Ciao Marco,
I assume the old process did not get the signal and continues to serve
requests with the old configuration.
Can you confirm the number of haproxy processes running is more than you
expect? Are you using nbproc or single process mode (the latter is the
default)?
Does the PID file contain the correct PID?
If the old haproxy instance is not getting the signal it will continue
to serve requests and the kernel will load-balance between the old and
the new instance, leading to the behavior you are describing.
We really need an no-reuseport knob to confirm those kinds of issues...
Regarding the actual problem, I would suggest to upgrade to latest
stable release 1.6.9 first of all. Then we can actually start
troubleshooting, but there are important bugfixes in those 4 releases.
cheers,
lukas