The abns_socket in seamless-reload regtest regularly fails in Travis-CI
on smaller machines only (typically the ppc64le and sometimes s390x).
The error always reports an incomplete HTTP header as seen from the
client. And this can occasionally be reproduced on the minicloud ppc64le
image when setting a huge file descriptors limit (1 million).
What happens in fact is the following: depending on the binding order,
some connections from the client might reach the TCP listener on the
old instance and be forwarded to the ABNS listener of the second
instance just being prepared to start up. But due to the huge number
of FDs, setting them up takes slightly more time and the 20ms server
timeout may expire before the new instance finishes its startup. This
can result in an occasional 504, except that since the client timeout
is the same as the server timeout, both sides are closed at the same
time and the client doesn't receive the 504.
In addition a second problem plugs onto this: by default http-reuse is
enabled. Some requests being forwarded to the older instance will be
sent over an already established connection. But the CPU used by the
starting process using many FDs will be taken away from the older
process, whose abns listener will not see a request for more than 20ms,
and will decide to kill the idle client connection. At the same moment
the TCP proxy forwards a request over this closing connection, it
detects the close and silently closes the other side to let the
client retry, which is detected by the vtest client as another case
of empty header. This is easier to reproduce in VMs with few CPUs
(2 or less) and some noisy neighbors such as a few spinning loops in
background.
Let's just increase this tests' timeout to avoid this. While a few
ms are close to the scheduler's granularity, this test is never
supposed to trigger the timeouts so it's safe to go higher without
impacts on the test execution time. At one second the problem seems
impossible to reproduce on the minicloud VMs.
ce6fc25b17