@OliverJAsh no need to create any unique method exactly for Eithers. Js provides a safe way to make custom protocols via Symbol interfaces. Therefore it is possible to store any type meta info in fields and methods inside symbol property to tell other modern functional libraries about your typeclasses in a conventional way

And we already have success story. Streaming libraries are uses symbol-observable to have interop with each other.

import { from } from 'most'
import { createStore } from 'redux'

const someStore = createStore(s => s) // typical redux store...
const storeUpdates$ =
  from(someStore) // ...also supports interconnection via Symbol.observable

So why not to try it too?

I already made an implementation of that concept and using it in production code https://github.com/zerobias/apropos/tree/develop/packages/signature
It contains symbol itself and several methods to use class type signatures in statically typed environment

from io-ts.

@OliverJAsh What do you mean with "adapter"? A fold called on the validation result is the simplest solution I can think of

import * as t from 'io-ts'
import { Either, Left, Right } from 'funfix-core'

type FunFixValidation<T> = Either<Array<t.ValidationError>, T>

function myvalidate<T>(v: any, type: t.Type<T>): FunFixValidation<T> {
  return t.validate(v, type).fold<FunFixValidation<T>>(Left, Right)
}

from io-ts.

@gcanti I think @OliverJAsh was maybe thinking about something without the overhead to write the call.. But I can't see that happening without all code using fp-ts Either to be abstracted over it, which seems to bring a lot more overhead..

from io-ts.

By "adapter" I was thinking of a way to provide an Either constructor to io-ts. E.g.

t.validate(value, type, {
  getEither: () => FunfixEither
})

Essentially I want to avoid depending on (and therefore bundling) fp-ts when using io-ts if I already have another Either type in my application. 🤔

from io-ts.

That's food for thought. Thanks for sharing @zerobias!

from io-ts.

But I can't see that happening without all code using fp-ts Either to be abstracted over it

@sledorze We can go even further, we might abstract over the effect returned by validate and use finally tagless

export type Errors = Array<ValidationError>

export interface MonadThrow<E, M> extends Monad<M> {
  throwError: <A>(e: E) => HKT<M, A>
}

export interface MonadType<M> extends MonadThrow<Errors, M> {
  zipWith: <A, B, C>(f: (a: A, b: B) => C) => (fa: HKT<M, A>, lazyfb: () => HKT<M, B>) => HKT<M, C>
}

export class Type<M, S, A> {
  readonly _A: A
  readonly _S: S
  readonly _M: M
  constructor(
    readonly name: string,
    readonly is: (v: any) => v is A,
    readonly validate: (s: S, c: Context) => HKT<M, A>,
    readonly serialize: (a: A) => S
  ) {}
}

export class StringType<M> extends Type<M, any, string> {
  readonly _tag: 'StringType' = 'StringType'
  constructor(M: MonadType<M>) {
    super(
      'number',
      (v): v is string => typeof v === 'string',
      (s, c) => (this.is(s) ? M.of(s) : M.throwError([{ value: s, context: c }])),
      identity
    )
  }
}

export const getStringType = <M>(M: MonadType<M>): StringType<M> => new StringType(M)

It would allow

• other Either implementations
• async validations
• "bail out" validations (i.e. compute and return only the first error)

from io-ts.

I wrote a POC supporting asynchronous validations (branch 78)

Basically it works like this: you provide an instance of MonadType<M> for some monad M and you get back a complete runtime type system which works in that monadic context.

// core.ts
export interface MonadThrow<E, M> extends Monad<M> {
  throwError: <A>(e: E) => HKT<M, A>
}

export interface MonadType<M> extends MonadThrow<Array<ValidationError>, M> {
  zipWith: <A, B, C>(f: (a: A, b: B) => C) => (fa: HKT<M, A>, lazyfb: () => HKT<M, B>) => HKT<M, C>
  attempt: <A>(fx: HKT<M, A>, lazyfy: () => HKT<M, A>) => HKT<M, A>
}

export type Is<A> = (v: any) => v is A
export type Validate<M, S, A> = (s: S, context: Context) => HKT<M, A>
export type Serialize<S, A> = (a: A) => S

export class Type<M, S, A> {
  readonly '_A': A
  readonly '_S': S
  readonly '_M': M
  constructor(
    readonly name: string,
    readonly is: Is<A>,
    readonly validate: Validate<M, S, A>,
    readonly serialize: Serialize<S, A>
  ) {}
}

export const getTypeSystem = <M>(M: MonadType<M>): TypeSystem<M> => {
  ...
}

So depending on the instance you pass, you can choose

• sync / async validations
• get all errors / only the first
• Either implementation

I wrote 5 instances:

• src/either.ts: (2 instances) for fp-ts Either: all errors / first error + synchronous validations
• src/taskEither.ts: (2 instances) for fp-ts TaskEither: all errors / first error + asynchronous validations
• test/sync/funxfix/interface.ts (1 instance) for funfix Either: all errors + synchronous validations

from io-ts.

@sledorze This is a first attempt to get a basic idea of performances (using benchmark.js)

Code

suite
  .add('io-ts', function() {
    const T = t.type({
      a: t.string,
      b: t.number,
      c: t.array(t.boolean),
      d: t.tuple([t.number, t.string])
    })
    t.validate({}, T)
  })

Results

# 0.8.2

io-ts x 140,895 ops/sec ±3.03% (79 runs sampled)
io-ts x 140,089 ops/sec ±3.37% (77 runs sampled)
io-ts x 142,319 ops/sec ±2.71% (80 runs sampled)

# 0.9.0

io-ts x 190,435 ops/sec ±0.77% (88 runs sampled)
io-ts x 190,476 ops/sec ±0.67% (87 runs sampled)
io-ts x 182,115 ops/sec ±0.65% (88 runs sampled)

# POC

io-ts x 153,537 ops/sec ±0.68% (86 runs sampled)
io-ts x 155,058 ops/sec ±0.82% (88 runs sampled)
io-ts x 152,557 ops/sec ±0.71% (84 runs sampled)

from io-ts.

@gcanti that's interesting, also wondering what happened between 0.8.2 and 0.9.0..
What is the platform tested? nodeJs? this brings so much questions
is the POC test doing a 'fail fast' path with the monadic 'only the first error' approach?

from io-ts.

what happened between 0.8.2 and 0.9.0

@sledorze it looks like creating a type is less expensive in 0.9.0+

node --version
v8.1.2

Type definitions only

Code

import * as Benchmark from 'benchmark'
import * as t from 'io-ts'

const suite = new Benchmark.Suite()

suite
  .add('io-ts', function() {
    t.type({
      a: t.string,
      b: t.number,
      c: t.array(t.boolean),
      d: t.tuple([t.number, t.string])
    })
  })
  .on('cycle', function(event: any) {
    console.log(String(event.target))
  })
  .on('complete', function(this: any) {
    console.log('Fastest is ' + this.filter('fastest').map('name'))
  })
  .run({ async: true })

Results

0.8.2

277,827 ops/sec ±2.50% (72 runs sampled)

0.9.0

431,908 ops/sec ±0.65% (88 runs sampled)

POC

424,371 ops/sec ±0.63% (88 runs sampled)

Validations only

Code

import * as Benchmark from 'benchmark'
import * as t from 'io-ts'

const suite = new Benchmark.Suite()

const T = t.type({
  a: t.string,
  b: t.number,
  c: t.array(t.boolean),
  d: t.tuple([t.number, t.string])
})
const payload = { c: [1], d: ['foo'] }

suite
  .add('io-ts', function() {
    t.validate(payload, T)
  })
  .on('cycle', function(event: any) {
    console.log(String(event.target))
  })
  .on('complete', function(this: any) {
    console.log('Fastest is ' + this.filter('fastest').map('name'))
  })
  .run({ async: true })

Results

0.8.2

210,008 ops/sec ±0.63% (87 runs sampled)

0.9.0

205,194 ops/sec ±0.66% (85 runs sampled)

POC (all errors)

158,106 ops/sec ±0.56% (87 runs sampled)

POC (first error)

745,731 ops/sec ±0.60% (88 runs sampled)

from io-ts.

@gcanti is there some differences on all/first error for successful matches? (no error generated?)

from io-ts.

@sledorze not that much

All
88,530 ops/sec ±1.46% (86 runs sampled)
First
94,668 ops/sec ±0.94% (89 runs sampled)

It looks like there are some low hanging fruits, for example this is the results using linked lists

Base reference (0.9.0)

invalid payload x 197,033 ops/sec ±0.84% (82 runs sampled)
valid payload (no errors) x 178,816 ops/sec ±1.22% (89 runs sampled)

Context and Errors as Arrays (POC all errors)

invalid payload x 154,399 ops/sec ±0.97% (85 runs sampled)
valid payload (no errors) x 91,946 ops/sec ±0.98% (87 runs sampled)

Context and Errors as LinkedLists + optimized zipEithers (POC all errors)

invalid payload x 509,834 ops/sec ±0.52% (89 runs sampled)
valid payload (no errors) x 148,614 ops/sec ±1.92% (85 runs sampled)

from io-ts.

@gcanti indeed, that's quite a lot of performance gains.
btw are your libs integrated into a CI ?

It may be interesting to track down that automatically, like so:
https://travis-ci.org/lloydmeta/frunk/jobs/172486500#L398

I may help on that (if we decide to integrate on a CI first).

from io-ts.