Kubernetes and AWS VPC Peering

Image courtesy Flickr user Alex Gretch

We’re making a bet with our next generation of services at Gridium. Our bet is that microservices will serve us better than monolithic services, and we’re in the process of moving our existing services to this kind of architecture too. Our chosen backing technologies are Docker, hosted in AWS, using Kubernetes.

These technologies are still pretty young, and there are minefields awaiting those who try anything but the happiest of paths. Here I’ll document a huge problem we ran into, in hopes that I can help someone else that runs into it.

Rainbows and Unicorns

Let me start by saying that Kubernetes is pretty cool. There are a lot of reasons why we chose it for hosting our Docker services. One of the best parts for us was the rolling update functionality that comes in the box, enabled by how services and replication controllers use the metadata system. Basically, you get something kind of like blue-green deployments for free, and none of your services ever have to have any downtime because you deployed a change.

Another nice part is the kube-up script, which with just a few small configuration knobs, can spin up a Kubernetes cluster in a variety of environments in about 10 minutes. Since the rest of our infrastructure is already in Amazon AWS, it seems natural to spin up our new cluster there.

So I did, and at first everything was rosy. The first little surprise: the Kubernetes script created a new VPC for the cluster to run in. A VPC is essentially an entirely separate, self-contained network, where all the machines inside can route to each other, and to the open Internet, but not to any other VPC. That sounds like a good thing, and it generally is – Kubernetes and Docker can do whatever crazy networking voodoo they want to, and it won’t mess with our already-running-and-with-customers services (which are in a pre-existing VPC).

But there’s a problem. We’re using RDS to host our data, because who runs their own database server in 2015? The problem is that our RDS instance belongs to our main VPC, so it won’t be visible from the new Kubernetes VPC.

Time for some peering. Peering lets you set up routing so that a certain range of IP addresses will route into another VPC. This sounds like exactly what we want. Our Kubernetes VPC’s IP address range is 172.20.*.*, and the original VPC’s address range is 10.0.*.*. So I set up the peering connection, and altered the route tables on either side, so that if a machine in VPC1 is looking for a machine in VPC2, it will route through the peering connection. I also made sure to change the RDS instance’s security group to allow connections from 172.20.*.*.

Now I can SSH into one of my Kubernetes minions, and psql -h <RDS host>asks me for a password. Perfect, so our network configuration is done, and we can start shipping containerized applications, right?

Not so fast. If I try that same command from inside a container, I get a timeout. Wait, I can connect to machine from a host, but try it from within a Docker container on that same host, it fails?


Here’s where I got roadblocked for two weeks.

I knew that Kubernetes gives each running container its own IP address, so I added those ranges to the security group, and even configured the security group to allow connections from any IP address. I checked, double-checked, and triple-checked all the routing tables. I tried everything I could think of, which admittedly isn’t much – I’m pretty green when it comes to networking. No dice. Nothing worked.

Then I started asking friends and family, and people I didn’t even know who had expertise in Docker and Kubernetes. Jeff Nickoloff and I puzzled over this with some coffee, and he was stumped too. He suggested I get in touch with Kelsey Hightower, since he’s involved with a lot of the internals of the systems I’m working with, and he lives in Portland.

So I did, and Kelsey was a real mensch. Even though he was in Japan at the time, and it was the middle of the night there, he spent 90 minutes on a hangout with me to figure out the problem.

The Problem

The problem here is a difference in expectations as to what IP addresses are owned by whom. Kelsey was explaining to me how Docker containers should be using masqueraded IP addresses for traffic outside the Docker daemon, when we ran this on one of the minions:

$ iptables -v -L -n -t nat
 Chain POSTROUTING (policy ACCEPT 78 packets, 4680 bytes)
 pkts bytes target prot opt in out source destination
 87316 5451K MASQUERADE all -- * eth0 !

What that MASQUERADE line means is “if the destination address isn’t a 10.*.*.* address (which Kubernetes reserves for itself), send it as though it were coming from the EC2 instance’s IP address.” That’s a great rule, and it allows containers to talk to the outside world, and the return traffic finds its way back to the EC2 instance, and then back to the container that initiated the connection.

However. Recall from earlier that our pre-existing VPC’s IP addresses are in the 10.0.*.* range, including the RDS instance we were trying to connect to. This iptables rule is turning off masquerading to those IP addresses, so the RDS instance was seeing packets that look like they come from a 10.something address. Trying to respond to such a packet just doesn’t work; the other VPC thinks 10.something addresses are inside its own VPC, but there’s no machine at that address.

As a quick test, we punched a very small hole through this rule:

$ sudo iptables -t nat -I POSTROUTING -d -o eth0 -j MASQUERADE
 $ iptables -v -L -n -t nat
 Chain POSTROUTING (policy ACCEPT 32 packets, 1920 bytes)
 pkts bytes target prot opt in out source destination
 2 120 MASQUERADE all -- * eth0
 87355 5454K MASQUERADE all -- * eth0 !

We inserted a rule at the top that says “if the destination is the RDS instance, make sure you masquerade as the host machine.” And it worked! Now the Docker containers can connect to the database. And I owe Kelsey a nice lunch.

The Solution Band-Aid

It’s possible this won’t be a problem for you, unless you’re using a peering connection to a VPC that uses those addresses, in which case it’s going to be a serious problem for you.

Unfortunately there isn’t a full-on fix right now. If you’re going to peer to a VPC whose IP addresses start with 10, you’re going to have to muck with your nodes manually, because that rule is hard coded.

So as of right now, the answer is to SSH into every one of your minions and do this:

$ sudo iptables -t nat -I POSTROUTING -d /32 -o eth0 -j MASQUERADE

Good luck, and I hope you found this post if this problem affects you.

UPDATE: A Note About DNS

Once we got this up and running, we noticed glitchy DNS response. When one container was trying to find a separate service, it would only sometimes work. I did a little digging, and here’s what I found in the /etc/resolv.conf on both of my minions:

search default.svc.cluster.local svc.cluster.local cluster.local us-west-1.compute.internal
options ndots:5

That first line is key. Since the foreign VPC was in the 10.0.0.* space, I had set up AWS routing to send traffic to those IPs through the peering connection. Kubernetes is expecting the SkyDNS service to respond at that address, but traffic was getting routed elsewhere. When I corrected this issue, DNS started working as expected.

The bottom line here is that Kubernetes really wants to have ownership of all IP addresses starting with 10. If you try to put other machines into that range, you’re going to have problems.

About Ben Straub

Ben Straub is a lifelong developer, and enthusiast of the craft of making great software. He enjoys reading, taking his kids on bike rides, chocolate, dogs, those little notebooks you carry around with you, photography, a good weekend hack, traveling, writing, food, craftsmanship, a great pen, Markdown, music, movies, and talking to amazing people. He can frequently be found exploring his hometown of Portland with his wife and two kids.

0 replies on “Kubernetes and AWS VPC Peering”

You may also be interested in...

Migrating to Aurora: easy except the bill

Pay for usage sounds good, but how you use the database is key. Migrating our production database from Postgres to Aurora was easy, until we noticed that our daily database costs more than doubled. Our increased costs were driven by storage IO, a type of usage we’d never before optimized or even measured. We resized… Read More

Fresh baked software

In the olden days, software used to come in a box. Now, we can make sure our users always have fresh baked software.