Traffic load-balancing issue on a target removal about meridio HOT 2 CLOSED

I made a test in ovl/forwarder-test that does;

1. Start cluster and a trench with 4 targets
2. Scale replicas 4 -> 8
3. Check communication.
4. Disconnect 3 targets (but the PODs lives on)
5. Check communication
6. Reconnect the 3 targets
7. Check communication
8. Repeat from 4 n times
9. Scale replicas 8 -> 4
10. Check communication

Check communication

1. Explicitly check number of targets in nfqlb (usually succesful fast)
2. Make 120 connects with mconnect (usually disturbances for ~1min)
3. Repeat 2 for ipv6

Runned 25 iterations (n=25) without any lasting problem (took ~40min). But it takes time and re-tries after a disconnect/reconnect before all 120 connects are succesful, usually ~1min.

from meridio.

The re-try problem solved. The large number (relatively) of fast connects saturated the conntrack table. A conntrack -F before the mconnects made the re-tries disappear. Ran 30 iterations without any problems at all. Closing the issue

from meridio.