2 ms·
> K8s developers are most definitely not primarily sponsored by Google - more than 60% of K8s devs are NOT Googlers. The key metric, in my view, is how many of
by otterley 10y ago
> K8s developers are most definitely not primarily sponsored by Google - more than 60% of K8s devs are NOT Googlers.
The key metric, in my view, is how many of them can approve PRs into master. How many of them are not Googlers?
Moreover, what would happen if a PR arrived that might make K8S incompatible with GCE, but be otherwise better for everyone else? I am certain it would be summarily rejected by Google.
> Overlay networks are definitely not a hack - many partners set them up with great benefit. The fact is networking is hard (tm), and unless you're just looking for a flat network, then you're going to have to use SOMETHING.
I respectfully disagree. The use of multiple IP addresses per node is a hard requirement of K8S, even though it is arguably unnecessary in environments whose servers can be bound to arbitrary ports.
By doing so, K8S made an explicit tradeoff that was better suited to Google's cloud offering than others'. I'm not saying it was the wrong decision; I'm simply saying that it forces complexity on those who operate in single-IP-per-host environments. (In case you think I'm unfairly laying blame, I point the finger equally at AWS and other cloud providers who refuse to make it simple to allocate a useful number of IP addresses per instance.)
I maintain my characterization that overlay networks are a hack: they make tracing more complex (tcpdump doesn't natively understand VXLAN or decapsulate their frames); they are compatible with few (if any) NetFlow analyzers (which are often used by orgs who use them for IDS and other purposes); and they add overhead to packet processing, particularly in virtualized environments that don't support VXLAN hardware offloading.
- smarterclayton 10y agoI originally objected to IP per pod given that it was so early in wide scale deployment of SDN on metal (at Red Hat, a year before Kube 1.0, we were already worried about how in the world we could support IP per pod as well as on gce). In light of everything, I think it was the right decision. While VXLAN is cumbersome, IP per entity allows very powerful integration into existing networks, fits well with almost all tools in the space, and makes apps much easier to run. I do still want to see more solid work done for programmable BGP (calico) and programmable iptables (Contiv) and even programmable routers (it's not that hard to program routers on the fly today, just incredibly specific to each technology). I also look forward to being able to exploit tools like ECMP more effectively within the cluster to do L3 load balancing and DSR - much of that would be a lot harder without being able to rely on endpoints everywhere.