Cucumber-js: Proposal: Programmatic API for running cucumber-js

Created on 28 Jun 2021  ยท  29Comments  ยท  Source: cucumber/cucumber-js

Problem

Currently we don't have a good way to programmatically run cucumber-js. The need is from two angles:

  • Testing cucumber-js - for the CCK and our acceptance tests in the project
  • Testing custom formatters and snippets
  • Projects that use cucumber-js as part of a framework (e.g. Serenity, Stryker)

As-is

What tends to happen at the moment is a new instance of Cli is created with strung-together argv input. It's obviously very unweildy and also isn't on the public API.

Sometimes (possibly due to perceived fragility of the above) frameworks will just rely on the cucumber-js CLI but struggle to find ways to integrate and have their own options.

The Runtime class is currently part of the public API but it's not useful in these contexts, depending on the pickles and support code to be provided by the caller.

Proposal

Two components in the project:

runCucumber

New async function that executes a test run in-process. Responsibilities:

  • Accept an options object with a nice interface
  • Do the "work" of resolving pickles, loading support code, running test cases, orchestrating formatters
  • Return a promise that resolves to a result

This would be part of the public API and we'd encourage framework maintainers to use it when "wrapping" cucumber-js. We'd also use it for our own testing.

As much as possible, it would avoid direct interaction with process, instead accepting normalised options and stream interfaces for output, and leaving it to the caller to decide how to exit based on the result or an unhandled error.

Also Runtime should come off the public API as it's really an internal thing.

CLI

Effectively a "client" of runCucumber. Responsibilities:

  • Aggregate options from various sources (argv, env vars, config files) (see comment)
  • Call runCucumber with the resolved options
  • Exit as appropriate based on the results

This would continue to not be on the public API. Also it would only use functions/interfaces that are on the public API, such that we could easily break it out into its own package at some point, as is a common pattern now with projects like Jest.

This decoupling also paves the way for some interesting new CLI features without having them bleed into the internals, e.g.:

  • --gui for the cucumber-electron stuff
  • --interactive for quick targeted reruns when TDD'ing

etc

We would also expose functions (consumable by the CLI and by others) for:

  • Getting the options
  • Handling i18nKeywords and i18nLanguages

Timescale

We'd target this at the upcoming 8.0.0 release. I'm ready to gets started on this today.

breaking change enhancement

Most helpful comment

How would this work for a reporter that doesn't need an output file name?

formats: {
    stdout: './my-awesome-stryker-formatter',
    files: {
      'report.html': './my/fancy-reporter.js',
      'other-report.html': '@me/reporter-package',
    }
  },

(only one formatter can use the stdout stream)

All 29 comments

Paging for initial feedback: @aslakhellesoy @charlierudolph @aurelien-reeves @mattwynne @nicojs @jan-molak

I love this proposal! We already an API like this in fake-cucumber's runCucumber

@davidjgoss - sounds great!

For your reference, here's how Serenity/JS invokes Cucumber at the moment - CucumberCLIAdapter
And here's the logic around converting configuration parameters to argv - CucumberOptions.

Being able to provide an options object instead of argv would be much nicer ๐Ÿ‘๐Ÿป

Love it!

While specifying that new public API, we may also consider what recently happened with issue #1489 and think about providing public APIs to have more and better interaction with the filters and the resulting features under test

Having a public API is better than nothing, so go ahead ๐Ÿ‘!

Preferably I would also have an API to load profiles using the same rules as cucumber-js does, so I can mimic the exact behavior of a normal cucumber-js call.

loadProfiles(directory = process.cwd()): Record<string, Profile>

StrykerJS will also rely heavily on the custom_formatters API and the events published by the eventBroadcaster. Could we add those to the public API as well? See: https://github.com/stryker-mutator/stryker-js/blob/03b1f20ed933d3a50b52022cfe363c606c2b16c5/packages/cucumber-runner/src/stryker-formatter.ts#L45-L69

Preferably I would also have an API to load profiles using the same rules as cucumber-js does, so I can mimic the exact behavior of a normal cucumber-js call.

This is a good point. Profiles (in their current form at least) are fundamentally coupled to the CLI so it feels right to keep them on that side of the boundary, but we could still expose a function to load them and generate a partial options object.

While specifying that new public API, we may also consider what recently happened with issue #1489 and think about providing public APIs to have more and better interaction with the filters and the resulting features under test.

I think we could include an option for providing a custom pickle filter when calling the API (in addition to the names, tags etc that drive the built-in filtering).

The current syntax for profiles if very command-line-y.

I would โค๏ธโค๏ธโค๏ธ to be able to specify profiles in a more generic format such as JSON, JavaScript, YAML or environment variables. In JSON it could look like this:

.cucumber.json

{
  "default": {
    "requireModule": ["ts-node/register"],
    "require": ["support/**/*./ts"],
    "worldParameters": {
      "appUrl": "http://localhost:3000/",
    },
    "format": ["progress-bar", "html:./cucumber-report.html"]
  },
  "ci": {
    "requireModule": ["ts-node/register"],
    "require": ["support/**/*./ts"],
    "worldParameters": {
      "appUrl": "http://localhost:3000/",
    },
    "format": ["html:./cucumber-report.html"],
    "publish": true
  }
}

Or, using JavaScript

.cucumber.js

const common = {
  "requireModule": ["ts-node/register"],
  "require": ["support/**/*./ts"],
  "worldParameters": {
    "appUrl": "http://localhost:3000/",
  }
}

module.exports = {
  default: {
    ...common,
    "format": ["progress-bar", "html:./cucumber-report.html"]
  },
  ci: {
    ...common,
    "format": ["html:./cucumber-report.html"],
    "publish": true
  }
}

Or even with environment variables (for example loaded with a tool like dotenv):

.cucumber.env

CUCUMBER_PROFILE_DEFAULT_REQUIREMODULE=ts-node/register
CUCUMBER_PROFILE_DEFAULT_REQUIRE=ts-node/register
CUCUMBER_PROFILE_DEFAULT_WORLDPARAMETERS_APPURL=http://localhost:3000/
CUCUMBER_PROFILE_DEFAULT_FORMAT=progress-bar,html:./cucumber-report.html
CUCUMBER_PROFILE_CI_REQUIREMODULE=ts-node/register
CUCUMBER_PROFILE_CI_REQUIRE=ts-node/register
CUCUMBER_PROFILE_CI_WORLDPARAMETERS_APPURL=http://localhost:3000/
CUCUMBER_PROFILE_CI_FORMAT=progress-bar,html:./cucumber-report.html
CUCUMBER_PROFILE_CI_PUBLISH=true

In fact, the config library does exactly this. We never ended up integrating it in Cucumber-JVM because other stuff got in the way, but maybe we could give it a go with a JavaScript implementation?

@aslakhellesoy agree that would be awesome! I'll try and get a POC going for this proposal so we have something a bit more concrete to talk around, and would love to make profiles right as part of it (4.5 years and counting on #751 ๐Ÿ˜„)

Refs. #1004

This is a good point. Profiles (in their current form at least) are fundamentally coupled to the CLI so it feels right to keep them on that side of the boundary, but we could still expose a function to load them and generate a partial options object.

Yes, that would be awesome and much appreciated from the point of view of plugin creators.

I would โค๏ธโค๏ธโค๏ธ to be able to specify profiles in a more generic format such as JSON, JavaScript, YAML or environment variables. In JSON it could look like this:

That sounds great! And is also exactly the reason I would appreciate an API to load them the same way as cucumberJS does. Loading a single cucumber.js file is trivial. Replicating a configuration file loading algorithm, including precedence, file format, etc AND maintaining it is something else entirely ๐Ÿ˜….

Q: Would I be able to run runCucumber twice in succession _without clearing the require cache_? This is important for the mutation testing use case.

We want to load the environment and run the tests multiple times in quick succession while changing a global variable in order to switch the active mutant.

Right now, we're using the cli private API and we need to clear the step definition files from the require.cache between each test run. This is not ideal for CommonJS and won't work at all for esm.

Pseudo code of our use case:

const profiles = await loadProfiles();
const options = {
  ...profiles.default,
  formatter: require.resolve('./our-awesomely-crafted-formatter'),
  some: 'other options we want to override',
}
const cucumber = new Cucumber(options);

// Allow cucumber to load the step definitions once. 
// This is `async`, so support for esm can be added without a breaking change
await cucumber.initialize();

// Initial test run ("dry run"), without mutants active
await cucumber.run();

collectMutantCoveragePerTestFromFormatter();

// Start mutation testing:

global.activeMutant = 1;
await cucumber.run({ features: ['features/a.feature:24']);
collectResultsFromFormatterToDetermineKilledOrSurvivedMutant()

global.activeMutant = 2;
await cucumber.run({ features: ['features/b.feature:24:25:26', 'features/c.feature:12']);
collectResultsFromFormatterToDetermineKilledOrSurvivedMutant()

// etc

@nicojs definitely agree we need this, it's come up a few times before with e.g. people wanting to run cucumber in a lambda, and I'd also like to add an interactive mode which would need it too.

What I had sketched out so far was in more of a functional style but same fundamental concept I think:

const runnerOptions = {
    support: {
        require: ['features/support/**/*.js']
    }
}

// initial run returns support code library
const { support } = await runCucumber(runnerOptions)

// subsequent run reuses support code library
await runCucumber({
    ...runnerOptions,
    support
})

That works for us ๐Ÿ‘

Works for us too ๐Ÿ‘๐Ÿป

I really think that something like this would be extremely useful as an alternative for the great mess we have today with testing tool integrations (Jest, Cypress), for example I found these problems (in the order of importance):

  • cypress-cucumber-preprocessor does not support tags on Examples (https://github.com/TheBrainFamily/cypress-cucumber-preprocessor/issues/196)
  • jest-cucumber does not support Cucumber JSON reporting (https://github.com/bencompton/jest-cucumber/issues/27)
  • cypress-cucumber-preprocessor generates multiple Cucumber JSON reports with no offical support for aggregation (https://github.com/TheBrainFamily/cypress-cucumber-preprocessor/issues/423)
  • jest-cucumber is not as convenient as jest-cucumber-fusion
  • there's also cucumber-jest...
  • Karma does not have a working implementation anymore (https://github.com/cucumber/cucumber-js/issues/1095)

I would rather see some minimal glue code between Jest/Karma/Cypress/etc. and cucumber-js so I don't have to suffer for all those missing features I need to use.

Great suggestion @davidjgoss ๐Ÿ‘

This separation of concerns between the command-line user interface and the "business logic" of parsing and executing scenarios as tests reminds me of the hexagonal architecture pattern.

In cucumber-ruby we actually split the core domain logic (or "inner hexagon") into a separate gem package, as we were rebuilding it from scratch in a "clean room". I realise that's not the context here, but it might be worth drawing some inspiration from, or feeding back innovations from this design into the Ruby API. There's an example in the cucumber-ruby-core gem's README of how to use that API.

Okay, here's a first pass at the API signature for the "run" bit. It's heavily based on the IConfiguration object we have internally (so shouldn't cause too much refactoring lower down) but just a little less "flat":

export interface IRunCucumberOptions {
  cwd: string
  features: {
    defaultDialect?: string
    paths: string[]
  }
  filters: {
    name?: string[]
    tagExpression?: string
  }
  support:
    | {
        transpileWith?: string[]
        paths: string[]
      }
    | ISupportCodeLibrary
  runtime: {
    dryRun?: boolean
    failFast?: boolean
    filterStacktraces?: boolean
    parallel?: {
      count: number
    }
    retry?: {
      count: number
      tagExpression?: string
    }
    strict: boolean
    worldParameters?: any
  }
  formats: {
    stdout: string
    files: Record<string, string>
    options: IParsedArgvFormatOptions
  }
}

export interface IRunResult {
  success: boolean
  support: ISupportCodeLibrary
}

export async function runCucumber(
  options: IRunCucumberOptions
): Promise<IRunResult> {
  // do stuff
}

And a very contrived example usage:

const result = await runCucumber({
  cwd: process.cwd(),
  features: {
    paths: ['features/**/*.feature'],
  },
  filters: {
    name: ['Acme'],
    tagExpression: '@interesting',
  },
  support: {
    transpileWith: ['ts-node'],
    paths: ['features/support/**/*.ts'],
  },
  runtime: {
    failFast: true,
    retry: {
      count: 1,
      tagExpression: '@flaky',
    },
    strict: true,
    worldParameters: {
      foo: 'bar',
    },
  },
  formats: {
    stdout: '@cucumber/pretty-formatter',
    files: {
      'report.html': 'html',
      'TEST-cucumber.xml': 'junit',
    },
    options: {
      printAttachments: false,
    },
  },
})

Feedback welcome! Note this doesn't cover the profile/config loading stuff which would be another function.

I think this looks good. Question: How would I configure a custom formatter?

@nicojs sorta like on the CLI

formats: {
    files: {
      'report.html': './my/fancy-reporter.js',
      'other-report.html': '@me/reporter-package',
    }
  },

Great to see progress on this @davidjgoss!

I don't want to slow down the progress on this, but at the same time I want to make sure we adopt a format that can work for other Cucumber implementations too.

Eventually a JSON schema, but while we discuss it I think TypeScript types are easier for us humans to parse.

I suggest we create a new issue about the proposed format in the cucumber/common monorepo and invite the core team to discuss there.

@aslakhellesoy will do.

What would you think about the programmatic API not being tied to the common options structure? Like we'd map from that to the runCucumber options. It adds maybe a little complexity but appeals because of things like having a support block that's either parameters to load, or a previously loaded support code library. Could do similar for features+pickles too. And there are various options we'd support on the CLI (e.g. --exit) that aren't appropriate on the programmatic API.

What would you think about the programmatic API not being tied to the common options structure?

I think that's fine, as long as we provide a function that converts from the options file contents to the data structure runCucumber wants.

Eventually a JSON schema, but while we discuss it I think TypeScript types are easier for us humans to parse.

Why do we need to choose? We're using JSON schema in StrykerJS to generate typescript using json-schema-to-typescript. We're not committing the TS output files to source control, instead we generate them on the fly using a prebuild step.

JSON schemas are still somewhat readable for humans IMO. We've already had PR's on the Stryker repo and people seem to know what to do ๐Ÿคทโ€โ™€๏ธ

sorta like on the CLI

formats: {
    files: {
      'report.html': './my/fancy-reporter.js',
      'other-report.html': '@me/reporter-package',
    }
  },

How would this work for a reporter that doesn't need an output file name? Like so:

formats: {
  files: {
    '': require.resolve('./my-awesome-stryker-formatter')
  }
}

Why do we need to choose?

I think we should use a JSON Schema as a single source of truth for the structure of the configuration. -And then generate TypeScript/Java/Whatever code from that schema.

But JSON Schema is a bit hard to read for humans, so while we are discussing the schema in a GitHub issue in cucumber/common I was suggesting TypeScript to facilitate the discussion.

See what I mean?

JSON schemas are still somewhat readable for humans IMO

Not to me :-) Too verbose.

How would this work for a reporter that doesn't need an output file name?

formats: {
    stdout: './my-awesome-stryker-formatter',
    files: {
      'report.html': './my/fancy-reporter.js',
      'other-report.html': '@me/reporter-package',
    }
  },

(only one formatter can use the stdout stream)

Was this page helpful?
0 / 5 - 0 ratings