I hit a strange behavior in which the clone connection error from one test is raised as the error for another test. The tests are not written using trial, but I have thousands of other tests and I have never seen this behaviour. I tried to put a self-contained example at https://gist.github.com/adiroiban/edc0776e3337d0bd3f093aa0f2819deb#file-test... you will need to run it in a virtulenv as it will pull quite a few of dependencies. I have traced the error to this code https://github.com/twisted/twisted/blob/03dcdfb5933c4f83ce6aad3f4bdf080cda65... if err.check(error.ConnectionDone): err.value = smtp.SMTPConnectError( -1, "Unable to connect to server.") self.result.errback(err.value) It looks to me like the initial failure (err) is hijacked here and its errors will never be cleared. If I am doing something like if err.check(error.ConnectionDone): value = smtp.SMTPConnectError( -1, "Unable to connect to server.") else: value = err.value err.cleanFailure() self.result.errback(value) then the test will pass. But I still don't understand how the error from one client request is passed to another client request when they are using different servers. Do you see anything suspicious with this code? Thanks! -- Adi Roiban