intermittent allmydata.test.test_introducer.NonV1Server.test_failure failure

[exarkun]

Traceback (most recent call last):
Failure: twisted.trial.util.DirtyReactorAggregateError: Reactor was unclean.
DelayedCalls: (set twisted.internet.base.DelayedCall.debug = True to debug)
<DelayedCall 0x10a5a0bd8 [240.085453033s] called=0 cancelled=0 Broker.keepaliveTimerFired()>
<DelayedCall 0x10a5a04d0 [240.08481288s] called=0 cancelled=0 Broker.keepaliveTimerFired()>

and

Traceback (most recent call last):
Failure: twisted.trial.util.DirtyReactorAggregateError: Reactor was unclean.
Selectables:
<Negotiation #0 on 55074>
<<class 'twisted.internet.tcp.Client'> to ('127.0.0.1', 55074) at 10a3a4b50>

ie, a leaked TCP client and a couple delayed calls.

[exarkun]

This was observed on the OS X travis-ci job. It may be that this test can only realistically fail on OS X due to distinctive ordering of TCP events on that platform.

[exarkun]

It looks like Tub.stopService doesn't actually wait for its (Foolscap) Broker connections to close. Instead, it closes them and waits about a reactor tick ("eventually"). Evidently this is good enough most of the time but it's not a guarantee the connections will have been closed in time.

Fortunately, Broker already has an API for being notified when its connection is actually lost. Unfortunately, for some reason, Tub.stopService explicitly disables that functionality when disconnecting its Brokers! I don't know why it does this.

[exarkun]

I think this will be fixed by a Foolscap change, ​https://github.com/warner/foolscap/pull/38

[warner]

Foolscap-0.12.7 has been released, with that change.