autoscaler

README

Autoscaler Microservice

---

📄 Contents

• Contents
  • Description
  • Deployment
  • How does the autoscaler work?
  • Directory and File description
  • Common questions

❕ Description

The autoscaler's goal is to scale scheduler replicas depending on cluster load to ensure performance is up to user requirement.

---

❕ Deployment of the autoscaler service

Before deploying the coordinator.yaml file, please configure the environment variables to the correct values used by the queue microservice.

  env:
  - name: PC_METRIC_URL
    value: "pod-coordinator.custom-scheduler.svc.cluster.local:8080/metrics"
  - name: MQ_HOST
    value: "sched-rabbitmq-0.sched-rabbitmq.custom-scheduler.svc.cluster.local"
  - name: MQ_MANAGE_PORT
    value: "15672"
  - name: MQ_USER
    value: "guest"
  - name: MQ_PASS
    value: "guest"
  - name: INTERVAL
    value: "300"
  - name: DEFAULT_QUEUE
    value: "epsilon.distributed"
  - name: POD_NAMESPACE

The **DEFAULT_QUEUE** is the queue used by the general-purpose scheduler. In Epsilon, atleast one scheduler service need to act as the default scheduler.
The **PC_METRIC_URL** is the hostname of the coodinator service (This can be ignored if the QueueTheory plugin is not enabled)

---

❕ How does the autoscaler operates?

[STEP 1]
The autoscaler will first get cluster metrics that the plugins require based on a specified time interval.

[STEP 2]
After getting the metrics the autoscaler will proceed to send the information to the different plugins and wait for their reply.

[STEP 3]
Once all the plugin's replies are consolidated the autoscaler will make a decision based on majority vote.

[STEP 4]
The autoscaler will not attempt to scale up or down the scheduler replicas if there is a tie and will try again on the next time interval.

---

❕ Directory and File Description

Directory Name	File name	Description
/	main.go	Implementation code of the main routine
/	helper.go	Contain helper methods use by the main routine
/interfaces	interface.go	Contains the auto scaler plugin interface definition
/plugins/linear_regression	plugin.go	Contains the linear regression plugin implementation
/plugins/queue_theory	plugin.go	Contains the queue theory plugin implementation
/plugins/rabbitmq	plugin.go	Contains the rabbitmq plugin implementation
/plugins/scheduler_prob	plugin.go	Contains the scheduler conflict probability plugin implementation
/yaml	autoscaler.yaml	Deployment file to deploy the scheduler in a Kubernetes cluster
/docker	Dockerfile	Used by docker to create a docker image

---

❕ Common questions

How to add a new autoscaler plugin

  1. Create a new folder in /plugins
  2. Write the plugin implementation and store the file in the new folder created in 1
  3. Open main.go and intialize the plugin. Intialization of plugins can be found at line 168. 

---