21 KiB

Raw Blame History

minipass

A very minimal implementation of a PassThrough stream

It's very fast for objects, strings, and buffers.

Supports pipe()ing (including multi-pipe() and backpressure transmission), buffering data until either a data event handler or pipe() is added (so you don't lose the first chunk), and most other cases where PassThrough is a good idea.

There is a read() method, but it's much more efficient to consume data from this stream via 'data' events or by calling pipe() into some other stream. Calling read() requires the buffer to be flattened in some cases, which requires copying memory.

There is also no unpipe() method. Once you start piping, there is no stopping it!

If you set objectMode: true in the options, then whatever is written will be emitted. Otherwise, it'll do a minimal amount of Buffer copying to ensure proper Streams semantics when read(n) is called.

objectMode can also be set by doing stream.objectMode = true, or by writing any non-string/non-buffer data. objectMode cannot be set to false once it is set.

This is not a through or through2 stream. It doesn't transform the data, it just passes it right through. If you want to transform the data, extend the class, and override the write() method. Once you're done transforming the data however you want, call super.write() with the transform output.

For some examples of streams that extend Minipass in various ways, check out:

Differences from Node.js Streams

There are several things that make Minipass streams different from (and in some ways superior to) Node.js core streams.

Please read these caveats if you are familiar with node-core streams and intend to use Minipass streams in your programs.

Timing

Minipass streams are designed to support synchronous use-cases. Thus, data is emitted as soon as it is available, always. It is buffered until read, but no longer. Another way to look at it is that Minipass streams are exactly as synchronous as the logic that writes into them.

This can be surprising if your code relies on PassThrough.write() always providing data on the next tick rather than the current one, or being able to call resume() and not have the entire buffer disappear immediately.

However, without this synchronicity guarantee, there would be no way for Minipass to achieve the speeds it does, or support the synchronous use cases that it does. Simply put, waiting takes time.

This non-deferring approach makes Minipass streams much easier to reason about, especially in the context of Promises and other flow-control mechanisms.

No High/Low Water Marks

Node.js core streams will optimistically fill up a buffer, returning true on all writes until the limit is hit, even if the data has nowhere to go. Then, they will not attempt to draw more data in until the buffer size dips below a minimum value.

Minipass streams are much simpler. The write() method will return true if the data has somewhere to go (which is to say, given the timing guarantees, that the data is already there by the time write() returns).

If the data has nowhere to go, then write() returns false, and the data sits in a buffer, to be drained out immediately as soon as anyone consumes it.

Hazards of Buffering (or: Why Minipass Is So Fast)

Since data written to a Minipass stream is immediately written all the way through the pipeline, and write() always returns true/false based on whether the data was fully flushed, backpressure is communicated immediately to the upstream caller. This minimizes buffering.

Consider this case:

const {PassThrough} = require('stream')
const p1 = new PassThrough({ highWaterMark: 1024 })
const p2 = new PassThrough({ highWaterMark: 1024 })
const p3 = new PassThrough({ highWaterMark: 1024 })
const p4 = new PassThrough({ highWaterMark: 1024 })

p1.pipe(p2).pipe(p3).pipe(p4)
p4.on('data', () => console.log('made it through'))

// this returns false and buffers, then writes to p2 on next tick (1)
// p2 returns false and buffers, pausing p1, then writes to p3 on next tick (2)
// p3 returns false and buffers, pausing p2, then writes to p4 on next tick (3)
// p4 returns false and buffers, pausing p3, then emits 'data' and 'drain'
// on next tick (4)
// p3 sees p4's 'drain' event, and calls resume(), emitting 'resume' and
// 'drain' on next tick (5)
// p2 sees p3's 'drain', calls resume(), emits 'resume' and 'drain' on next tick (6)
// p1 sees p2's 'drain', calls resume(), emits 'resume' and 'drain' on next
// tick (7)

p1.write(Buffer.alloc(2048)) // returns false

Along the way, the data was buffered and deferred at each stage, and multiple event deferrals happened, for an unblocked pipeline where it was perfectly safe to write all the way through!

Furthermore, setting a highWaterMark of 1024 might lead someone reading the code to think an advisory maximum of 1KiB is being set for the pipeline. However, the actual advisory buffering level is the sum of highWaterMark values, since each one has its own bucket.

Consider the Minipass case:

const m1 = new Minipass()
const m2 = new Minipass()
const m3 = new Minipass()
const m4 = new Minipass()

m1.pipe(m2).pipe(m3).pipe(m4)
m4.on('data', () => console.log('made it through'))

// m1 is flowing, so it writes the data to m2 immediately
// m2 is flowing, so it writes the data to m3 immediately
// m3 is flowing, so it writes the data to m4 immediately
// m4 is flowing, so it fires the 'data' event immediately, returns true
// m4's write returned true, so m3 is still flowing, returns true
// m3's write returned true, so m2 is still flowing, returns true
// m2's write returned true, so m1 is still flowing, returns true
// No event deferrals or buffering along the way!

m1.write(Buffer.alloc(2048)) // returns true

It is extremely unlikely that you don't want to buffer any data written, or ever buffer data that can be flushed all the way through. Neither node-core streams nor Minipass ever fail to buffer written data, but node-core streams do a lot of unnecessary buffering and pausing.

As always, the faster implementation is the one that does less stuff and waits less time to do it.

Immediately emit `end` for empty streams (when not paused)

If a stream is not paused, and end() is called before writing any data into it, then it will emit end immediately.

If you have logic that occurs on the end event which you don't want to potentially happen immediately (for example, closing file descriptors, moving on to the next entry in an archive parse stream, etc.) then be sure to call stream.pause() on creation, and then stream.resume() once you are ready to respond to the end event.

Emit `end` When Asked

One hazard of immediately emitting 'end' is that you may not yet have had a chance to add a listener. In order to avoid this hazard, Minipass streams safely re-emit the 'end' event if a new listener is added after 'end' has been emitted.

Ie, if you do stream.on('end', someFunction), and the stream has already emitted end, then it will call the handler right away. (You can think of this somewhat like attaching a new .then(fn) to a previously-resolved Promise.)

To prevent calling handlers multiple times who would not expect multiple ends to occur, all listeners are removed from the 'end' event whenever it is emitted.

Impact of "immediate flow" on Tee-streams

A "tee stream" is a stream piping to multiple destinations:

const tee = new Minipass()
t.pipe(dest1)
t.pipe(dest2)
t.write('foo') // goes to both destinations

Since Minipass streams immediately process any pending data through the pipeline when a new pipe destination is added, this can have surprising effects, especially when a stream comes in from some other function and may or may not have data in its buffer.

// WARNING! WILL LOSE DATA!
const src = new Minipass()
src.write('foo')
src.pipe(dest1) // 'foo' chunk flows to dest1 immediately, and is gone
src.pipe(dest2) // gets nothing!

The solution is to create a dedicated tee-stream junction that pipes to both locations, and then pipe to that instead.

// Safe example: tee to both places
const src = new Minipass()
src.write('foo')
const tee = new Minipass()
tee.pipe(dest1)
tee.pipe(dest2)
src.pipe(tee) // tee gets 'foo', pipes to both locations

The same caveat applies to on('data') event listeners. The first one added will immediately receive all of the data, leaving nothing for the second:

// WARNING! WILL LOSE DATA!
const src = new Minipass()
src.write('foo')
src.on('data', handler1) // receives 'foo' right away
src.on('data', handler2) // nothing to see here!

Using a dedicated tee-stream can be used in this case as well:

// Safe example: tee to both data handlers
const src = new Minipass()
src.write('foo')
const tee = new Minipass()
tee.on('data', handler1)
tee.on('data', handler2)
src.pipe(tee)

USAGE

It's a stream! Use it like a stream and it'll most likely do what you want.

const Minipass = require('minipass')
const mp = new Minipass(options) // optional: { encoding, objectMode }
mp.write('foo')
mp.pipe(someOtherStream)
mp.end('bar')

OPTIONS

encoding How would you like the data coming out of the stream to be encoded? Accepts any values that can be passed to Buffer.toString().
objectMode Emit data exactly as it comes in. This will be flipped on by default if you write() something other than a string or Buffer at any point. Setting objectMode: true will prevent setting any encoding value.

API

Implements the user-facing portions of Node.js's Readable and Writable streams.

Methods

write(chunk, [encoding], [callback]) - Put data in. (Note that, in the base Minipass class, the same data will come out.) Returns false if the stream will buffer the next write, or true if it's still in "flowing" mode.
end([chunk, [encoding]], [callback]) - Signal that you have no more data to write. This will queue an end event to be fired when all the data has been consumed.
setEncoding(encoding) - Set the encoding for data coming of the stream. This can only be done once.
pause() - No more data for a while, please. This also prevents end from being emitted for empty streams until the stream is resumed.
resume() - Resume the stream. If there's data in the buffer, it is all discarded. Any buffered events are immediately emitted.
pipe(dest) - Send all output to the stream provided. There is no way to unpipe. When data is emitted, it is immediately written to any and all pipe destinations.
on(ev, fn), emit(ev, fn) - Minipass streams are EventEmitters. Some events are given special treatment, however. (See below under "events".)
promise() - Returns a Promise that resolves when the stream emits end, or rejects if the stream emits error.
collect() - Return a Promise that resolves on end with an array containing each chunk of data that was emitted, or rejects if the stream emits error. Note that this consumes the stream data.
concat() - Same as collect(), but concatenates the data into a single Buffer object. Will reject the returned promise if the stream is in objectMode, or if it goes into objectMode by the end of the data.
read(n) - Consume n bytes of data out of the buffer. If n is not provided, then consume all of it. If n bytes are not available, then it returns null. Note consuming streams in this way is less efficient, and can lead to unnecessary Buffer copying.
destroy([er]) - Destroy the stream. If an error is provided, then an 'error' event is emitted. If the stream has a close() method, and has not emitted a 'close' event yet, then stream.close() will be called. Any Promises returned by .promise(), .collect() or .concat() will be rejected. After being destroyed, writing to the stream will emit an error. No more data will be emitted if the stream is destroyed, even if it was previously buffered.

Properties

bufferLength Read-only. Total number of bytes buffered, or in the case of objectMode, the total number of objects.
encoding The encoding that has been set. (Setting this is equivalent to calling setEncoding(enc) and has the same prohibition against setting multiple times.)
flowing Read-only. Boolean indicating whether a chunk written to the stream will be immediately emitted.
emittedEnd Read-only. Boolean indicating whether the end-ish events (ie, end, prefinish, finish) have been emitted. Note that listening on any end-ish event will immediateyl re-emit it if it has already been emitted.
writable Whether the stream is writable. Default true. Set to false when end()
readable Whether the stream is readable. Default true.
buffer A yallist linked list of chunks written to the stream that have not yet been emitted. (It's probably a bad idea to mess with this.)
pipes A yallist linked list of streams that this stream is piping into. (It's probably a bad idea to mess with this.)
destroyed A getter that indicates whether the stream was destroyed.
paused True if the stream has been explicitly paused, otherwise false.
objectMode Indicates whether the stream is in objectMode. Once set to true, it cannot be set to false.

Events

data Emitted when there's data to read. Argument is the data to read. This is never emitted while not flowing. If a listener is attached, that will resume the stream.
end Emitted when there's no more data to read. This will be emitted immediately for empty streams when end() is called. If a listener is attached, and end was already emitted, then it will be emitted again. All listeners are removed when end is emitted.
prefinish An end-ish event that follows the same logic as end and is emitted in the same conditions where end is emitted. Emitted after 'end'.
finish An end-ish event that follows the same logic as end and is emitted in the same conditions where end is emitted. Emitted after 'prefinish'.
close An indication that an underlying resource has been released. Minipass does not emit this event, but will defer it until after end has been emitted, since it throws off some stream libraries otherwise.
drain Emitted when the internal buffer empties, and it is again suitable to write() into the stream.
readable Emitted when data is buffered and ready to be read by a consumer.
resume Emitted when stream changes state from buffering to flowing mode. (Ie, when resume is called, pipe is called, or a data event listener is added.)

Static Methods

Minipass.isStream(stream) Returns true if the argument is a stream, and false otherwise. To be considered a stream, the object must be either an instance of Minipass, or an EventEmitter that has either a pipe() method, or both write() and end() methods. (Pretty much any stream in node-land will return true for this.)

EXAMPLES

Here are some examples of things you can do with Minipass streams.

simple "are you done yet" promise

mp.promise().then(() => {
  // stream is finished
}, er => {
  // stream emitted an error
})

collecting

mp.collect().then(all => {
  // all is an array of all the data emitted
  // encoding is supported in this case, so
  // so the result will be a collection of strings if
  // an encoding is specified, or buffers/objects if not.
  //
  // In an async function, you may do
  // const data = await stream.collect()
})

collecting into a single blob

This is a bit slower because it concatenates the data into one chunk for you, but if you're going to do it yourself anyway, it's convenient this way:

mp.concat().then(onebigchunk => {
  // onebigchunk is a string if the stream
  // had an encoding set, or a buffer otherwise.
})

iteration

You can iterate over streams synchronously or asynchronously in platforms that support it.

Synchronous iteration will end when the currently available data is consumed, even if the end event has not been reached. In string and buffer mode, the data is concatenated, so unless multiple writes are occurring in the same tick as the read(), sync iteration loops will generally only have a single iteration.

To consume chunks in this way exactly as they have been written, with no flattening, create the stream with the { objectMode: true } option.

const mp = new Minipass({ objectMode: true })
mp.write('a')
mp.write('b')
for (let letter of mp) {
  console.log(letter) // a, b
}
mp.write('c')
mp.write('d')
for (let letter of mp) {
  console.log(letter) // c, d
}
mp.write('e')
mp.end()
for (let letter of mp) {
  console.log(letter) // e
}
for (let letter of mp) {
  console.log(letter) // nothing
}

Asynchronous iteration will continue until the end event is reached, consuming all of the data.

const mp = new Minipass({ encoding: 'utf8' })

// some source of some data
let i = 5
const inter = setInterval(() => {
  if (i-- > 0)
    mp.write(Buffer.from('foo\n', 'utf8'))
  else {
    mp.end()
    clearInterval(inter)
  }
}, 100)

// consume the data with asynchronous iteration
async function consume () {
  for await (let chunk of mp) {
    console.log(chunk)
  }
  return 'ok'
}

consume().then(res => console.log(res))
// logs `foo\n` 5 times, and then `ok`

subclass that `console.log()`s everything written into it

class Logger extends Minipass {
  write (chunk, encoding, callback) {
    console.log('WRITE', chunk, encoding)
    return super.write(chunk, encoding, callback)
  }
  end (chunk, encoding, callback) {
    console.log('END', chunk, encoding)
    return super.end(chunk, encoding, callback)
  }
}

someSource.pipe(new Logger()).pipe(someDest)

same thing, but using an inline anonymous class

// js classes are fun
someSource
  .pipe(new (class extends Minipass {
    emit (ev, ...data) {
      // let's also log events, because debugging some weird thing
      console.log('EMIT', ev)
      return super.emit(ev, ...data)
    }
    write (chunk, encoding, callback) {
      console.log('WRITE', chunk, encoding)
      return super.write(chunk, encoding, callback)
    }
    end (chunk, encoding, callback) {
      console.log('END', chunk, encoding)
      return super.end(chunk, encoding, callback)
    }
  }))
  .pipe(someDest)

subclass that defers 'end' for some reason

class SlowEnd extends Minipass {
  emit (ev, ...args) {
    if (ev === 'end') {
      console.log('going to end, hold on a sec')
      setTimeout(() => {
        console.log('ok, ready to end now')
        super.emit('end', ...args)
      }, 100)
    } else {
      return super.emit(ev, ...args)
    }
  }
}

transform that creates newline-delimited JSON

class NDJSONEncode extends Minipass {
  write (obj, cb) {
    try {
      // JSON.stringify can throw, emit an error on that
      return super.write(JSON.stringify(obj) + '\n', 'utf8', cb)
    } catch (er) {
      this.emit('error', er)
    }
  }
  end (obj, cb) {
    if (typeof obj === 'function') {
      cb = obj
      obj = undefined
    }
    if (obj !== undefined) {
      this.write(obj)
    }
    return super.end(cb)
  }
}

transform that parses newline-delimited JSON

class NDJSONDecode extends Minipass {
  constructor (options) {
    // always be in object mode, as far as Minipass is concerned
    super({ objectMode: true })
    this._jsonBuffer = ''
  }
  write (chunk, encoding, cb) {
    if (typeof chunk === 'string' &&
        typeof encoding === 'string' &&
        encoding !== 'utf8') {
      chunk = Buffer.from(chunk, encoding).toString()
    } else if (Buffer.isBuffer(chunk))
      chunk = chunk.toString()
    }
    if (typeof encoding === 'function') {
      cb = encoding
    }
    const jsonData = (this._jsonBuffer + chunk).split('\n')
    this._jsonBuffer = jsonData.pop()
    for (let i = 0; i < jsonData.length; i++) {
      try {
        // JSON.parse can throw, emit an error on that
        super.write(JSON.parse(jsonData[i]))
      } catch (er) {
        this.emit('error', er)
        continue
      }
    }
    if (cb)
      cb()
  }
}

21 KiB Raw Blame History