Skip to main content

eitherway

Yet Another Option and Result Implementation (YAORI)

Safe abstractions for fallible flows inspired by F# and Rust.

Disclaimer

This is still experimental software.

Motivation

Let’s be honest: The community already nurtured a bunch of great projects, providing similar abstractions. The goal has always been to make it easier and more explicit to handle error cases under the “errors are values” premise. After having worked with quite a few of these existing abstractions, a couple of issues arose/persisted:

  • Overly strict: Some of the existing implementations don’t account for the variance emerging naturally in a structural type system.
  • Onboarding is hard: Most of the existing projects provide either very little or scattered documentation. Forcing new users to constantly switch context in order to understand how they can achieve their goal or if their chosen operation is suitable for their use case.
  • Lack of async support: Very few existing projects offer abstractions for working in an async context, none of them really being first class citizens.

The goal here really is to make the abstractions provided by eitherway the most safe, productive and overall enjoyable to work with. Irrespective of experience or employed context.

Design Goals

eitherway is trying to close the gap between type-safety, idiom and productivity by focusing on the following goals:

  • Pragmatism: There is no need to strictly port something from another language with very different semantics. When decisions arrive, which call for a trade-off, eitherway will always try to offer a solution geared towards the constraints and idiom of Typescript.
  • Compatibility: Interacting with one of the structures defined by eitherway should be painless in the sense that things do what you would expect them to do in Typescript and are compatible with inherent language protocols (e.g. Iterator protocol).
  • Safety: Once an abstraction is instantiated, no inherent operation should ever panic.
  • Performance: All abstractions provided here should strive to amortize the cost of usage and be upfront about these costs.
  • Documentation: All structures come with full, inline documentation. Making it easy to understand what is currently happening and if the chosen operation is suitable for the desired use case. (Still very much in progress)

eitherway in Action

function toUpperCase(input: string | undefined): Task<string, TypeError> {
  return Option(input) // All nullish values are None
    .okOrElse(() => TypeError("Input is undefined")) // Convert to Result<string, TypeError>
    .into((res) => Task.of(res)) // Push the result into an async context
    .map(async (str) => {
      await sleep(1);
      return str.toUpperCase();
    });
}

function stringToLength(input: string): Task<number, TypeError> {
  return Option.fromCoercible(input) // All falsy types are None
    .okOrElse(() => TypeError("Input string is empty"))
    .into((res) => Task.of(res))
    .map(async (str) => {
      await sleep(1);
      return str.length;
    });
}

function powerOfSelf(input: number): Task<number, TypeError> {
  return Option.fromCoercible(input)
    .okOrElse(() =>
      TypeError("Cannot perform computation with NaN, Infinity or 0")
    )
    .into((res) => Task.of(res))
    .andThen(async (n) => { // Promise<Result<T, E>> and Task<T, E> can be used interchangeably for async composition
      await sleep(1);
      return Option.fromCoercible(Math.pow(n, n))
        .okOrElse(() => TypeError("Cannot calculate result"));
    });
}

function processString(input: string | undefined): Task<number, TypeError> {
  return toUpperCase(input) // Synchronous and asynchronous composition work the same
    .andThen(stringToLength)
    .andThen(powerOfSelf);
}

async function main(): Promise<Result<number, TypeError>> {
  const result = await processString("foo"); // Task is of course awaitable

  return Task.of(result); // You can return Task<T, E> as Promise<Result<T, E>>
}

main()
  .then((result) => {
    result // Result<number, TypeError>
      .inspect(console.log)
      .inspectErr(console.error);
  })
  .catch((e) => "Unreachable!")
  .finally(() => console.log("DONE!"));

Installation

Getting started

API

Option

Result

Task

Best Practices

FAQ

Q: Why should I even use something like this? A: It’s nice. Really.

Explicit error types and built-in happy/error path selectors lead to expressive code which is often even more pleasant to read.

Compare these examples, taken from the benchmark suite: 
/**
 * ==================
 *    SYNCHRONOUS
 * ==================
 */

/* Classic exception style */
function processString(input: string | undefined): number {
  try {
    const upperCased = toUpperCase(input);
    const length = stringToLength(upperCased);
    return powerOfSelf(length);
  } catch (error: unknown) {
    if (error instanceof TypeError) {
      console.error(error.message);
      throw error;
    }
    throw new TypeError("Unknown error", { cause: error });
  }
}

/* Equivalent Result flow */
function processString(input: string | undefined): Result<number, TypeError> {
  return toUpperCase(input)
    .andThen(stringToLength)
    .andThen(powerOfSelf)
    .inspectErr((e) => console.error(e.message));
}

/**
 * ==================
 *    ASYNCHRONOUS
 * ==================
 */

/* Classic exception style */
async function processString(input: string | undefined): Promise<number> {
  try {
    const upperCased = await toUpperCase(input);
    const length = await stringToLength(upperCased);
    return await powerOfSelf(length);
  } catch (error: unknown) {
    if (error instanceof TypeError) {
      console.error(error.message);
      throw error;
    }
    throw new TypeError("Unknown error", { cause: error });
  }
}

/* Equivalent Task flow */
function processString(input: string | undefined): Task<number, TypeError> {
  return toUpperCase(input)
    .andThen(stringToLength)
    .andThen(powerOfSelf)
    .inspectErr((e) => console.error(e.message));
}

Apart from making error cases explicit, the abstractions provided here foster a code style, which naturally builds up complex computations via composition of small, focused functions/methods, where boundaries are defined by values. Thus leading to a highly maintainable and easily testable code base.

Even better: These abstractions come with practically no overhead (see the next section).

Here are a couple of videos, explaining the general benefits in more detail:

Q: What is the performance impact of using this? A: Practically none.

You can run the benchmark suite yourself with $ deno bench.

The benchmark results suggest, that for nearly all practical considerations there is no or virtually no overhead of using the abstractions provided by eitherway vs. a classic exception propagation style.

Although the result and task flows were slightly faster in the runs below, it’s important not to fall into a micro-optimization trap. The conclusion should not necessarily be “use eitherway, it’s faster”, but rather “use eitherway, it’s practically free”.

The overall performance thesis is that by returning errors instead of throwing, catching and re-throwing exceptions, the instantiation costs of the abstractions provided here are amortized over call-stack depth & it’s size, as well as the optimizations the linear return path allows, sometimes even leading to small performance improvements. This sounds plausible, and the results are not refuting the null hypothesis here, but benchmarking is hard and for most use cases, the difference really won’t matter.

Synchronous exception propagation vs. result chaining 
cpu: Intel(R) Core(TM) i9-9880H CPU @ 2.30GHz runtime: deno 1.33.2
(x86_64-apple-darwin)

## file:///projects/eitherway/bench/sync_bench.ts benchmark time (avg) (min … max) p75 p99 p995

SyncExceptions 29.15 µs/iter (20.54 µs … 472 µs) 31.34 µs 38.22 µs 49.28 µs
SyncResultFlow 15.49 µs/iter (11.07 µs … 441.17 µs) 15.44 µs 31.69 µs 43.37 µs

summary SyncResultFlow 1.88x faster than SyncExceptions
Asynchronous exception propagation vs. task chaining 
cpu: Intel(R) Core(TM) i9-9880H CPU @ 2.30GHz runtime: deno 1.33.2
(x86_64-apple-darwin)

## file:///projects/eitherway/bench/async_bench.ts benchmark time (avg) (min … max) p75 p99 p995

AsyncExceptions 24.78 ms/iter (22.08 ms … 25.55 ms) 25.46 ms 25.55 ms 25.55 ms
TaskInstanceFlow 23.88 ms/iter (21.28 ms … 25.8 ms) 24.57 ms 25.8 ms 25.8 ms
TaskOperatorFlow 24.21 ms/iter (21.33 ms … 25.73 ms) 25.36 ms 25.73 ms 25.73 ms
TaskEarlyReturnFlow 24.04 ms/iter (20.36 ms … 25.47 ms) 25.42 ms 25.47 ms 25.47
ms

summary TaskInstanceFlow 1.01x faster than TaskEarlyReturnFlow 1.01x faster than
TaskOperatorFlow 1.04x faster than AsyncExceptions
Micro benchmarks If you have a highly performance sensitive use case, you should be using a different language. On a more serious note, also small costs can add up and as a user, you should know how high the costs are. So here are a few micro benchmarks: 
cpu: Intel(R) Core(TM) i9-9880H CPU @ 2.30GHz runtime: deno 1.33.2
(x86_64-apple-darwin)

## file:///projects/eitherway/bench/micro_bench.ts benchmark time (avg) (min … max) p75 p99 p995

Promise.resolve(Ok) 44.33 ns/iter (35.81 ns … 106.41 ns) 44.6 ns 62.58 ns 72.56
ns Task.succeed 105.43 ns/iter (88.44 ns … 227.26 ns) 108.97 ns 204.75 ns 212.54
ns Promise.resolve(Err) 3.11 µs/iter (3.06 µs … 3.27 µs) 3.13 µs 3.27 µs 3.27 µs
Task.fail 2.94 µs/iter (2.71 µs … 3.35 µs) 3.25 µs 3.35 µs 3.35 µs

summary Promise.resolve(Ok) 2.38x faster than Task.succeed 66.41x faster than
Task.fail 70.14x faster than Promise.resolve(Err)

## file:///projects/eitherway/bench/micro_bench.ts benchmark time (avg) (min … max) p75 p99 p995

Ok 5.1 ns/iter (4.91 ns … 22.27 ns) 5.02 ns 8.62 ns 11.67 ns Err 4.88 ns/iter
(4.7 ns … 17.93 ns) 4.81 ns 8.18 ns 10.52 ns Option 90.39 ns/iter (83.63 ns …
172.61 ns) 93.31 ns 135.19 ns 146.79 ns

summary Err 1.05x faster than Ok 18.52x faster than Option

## file:///projects/eitherway/bench/micro_bench.ts benchmark time (avg) (min … max) p75 p99 p995

Async Exception Propagation 9.08 µs/iter (8.95 µs … 9.26 µs) 9.18 µs 9.26 µs
9.26 µs Async Error Propagation 6.32 µs/iter (6.24 µs … 6.52 µs) 6.37 µs 6.52 µs
6.52 µs

summary Async Error Propagation 1.44x faster than Async Exception Propagation
Q: Why can’t I use Task as the return type of an async function? A: That’s a general restriction of JavaScript.

A function defined with the async keyword, must return a “system” Promise. Although Task<T, E> (currently) is a proper subclass of Promise, it cannot be used in the Return Type Position of an async function, because it’s NOT a “system” promise (for lack of a better word).

Since Task<T, E> is a subclass of Promise<Result<T, E>>, it’s possible to return it as such from an async function though or just await it.

async function toTask(str: string): Promise<Result<string, never>> {
  return Task.succeed(str);
}

Furthermore, Task<T, E> is merely a composability extension for Promise<Result<T, E>>. As such, you can cheaply convert every Promise<Result<T, E> via the Task.of() constructor, or use the promise operators to compose your pipeline.

Q: Why subclassing Promises instead of just providing a PromiseLike abstraction? A: For compatibility reasons.

The drawback of the current implementation is that we cannot evaluate Task<T, E> lazily. On the other hand, a lot of framework or library code is still (probably needlessly) invariant over PromiseLike types. Therefore subclassing the native Promise and allowing the users to treat Promise<Result<T, E>> and Task<T, E> interchangeably in most situations, was the preferred solution.

Prior Art

License & Contributing

Contributing

Please see CONTRIBUTING for more information.

License

Licensed under MIT license.