Custom Pod Autoscaler Framework
What is it?
The Custom Pod Autoscaler Framework is a way to allow people to create and use custom scalers, similar to the Horizontal Pod Autoscaler, in Kubernetes.
Why would I use it?
Kubernetes provides the Horizontal Pod Autoscaler, which allows automatic scaling of the number of replicas in a resource (Deployment, ReplicationController, ReplicaSet, StatefulSet) based on metrics that you feed it. Mostly the metrics used are CPU/memory load, which is sufficient for most applications. You can specify custom metrics to feed into it through the metrics API also.
The limitation in the Horizontal Pod Autoscaler is that it has a hard-coded algorithm for assessing these metrics:
desiredReplicas = ceil[currentReplicas * ( currentMetricValue / desiredMetricValue )]
If you need more flexibility in your scaling, beyond this algorithm, Horizontal Pod Autoscaler doesn't meet your needs, you need to write your own scaling logic.
What application would need custom scaling?
Taking an example from Google Cloud's tutorial for hosting game servers on Kubernetes, there is a section discussing autoscaling:
The autoscaler can currently only scale the instance group based on CPU usage, which can be a misleading indicator of DGS load. Many DGSs are designed to consume idle cycles in an effort to optimize the game's simulation.
As a result, many game developers implement a custom scaling manager process that is DGS aware to deal with the specific requirements of this type of workload.
The crux of the issue here is that for game servers, it doesn't make sense to scale on CPU load or memory usage, and even if you implemented custom metrics the scaling algorithm wouldn't scale with these in a sensible way. The game servers should scale on number of players on the servers, or number of players looking to join a server - trying to ensure there are always positions available.
How does it work?
A Custom Pod Autoscaler has a base program that handles interacting with user logic, for example by using shell commands and piping data into them.
When developing a Custom Pod Autoscaler you define logic for two stages:
- Metric gathering - collecting or generating metrics; can be calling metrics APIs, running calculations locally, making HTTP requests.
- Evaluating metrics - taking these gathered metrics and using them to decide how many replicas a resource should have.
These two pieces of logic are all the custom logic required to build a Custom Pod Autoscaler, the base program will handle all Kubernetes API interactions for scaling/retrieving resources.