parallel and corrupted data?

[ronag]

I've noticed a strange occurrence. I'm reading raw compressed data from multiple files and then concatenating them:

   .map(url => _(request.get(url))
   .parallel(8)
   .through(decoder())

However, this results sometimes in corrupted data.

While the following works just fine:

   .map(url => _(request.get(url))
   .parallel(1)
   .through(decoder())

I don't have any reproducible example yet. The problem is somewhat intermittent.

[vqvu]

Can you try testing against the latest master? parallel had a buffering bug where it would drop data in certain circumstances (see #234). This was fixed in #302 but has not yet been released.

[ronag]

Will need to do some more testing but so far that seems to have solved it.

Looking forward to the next release :).

[ronag]

Nope. Still a problem.

[vqvu]

Do you know why the data is corrupted? Are the Buffers being dropped or reordered?

[ronag]

It's very difficult to tell. It's H264 video buffers that are sent to ffplay. Analysing a H264 bitstream is a bit complicated.

[vqvu]

Hmm, I tried reading a bunch of local files via parallel and couldn't find any issue.

I've pushed a debug branch here: https://github.com/vqvu/highland/tree/parallel-debug (vqvu/highland#parallel-debug).

Can you run this code against that branch? It should track the order of buffers that request emits and tell you if anything is wrong.

Don't forget to npm update --dev first!

var counter = [];
var i = 0;
var checkState = {
    urlIndex: 0,
    bufIndex: 0
};

// Check correct version
_.whatAmI();
// => Highland 2.x. debug-parallel edition

stream.map(url => _(request.get(url)))
    .map(function (s) {
        var urlIndex = i++;
        counter[urlIndex] = 0;
        return s.map(function (buf) {
            return {
                index: [urlIndex, counter[urlIndex]++],
                buf: buf,
            };
        });
    })
    .parallel(8)
    .map(function (obj) {
        if (obj.index[0] === checkState.urlIndex + 1) {
            checkState.urlIndex++;
            checkState.bufIndex = 0;
        }

        if (obj.index[0] === checkState.urlIndex) {
            if (obj.index[1] !== checkState.bufIndex) {
                console.log('Corrupt data.', checkState, obj.index);
            }
            checkState.bufIndex++;
        }
        else {
            console.log('Corrupt data.', checkState, obj.index);
        }

        return obj.buf;
    })
   .through(decoder())

[ronag]

How do I checkout a branch with npm?

http+git@github.com:vqvu/highland.git...something

[ronag]

See also, #330, with which I have no issues, so far.

[apaleslimghost]

@ronag you can run npm install vqvu/highland#parallel-debug

[apaleslimghost]

(depending on your shell, you might need to escape the #)

[vqvu]

Ok, I think I have a broken test case. Will type up the fix soon-ish. Though I'm not sure if this is the problem you're encountering or not.

var s1 = _([1, 2, 3]);
var s2 = _([11, 12, 13]);
_([s1.fork(), s2, s1.fork()])
    .parallel(3)
    .consume(function (err, x, push, next) {
        push(err, x);
        if (x.buf === 21) {
            // Pause for a while.
            setTimeout(next, 1000);
        }
        else if (x !== _.nil) {
            next();
        }
    })
    .toArray(_.log);
// => [ 1, 2, 3, 11, 12, 13 ]
// should be
// => [ 1, 2, 3, 11, 12, 13, 1, 2, 3 ]

Basically, when s1 ends, if

• s2 has already ended, and
• the parallel stream is paused (caused by the decoder)

then anything that s3 has already emitted gets dropped on the floor.

[vqvu]

Can you try testing against the code in #331?

vqvu/highland#parallel-issue-328