Connection Problems with Keep-Alive

[xsc]

Hey there,

it seems to me that the HTTP keep-alive feature in aleph servers might be broken. I created a minimal project to illustrate the problem:

https://github.com/xsc/aleph-stream-testcase

I tested against 0.4.1 and 0.4.2-alpha4, by setting up a server that returns either an InputStream or a byte array as :body, then used Apache Bench with the -k option to reuse connections during the test run, e.g.:

$ ab -n 500 -c 4 -k http://localhost:9877/bytes
This is ApacheBench, Version 2.3 <$Revision: 655654 $>
Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
Licensed to The Apache Software Foundation, http://www.apache.org/

Benchmarking localhost (be patient)

Test aborted after 10 failures

Total of 12 requests completed

Note that a run without -k works perfectly fine:

$ ab -v1 -n 500 -c 4 http://localhost:9877/bytes
This is ApacheBench, Version 2.3 <$Revision: 655654 $>
Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
Licensed to The Apache Software Foundation, http://www.apache.org/

Benchmarking localhost (be patient)
Completed 100 requests
Completed 200 requests
Completed 300 requests
Completed 400 requests
Completed 500 requests
Finished 500 requests

Server Software:        Aleph/0.4.1
Server Hostname:        localhost
Server Port:            9877
...

For the InputStream version there are some Exceptions (e.g. error in manifold.utils/future-with java.io.IOException: Stream closed and error in message propagation java.io.IOException: Broken pipe). I can provide complete stacktraces but they should be produced by the test project I provided.

If you need any more information, let me know! Yannick

---

$ lein version
Leiningen 2.6.1 on Java 1.8.0_45 Java HotSpot(TM) 64-Bit Server VM

[ztellman]

I'm able to reproduce your failure, but this, for instance, works fine:

 httperf --num-conns=16 --rate=16 --num-calls=10000 --port=8=9877 --uri=/stream

Also, Aleph has successfully run similar tests on the TechEmpower benchmarks without issue. I noticed that the throughput of the tests improved significantly when I added this to the project.clj:

:jvm-opts ["-server" "-Xmx1g"]

I also noticed that when I returned a shorter byte array via /bytes, ApacheBench works fine. I'm tempted at this point to blame ApacheBench for handling large responses via keep-alive improperly. Thoughts?

[xsc]

We experienced problems (i.e. connections being closed before the data was fully transmitted) with aleph behind proxies like CloudFront or HAProxy which prompted me to investigate the matter, especially the keep-alive behaviour between the two.

But, seeing as ab is seemingly drunk, those might just have been instances of #191, so I'll close the issue for now and see what I can gather over the course of next week.

Thanks for the quick response!