Re: Numen with reference to vmap

[Phil Sutter <phil@nwl.cc>]

Hi,

On Wed, Dec 04, 2019 at 01:47:47PM +0000, Serguei Bezverkhi (sbezverk) wrote:
> Thank you for your reply. It is very unfortunate indeed. Here is the scenario where I thought to use a non-anonymous vmap.
> 
> Each k8s service can have 0, 1 or more associated endpoints, backends (pods providing this service). 0 endpoint already taken care of in filter prerouting hook.  When there are 1 or more, proxy needs to load balance incoming connections between endpoints.I thought to create vmap per service with 1 rule per service . When an endpoint gets updated (add/deleted) which could happen anytime then the only vmap get corresponding update and my hope was that automagically load balancing will be adjusted to use updated endpoints list.
> 
> With what you explained, I am not sure if dynamic load balancing is possible at all. If numgen work only with static anonymous vmap and fixed modulus , the only way to address this dynamic nature of endpoints is to recreate service rule everytime when number of endpoints changes (recalculate modulus and entries in vmap). I suspect it is way less efficient.

Well, if you have a modulus of, say, 5 and your vmap contains only
entries 0 to 3 your setup is broken anyway. So I guess you will need to
adjust modulus along with entries in vmap at all times.

What is the iptables-equivalent you want to replace? Maybe that serves
as inspiration for how to solve it in nftables.

> What will happen to dataplane and packets in transit when the rule will be deleted and then recreated? I suspect it might result in dropped packets, could you please comment on the possible impact?

Well, you could replace the rule in a single transaction, that would
eliminate the timespan the rule doesn't exist. AFAICT, this is RCU-based
so packets will either hit the old or the new rule then.

Cheers, Phil