vercel/DEVELOPING_A_RUNTIME.md

Runtime Developer Reference

The following page is a reference for how to create a Runtime by implementing the Runtime API interface. It's a way to add support for a new programming language to Vercel.

Note: If you're the author of a web framework, please use the Build Output API instead to make your framework compatible with Vercel.

A Runtime is an npm module that implements the following interface:

interface Runtime {
  version: number;
  build: (options: BuildOptions) => Promise<BuildResult>;
  prepareCache?: (options: PrepareCacheOptions) => Promise<CacheOutputs>;
  shouldServe?: (options: ShouldServeOptions) => Promise<boolean>;
  startDevServer?: (
    options: StartDevServerOptions
  ) => Promise<StartDevServerResult>;
}

The version property and the build() function are the only required fields. The rest are optional extensions that a Runtime may implement in order to enhance functionality. These functions are documented in more detail below.

Official Runtimes are published to the npm registry as a package and referenced in the use property of the vercel.json configuration file.

Note: The use property in the builds array will work with any npm install argument such as a git repo URL, which is useful for testing your Runtime. Alternatively, the functions property requires that you specify a specific tag published to npm, for stability purposes.

See the Runtimes Documentation to view example usage.

Runtime Exports

version

A required exported constant that decides which version of the Runtime API to use.

The latest and suggested version is 3.

Example:

export const version = 3;

build()

A required exported function that returns a Serverless Function.

What's a Serverless Function? Read about Serverless Functions to learn more.

Example:

import { BuildOptions, createLambda } from '@vercel/build-utils';

export async function build(options: BuildOptions) {
  // Build the code here…

  const lambda = createLambda(/* … */);
  return {
    output: lambda,
    routes: [
      // If your Runtime needs to define additional routing, define it here…
    ],
  };
}

prepareCache()

An optional exported function that is executed after build() is completed. The implementation should return an object of Files that will be pre-populated in the working directory for the next build run in the user's project. An example use-case is that @vercel/node uses this function to cache the node_modules directory, making it faster to install npm dependencies for future builds.

Note: Only files within the repo root directory can be cached.

Example:

import { relative } from 'path';
import { glob, PrepareCache } from '@vercel/build-utils';

export const prepareCache: PrepareCache = async ({
  workPath,
  repoRootPath,
}) => {
  // Create a mapping of file names and `File` object instances to cache here…
  const rootDirectory = relative(repoRootPath, workPath);
  const cache = await glob(`${rootDirectory}/some/dir/**`, repoRootPath);
  return cache;
};

shouldServe()

An optional exported function that is only used by vercel dev in Vercel CLI and indicates whether a Runtime wants to be responsible for responding to a certain request path.

Example:

import { ShouldServeOptions } from '@vercel/build-utils';

export async function shouldServe(options: ShouldServeOptions) {
  // Determine whether or not the Runtime should respond to the request path here…

  return options.requestPath === options.entrypoint;
}

If this function is not defined, Vercel CLI will use the default implementation.

startDevServer()

An optional exported function that is only used by vercel dev in Vercel CLI. If this function is defined, Vercel CLI will not invoke the build() function, and instead invoke this function for every HTTP request. It is an opportunity to provide an optimized development experience rather than going through the entire build() process that is used in production.

This function is invoked once per HTTP request and is expected to spawn a child process which creates an HTTP server that will execute the entrypoint code when an HTTP request is received. This child process is single-serve (only used for a single HTTP request). After the HTTP response is complete, vercel dev sends a shut down signal to the child process.

The startDevServer() function returns an object with the port number that the child process' HTTP server is listening on (which should be an ephemeral port) as well as the child process' Process ID, which vercel dev uses to send the shut down signal to.

Hint: To determine which ephemeral port the child process is listening on, some form of IPC is required. For example, in @vercel/go the child process writes the port number to file descriptor 3, which is read by the startDevServer() function implementation.

It may also return null to opt-out of this behavior for a particular request path or entrypoint.

Example:

import { spawn } from 'child_process';
import { StartDevServerOptions } from '@vercel/build-utils';

export async function startDevServer(options: StartDevServerOptions) {
  // Create a child process which will create an HTTP server.
  //
  // Note: `my-runtime-dev-server` is an example dev server program name.
  // Your implementation will spawn a different program specific to your runtime.
  const child = spawn('my-runtime-dev-server', [options.entrypoint], {
    stdio: ['ignore', 'inherit', 'inherit', 'pipe'],
  });

  // In this example, the child process will write the port number to FD 3…
  const portPipe = child.stdio[3];
  const childPort = await new Promise(resolve => {
    portPipe.setEncoding('utf8');
    portPipe.once('data', data => {
      resolve(Number(data));
    });
  });

  return { pid: child.pid, port: childPort };
}

Execution Context

• Runtimes are executed in a Linux container that closely matches the Servereless Function runtime environment.
• The Runtime code is executed using Node.js version 12.x.
• A brand new sandbox is created for each deployment, for security reasons.
• The sandbox is cleaned up between executions to ensure no lingering temporary files are shared from build to build.

All the APIs you export (analyze(), build(), prepareCache(), etc.) are not guaranteed to be run in the same process, but the filesystem we expose (e.g.: workPath and the results of calling getWritableDirectory ) is retained.

If you need to share state between those steps, use the filesystem.

Directory and Cache Lifecycle

When a new build is created, we pre-populate the workPath supplied to analyze with the results of the prepareCache step of the previous build.

The analyze step can modify that directory, and it will not be re-created when it's supplied to build and prepareCache.

Accessing Environment and Secrets

The env and secrets specified by the user as build.env are passed to the Runtime process. This means you can access user env via process.env in Node.js.

Supporting Large Environment

We provide the ability to support more than 4KB of environment (up to 64KB) by way of a Lambda runtime wrapper that is added to every Lambda function we create. These are supported by many of the existing Lambda runtimes, but custom runtimes may require additional work.

The following Lambda runtime families have built-in support for the runtime wrapper:

• nodejs
• python (>= 3.8)
• ruby
• java11
• java8.al2 (not java8)
• dotnetcore

If a custom runtime is based on one of these Lambda runtimes, large environment support will be available without further configuration. Custom runtimes based on other Lambda runtimes, including those that provide the runtime via provided and provided.al2, must implement runtime wrapper support and indicate it via the supportsWrapper flag when calling createLambda.

To add support for runtime wrappers to a custom runtime, first check the value of the AWS_LAMBDA_EXEC_WRAPPER environment variable in the bootstrap script. Its value is the path to the wrapper executable.

The wrapper must be passed the path to the runtime as well as any parameters that the runtime requires. This is most easily done in a small bootstrap script.

In this simple bash example, the runtime is called directly if AWS_LAMBDA_EXEC_WRAPPER has no value, otherwise the wrapper is called with the runtime command as parameters.

#!/bin/bash

exec $AWS_LAMBDA_EXEC_WRAPPER path/to/runtime param1 param2

If the bootstrap file is not a launcher script, but the entrypoint of the runtime itself, replace the bootstrap process with the wrapper. Pass the path and parameters of the executing file, ensuring the AWS_LAMBDA_EXEC_WRAPPER environment variable is set to blank.

This bash example uses exec to replace the running bootstrap process with the wrapper, passing its own path and parameters.

#!/bin/bash

if [[ -n $AWS_LAMBDA_EXEC_WRAPPER ]]
  __WRAPPER=$AWS_LAMBDA_EXEC_WRAPPER
  AWS_LAMBDA_EXEC_WRAPPER=""
  exec $__WRAPPER "$0" "${@}"
fi

# start the actual runtime functionality

Note that unsetting the variable may not have the desired effect due to the way Lambda spawns runtime processes. It is better to explicitly set it to blank.

The best way to replace the existing bootstrap process is with the execve syscall. This is achieved by using exec in bash to replace the running process with the wrapper, maintaining the same PID and environment.

Once support for runtime wrappers is included, ensure supportsWrapper is set to true in the call to createLambda. This will inform the build process to enable large environment support for this runtime.

Utilities as peerDependencies

When you publish your Runtime to npm, make sure to not specify @vercel/build-utils (as seen below in the API definitions) as a dependency, but rather as part of peerDependencies.

@vercel/build-utils Types

Files

import { File } from '@vercel/build-utils';
type Files = { [filePath: string]: File };

This is an abstract type that is implemented as a plain JavaScript Object. It's helpful to think of it as a virtual filesystem representation.

When used as an input, the Files object will only contain FileRefs. When Files is an output, it may consist of Lambda (Serverless Functions) types as well as FileRefs.

An example of a valid output Files object is:

{
  "index.html": FileRef,
  "api/index.js": Lambda
}

File

This is an abstract type that can be imported if you are using TypeScript.

import { File } from '@vercel/build-utils';

Valid File types include:

• FileRef
• FileFsRef
• FileBlob

FileRef

import { FileRef } from '@vercel/build-utils';

This is a class that represents an abstract file instance stored in our platform, based on the file identifier string (its checksum). When a Files object is passed as an input to analyze or build, all its values will be instances of FileRef.

Properties:

• mode: Number file mode
• digest: String a checksum that represents the file

Methods:

• toStream(): Stream creates a Stream of the file body

FileFsRef

import { FileFsRef } from '@vercel/build-utils';

This is a class that represents an abstract instance of a file present in the filesystem that the build process is executing in.

Properties:

• mode: Number file mode
• fsPath: String the absolute path of the file in file system

Methods:

• static async fromStream({ mode: Number, stream: Stream, fsPath: String }): FileFsRef creates an instance of a FileFsRef from Stream, placing file at fsPath with mode
• toStream(): Stream creates a Stream of the file body

FileBlob

import { FileBlob } from '@vercel/build-utils';

This is a class that represents an abstract instance of a file present in memory.

Properties:

• mode: Number file mode
• data: String | Buffer the body of the file

Methods:

• static async fromStream({ mode: Number, stream: Stream }): FileBlob creates an instance of a FileBlob from Stream with mode
• toStream(): Stream creates a Stream of the file body

Lambda

import { Lambda } from '@vercel/build-utils';

This is a class that represents a Serverless Function. An instance can be created by supplying files, handler, runtime, and environment as an object to the createLambda helper. The instances of this class should not be created directly. Instead, invoke the createLambda helper function.

Properties:

• files: Files the internal filesystem of the lambda
• handler: String path to handler file and (optionally) a function name it exports
• runtime: LambdaRuntime the name of the lambda runtime
• environment: Object key-value map of handler-related (aside of those passed by user) environment variables
• supportsWrapper: Boolean set to true to indicate that Lambda runtime wrappers are supported by this runtime

LambdaRuntime

This is an abstract enumeration type that is implemented by one of the following possible String values:

• nodejs20.x
• nodejs18.x
• nodejs16.x
• java11
• python3.9
• dotnet6
• dotnetcore3.1
• ruby2.7
• provided.al2
• provided.al2023

@vercel/build-utils Helper Functions

The following is exposed by @vercel/build-utils to simplify the process of writing Runtimes, manipulating the file system, using the above types, etc.

createLambda()

Signature: createLambda(Object spec): Lambda

import { createLambda } from '@vercel/build-utils';

Constructor for the Lambda type.

const { createLambda, FileBlob } = require('@vercel/build-utils');
await createLambda({
  runtime: 'nodejs8.10',
  handler: 'index.main',
  files: {
    'index.js': new FileBlob({ data: 'exports.main = () => {}' }),
  },
});

download()

Signature: download(): Files

import { download } from '@vercel/build-utils';

This utility allows you to download the contents of a Files data structure, therefore creating the filesystem represented in it.

Since Files is an abstract way of representing files, you can think of download() as a way of making that virtual filesystem real.

If the optional meta property is passed (the argument for build()), only the files that have changed are downloaded. This is decided using filesRemoved and filesChanged inside that object.

await download(files, workPath, meta);

glob()

Signature: glob(): Files

import { glob } from '@vercel/build-utils';

This utility allows you to scan the filesystem and return a Files representation of the matched glob search string. It can be thought of as the reverse of download.

The following trivial example downloads everything to the filesystem, only to return it back (therefore just re-creating the passed-in Files):

const { glob, download } = require('@vercel/build-utils');

exports.build = ({ files, workPath }) => {
  await download(files, workPath);
  return glob('**', workPath);
};

getWritableDirectory()

Signature: getWritableDirectory(): String

import { getWritableDirectory } from '@vercel/build-utils';

In some occasions, you might want to write to a temporary directory.

rename()

Signature: rename(Files, Function): Files

import { rename } from '@vercel/build-utils';

Renames the keys of the Files object, which represent the paths. For example, to remove the *.go suffix you can use:

const rename = require('@vercel/build-utils')
const originalFiles = { 'one.go': fileFsRef1, 'two.go': fileFsRef2 }
const renamedFiles = rename(originalFiles, path => path.replace(/\.go$/, '')