Rolling Upgrade and HA Proxy Draining - not working?

#1

Can anybody confirm, what is their working configuration to achieve successful draining.

I am on Rancher 1.6.25 and HaProxy with 0.9.6, and the recommended configuration in place from Rancher documentation.

My initial observation is: not working:
Test procedure:

  1. jmeter makes small requests 5/s to a rancher loadbalancer with service that has 2 containers
  2. Manually upgrade the service (StartBeforeStop, Drain 25 secs)

By the time I upgrade the service i have at least 2 broken connections, sometimes more:

On loadbalancer:

Anybody can share their experience?