and adds an error message on nodes that expect a secure string.

When the spec is applied to a build configuration three things happen:

  • The config is turned into an object oriented representation as well by the way of calling This method uses classes in Doc::Value in order to build a tree of nodes that maps to the given build config hash.

  • The config structure is normalized by the way of calling Doc::Change.apply. This method applies various strategies in order to attempt to fix potential problems with the given structure, such as typos, misplaced keys, wrong section types. In some cases it will store messages on the resulting tree. Change strategies are determined based on the type of the given node. Some strategies can be required by the spec for certain sections that need very specific normalizations, such as cache, env, notifications.

  • The resulting config is validated by the way of calling Doc::Validate.apply. This method applies various validations, and sets default values. It also stores (most of the) messages on the resulting tree. Sections also can require specific validations. The only section specific validation currently is template (which validates used var names in notification templates).

Examples of type specific change strategies:

  • alert: unset the value if the given value is a string while a secure string is expected
  • cast: try casting the value to another type if required by the spec (e.g. the string "true" to the boolean true)
  • downcase: downcase a string if required by the spec
  • keys: add required keys, and attempt to fix an unknown key by removing special chars and finding typos (uses a dictionary, as well levenshtein and similar simple strategies)
  • merge: merge an sequence of maps into a single maps if the spec requires a map (common mistake)
  • migrate: move the section to either a child or parent section if the key is known on that section and the resulting section validates without errors
  • pick: pick the first value of a given sequence for a scalar node
  • repair: repair bash command strings that have been broken up into maps by the YAML parser (e.g. script strings that contain an unescaped colon)
  • value: de-alias fixed node values, and try fixing typos in unknown values
  • wrap: wrap the given node into a sequence if required by the spec (e.g. os needs to result in an array)

Section specific change strategies:

  • enabled: normalize enabled and disabled values, set enabled if missing (used by, for example, notifications)
  • inherit: inherit certain keys from the parent node if present (used by, for example, notifications)

Examples of the validations:

  • alias: use the other value if the given value is an alias for another value
  • default: use a default value as required by the spec if the node does not have a value
  • edge: add an info level message if this node type is used (experimental features)
  • empty: drop an empty section
  • flags: add an info level message if this node type is used (features that require a feature flag)
  • format: unset the value and if the given value does not conform with the format as required by the spec
  • invalid_type: unset the value if the given value's type is not compatible with the spec's node type
  • required: drop an empty node if the node is required by the spec
  • template: drop the value if the given template string uses unknown variables
  • unknown_keys: drop unknown keys from a strict map
  • unknown_value: drop unknown values from a fixed type
  • unsupported_key: unset the value if the given key is not compatible with the current language or os, as defined by the spec
  • unsupported_value: unset the value if the given fixed value is not compatible with the current language or os, as defined by the spec

Section specific validations:

  • template: unset the value if unknown variable names are used in a notification template


There are three sets of classes that are used to build trees:

  • Travis::Yaml::Spec builds the static, unexpanded Ruby hash that can be served as JSON.
  • Travis::Yaml::Doc::Spec is used to build an object oriented tree of nodes from the expanded spec.
  • Travis::Yaml::Doc::Value is used to build an object oriented tree of nodes that represent the given build config.

Only the last one, Doc::Value, is re-used at runtime, i.e. only for the given build configs we build new trees. The Doc::Spec representation is kept in memory, is static, and remains unchanged.

For each build config we then apply all relevant changes (Doc::Change) and all relevant validations (Doc::Validate) to each node.

Expanding a build matrix

A given build configuration can be expanded to job matrix configurations by:


E.g. a build config like:

  language: 'ruby',
  ruby: ['2.2', '2.3'],
  env: ['FOO=foo', 'BAR=bar']

will be expanded to:

  { language: 'ruby', ruby: '2.2', env: 'FOO=foo' },
  { language: 'ruby', ruby: '2.2', env: 'BAR=bar' },
  { language: 'ruby', ruby: '2.3', env: 'FOO=foo' },
  { language: 'ruby', ruby: '2.3', env: 'BAR=bar' },


Tests can be run against YAML examples contained in the documentation. These are not executed by default, but need to be opted in to either by defining an env var DOCS or running the file spec_docs_spec.rb only.

cd ..
git clone
cd -

# run all tests
DOCS=true rspec

# run documentation tests
rspec spec/docs_spec.rb


(TODO join with the above?)

The format supports the concept of includes. I.e. a node can list certain global sections that it wants to include. Clients are then supposed to merge the included sections to the given node.

We are using this for language support keys and keys (such as gemfile and ruby that are allowed only on the language ruby), and other keys that are shared between the root node and the nodes matrix.include, matrix.exclude, and matrix.allow_failures.

The reason we are doing this is to keep the spec at a reasonable size. If we would not share these keys the spec size would explode from ~600K to over 14MB (when formatted as pretty printed JSON, i.e. including lots of whitespace).

The specification for language support keys like gemfile and ruby takes up some space as follows:

We have 29 languages. They provide 63 such keys in total. Each of these keys takes up at ~300 to 800 bytes (when pretty formatted as JSON). The full support node that holds these keys, including whitespace, is 20238 bytes big.

Now these keys are not only relevant on the root node, but also relevant on the matrix.include, matrix.exclude, and matrix.allow_failures nodes, as well as the deployment conditions on node. Also, again, the deployment provider nodes are not only relevant on the root node, but on the 3 matrix.* nodes mentioned. We have 41 deployment providers at the moment.

So we'd need to include these 63 keys 4 + 4 * 41 times, taking up ~3.4MB space.

On top of that, a couple other huge keys are shared between the root node, and the 3 matrix.* nodes mentioned, such as the addons node, which is ~240K big, and the deploy node, which is ~170K big (not counting the language support keys discussed before).

Because of the deep nesting that these keys then have, a huge amount of additional whitespace is added on top of that, resulting in a total spec size of ~20MB, making the spec pretty much impossible to read for humans, and expensive to download (in case we'd want to use it in JS client for example).


This gem also contains a web API for parsing and expanding a build config, which is used by travis-gatekeeper when processing a build request.

To start the server:

$ foreman start web

The API contains three endpoints:


$ curl -X GET /v1 | jq .
  "version": "v1"


The body of the request should be raw YAML. The response contains parsing messages and the validated and normalised config.

$ curl -X POST --data-binary @travis.yml /v1/parse | jq .
  "version": "v1",
  "messages": [
      "level": "info",
      "key": "language",
      "code": "default",
      "args": {
        "key": "language",
        "default": "ruby"
  "full_messages": [
    "[info] on language: missing :language, defaulting to: ruby"
  "config": {
    "language": "ruby",
    "os": [
    "dist": "trusty",
    "sudo": false


The body of the request should be the JSON config found in the response from /v1/parse. The response will contain the expanded matrix of jobs.

$ curl -X POST --data-binary @config.json /v1/expand | jq .
  "version": "v1",
  "matrix": [
      "os": "linux",
      "language": "ruby",
      "dist": "trusty",
      "sudo": false