This test seems destined to be flaky in TSAN environments.
The initial sleep is there so that the stop signal to EchoServer is sent while it's blocked inside the echo loop. That appears to be how we can safely assert that one write and one recv both succeed, while the second recv fails.
However, it's possible for EchoServer to be so slow to start that 100 ms isn't enough, and the stop signal reaches it before it enters the loop. Then the first write will fail like this:
Alexey said he'd take a look at this.