Thrift TypeScript

Generate TypeScript from Thrift IDL files.

A modify version of @creditkarma/thrift-typescript.

Use --target apachem to generate Apache Thrift compatible code with context support.

With a modify version of createServer.

/* eslint-disable @typescript-eslint/no-explicit-any */
/* eslint-disable @typescript-eslint/ban-ts-comment */
import type { ServerOptions, TProcessorConstructor, TTransport } from 'thrift';
import { TBufferedTransport, TBinaryProtocol } from 'thrift';
import * as tls from 'tls';
import * as net from 'net';
import * as constants from 'constants';

const {
  THeaderProtocol,
} = require('thrift/lib/nodejs/lib/thrift/header_protocol.js');
const InputBufferUnderrunError = require('thrift/lib/nodejs/lib/thrift/input_buffer_underrun_error.js');

export function createMultiplexServer<TProcessor, THandler>(
  processor: TProcessor,
  options?: ServerOptions<TProcessor, THandler>,
) {
  const transport =
    options && options.transport ? options.transport : TBufferedTransport;
  const protocol =
    options && options.protocol ? options.protocol : TBinaryProtocol;

  function serverImpl(this: any, stream: net.Socket) {
    const self = this as any;
    stream.on('error', function (err) {
      self.emit('error', err);
    });
    stream.on(
      'data',
      // @ts-ignore
      transport.receiver(function (transportWithData: TTransport) {
        const input = new protocol(transportWithData);
        const outputCb = function (buf?: Buffer) {
          try {
            if (buf) {
              stream.write(buf);
            }
          } catch (err) {
            self.emit('error', err);
            stream.end();
          }
        };

        let output = new protocol(new transport(undefined, outputCb));
        // Read and write need to be performed on the same transport
        // for THeaderProtocol because we should only respond with
        // headers if the request contains headers
        if (protocol === THeaderProtocol) {
          output = input;
          // @ts-ignore
          output.trans.onFlush = outputCb;
        }

        try {
          do {
            // @ts-ignore
            processor.process(input, output, stream);
            transportWithData.commitPosition();
            // eslint-disable-next-line no-constant-condition
          } while (true);
        } catch (err) {
          if (err instanceof InputBufferUnderrunError) {
            //The last data in the buffer was not a complete message, wait for the rest
            transportWithData.rollbackPosition();
          } else if ((err as Error).message === 'Invalid type: undefined') {
            //No more data in the buffer
            //This trap is a bit hackish
            //The next step to improve the node behavior here is to have
            //  the compiler generated process method throw a more explicit
            //  error when the network buffer is empty (regardles of the
            //  protocol/transport stack in use) and replace this heuristic.
            //  Also transports should probably not force upper layers to
            //  manage their buffer positions (i.e. rollbackPosition() and
            //  commitPosition() should be eliminated in lieu of a transport
            //  encapsulated buffer management strategy.)
            transportWithData.rollbackPosition();
          } else {
            //Unexpected error
            self.emit('error', err);
            stream.end();
          }
        }
      }),
    );

    stream.on('end', function () {
      stream.end();
    });
  }

  if (options && options.tls) {
    if (
      !('secureProtocol' in options.tls) &&
      !('secureOptions' in options.tls)
    ) {
      options.tls.secureProtocol = 'SSLv23_method';
      options.tls.secureOptions =
        constants.SSL_OP_NO_SSLv2 | constants.SSL_OP_NO_SSLv3;
    }
    return tls.createServer(options.tls, serverImpl);
  } else {
    return net.createServer(serverImpl);
  }
}

export function createServer<TProcessor, THandler>(
  processor: TProcessorConstructor<TProcessor, THandler>,
  handler: THandler,
  options?: ServerOptions<TProcessor, THandler>,
): net.Server | tls.Server {
  const pc = 'Processor' in processor ? processor.Processor : processor;
  return createMultiplexServer(new pc(handler), options);
}

Installation

$ npm install --save @seraphli/thrift-typescript

Usage

Thrift TypeScript provides both a JavaScript and a command line API.

Given the following files

thrift/simple.thrift

struct MyStruct {
    1: required i32 id,
    2: required bool field1,
    # 3: required string field,
    4: required i16 field,
}

You can generate TypeScript via the command line:

$ thrift-typescript --target apache --rootDir . --sourceDir thrift --outDir codegen simple.thrift

The available options are:

• --rootDir: This is used to resolve out and source directories. Defaults to current directory.
• --outDir: The directory to save generated files to. Will be created if it doesn't exist. Defaults to 'codegen'.
• --sourceDir: The directory to search for source Thrift files. Defaults to 'thrift'.
• --target: The core library to generate for, either 'apache' or 'thrift-server'. Defaults to 'apache'.
• --strictUnions: Should we generate strict unions (Only available for target = 'thrift-server'. More on this below). Defaults to undefined.
• --fallbackNamespace: The namespace to fallback to if no 'js' namespace exists. Defaults to 'java'. Set to 'none' to use no namespace.
• --withNameField: Should we generate a __name field on each struct-like object that contains its Thrift name (Only available for target = 'thrift-server'). Defaults to undefined.

All other fields are assumed to be source files.

If no explicit list of files is provided all files ending in '.thrift' found in the sourceDir will be used.

You can gen code from more than one Thrift file:

$ thrift-typescript one.thrift two.thrift three.thrift

JavaScript API

You can also generate files using the JavaScript API:

import { generate } from '@creditkarma/thrift-typescript'

// Generates TypeScript and saves to given outDir
generate({
    rootDir: '.',
    sourceDir: 'thirft',
    outDir: 'codegen',
    target: 'thrift-server',
    files: [
        'simple.thrift'
    ],
    fallbackNamespace: 'java',
})

Thrift to String

You can also generate TypeScript from a string of Thrift without saving to file.

Note: This method of code generation does not support includes. The Thrift generator must be able to resolve all identifiers which it can't do without a description of the file structure.

import { readFileSync } from 'fs'
import { make } from '@creditkarma/thrift-typescript'

const rawThrift: string = readFileSync('./thrift/simple.thrift', 'utf-8')
const generatedCode: string = make(rawThrift)

Thrift Server

v2.x of Thrift TypeScript equires @creditkarma/thrift-server v0.7.0 or higher

While Thrift TypeScript can be used to generate code comaptible with the Apache Thrift Library, it is recommended to use with Thrift Server. Details on the Apache usage are below.

Thrift Server adds Thrift support to Express or Hapi with plugins or middleware. The other advantange of using the codegen with Thrift Server is the addition of context to service clients and service handlers. Context can be used to do things like auth or tracing in Thrift service methods. Context is an optional final parameter to all service handler methods and all service client methods.

Install the Thrift Server implementation for your server of choice. For this example we will be using express middleware and the request http client library.

$ npm install --save @creditkarma/thrift-server-core
$ npm install --save @creditkarma/thrift-server-express
$ npm install --save @creditkarma/thrift-client
$ npm install --save express
$ npm install --save request
$ npm install --save @types/express
$ npm install --save @types/request

Given this service let's build a client and server based on our generated code.

service Caluculator {
    i32 add(1: i32 left, 2: i32 right)
    i32 subtract(1: i32 left, 2: i32 right)
}

Run codegen for your Thrift service. The target option is required here, otherwise the generated code will only work with the Apache libs.

$ thrift-typescript --target thrift-server --rootDir . --sourceDir thrift --outDir codegen

Client

In this example we are using the Request library as our underlying connection instance. The options for Request (CoreOptions) are our request context.

You'll notice that the Client class is a generic. The type parameter represents the type of the context. This is usually going to be of type CoreOptions from the Request library.

import {
    createHttpClient,
    HttpConnection,
} from '@creditkarma/thrift-client'

import * as request from 'request'
import { CoreOptions } from 'request'

import { Calculator } from './codegen/calculator'

const CONFIG = {
    hostName: 'localhost',
    port: 8045
}

const thriftClient: Calculator.Client<CoreOptions> = createHttpClient(Calculator.Client, CONFIG)

thriftClient.add(5, 7, { headers: { 'X-Trace-Id': 'xxxxxx' } })
    .then((response: number) => {
        expect(response).to.equal(12)
        done()
    })

Server

In the server we can then inspect the headers we set in the client.

import * as bodyParser from 'body-parser'
import * as express from 'express'
import { ThriftServerExpress } from '@creditkarma/thrift-server-express'

import {
    Calculator,
} from './codegen/calculator'

// express.Request is the context for each of the service handlers
const serviceHandlers: Calculator.IHandler<express.Request> = {
    add(left: number, right: number, context?: express.Request): number {
        if (context && context.headers['x-trace-id']) {
            // You can trace this request, perform auth, or use additional middleware to handle that.
        }
        return left + right
    },
    subtract(left: number, right: number, context?: express.Request): number {
        return left - right;
    },
}

const PORT = 8090

const app = express()

app.use(
    '/thrift',
    bodyParser.raw(),
    ThriftServerExpress(Calculator.Processor, serviceHandlers),
)

app.listen(PORT, () => {
    console.log(`Express server listening on port: ${PORT}`)
})

Generated Data Types

When generating TypeScript from Thrift source what data types are generated?

Simple Types

These are: booleans, strings, numbers, sets, maps, lists, enums and typedefs. All of these translate almost directly to TypeScript.

Given Thrift:

const bool FALSE_CONST = false
const i32 INT_32 = 32
const i64 INT_64 = 64
const list<string> LIST_CONST = ['hello', 'world', 'foo', 'bar']
const set<string> SET_CONST = ['hello', 'world', 'foo', 'bar']
const map<string,string> MAP_CONST = { 'hello': 'world', 'foo': 'bar' }

enum Colors {
    RED,
    GREEN,
    BLUE,
}

typedef string name

Generated TypeScript:

export const FALSE_CONST: boolean = false;
export const INT_32: number = 32;
export const INT_64: thrift.Int64 = new thrift.Int64(64);
export const LIST_CONST: Array<string> = ["hello", "world", "foo", "bar"];
export const SET_CONST: Set<string> = new Set(["hello", "world", "foo", "bar"]);
export const MAP_CONST: Map<string, string> = new Map([["hello", "world"], ["foo", "bar"]]);

export enum Colors {
    RED,
    GREEN,
    BLUE
}

export type name = string;

The only interesting thing here is the handling of i64. JavaScript doesn't support a full 64-bits of integer percision, so we wrap the value in an Int64 object. You will notice that this doesn't really help in cases where you define a constant or default value in your Thrift file, but it does allow 64-bit integers received from outside of JS to be handled correctly. The object is exported from @creditkarma/thrift-server-core and extends node-int64.

Struct

A struct is intuitively analogous to an interface.

Given Thrift:

struct User {
    1: required string name
    2: string email
    3: required i32 id
}

Generated TypeScript:

export interface IUser {
    name: string
    email?: string
    id: number
}

Note: We adopt the convention of prefixing interfaces names with a capital 'I'.

Only fields that are explicitly required loose the ?.

The __name Field

You can pass an option to generate an additional property on each struct-like (structs, unions, exceptions) object that is its literal name in the Thrift file.

For example if we rendered the User struct with the --withNameField option the generated TypeScript would change:

export interface IUser {
    __name: "User"
    name: string
    email?: string
    id: number
}

When generating with this option any data types you create and pass into a client method do not need the __name field. However, any data you get back from a service will contain this additional property.

Union

Unions in Thrift are very similar to structs. The difference is they only allow one field to be set. They also require that one field is set. Implicitly all fields are optional, but one field must be set.

So, this translates into a struct with all optional fields:

Given Thrift:

union MyUnion {
    1: string option1
    2: i32 option2
}

Generated TypeScript (without strict unions):

export interface IMyUnion = {
    option1?: string
    option2?: undefined
}

Note: The difference here is that a runtime error will be raised if one of the fields isn't set or if more than one of the fields is set.

Exception

Exceptions are errors that can be thrown by service methods. It is more natural in JS/TS to create and throw new errors. So our defined exceptions will become JS classes.

Given Thrift:

exception MyException {
    1: string message
    2: i32 code
}

Generated TypeScript:

export class MyException extends thrift.StructLike implements IMyException {
    public message: string
    public code?: number
    constructor(args?: { message?: string, code?: number }) {
        // ...
    }
}

Then in your service client you could just throw the exception as you would any JS error throw new MyException({ message: 'whoops', code: 500 });

Service

Services are a little more complex. There are two parts to a service. There is the Client for sending service requests and the Processor for handling service requests. The service Client and the service Processor are each generated classes. They are wrapped, along with some other internal objects, in a namespace that has the name of your service.

Given Thrift:

service MyService {
    User getUser(1: i32 id) throws (1: MyException exp);
}

Generated TypeScript:

export namespace MyService {
    export class Client<Context> {
        constructor(connection: thrift.IThriftConnection<Context>) {
            // ...
        }
        getUser(id: number): Promise<User> {
            // ...
        }
    }
    export interface IHandler<Context> {
        getUser(id: number, context?: Context): User | Promise<User>
    }
    export class Processor {
        constructor(handler: IHandler<Context>) {
            // ...
        }
        public process(input: thrift.TProtocol, output: thrift.TProtocol, context: Context): Promise<Buffer> {
            // ...
        }
    }
}

The Client is pretty straight forward. You create a Client instance and you can call service methods on it. The inner-workings of the Processor aren't something consumers need to concern themselves with. The more interesting bit is IHandler. This is the interface that service teams need to implement in order to meet the promises of their service contract. Create an object that satisfies <service-name>.IHandler and pass it to the construction of <service-name>.Processor and everything else is handled for you.

Loose Types

Given these two structs:

struct User {
    1: required i64 id
}

struct Profile {
    1: required User user
    2: binary data
    3: i64 lastModified
}

There is something of a difference between how we want to handle things in TypeScript and how data is going to be sent over the wire. Because of this when we generate interfaces for these structs we generate two interfaces for each struct, one is an exact representation of the Thrift, the other is something looser that provides more flexibility to working with the data in JavaScript.

The main difference is that fields marked as i64 can be represented as a number, as string or an Int64 object and binary can be represented as either a string or a Buffer object.

JavaScript traditionally (bigint is new and not fully supported yet) does not support 64-bit integers. This means we need to wrap the Thrift i64 type in the Int64 object to maintain precision. In your TypeScript code you may be working with these just as number (confident JavaScript's 53-bit precision is good enough for you) or as a string. These loose types allow you to do that and the generated code will handle the conversions to Int64 for you.

Generated TypeScript:

interface IUser {
    id: thrift.Int64
}
interface IUserArgs {
    id: number | string | thrift.Int64
}
interface IProfile {
    user: IUser
    data?: Buffer
    lastModified?: thrift.Int64
}
interface IProfileArgs {
    user: IUserArgs
    data?: string | Buffer
    lastModified?: number | string | thrift.Int64
}

The names of loose interfaces just append Args onto the end of the interface name. The reason for this is these interfaces will most often be used as arguments in your code.

Where are the loose interfaces used? The loose interfaces can be used anywhere you, the application developer, are giving data to the generated code, either as the arguments to a client method or the return value of a service handler.

If we had this service:

service ProfileService {
    Profile getProfileForUser(1: User user)
    User getUser(1: i64 id)
}

And generated TypeScript:

namespace ProfileService {
    export class Client<Context> {
        constructor(connection: thrift.IThriftConnection<Context>) {
            // ...
        }
        getProfileForUser(user: IUserArgs, context?: Context): Promise<IProfile> {
            // ...
        }
        getUser(id: number | string | Int64): Promise<IUser> {
            // ...
        }
    }
    export interface IHandler<Context> {
        getProfileForUser(user: IUser, context: Context): Promise<IProfileArgs>
        getUser(id: Int64, context: Context): Promise<IUserArgs>
    }
}

As you can see from this sketch of generated types when data leave application code and crossed the boundary into the generated code you can pass loose values, when the data comes from generated code it will always be of the strict types.

We can use a User object where the id is a number or a string without having to wrap it in Int64. These conversions are handled for us. A number passed in is wrapped in Int64 by using the Int64 constructor: new Int64(64). A string passed in place of an Int64 is converted using the static fromDecimalString method: Int64.fromDecimalString('64'). Similarly string data can be passed to a binary field and the conversion to Buffer is handled under the hood. This are just convinience interfaces to make handling the Thrift objects in TypeScript a little easier. You will notice service methods always return an object of the more strict interface. Also, the more strict interface can always be passed where the loose interface is expected.

Sending Data Over the Wire

When it comes to struct-like data types (struct, union and exception) usually you don't need to know much more than what data types are generated. However, in addition to the generated interface/union/class the code generator also creates a companion object that knows how to send the given object over the wire.

Looking back at the User object from our struct example, in addition to the interface, the code generator creates a codec object like this:

export const UserCodec: thrift.IStructCodec<IUserArgs, IUser> {
    encode(obj: IUserArgs, output: thrift.TProtocol): void {
        // ...
    },
    decode(input: thrift.TProtocol): IUser {
        // ...
    }
}

It's just an object that knows how to read the given object from a Thrift Protocol or write the given object to a Thrift Protocol.

The codec will always follow this naming convention, just appending Codec onto the end of your struct name.

Strict Unions

Note: Strict unions require thrift-server version 0.13.x or higher.

This is an option only available when generating for thrift-server. This option will generate Thrift unions as TypeScript unions. This changes the codegen in a few significant ways.

Back with our example union definition:

union MyUnion {
    1: string option1
    2: i32 option2
}

When compiling with the --strictUnions flag we now generate TypeScript like this:

enum MyUnionType {
    MyUnionWithOption1 = "option1",
    MyUnionWithOption2 = "option2"
}
type MyUnion = IMyUnionWithOption1 | IMyUnionWithOption2
interface IMyUnionWithOption1 {
    __type: MyUnionType.MyUnionWithOption1
    option1: string
    option2?: undefined
}
interface IMyUnionWithOption2 {
    __type: MyUnionType.MyUnionWithOption2
    option1?: undefined
    option2: number
}
type MyUnionArgs = IMyUnionWithOption1Args | IMyUnionWithOption2Args
interface IMyUnionWithOption1Args {
    option1: string
    option2?: undefined
}
interface IMyUnionWithOption2Args {
    option1?: undefined
    option2: number
}

The enum represents all potential values of the __type property attached to each variation of our union. Instead of generating one interface with optional properties we generate one interface for each field where that field is required. Our resulting type is then the union of multiple interfaces each with only one property. This provides compile-time guarantees that we are setting one and only one field for the union.

The loose interfaces, the Args interfaces, behave much like the loose interfaces for structs. They allow you to use number in place of Int64 or allow you to pass either string or Buffer for binary types. In addition, they also forgo the __type property. In the codegen we can tell what you are passing by the fields you set. This means in most instances you don't need to provide the __type property. You can use the loose interfaces as the return value for service functions or as the arguments for client methods.

This output is more complex, but it allows us to do a number of things. The most significant of which may be that it allows us to take advantage of discriminated unions in our application code:

function processUnion(union: MyUnion) {
    switch (union.__type) {
        case MyUnionType.MyUnionWithOption1:
            // Do something
        case MyUnionType.MyUnionWithOption2:
            // Do something
        default:
            const _exhaustiveCheck: never = union
            throw new Error(`Non-exhaustive match for type: ${_exhaustiveCheck}`)
    }
}

The fact that each interface we generate defines one required field and some n number of optional undefined fields we can do things like check union.option2 !== undefined without a compiler error, but we will get a compiler error if you try to use a value that shouldn't exist on a given union. This expands the ways you can operate on unions to be more general.

Using this form will require that you prove to the compiler that one (and only one) field is set for your unions.

In addition to the changed types output, the --strictUnions flag changes the output of the Codec object. The Codec object will have one additional method create. The create method takes one of the loose interfaces and coerces it into the strict interface (including the __type property).

For the example MyUnion that would be defined as:

const MyUnionCodec: thrift.IStructToolkit<IUserArgs, IUser> { = {
    create(args: MyUnionArgs): MyUnion {
        // ...
    },
    encode(obj: IUserArgs, output: thrift.TProtocol): void {
        // ...
    },
    decode(input: thrift.TProtocol): IUser {
        // ...
    }
}

Note: In a future breaking release all the Codec objects will be renamed to Toolkit as they will provide more utilities for working with defined Thrift objects.

Apache Thrift

The generated code can also work with the Apache Thrift Library.

$ npm install --save thrift
$ npm install --save @types/thrift

Given this service let's build a client and server based on our generated code.

service Calculator {
    i32 add(1: i32 left, 2: i32 right)
    i32 subtract(1: i32 left, 2: i32 right)
}

Run codegen for your Thrift service. Here the --target option isn't needed as apache is the default build target.

$ thrift-typescript --rootDir . --sourceDir thrift --outDir codegen

Client

import {
    createHttpConnection,
    createHttpClient,
    HttpConnection,
} from 'thrift'

import { Calculator } from './codegen/calculator'

// The location of the server endpoint
const CONFIG = {
    hostName: 'localhost',
    port: 8045
}

const options = {
    transport: TBufferedTransport,
    protocol: TBinaryProtocol,
    https: false,
    headers: {
        Host: config.hostName,
    }
}

const connection: HttpConnection = createHttpConnection(CONFIG.hostName, CONFIG.port, options)
const thriftClient: Calculator.Client = createHttpClient(Calculator.Client, connection)

// All client methods return a Promise of the expected result.
thriftClient.add(5, 6).then((result: number) =>{
    console.log(`result: ${result}`)
})

Server

import {
    createWebServer,
    TBinaryProtocol,
    TBufferedTransport,
} from 'thrift'

import { Calculator } from './codegen/calculator'

// Handler: Implement the Calculator service
const myServiceHandler = {
    add(left: number, right: number): number {
        return left + right
    },
    subtract(left: number, right: number): number {
        return left - right
    },
}

// ServiceOptions: The I/O stack for the service
const myServiceOpts = {
    handler: myServiceHandler,
    processor: Calculator,
    protocol: TBinaryProtocol,
    transport: TBufferedTransport
}

// ServerOptions: Define server features
const serverOpt = {
    services: {
        '/': myServiceOpts
    }
}

// Create and start the web server
const port: number = 8045;
createWebServer(serverOpt).listen(port, () => {
    console.log(`Thrift server listening on port ${port}`)
})

Notes

The gererated code can be used with many of the more strict tsc compiler options.

{
    "compilerOptions": {
        "noImplicitAny": true,
        "noImplicitThis": true,
        "strictNullChecks": true,
        "strictFunctionTypes": true,
        "noUnusedLocals": true
    }
}

Development

Install dependencies with

npm install

Build

npm run build

Run test in watch mode

npm run test:watch

Contributing

For more information about contributing new features and bug fixes, see our Contribution Guidelines. External contributors must sign Contributor License Agreement (CLA)

License

This project is licensed under Apache License Version 2.0