PYTHON-973: Use global variable for libev loops so it can be subclassed

[beltran]

This PR is very similar to #941, so some of the review comments there might apply here as well @mambocab

[beltran]

Seems like the import has to happen here so the it happens after _global_loop is initialized, if it's imported at the beginning, _global_loop is None

[beltran]

Same here

[mambocab]

Code looks good! 2 things:

• I should have asked this earlier, but could you add a basic test of the subclassing behavior?
• Any idea what's up with pypy on Travis? I can look into it if you like, I just haven't seen it until now.

[beltran]

It's hard to add in a test that fails on master and passes here because of the following:

Case 1

class C1(LibevConnection):
    pass


class C2(LibevConnection):
    pass

clusterC1 = Cluster(connection_class=C1)
session1 = clusterC1.connect(wait_for_all_pools=True)
clusterC2 = Cluster(connection_class=C2)
session2 = clusterC2.connect(wait_for_all_pools=True)

# Two threads in master now, 1 in this PR

Case 2

class C1(LibevConnection):
    pass


class C2(LibevConnection):
    pass

cluster = Cluster(connection_class=LibevConnection)
session = clusterC1.connect(wait_for_all_pools=True)

clusterC1 = Cluster(connection_class=C1)
session1 = clusterC1.connect(wait_for_all_pools=True)
clusterC2 = Cluster(connection_class=C2)
session2 = clusterC2.connect(wait_for_all_pools=True)

# 1 thread in both master and this PR. Because the subclasses use the father class loop

In our tests in tests/integration/__init__.py we do cluster = Cluster(connection_class=LibevConnection)
Therefore the subclasses will use the event loop from the father.

I think we can test that _global_loop works fine but it's harder to test that it doesn't create more than one event loop.

As for the pypy issue looks like gevent is not being installed in pypy (this is happening in master as well). There was a release 4 days ago so it's probably related to this.

[mambocab]

Ohhhh, interesting -- I see what's up with the test. So does

cassandra.io.libevreactor._global_loop._cleanup()
cassandra.io.libevreactor._global_loop = None

not set things up to avoid the "subclasses use parent event loop" problem? I don't know that I follow that.

[beltran]

Yeah that's correct. That stops the libev loop and the tests verifies:

• If there's NO event loop -> Two subclasses are used in different clusters -> Only one new event loop is created

Before:

• If there's NO event loop -> Two subclasses are used in different clusters -> Two new event loop is created

But before also happened:

• If there's an event loop -> Two subclasses are used in different clusters -> Only ONE new event loop is created

[beltran]

I'll add a similar test to the asyncore one here

[mambocab]

Rebased c8728b9 onto master at 29e7f8f.

[mambocab]

Tests ran clean using libev; merging.