This reason for the flaky is is as follows:

• Travis has limited resource (2 cores)
• When the Queue actor initialization took too long, all the enqueue tasks will be blocked by the initialization task. Workers running ray.get on enqueue tasks are blocked and raylet will mark the resource as available, therefore starting more workers (up to 100).
• When many python process starts, the operating system will start multiplexing the processes. The actor initialization might never able to went.
• Eventually, we run out of memory by starting many python processes.

Fix:

• Make sure the queue actor is initialized before sending enqueue task.

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/16911/
Test PASSed.

Reran 3 times on Travis, test_fork all passed.