ping @jonathanong @fengmk2

LGTM

@juliangruber all fixed

hmmm, this seems more like an edge case to me. are there other cases that would benefit from this change?

The most serious problem now is http keepalive sockets will be destroyed, and we must fix this issue. Not sure any other side effect if we destroy the streams.

are you sure destroying a http stream will affect other requests on the same socket? sounds more like a node bug to me

If you destroy an IncommingMessage, it will destroy the socket behind: https://github.com/nodejs/node/blob/master/lib/_http_incoming.js#L92

That will affect other IncommingMessage's that reusing the same socket and will cause socket hang up error. This is reported in ali-sdk/ali-oss#40 (sorry this issue is in chinese).

any suggestions? I'll merge this PR if no response in the next day. :)

Hey man, no need to put a deadline on it. If you need it urgently use a fork or find a userland solution.

I think this is just a bug fix. so if anyone doubt this patch, I'll leave it here and fix in our own framework for now. :)

this patch fails if the body stream is endless, like for example:

this.body = fs.createReadStream('/dev/urandom');

or more real-worldy

this.body = serverSentEvents(createUpdateStream());

in the cases above it's required to destroy the stream.

and trying to read it all out will never finish

although i don't think pipe an endless stream to response meaningful. but maybe there is a solution has less impact. detect the stream in destroy and don't destroy IncommingMessage(a breaking change in destroy).

@juliangruber how about this?

too edge-casy. koa should not need to worry about this.

Just to be sure, this would work for you, right?

httpStream.destroy = blackHole.bind(null, httpStream);
this.body = httpStream;

it is a common usage that use koa as a http proxy, and anybody who use this.body = yield request('http://foo.bar/'); //keep-alive will hit this "edge case", and it is really hard to figure out why. so I think we should fix it in framework level, because it is a bug in koa, and we can't let everyone to use httpStream.destroy = blackHole.bind(null, httpStream); hack.

maybe a better solution is to just wrap the http stream in another stream so you never set the http stream as the actual .body=.

const PassThrough = require('stream').PassThrough

app.use(function * (next) {
  this.body = httpStream.on('error', this.onerror).pipe(PassThrough())
})

would need docs on this though... super edge case

would need docs on this though... super edge case

what about documenting that a body stream can be .destroyed and showing how to guard against that

what about documenting that a body stream can be .destroyed and showing how to guard against that
Show all checks

+1 @juliangruber

anyways prefer going the documentation route. will close this when i or someone else adds documentation

It's ok to add some warning in documentation.

Didn't understand a thing from the documentation on what's happening here.
It's just like "warning, never set a stream as ctx.body".

@catamphetamine, they were warning not to set a http-request-alike-stream which has keep-alive enabled. Because the framework will close the stream by default, which will break keep-alive feature.
Keep-alive means to reuse tcp socket connection for two or more http connections.

Generally, it's OK with ctx.body = fs-read-stream.
Just be careful with ctx.body = request('host:port/path').

@dead-horse
cnpmjs.org中目前用的koa的版本是"koa": "^1.2.0", 如果使用oss作为nfs， 在从oss上get的时候会有一堆的sockets hang up。

有什么好的办法能避免这类错误吗？

cnpm/cnpmjs.org#1401

谢谢

@dead-horse 马哥好，昨天夜里见到这个 issue 忘了跟你打招呼了，白天来踩一下。

----- 骄傲的用中文交流