Kubernetes Service Discovery
I am helping my client transition from a few large monoliths hosted on AWS to micro-services hosted on Kubernetes. It is my first experience with Kubernetes and I must say that I am impressed so far.
Of course, micro-services work best when it is easy to find a service’s dependencies. This is one area where Kubernetes shines as it has a well designed service abstraction.
You can find services through two approaches:
- Using the environment variables that use the same conventions as those created by Docker links.
- Using DNS to resolve the service names to the service’s IP address.
Kubernetes injects environment variables for each service and each port exposed by the service. This makes it easy to deploy containers that use Docker links to find their dependencies. For example, if we are exposing a RabbitMQ service, we can locate it using the
RABBIT_MP_SERVICE_SERVICE_PORT variables. Other environment variables are also exposed to support this.
The easiest way to find out what environment variables are exposed are to exec the
env command within a pod:
1 2 3 4 5 6 7 8 9 10 11 12 13 14
kubectl exec memcached-rm58b env | grep RABBITMQ RABBITMQ_SERVICE_SERVICE_HOST=10.0.143.172 RABBITMQ_SERVICE_SERVICE_PORT_A=5672 RABBITMQ_SERVICE_PORT_5672_TCP_ADDR=10.0.143.172 RABBITMQ_SERVICE_PORT_15672_TCP_ADDR=10.0.143.172 RABBITMQ_SERVICE_PORT_15672_TCP_PORT=15672 RABBITMQ_SERVICE_PORT_15672_TCP=tcp://10.0.143.172:15672 RABBITMQ_SERVICE_PORT_5672_TCP_PORT=5672 RABBITMQ_SERVICE_SERVICE_PORT_B=15672 RABBITMQ_SERVICE_PORT_5672_TCP_PROTO=tcp RABBITMQ_SERVICE_PORT=tcp://10.0.143.172:5672 RABBITMQ_SERVICE_PORT_5672_TCP=tcp://10.0.143.172:5672 RABBITMQ_SERVICE_PORT_15672_TCP_PROTO=tcp RABBITMQ_SERVICE_SERVICE_PORT=5672
Kubernetes has a kube-dns addon that exposes the service’s name as a DNS entry. As a result, you can tell your application to connect to a host name. The advantage of this approach is that you do not need to do anything different than you would otherwise.
The service names are scoped within namespaces. This allows you to run different deployment of a service for each namespace (for example, one per developer or one per environments) without having to edit configuration files.
You can use either approach to write services that adapt to the environment in which they are deployed without having to change it’s configuration.
But What About the API?
Kubernetes provides a powerful API that allows you to inspect and discover services, replication controllers, pods and other component of a cluster. This would allow you to build deeper service discovery mechanisms where appropriate. But they are generally overkill for the needs of an application that needs to connect to it’s dependencies. I’ll discuss a few scenarios where this can be useful in a future post.