Skip to content

Latest commit

 

History

History
157 lines (147 loc) · 5.36 KB

README.md

File metadata and controls

157 lines (147 loc) · 5.36 KB

Kubernetes Workload Scaler

This is a controller to scale the specific deployment with various way. This controller can work for multi-cluster.

  • If cluster name is given within metric labels, it calculates the rate and scale the workload by cluster name
  • If cluster name is given alarm labels, it scales the workload on the specified cluster
  • If cluster name is not given, it will only calculate and scale in the same cluster controller lives

Organizing multi-cluster:

Supported methods to scale workload and Usage

1- Prometheus Alert API:

Constantly, watches the Prometheus API (http://host:port/api/v1/alerts) to catch the firing alarms. We must have two different alarm rule for scaling out and scaling in with the following labels.
For example:

  prometheus.rules: |-
    groups:
    - name: Pod Memory
      rules:
      - alert: php-apache-scaling-out # This must be provided
        expr: sum(container_memory_usage_bytes{container="php-apache", namespace="default"}) / (count(container_memory_usage_bytes{container="php-apache", namespace="default"})) > 11000000
        for: 1m
        labels:
          severity: slack
          scaling: out # Must include
          cluster_name: <cluster_name> # If not specified, it load incluster config
        annotations:
          summary: High Memory Usage
      - alert: php-apache-scaling-in # This must be provided
        expr: sum(container_memory_usage_bytes{container="php-apache", namespace="default"}) / (count(container_memory_usage_bytes{container="php-apache", namespace="default"})) < 11000000
        for: 1m
        labels:
          severity: slack
          scaling: in # Must include
          cluster_name: <cluster_name> # If not specified, it load incluster config
        annotations:
          summary: Low Memory Usage

Before deploying the pod, define the parameters

      env:
        - name: workload
          value: "Deployment"
        - name: scale-name
          value: "php-apache"
        - name: namespace
          value: "default"
        - name: scaling-number
          value: "1"
        - name: max-pod-number
          value: "10"
        - name: min-pod-number
          value: "2"
        - name: time-interval
          value: "60"
        - name: kube-config
          value: "/etc/kube/config"
        - name: managment-type
          value: "prometheus_alert_api"
        - name: prometheus-host
          value: "prometheus-service"
        - name: prometheus-port
          value: "8080"
        - name: scaling-out-name
          value: "php-apache-scaling-out"
        - name: scaling-in-name
          value: "php-apache-scaling-in"

If you set the alarms, you can simply run this deployment to scale out/in

kubectl apply -f https://raw.githubusercontent.com/eminaktas/k8s-workload-scaler/main/examples/php-apache-sample.yaml

Then,

kubectl create secret generic kube-config-file --from-file=config=$HOME/.kube/config
kubectl apply -f https://raw.githubusercontent.com/eminaktas/k8s-workload-scaler/main/examples/k8s-prometheus-sample.yaml

It will simply chekcs the Prometheus API and if receives a firing alarm it will trigger regarding the scaling type (we must define scaling: in and scaling: out labels)

2- Prometheus Metric API:

Reads, calculates and checks for any violation of thresholds to scale out or scale in

      env:
        - name: workload
          value: "Deployment"
        - name: scale-name
          value: "php-apache"
        - name: namespace
          value: "default"
        - name: scaling-number
          value: "1"
        - name: max-pod-number
          value: "10"
        - name: min-pod-number
          value: "2"
        - name: time-interval
          value: "0"
        - name: kube-config
          value: "/etc/kube/config"
        - name: management-type
          value: "prometheus_metric_api"
        - name: prometheus-host
          value: "prometheus-service"
        - name: prometheus-port
          value: "8080"
        - name: metric-name
          value: "apache_accesses_total"
        - name: label-1
          value: "kubernetes_name=apache-exporter"
        - name: label-2
          value: "run=php-apache"
        - name: scaling-out-threshold-value
          value: "0.8"
        - name: scaling-in-threshold-value
          value: "0.2"
        - name: rate_value
          value: "300" # 5 min

You can simply run this deployment to scale out/in. This includes an apache exporter. After deploying, make sure your Prometheus collecting metrics.

kubectl apply -f https://raw.githubusercontent.com/eminaktas/k8s-workload-scaler/main/examples/php-apache-sample.yaml

Then,

kubectl create secret generic kube-config-file --from-file=config=$HOME/.kube/config
kubectl apply -f https://raw.githubusercontent.com/eminaktas/k8s-workload-scaler/main/examples/k8s-prometheus-metric-sample.yaml

Supported Workloads

SUPPORTED_WORKLOAD = [
    'Deployment',
    'StatefulSet',
    'ReplicaSet',
    'ReplicationController',
]

Features

  • Auto-Scaling
  • Multi-cluster support
  • Prometheus alert api and metric api support

Docker image