Strictly on the network side, there is no blocker one way or the other.

I think I miss some context, what's the current low-traffic setup ? What's the downside of only using the source IP in the hashing? Not enough clients for proper balancing ?

In T368545#9929335, @ayounsi wrote:

I think I miss some context, what's the current low-traffic setup ?

Usually services use wrr (weighted round robin) to balance traffic across nodes

In T368545#9929335, @ayounsi wrote:

What's the downside of only using the source IP in the hashing? Not enough clients for proper balancing ?

Yep, that's what I suspect, that using source IP only would lead to some realservers getting significantly more traffic than others given the relatively small pool of source IPs. And it looks like that behavior from maglev is already being (ab)used by thanos-web low-traffic service:

# Needed for SSO sessions to stick. As of Dec 2022 backend
# selection from varnish -> ATS is always random for pass traffic
# therefore even with this we need at most one thanos-fe host
# pooled at a time (per site).
scheduler: mh

It is pretty clear to me that the only way to have fair load balancing with maglev is if we do the consistent hashing using the remote port as well.

We're in the situation where we have less client IPs than backends for virutally every service, so source IP hashing would never be enough to get a fair distribution.