avoid loop without yield when count of sockets is too high

[gawen]

For Python 2.x, sleep_interval is 0 if len(self.sockets) is above
self.ping_timeout (default: 60). With eventlet, self.sleep(0) looks it does not
yield to the event loop, queueing work which is not processed. In this
condition, the symptom is the server using 100% of CPU after some time, even
when count of sockets drop below len(self.sockets).

This fix forces sleep_interval to be a float, not equal to 0, forcing eventlet
to yield, processing the queued work and avoiding CPU to be at 100%.

This fix does nothing for Python 3.x.

[miguelgrinberg]

I'll test this. I find it odd that eventlet does not release the CPU with sleep(0), even if it is only Python 2.

[maximebf]

I can confirm that this fixes the issue we were having with 100% CPU usage.

Here is our load graph before and after applying the patch:

(The load was 100% confirmed to be from the python socketio process)

[miguelgrinberg]

So how is this not a bug with eventlet then?

[gawen]

I find it odd that eventlet does not release the CPU with sleep(0), even if it is only Python 2.

To be fair, this is my interpretation, and this might be incorrect.

So how is this not a bug with eventlet then?

It is my understanding that sleep_interval should be a float. Making so for Python 2 fixes the 100% CPU loop which never ceases when len(self.sockets) is too high.

[miguelgrinberg]

I thought about this and even though I think there is something fishy going on with eventlet, this is a good change, so I'm merging.