Self congestion, fq_codel, bufferbloat, network namespaces and containers

While in barcelona, I went to the enormous containercon conference, 5000+ attendance, tons and tons of people using docker, and I learned for the first time the enormous size of the kubernetes deployment. The website claims several billion containers are run every week, which I do not doubt.

So I sat down and said… hmmm… what if I had a netperf/irtt container I could toss around for testing bufferbloat issues? It’s not clear to me, either, how often network namespaces are used in the field(?) out of that billions deployment, but on this bug report, at least, I showed the issues network namespaces had (at that point in time) with self-congesting on overload:

So my questions are:

How often are network namespaces used?
What’s the usual underlying qdisc?
Is there a way to monitor local packet loss or marking?

  1. Network namespaces are almost always used on every pod.

2,3 really depend on your networking overlay, I guess. If you are using no network overlay (like using kubenet), I don’t remember the name by hard but it was something like noqdisc if the bridge interface IIRC. And, in that case, an htb to limit pods egress too.