Storage: repeatedly streaming a readable from storage "freezes" after a a few times

[cristiano-belloni]

Hi all,
I'm trying to stream a file from my server to the client. Waiting from #318 to be closed, I'm using @stephenplusplus 's https://github.com/stephenplusplus/range-stream. By the way, this doesn't change the issue I'm seeing, which seems related to createReadStream().

First of all, what I do on the server on a ranged request is extracting the range from the request in the start and end variables, then get a file:

var file = resourceBucket.file(fileName);

create a readable:

var rStream = file.createReadStream()

then write a 206 on res and pipe the readable into a range request:

res.writeHead(206, header);
rStream.pipe(rangeStream(start, end)).pipe(res);

rangeStream is @stephenplusplus module, here modified with a few console.log statements to see what's happening:

var through = require("through2");

module.exports = function (start, end) {
    end += 1;
    var bytesReceived = 0;

    console.log ('---------------------------------------------------');
    console.log ("Sending a range from: " + start + " to: " + end);

    return through(function (chunk, enc, next) {
        console.log ('start, end: ' + start, end);
        console.log ('chunk.length: ' + chunk.length);
        bytesReceived += chunk.length;
        console.log ('bytesReceived: ' + bytesReceived);

        if (bytesReceived >= start) {

            if (start - (bytesReceived - chunk.length) > 0) {
                console.log ('slicing chunk');
                chunk = chunk.slice(start - (bytesReceived - chunk.length));
            }

            if (end <= bytesReceived) {
                console.log ('calling end');
                console.log ('---------------------------------------------------');
                this.push(chunk.slice(0, chunk.length - (bytesReceived - end)));
                this.end()
            } else {
                console.log ('pushing chunk');
                this.push(chunk)
            }
        }
        console.log ('calling next');
        next();
    });
};

Initially everything works fine, the readable streams, this.end gets called and the file streams OK.
But after a few times the client does ranged requests (specifically via an <audio> HTML element), rStream seems to freeze and never send data to rangeStream again. Here is the log:

Sending a range from: 44 to: 6545454
start, end: 44 6545454
chunk.length: 10386
bytesReceived: 10386
slicing chunk
pushing chunk
calling next
start, end: 44 6545454
chunk.length: 2710
bytesReceived: 13096
pushing chunk
calling next
[... lots of chunks coming through ...]
start, end: 44 6545454
chunk.length: 16384
bytesReceived: 629918
pushing chunk
calling next 
[ hanging forever ]

As you can see, the code calls next(), but no data comes into it at all. This seems an issue with rStream, and specifically with the readable. I also have an error handler on rStream, which never gets called:

rStream.on('error', function (err) {
                        log.error({error: err}, 'downloadLayer - ERROR ON rstream');
                        rStream.end();
                    });

Oddly enough, this seems to happen only when the range is 44 - end. But 44 being the size of the wav header, it's a typical range value, and that could be a red herring.

[stephenplusplus]

Are you creating a new instance of a Readable Stream every time, or reusing one (eg do you call file.createReadStream once or for each request)? I think things get messy if you reusing streams that already consumed data.

[cristiano-belloni]

I call it for each request. My exact code is:

reqHandler = function (req, res) {
   // [...] hit the db, construct pathName
   gcloud.readFile(encodeURIComponent(pathName), function (err, file) {
       // [...] get the range from req
       res.writeHead(206, header);
       rStream = file.readable;
       rStream.pipe(rangeStream(start, end)).pipe(res);
   })

And my small gcloud module has:

exports.readFile = function (fileName, cb) {
    var file = resourceBucket.file(fileName);
    file.getMetadata (function(err, metadata) {
        if (err) {
            return cb (err, null);
        }
        cb (null, {meta: metadata, readable: file.createReadStream()});
    });
   };
}

So a new file is created for every request, if I'm not missing something obvious.

[stephenplusplus]

No, you're right in that you're creating a new file instance each request.

Does every range request fail?

When you see that it stops sending data, are there any similarities between other failed instances? -- as an example, is the byte count received always the same?

[cristiano-belloni]

No similarities, unluckily. It took me 5-6 minutes to get a freezing now (I'm deactivating caching in the browser to make it happen quicker), and the final log is completely different, in terms of values:

calling next
start, end: 919424 6545454
chunk.length: 32768
bytesReceived: 4734141
pushing chunk
calling next

Worth mentioning that when it happens, it completely freezes my server (it doesn't answer any request at all, even unrelated ones). ps aux doesn't show high cpu use on the node process, btw.
I'm using node v0.10.29 (installed from Wheezy apt-get).

Let me know if some more tests would be useful, and I'll try them.

[cristiano-belloni]

And, to answer your first question: it doesn't happen on every ranged request. The first n ranged requests work perfectly fine, it freezes at the n+1, where n seems random.

[stephenplusplus]

So far, I can't think of a reason the stream would fail. If anything, it sounds like it could be a caching issue somewhere upstream: request or the API itself. All that's returned from 'createReadStream' is a request instance.

Is there any way you can produce a headless test case to recreate the behavior of the ranged requests and get it to reproduce the bug?

[cristiano-belloni]

Yep, here's to you. This kills (freezes) my server after a couple of minutes: https://gist.github.com/janesconference/22afe04aba3dde02814d

[cristiano-belloni]

Updated it with a deterministic version, still freezes it.

[cristiano-belloni]

Update: the deterministic gist freezes it after exactly 5 requests.

[stephenplusplus]

I started off by plugging in "yahoo.com" as an endpoint, and everything ran smoothly. I then started testing directly with file.createReadStream() and everything performed as expected. After that, I integrated range-stream, and immediately saw it stop logging "done" after 5 requests. However, I continued running the tests and noticed very slow responses from the server, sometimes making it seem like I would never get a done, but if I waited long enough, it did come. I assume that was the same thing that happened in the first case where I thought I had reproduced the behavior. Also worth noting, my test file is quite small (~3mb).

Still no answers, unfortunately, but I will try to play with it more tomorrow.

[cristiano-belloni]

You reproduced it.
My file is ~6 megabytes, so not so big. I get the same result: it stops logging after 5 requests. The server side freezes on the last call to next. It's possible that it doesn't completely freeze, though, but just becomes extremely unresponsive and slow - I kill it after a minute or so, will try to wait as you did.

[cristiano-belloni]

(Update: Tried to wait for a long time, but the server does not respond anymore after 5 iterations)

[stephenplusplus]

Can you paste the exact code you're using (fill in the TODO part of your gist). I'm not able to reproduce.

[cristiano-belloni]

@stephenplusplus, do you mean I should send you the API endpoint I'm using on my server or send over the server code? I can probably semplify the code, in the second case, because it's a complex server with integrations with Mongodb and JWT.

Btw, didn't you say yesterday that by plugging in range-stream you could observe your server stopping after 5 requests?