Skip to content

Build large Kubernetes clusters in AWS with the performance and visibility of native VPC networking

License

Notifications You must be signed in to change notification settings

romana/vpc-router

Repository files navigation

VPC-router

Introduction

Summary

The vpc-router implements automatic route failover and backup routes for Amazon VPC environments.

vpc-router lets users avoid route table limitations and build large Kubernetes clusters with the performance and visibility of native VPC networking.

It can also be used independently of Kubernetes whenever you need to manage routes, backup routes and route failover in VPC environments.

It provides a plugin architecture for the easy integration with other cloud orchestration systems.

Details

vpc-router is a utility for setting and deleting routes in Amazon EC2 VPC route tables and specifically for automatically managing route failover.

Each route is specified by a destination CIDR as well as a list of IP addresses of EC2 instances, which are eligible to receive packets for the CIDR. An instance from the list is chosen and a route is set. vpc-router continuously monitors instance health and performs an immediate route failover to another instance in the set in case of a detected instance failure.

Routes can be configured in different ways, but most commonly, vpc-router takes route configs from storage (a config file, in the future also a KV store) or via HTTP requests. It will make sure that routes in the VPC route table are updated as needed with every detected change to the route config.

By default, it applies all route updates to all the route tables it can find within a specified VPC.

Project origin

This program was developed for the Romana project to overcome the limit imposed on VPC route table entries (50 by default), which constricts the size of clusters. Avoiding this limit typically required running an overlay network, which does not offer the performance and visibility of native VPC networking.

Some users prefer to run CNI network providers that support more advance network policy APIs. However, most CNI pod networks require an overlay when clusters are split across Availability Zones (AZs), preventing HA clusters from delivering native VPC network performance. Romana, using vpc-router can build CNI pod networks across zones without an overlay.

While vpc-router was specifically designed for use with Romana and to take advantage of its topology aware IPAM in these Kubernetes deployment scenarios, it does not depend on either project and can also be used stand-alone.

Plugins for integration with different environments are provided. For example, a plugin to integrate with Romana.

Health-checks are also implemented via plugins. This means that vpc-router may either directly contact EC2 instances to check their health, or it may instead connect to AWS status and alert information, or use the node status provided by orchestration systems, such as Kubernetes.

Installation and running

You can either run vpc-router out of the source directory, or perform a full install, it can run outside of the VPC or within, on a cluster node or not.

Installation via pip

The vpc-router is in the Python Package Index (PyPi). Therefore, the simplest way to install it is just:

pip install vpcrouter

Run vpc-router out of the source directory

If you wish to work with the vpc-router sources, or contribute to the project, you might want to run vpc-router directly from the sources.

After downloading the code, create a virtual environment, activate it and install the required libraries. You can then use the vpcrouter-runner.py helper script to run vpc-router without a full install:

$ git clone [email protected]:paninetworks/vpc-router.git
$ virtualenv vpcrouter
$ source vpcrouter/bin/activate
$ cd vpc-router
$ pip install -r requirements/deploy.txt
$ ./vpcrouter-runner.py .....

Deploying in production environment

Please see our documentation on how to deploy vpc-router in production, which covers issues such as:

  • Performing a proper installation
  • IaM permissions for running on EC2 instances
  • Running vpc-router in a container

Contributing

Feedback, bug reports, issue tracker

We welcome any contributions, bug reports or feedback. Please use our issue tracker to file bugs or request additional features.

Developing vpc-router

In order to develop or extend vpc-router, please read the developer documentation for information that might be useful to get you started.

Built-in HTTP server to see internal state and config

vpc-router comes with a built-in HTTP server. By default it listens on localhost:33289. Send a GET request (with a web-browser, curl or wget, or any client you wish) to http://localhost:33290/ to receive a JSON formatted output with the current internal configuration of vpc-router.

The listen address and port can be modified with the -a (address) and -p (port) command line options.

Configuration

The route spec

vpc-router requires a route spec configuration in JSON format. It looks something like this:

{
    "10.55.16.0/24" : [ "10.33.20.142" ],
    "10.66.17.0/24" : [ "10.33.20.93", "10.33.30.22" ]
}

Each entry in the dictionary is keyed on the route's CIDR and then lists a number of eligible hosts, which can act as the target/router for this route. vpc-router randomly chooses an instance from a route's set of hosts.

If a route to a specified CIDR does not exist in the VPC's route tables, or if it does not specify a target that's contained in the specified host list, vpc-router creates or updates the route.

The health of those hosts is continuously monitored. If a host acting as router fails, vpc-router immediately switches the route to a different host from the set, if an alternative is available and healthy.

Note that vpc-router takes control of the routing tables and removes any entries of this type (interfaces on instances as target) if they are not part of the route spec.

Modes of operation

The modes for the detection of configuration updates are implemented via plugins. It is therefore easy to extend vpc-router to integrate with various orchestration systems. How to write plugins is documented separately.

A plugin may either accept a route spec in the format described above, or it may be used to translate other information into the properly formatted route spec.

Internal mode plugins

Out of the box, plugins for three different configuration update modes are included in the vpc-router source:

  • configfile: Continuosly monitor a route spec configuration file for any changes. The config file should contain the route spec in exactly the format described above.
  • http: Receive updated route specs via HTTP POSTs. The POSTed data should be the route-spec in exactly the format described above.
  • fixedconf: With this a static config can be provided on the command line. It is mostly used as a simple example for plugin developers. It does work and might be useful in some cases, but is not commonly going to be used in production. It translates the command line parameters into a route spec of the required format.

External mode plugins

It is also possible to write external plugins, which live in their own repository. There is currently one example of this:

  • romana: The Romana plugin provides seamless integration with the Romana project, for the creation of Kubernetes and OpenStack clusters without overlays, just natively routed network traffic. The vpc-router with the Romana plugin then allows those clusters to span multiple VPC Availability Zones, all while using native VPC networking and without being hindered by VPC's 50 route limit. The Romana plugin watches the network topology knowledge from Romana and translates it into the required route spec for vpc-router.

Mode 'configfile'

The following command starts vpc-router as a service daemon in 'configfile' mode:

$ vpcrouter -m configfile -f route-spec.conf -r us-east-1 -v vpc-350d6a51

The used options are:

  • -m configfile tells vpc-router to take config changes from a specified route spec file.
  • -f specifies the name of the route spec config file.
  • -r specifies the AWS region to which vpc-router should connect. Note: This can be omitted if vpc-router is run on an instance in the region.
  • -v specifies the VPC for which vpc-router should perform route updates. Note: This can be omitted if vpc-router is run on an instance within the VPC.

In 'configfile' mode the -f option must be used to specify the route spec config file. It must exist when the server is started. The server then continuously monitors this file for changes.

You can see an example route spec file in examples/route_spec_1.conf.

Mode 'http'

The following command starts vpc-router as a service daemon in the 'http' mode. It utilizes the built-in HTTP server to listen for new route specs:

$ vpcrouter -m http -r us-east-1 -v vpc-350d6a51

The used options are:

  • -m http tells vpc-router to start listening on a certain address and port for HTTP POST requests containing new route specs.
  • -r specifies the AWS region to which vpc-router should connect. Note: This can be omitted if vpc-router is run on an instance in the region.
  • -v specifies the VPC for which vpc-router should perform route updates. Note: This can be omitted if vpc-router is run on an instance within the VPC.

A new route spec can be POSTed to the /route_spec URL. The current route spec can be retrieved with a GET to that URL.

For example:

$ curl -X "POST" -H "Content-type:application/json" "http://localhost:33289/route_spec" -d '{"10.55.0.0/16" : [ "10.33.20.142" ], "10.66.17.0/24" : [ "10.33.20.93", "10.33.30.22" ]}'

Mode 'romana'

For integration with the Romana project, please see the vpc-router Romana plugin.

Continuous monitoring

Continuos monitoring is performed for all hosts listed in the route spec. If an instance does not appear healthy anymore and it is a current target for a route then the route will be automatically updated to point to an alternate target, if a healthy one is available.

The health-check itself is implemented via plugins, which gives vpc-router the flexibility to use a wide variety of information to determine whether an EC2 routing instance is healthy. By default, it uses the 'icmpecho' plugin, which utilizes an ICMPecho ('ping') request to actively check the responsiveness of instances. A 'tcp' plugin, which attempts TCP connection attempts on a specified port, is also provided.

Use the --health option to select the health monitor plugin, for example:

$ vpcrouter --health tcp --tcp_check_port 22 --tcp_check_interval 5 ...

or:

$ vpcrouter --health icmpecho --icmp_check_interval 5 ...

The 'multi' plugin: Combining more than one health monitor plugin

A multi plugin is also provided with vpc-router. This plugin allows the combining of multiple, simpler health-monitoring plugins into complex or multi-layered instance health monitoring. Use the --health option to select the multi plugin. Then use the --multi_plugins option to specify the list of simple health plugins that should be used simultaneously (a list of column separated health-monitor plugin names). Any additional command line options added by those plugins can then also be specified.

As an example, let's configure vpc-router to use the icmpecho plugin (with a monitoring interval of 2 seconds) as well as the tcp plugin, which should monitor port 80.

$ vpcrouter --health multi --multi_plugins icmpecho:tcp \
                    --icmp_check_interval 2 --tcp_check_port 80 ...

An instance is considered 'failed' if ANY of the specified sub-plugins reports the instance as failed.

Considering an instance as 'healthy' again

If a health monitoring plugin reports an instance as 'failed', it will be considered 'failed' for some amount of time (the exact time depends on the plugin, but usually it's 10 times the selected monitoring interval, when using the 'multi' plugin, it will be 20 times the largest interval of the specified sub-plugins).

The health monitoring occasionally attempts to re-check failed instances to see if they have recovered. If not, it will report them as failed again.

If there has not been a 'failed' report about an instance within that time window, the instance will automatically be considered as 'healthy' again. This does not mean that routes are failing back to that instance, it just means that this instance becomes eligible to be a target for routes again.

TODO

A 'todo' list is maintained in the issue tracker of the project.

About

Build large Kubernetes clusters in AWS with the performance and visibility of native VPC networking

Resources

License

Stars

Watchers

Forks

Packages

No packages published