Down the bottom, which is where 'things we probably will never do' is when IPv6 comes in the door.
Azure (for instance) is a fully IPv6 enabled fabric. Microsoft "get" IPv6. They are all over it. They understand it, its baked into the DNA. So how come K8s people just kind of think "yea.. nah.. not right now"?
Because proxy Ipv6 at the edge is really sucky. We should be using native IPv6, preserve e2e under whatever routing model we need for reliability, and gateway the V4 through proxies in the longer term.
(serious Q btw)
The issue  has existed for over 3 years, so it's not a new suggestion.
Not to diminish the very real challenges in getting IPv6 implemented, but this is an interesting turn of phrase.. especially because rolling out IPv6 would actually solve a whole class of problems (and I'm not even a particularly big advocate of the need for IPv6, since most things should still be NATed anyway.)
(And especially considering parent's phrase "baked into the DNA" at Azure.)
S/Azure/Microsoft/ -Azure has IPv6 but I think its Immature.
https://azure.microsoft.com/en-au/updates/ipv6-for-azure-vms... Is about the underlying VM architecture, not support for Kubernetes network models.
I guess even today many people have problems getting more than a /64 in the office or home network (edit: it's supported usually with the prefix delegation option in DHCPv6 by most ISPs), so it's not frictionless in the dev environment.
Everything I've seen in their networking configuration screens and APIs appears to only allow IPv4 addresses.
Alas no. When I looked at k8s/Azure the IPv6 support was new.
My comment about IPv6 'baked into the DNA' of Microsoft is about Microsoft not Azure -A lot of the work on privacy addresses, the deployment of Teredo, adoption of ULA addresses, comes from people inside Microsoft, And they have been presenting recently at NANOG and IETF on IPv6 only deployments in the Redmond campus.
All containers can communicate with all other containers without NAT.
All nodes can communicate with all containers (and vice-versa) without NAT.
The IP that a container sees itself as is the same IP that others see it as.
When using Docker by itself, you get into all sorts of complicated situations because most running containers have an IP address that's host-specific and not routable for any other machines. This makes networking across hosts a giant pain. Kubernetes takes that away by making things behave exactly how you'd hope they'd behave. My IP as I see it is reachable by anybody in the cluster who has it (policy permitting).
The simplicity of working in this networking model means that there's a little more work for the networking infrastructure to handle, making sure that IPs are allocated without collision and that routes are known across many hosts. Several technologies exist to build these bridges, including old-school tech that has solved these exact problems for decades like BGP (see Calico/canal).
Ultimately, there's no silver bullet. I'd recommend giving the k8s networking page a read. 
The main difference is that Kunernetes assumes that all IP's are routable and Docker does not. When using bridge networking this means the admin must ensure routes are properly configured in the host for cross-host communication on Kunernetes.
Docker does not provide cross-host service discovery for bridge networking out of the box. This does not prevent admins from setting this up themselves.
For overlay networking solutions (e.g. Weave), the cross-host networking is handled for you and typically still even uses bridge networking to provide container connectivity, with service discovery also working cross-host.
ipvlan and macvlan are "underlay" solutions (i.e. attached directly to the host networking interfaces). For these it is expected that the admin has configured the networking and that containers on different hosts are routable. Service discovery should work across hosts with these solutions, but actual networking is dependent on the how the host networking is setup because the containers will be assigned IP's from the host's network and are bound to a particular host network interface.
When using ipvlan or macvlan (or overlay networking for that matter), Docker effectively makes the same assumptions as Kunernetes does for its networking.
I notice that you conveniently left out the "ingress" component. Stuff in K8s talking with other K8s stuff is easy. Getting the flows into K8s apps from outside the K8 network is amazingly clunky in its current state.
But outside the container, the strategy is still much simpler than other solutions (most of which predate Kubernetes). Kubernetes chooses to give every pod its own IP. This means choosing an internal network such as 10.x.x.x, and giving each machine a slice of it. This way, one single cluster shares the same big, flat space of IP addresses; not only do pods have the same IP inside the container, but they can talk to other pods using the other pod's IP, too.
But a key point is that Kubernetes is designed to take care of most of it. One part of it is the iptables proxy magic that it does to allow services to have dynamically assigned IPs, too, with simple load-balancing between them. The second part is the many built-in plugins for different, more complicated overlay strategies. Kubernetes' automatic configuration works out of the box on, say, AWS, without anything magical — Kubernetes natively talks to AWS to set up a routing table so that packets end up where they should. You don't need more complex overlay networking stacks such as Calico, Flannel or Weave right away.
As for ingress, it has absolutely been Kubernetes' weakest point for several years, and the Kubernetes team knows this perfectly well. That said, it's not complicated, thanks to the above. Once you have, say, Nginx listening on a port, routing traffic into the cluster is a matter of setting up a load balancer (at least on clouds like GCP, DigitalOcean and AWS), something which Kubernetes even can do automatically for you. The weak links are the ingress controller — the Nginx one is popular because it's stable and supports common features such as TLS, whereas others such as Voyager and Traefik are lagging — as well as the impedance mismatch with cloud LBs such as the Google Load Balancer.
So far, Kubernetes' ingress support has been generic: One ingress object can be used to "drive" different HTTP servers. The problem being, of course, that all HTTP implementations which have different settings (timeouts, TLS certs, CDN functionality) and concerns that the current, simple ingress format cannot support. I'm expecting this to change soon. Ingress portability really isn't an important concern, and the generic ingress format is a bottleneck for the ingress functionality to mature.
Why is having a big, flat namespace important? Routers route. Clos L3 networks are no longer a fancy thing. They're commonplace now. I don't see any advantage of having a flat network.
> One part of it is the iptables proxy magic that it does to allow services to have dynamically assigned IPs, too, with simple load-balancing between them.<
Ah yes, the iptables "magic". We call this, slowness and obfuscation. People who understand how to run networks don't like handwavy magic. We like simple, elegant concepts. Kubernetes networking is very far from simple and elegant. It's blackbox "magic".
>You don't need more complex overlay networking stacks such as Calico, Flannel or Weave right away.<
I run a native L3 network so have no need for an overlay network on top of it. That said, I'd argue that the overlay junk is probably easier for non-networking-fluent developers to setup and run compared to routing in AWS.
Kubernetes networking can be summed up thusly:
Great for developers who know nothing about networking but want to run at hyperscale. Terrible for people who actually know how to run networks properly.
Kubernetes ingress is garbage. Stop apologizing for it.
IPv6 would also get rid of 99% of these overly complicated hand-wavy solutions that Kubernetes proponents constantly tout as strong points. Give each node a /64, and you're set.
You're arguing against the value of a "big, flat namespace", yet you're also arguing for IPv6, which itself is a big, flat namespace? Do you see the contradiction, perhaps?
Dedicated CIDR for pods is important because it's simple. The symmetry is simple to explain, simple to understand; the same simplicity you'd get from IPv6.
Moreover, it's an abstraction that can be implemented however you want (custom routing on L3, SDN overlay, BGP). Not everone has a native L3 network. If you're on Google Cloud Platform, you get a virtual L3, but with other clouds, the networking is a bit more old hat. So again, simplicity and convenience. As for "overlay junk", the entirety of the Google Cloud itself is virtualized over what is probably the world's most sophisticated SDN overlay, so, well, some people's junk is other people's ragingly successful business, I suppose.
I'm not sure why you categorize the automatic iptables rules that Kubernetes set up as slow or obfuscated. It's only magical in the sense that Kubernetes automatically makes its cluster IPs load-balanced, a convenient system that you are in no way forced to use. If you have a better setup, feel free to use it instead.
We use Kubernetes ingress. It works. It could be better, but it's not "garbage". I really recommend against putting everything in such categorical terms. Everything in your comment is "junk" and "garbage", and the people who designed it (Google!) are morons who don't understand networking, somehow. That kind of arrogance on HN just makes you look foolish.
Can't each container be bound to a virtual network interface(macvlan) and use DHCP? That allows the network to configure and manage the address pool.
No fiddling with routing tables (well not for each node) and it allows peering of VPCs simply
The simplest option is to just use routing . You don't have to use an SDN. Not sure if DHCP is one of the officially supported options.
I know there are people out there who use MACvlan/IPvlan. Some people discourage these types of virtualized networks because the packet manipulation can be inefficient (unless the NIC explicitly supports it; I believe some support VXLAN?) and can hamper the kernel's scheduling.
Firstly _statically_ assigning an address range to each node is utter madness, firstly it limits the containers you can have. Secondly its terribly inflexible, its perfectly possible to have a beefy server have more than 254 containers running.
Thirdly it ties up a huge address ranges with _no_ flexibility. If you have nodes assigned to certain duties (like DB pods) then it can only realistically have a few containers. So the rest of the address range is wasted.
What is so frustrating is that all of this is automatically taken care of using DHCP and macvlan.
In the example thats linked, why isn't there a second adaptor on a different VLAN? Thats a far more simple and visible way of linking things together. I just don't see why you'd want to willingly fiddle with routing table when on a normal flat network its done for you, automatically.
This is a config value; if you want more containers per node, use a /23 or a /22 instead. It's entirely up to the operator, there's nothing magical about the default choice of /24 (except for it being easier to perform arithmetic on).
> Thirdly it ties up a huge address ranges with _no_ flexibility.
If you're using 10/8, then you have 16 bits' worth of /24 subnets, so 65k nodes by default. It's true that there are some companies in the world that have to worry about this limit, but for almost everybody I don't think this is a real problem.
Indeed, but its something extra that _you_ have think about after you've setup your VPC (if youre on AWS) not only does it mean you can deploy/configure un routable IPs by accident, its using a mechanism that _slows down_ your VPC, and adds a minefield of confinguration errors. its just madness.
It's just a LAN, why would you ever statically assign IPs? especally at scale, especially if you have a dynamic ever changing workload. Deploy a pod, two network interfaces, macvlan & AWS does the rest. Put a cloudwatch alert for DHCP exhaustion, or put a resource limit in for each AZ.
Put it this way: Why do you want to have to think about subnets _after_ you've created your VPC? (unless you've reached a limit...)
But this is the problem that most designs are trying to solve. Large L2s are notoriously fragile. 1,000 nodes, 50-100 pods/node is a lot of ARPs. And sometimes you want partitions between endpoints for security/isolation.
I agree with you about static assignment of addresses. But that's why (most) CNIs work with a controller of some kind for IPAM.
IMO, the problem complexity is hard to compress. You need to distribute/manage MAC addresses, routes, and/or state. Different designs would favor one over another.
In this case I think the traditional model works well, has excellent documentation, and scale much better than the alternatives, especailly in AWS.
Or maybe I'm wrong. :)
There is no difference between this and VM hosts.
Ingress components are maturing, and I think will seem more natural once they've matured, stabilized, and become more visible in the documentation.
At present though it's a bit of a hump... Running my own K8s cluster on-premise I suspect like a lot of that clunkiness ie because in its natural environments (Google and Googles Cloud), it has direct access to nice load balancing and routing systems/services.
This is the first of a two-part series, the first dealing with pod networking and the second with services. I plan a third on ingress after kubecon. It's a little GKE-specific in the implementation details, and the whole thing is pluggable and can be configured in different ways (as the OP shows), but I think it covers the fundamentals pretty well.
And this: https://www.youtube.com/watch?v=y2bhV81MfKQ
* Part 1: https://medium.com/@ApsOps/an-illustrated-guide-to-kubernete...
* Part 2: https://medium.com/@ApsOps/an-illustrated-guide-to-kubernete...
> This is especially problematic where the connected next-hop e.g. switch is expecting frames from a specific mac from a specific port.
e.g.: if the host is attached to a managed switch with a strict security policy, macvlan would not work.
Obviously it needs a switch at the otherside that can handle a huge and quick changing arp table. Also if you have mac address limiting typical on edge switches, its a non flyer
If you do something like advertise a /32 for each container you can very quickly fill up TCAMs on your network hardware (in particular cheap top of rack switches that are pervasive in data centers).
The entire v4 internet is something like 600k prefixes right now and the routers that can handle that many prefixes at line rate are irritatingly expensive. ToRs as of a couple of years ago when I last tested this would fall over at 1-10k prefixes.
So be careful when looking at BGP solutions because it's very easy to have a BGP topology that doesn't scale, despite it being the exchange protocol for the Internet.
Assuming everything is nice and hierarchical, you can easily aggregate an entire rack to a single prefix. Even the shitty ToR switches can usually handle a couple thousand prefixes, which should be plenty if done correctly.
Obviously you shouldn't be advertising /32s.
> The entire v4 internet is something like 600k prefixes right now ...
Just checked my edge routers and it looks like we're up to ~671k prefixes here and that number is still increasing everyday.
At least, that's what you do if you use Calico and want to be able to use hypervisor migration when using it with OpenStack.
Just make sure your TOR's can handle the amount of routes necessary, have a default route from the hypervisor to the internet, and from the TOR to the spine, and have the spine advertise a 0's route. So now the spine is the only place where you need beefier routers that can support more than the TOR's in terms of routes.
With some intelligence in the IPAM solution your host will get a /26 (or a /64) and will advertise that entire range, and only a single /32 is advertised if the VM/container moves to another hypervisor host (to support things like live-migration).
That being said, most TOR's can handle quite a large amount of routes these days. When I was at a telco we had some gear that did up to 128k routes, so splitting between IPv4/IPv6 we had 64k routes each. Which is plenty, even for larger clusters.
Yeah, that's what I meant. I thought that was clear from "aggregate an entire rack to a single prefix" but perhaps not.
When I said don't advertise /32s I meant "globally" and as a general rule (I didn't realize what I said would be taken so literally). There can be exceptions, of course, such as in the live-migration case you mentioned.
Easier said than done. Most datacenters are bit more deliberate about allocating addresses and hand them out in non-contiguous CIDRs. The VLAN mentality is still very prevalent. Getting a /20 at a time is pretty common.
Using overlapping IPs puts you right back into the overlay model.
>Assuming everything is nice and hierarchical, you can easily aggregate an entire rack to a single prefix.
Yes, exactly. The trick then becomes how to you ensure that endpoints that get created within the rack get an IP from the prefix? Romana (the project I work on) does this. It lets you capture your network topology for exactly this reason. This is especially important if/when you must filter routes at ToR.
I read your comment as, "Don't use technology that you can misconfigure, because you can misconfigure it!". Well yeah, the same can be said with anything networking related.
Rolls right off the tongue, doesn't it?
At least these terminate when resolved.
To the extent your requirement match theirs, this could be a good alternative. The most significant in my mind is that it's meant to be used in conjunction with Envoy. Envoy itself has its own set of design tradeoffs as well.
For example, Lyft currently uses 'service-assigned EC2 instances'. Not hard to see how this starting point would influence the design. The Envoy/Istio model of proxy per pod also reflects this kind of workload partitioning. Obviously, a design for a small number of pods (each with their own proxy) per instance is going to be very different from one that needs to handle 100 pods (and their IPs), or more, per instance.
Another is that k8s network policy can't be applied since the 'Kubernetes Services see connections from a node’s source IP instead of the Pod’s source IP'. But I don't think this CNI is intended to work with any other network policy API enforcement mechanism. Romana (the project I work on) and the other CNI providers that use iptables to enforce network policy rely on seeing the pod's source IP.
Again, this might be fine if you're running Envoy. On the other hand, L3 filtering on the host might be important.
Also, this design requires that 'CNI plugins communicate with AWS networking APIs to provision network resources for Pods'. This may or may not be something you want your instances to do.
FWIW, Romana lets you build clusters larger than 50 nodes without an overlay or more 'exotic networking techniques' or 'massive' complexity. It does it via simple route aggregation, completely standard networking.
>"Unfortunately, AWS’s VPC product has a default maximum of 50 non-propagated routes per route table, which can be increased up to a hard limit of 100 routes at the cost of potentially reducing network performance."
Could someone explain why increasing from 50 to 100 non-propagated routes in a VPC results in network performance degradation?
> Lincoln Stoll’s k8s-vpcnet, and more recently, Amazon’s amazon-vpc-cni-k8s CNI stacks use Elastic Network Interfaces (ENIs) and secondary private IPs to achieve an overlay-free AWS VPC-native solutions for Kubernetes networking. While both of these solutions achieve the same base goal of drastically simplifying the network complexity of deploying Kubernetes at scale on AWS, they do not focus on minimizing network latency and kernel overhead as part of implementing a compliant networking stack.
There they clearly state:
>"To run Kubernetes over AWS VPC, we would like to reach following additional goals:
Networking for Pods must support high throughput and availability, low latency and minimal jitter comparable to the characteristics a user would get from EC2 networking"
The net effect is that you can build large clusters without running out of VPC routes and no overlay is needed when traffic crosses AZs.
When a route is used to forward traffic for multiple instances, the target instance acts as router and forwards traffic to the final destination instance. This works because instances within an AZ have routes installed on them to the pod CIDRs on the other instances in the zone, so any one of them can perform this forwarding function.
Romana only piggybacks routes when there are no more VPC routes available, so for small cluster it's just like kubenet. For large clusters routes it uses all the instances to forward traffic so that none of them become a bottleneck.