Horizontal Pod Autoscaler¶

Scaling an application¶

You can scale an application in two ways:

Vertically: by adding more resources (RAM/CPU/Disk IOPS) to the same instance,
Horizontally: by adding more instances (replicas) of the same application.

The problem with vertical scaling is that either the required hardware (RAM, CPU, Disk IOPS) in a single machine costs too much, or the cloud provider cannot provision a machine with enough resources.

We use replica sets in Kubernetes to achieve horizontal scaling. The Horizontal Pod Autoscaler allows automating the process of maintaining the replica count proportionally to the application load.

Horizontal scaling strategy¶

The horizontal scaling strategy involves adding (or removing) the additional replicas of the same application. The problem here lies in the fact that most application load patterns can have spikes that are not predictable. This renders manual scaling nearly impossible. Luckily, we can automate this process.

The HPA & KEDA¶

Kubernetes has the Horizontal Pod Autoscaler (HPA) functionality. It can scale up (add more replicas) or down (remove idling replicas) based on some metrics. However, HPA does not have the metrics' source by default. CAST AI offers you a solution with the KEDA addon.

How does it work¶

KEDA consists of two components:

operator - watches k8s for ScaledObject resources and configures HPA accordingly
metrics-apiserver - a bridge between Kubernetes and various scaling sources (including Prometheus)

These components configure Kubernetes HPA and set up the custom metric sources. This enables us to autoscale almost any workload: Deployment, ReplicaSet, ReplicationController, or StatefulSet. KEDA supports autoscaling Jobs as well.

Examples¶

Autoscale Based on CPU and/or Memory usage¶

Let's create a Deployment and a Service that we will Autoscale:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: sample-app
  labels:
    app: sample-app
spec:
  # Note that we omit the replica count so
  # when we redeploy, we wouldn't override
  # replica count set by the autoscaler
  #replicas: 1
  selector:
    matchLabels:
      app: sample-app
  template:
    metadata:
      labels:
        app: sample-app
    spec:
      containers:
      - image: luxas/autoscale-demo:v0.1.2
        name: sample-app
        ports:
        - containerPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  name: sample-app
  labels:
    app: sample-app
spec:
  ports:
  - port: 8080
    name: http
    targetPort: 8080
    protocol: TCP
  selector:
    app: sample-app

Note: We do not specify the ReplicaCount ourselves

Now set up a CPU-based Autoscaler

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: sample-app
spec:
  scaleTargetRef:
    name:          sample-app
  minReplicaCount: 1        # Optional. Default: 0
  maxReplicaCount: 10       # Optional. Default: 100
  triggers:
    # Either of the triggers can be omitted.
    - type: cpu
      metadata:
        # Possible values: `Value`, `Utilization`, or `AverageValue`.
        # More info at: https://keda.sh/docs/2.0/scalers/cpu/#trigger-specification
        type: "Value"
        value: "30"
    - type: memory
      metadata:
        # Possible values: `Value`, `Utilization`, or `AverageValue`.
        # More info at: https://keda.sh/docs/2.0/scalers/memory/
        type: "Value"
        value: "512"

Now our Deployment autoscaling will be triggered either by CPU or Memory usage. We could use any other trigger, or remove either of those if we want (i.e. to autoscale only on the CPU basis and remove the Memory trigger, or vice-versa).

Autoscale based on the Prometheus metric¶

It is possible to autoscale based on the result of an arbitrary Prometheus query. CAST AI k8s clusters come with Prometheus deployed out-of-the-box.

Let's deploy the sample application again and instruct Prometheus to scrape metrics:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: sample-app
  labels:
    app: sample-app
spec:
  selector:
    matchLabels:
      app: sample-app
  template:
    metadata:
      labels:
        app: sample-app
      annotations:
        # These annotations the main difference!
        prometheus.io/path: "/metrics"
        prometheus.io/port: "8080"
        prometheus.io/scrape: "true"
    spec:
      containers:
      - image: luxas/autoscale-demo:v0.1.2
        name: sample-app
        ports:
        - containerPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  name: sample-app
  labels:
    app: sample-app
spec:
  ports:
  - port: 8080
    name: http
    targetPort: 8080
    protocol: TCP
  selector:
    app: sample-app

Now let's deploy the Autoscaler.

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: sample-app
spec:
  scaleTargetRef:
    name:          sample-app
  minReplicaCount: 1        # Optional. Default: 0
  maxReplicaCount: 10       # Optional. Default: 100
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prom.castai:9090
        metricName: http_requests_total_sample_app
        threshold: '1'
        # Note: query must return a vector/scalar single element response
        query: sum(rate(http_requests_total{app="sample-app"}[2m]))

Now let's generate some load and observe that the replica count is increased:

# Deploy busybox image
kubectl run -it --rm load-generator --image=busybox /bin/sh

# Hit ENTER for command prompt

# trigger infinite requests to the php-apache server
while true; do wget -q -O- http://sample-app:8080/metrics; done

# in order to cancel, hold CTRL+C
# in order to quit, initiate CTRL+D sequence

Troubleshooting¶

Verify that KEDA is scheduled and running (the suffixes might be different):

$ kubectl get pods -n keda
NAME                                      READY   STATUS    RESTARTS   AGE
keda-metrics-apiserver-59679c9f96-5lfr5   1/1     Running   0          74m
keda-operator-66744fc69d-7njdd            1/1     Running   0          74m

Describe ScaledObject for clues. In this case, scaledObjectRef points to nonexistent object:

$ kubectl describe scaledobjects.keda.sh sample-app
Name:         sample-app
Namespace:    default
Labels:       scaledObjectName=sample-app
Annotations:  API Version:  keda.sh/v1alpha1
Kind:         ScaledObject
Metadata:
  Creation Timestamp:  2020-11-10T10:12:38Z
  Finalizers:
    finalizer.keda.sh
  Generation:  1
  Managed Fields:
    <... snip ...>
  Resource Version:  394466
  Self Link:         /apis/keda.sh/v1alpha1/namespaces/default/scaledobjects/sample-app
  UID:               9394d57a-ae66-4e80-baf4-8d6bb7fd36f9
Spec:
  Advanced:
    Horizontal Pod Autoscaler Config:
      Behavior:
        Scale Down:
          Policies:
            Period Seconds:              15
            Type:                        Percent
            Value:                       100
          Stabilization Window Seconds:  300
    Restore To Original Replica Count:   true
  Cooldown Period:                       300
  Max Replica Count:                     10
  Min Replica Count:                     1
  Polling Interval:                      30
  Scale Target Ref:
    API Version:  apps/v1
    Kind:         Deployment
    Name:         sample-app
  Triggers:
    Metadata:
      Metric Name:     http_requests_total
      Query:           sum(rate(http_requests_total{app="sample-app"}[2m]))
      Server Address:  http://prom.castai:9090
      Threshold:       1
    Type:              prometheus
Status:
  Conditions:
    Message:  ScaledObject doesn't have correct scaleTargetRef specification
    Reason:   ScaledObjectCheckFailed
    Status:   False         <--------- This means that this check didn't pass
    Type:     Ready
    Message:  ScaledObject check failed
    Reason:   UnkownState
    Status:   Unknown
    Type:     Active
Events:       <none>

Inspect KEDA operator logs:

kubectl logs -n keda $(kubectl get pods -n keda -o name | grep operator)