Self-managed Kubernetes in Amazon Web Services (AWS)
4 MINUTE READ
Big picture
Use Calico with a self-managed Kubernetes cluster in Amazon Web Services (AWS).
Value
Managing your own Kubernetes cluster (as opposed to using a managed-Kubernetes service like EKS), gives you the most flexibility in configuring Calico and Kubernetes. Calico combines flexible networking capabilities with “run-anywhere” security enforcement to provide a solution with native Linux kernel performance and true cloud-native scalability.
Concepts
Kubernetes Operations (kops) is a cluster management tool that handles provisioning cluster VMs and installing Kubernetes. It has built-in support for using Calico as the Kubernetes networking provider.
There are many ways to install and manage Kubernetes in AWS. Using Kubernetes Operations (kops) is a good default choice for most people, as it gives you access to all of Calico’s flexible and powerful networking features. However, there are other options that may work better for your environment.
Or, you you can add calico to your cluster config. Run kops edit cluster and set the following networking configuration.
networking:
calico: {}
The geeky details of what you get:
Policy
Calico
Calico Policy
Kubernetes network policies are implemented by network plugins rather than Kubernetes itself. Simply creating a network policy resource without a network plugin to implement it, will have no effect on network traffic.
The Calico plugin implements the full set of Kubernetes network policy features. In addition, Calico supports Calico network policies, providing additional features and capabilities beyond Kubernetes network policies. Kubernetes and Calico network policies work together seamlessly, so you can choose whichever is right for you, and mix and match as desired.
IPAM
Calico
Calico IPAM
How Kubernetes assigns IP address to pods is determined by the IPAM (IP Address Management) plugin being used.
The Calico IPAM plugin dynamically allocates small blocks of IP addresses to nodes as required, to give efficient overall use of the available IP address space. In addition, Calico IPAM supports advanced features such as multiple IP pools, the ability to specify a specific IP address range that a namespace or pod should use, or even the specific IP address a pod should use.
CNI
Calico
Calico CNI
The CNI (Container Network Interface) plugin being used by Kubernetes determines the details of exactly how pods are connected to the underlying network.
The Calico CNI plugin connects pods to the host networking using L3 routing, without the need for an L2 bridge. This is simple and easy to understand, and more efficient than other common alternatives such as kubenet or flannel.
Overlay
IPIP
IPIP Overlay
An overlay network allows pods to communicate between nodes without the underlying network being aware of the pods or pod IP addresses.
Packets between pods on different nodes are encapsulated using IPIP, wrapping each original packet in an outer packet that uses node IPs, and hiding the pod IPs of the inner packet. This can be done very efficiently by the Linux kernel, but it still represents a small overhead, which you might want to avoid if running particularly network intensive workloads.
For completeness, in contrast, operating without using an overlay provides the highest performance network. The packets that leave your pods are the packets that go on the wire.
Routing
BGP
BGP Routing
BGP (Border Gateway Protocol) is used to dynamically program routes for pod traffic between nodes.
BGP is a standards-based routing protocol used to build the internet. It scales exceptionally well, and even the largest Kubernetes clusters represent a tiny amount of load compared to what BGP can cope with.
Calico can run BGP in three modes:
Full mesh - where each node talks BGP to each other, easily scaling to 100 nodes, on top of an underlying L2 network or using IPIP overlay
With route reflectors - where each node talks to one or more BGP route reflectors, scaling beyond 100 nodes, on top of an underlying L2 network or using IPIP overlay
Peered with TOR (Top of Rack) routers - in a physical data center where each node talks to routers in the top of the corresponding rack, scaling to the limits of your physical data center.
Datastore
etcd
etcd Datastore
Calico stores the operational and configuration state of your cluster in a central datastore. If the datastore is unavailable, your Calico network continues operating, but cannot be updated (no new pods can be networked, no policy changes can be applied, etc.).
Calico has two datastore drivers you can choose from:
etcd - for direct connection to an etcd cluster
Kubernetes - for connection to a Kubernetes API server
The advantages of using etcd as the datastore are:
Allows you to run Calico on non-Kubernetes platforms (e.g. OpenStack)
Allows separation of concerns between Kubernetes and Calico resources, for example allowing you to scale the datastores independently
Allows you to run a Calico cluster that contains more than just a single Kubernetes cluster, for example, bare metal servers with Calico host protection interworking with a Kubernetes cluster or multiple Kubernetes clusters.
For completeness, the advantages of using Kubernetes as the datastore are:
It doesn’t require an extra datastore, so is simpler to install and manage
You can use Kubernetes RBAC to control access to Calico resources
You can use Kubernetes audit logging to generate audit logs of changes to Calico resources.
?
Calico Deployment Options
Calico’s flexible modular architecture supports a wide range of deployment options, so you can select the best networking and network policy options for your specific environment. This includes the ability to run with a variety of CNI and IPAM plugins, and underlying networking options.
The Calico Getting Started guides default to the options most commonly used in each environment, so you don’t have to dive into the details unless you want to.
You can click on any deployment option to learn more.
As an alternative to Calico for both networking and network policy, you can use Amazon’s VPC CNI plugin for networking, and Calico for network policy. The advantage of this approach is that pods are assigned IP addresses associated with Elastic Network Interfaces on worker nodes. The IPs come from the VPC network pool and therefore do not require NAT to access resources outside the Kubernetes cluster.
Set your kops cluster configuration to:
networking:
amazonvpc: {}
Then install Calico for network policy only after the cluster is up and ready.
The geeky details of what you get:
Policy
Calico
Calico Policy
Kubernetes network policies are implemented by network plugins rather than Kubernetes itself. Simply creating a network policy resource without a network plugin to implement it, will have no effect on network traffic.
The Calico plugin implements the full set of Kubernetes network policy features. In addition, Calico supports Calico network policies, providing additional features and capabilities beyond Kubernetes network policies. Kubernetes and Calico network policies work together seamlessly, so you can choose whichever is right for you, and mix and match as desired.
IPAM
AWS
AWS IPAM
How Kubernetes assigns IP address to pods is determined by the IPAM (IP Address Management) plugin being used.
The AWS IPAM plugin dynamically allocates small blocks of IP addresses to nodes as required, using IP addresses from the underlying VPC (Virtual Private Cloud). The AWS IPAM plugin is used in conjunction with the Amazon VPC CNI plugin to provide VPC native pod networking.
CNI
AWS
AWS CNI
The CNI (Container Network Interface) plugin being used by Kubernetes determines the details of exactly how pods are connected to the underlying network.
The AWS Amazon VPC CNI and IPAM plugins provide pods with IP addresses from the underlying VPC (Virtual Private Cloud) to provide a VPC-Native pod network. The AWS VPC is used to route pod traffic between nodes, and understands which pod IP address are located on which nodes. This avoids the need for an overlay, and typically has good network performance characteristics.
In addition, pod IPs are understood by the broader AWS network, so for example, VMs outside of the cluster can connect directly to any pod without going via a Kubernetes service if desired.
Overlay
No
No Overlay
Operating without using an overlay provides the highest performance network. The packets that leave your pods are the packets that go on the wire.
For completeness, in contrast, with an overlay network, packets between pods on different nodes are encapsulated using a protocol such as VXLAN or IPIP, wrapping each original packet in an outer packet that uses node IPs, and hiding the pod IPs of the inner packet. This can be done very efficiently by the Linux kernel, but it still represents a small overhead, which you might want to avoid if running particularly network intensive workloads.
Routing
VPC Native
VPC Native Routing
The underlying cloud VPC (Virtual Private Cloud) is used to route pod traffic between nodes, and understands which pod IP address are located on which nodes. This avoids the need for an overlay, and typically has good performance characteristics.
In addition, pod IPs are understood by the broader cloud network, so for example, VMs outside of the cluster can connect directly to any pod without going via a Kubernetes service if desired.
Datastore
Kubernetes
Kubernetes Datastore
Calico stores the operational and configuration state of your cluster in a central datastore. If the datastore is unavailable, your Calico network continues operating, but cannot be updated (no new pods can be networked, no policy changes can be applied, etc.).
Calico has two datastore drivers you can choose from
etcd - for direct connection to an etcd cluster
Kubernetes - for connection to a Kubernetes API server
The advantages of using Kubernetes as the datastore are:
It doesn’t require an extra datastore, so is simpler to install and manage
You can use Kubernetes RBAC to control access to Calico resources
You can use Kubernetes audit logging to generate audit logs of changes to Calico resources
For completeness, the advantages of using etcd as the datastore are:
Allows you to run Calico on non-Kubernetes platforms (e.g. OpenStack)
Allows separation of concerns between Kubernetes and Calico resources, for example allowing you to scale the datastores independently
Allows you to run a Calico cluster that contains more than just a single Kubernetes cluster, for example, bare metal servers with Calico host protection interworking with a Kubernetes cluster or multiple Kubernetes clusters.
?
Calico Deployment Options
Calico’s flexible modular architecture supports a wide range of deployment options, so you can select the best networking and network policy options for your specific environment. This includes the ability to run with a variety of CNI and IPAM plugins, and underlying networking options.
The Calico Getting Started guides default to the options most commonly used in each environment, so you don’t have to dive into the details unless you want to.
You can click on any deployment option to learn more.
As discussed here, AWS CNI provisions multiple ENIs per node as the number of pods on the node increases.
AWS CNI will add entries for the primary ENI into the main routing table, and will then create routing tables for each additional ENI, starting at index 2. Additionally, if VLANs are being used, it appears that AWS CNI will use tables from 100 onwards.
By default, Felix considers routing table indexes from 1-250 to be under its control, and hence will remove the routing tables created by AWS CNI. This can cause loss of connectivity between pods if they are not on the primary ENI.
Note: The following steps will result in loss of connectivity between some pods. It is recommended to only make such changes during a maintenance window.
To ensure that AWS CNI and Felix manage separate ranges of routing tables, you must do the following:
Configure Felix to manage a routing table range which is distinct from the range used by AWS CNI:
Delete any routing rules and tables in the range 1-64 as they could be damaged or incomplete
Kill all the aws-node pods, which will force AWS CNI to recreate its routing rules and tables.
Kubespray
Kubespray is a tool for provisioning and managing Kubernetes clusters with support for multiple clouds including Amazon Web Services. Calico is the default networking provider, or you can set the kube_network_plugin variable to calico. See the Kubespray docs for more details.