Skip to content

Latest commit

 

History

History
113 lines (77 loc) · 8.65 KB

CREATE-NEW-SCALER.md

File metadata and controls

113 lines (77 loc) · 8.65 KB

Creating a new scaler

Developing a scaler

In order to develop a scaler, a developer should do the following:

  1. Download KEDA's code.
  2. Define the main pieces of data that you expect the user to supply so the scaler runs properly. For example, if your scaler needs to connect to an external source based on a connection string, you expect the user to supply this connection string in the configuration within the ScaledObject under trigger. This data will be passed to your constructing function as map[string]string.
  3. Create the new scaler struct under the pkg/scalers folder.
  4. Implement the methods defined in the scaler interface section.
  5. Create a constructor according to this.
  6. Change the buildScaler function in pkg/scaling/scalers_builder.go by adding another switch case that matches your scaler. Scalers in the switch are ordered alphabetically, please follow the same pattern.
  7. Run make build from the root of KEDA and your scaler is ready.

If you want to deploy locally:

  1. Open the terminal and go to the root of the source code.
  2. Run IMAGE_REGISTRY=docker.io IMAGE_REPO=johndoe make publish, where johndoe is your Docker Hub repo, this will create and publish images with your build of KEDA into your repo. Please refer the guide for local deployment for more details.
  3. Run IMAGE_REGISTRY=docker.io IMAGE_REPO=johndoe make deploy, this will deploy KEDA to your cluster.

Scaler interface

The scalers in KEDA are implementations of a KEDA Scaler Go interface declared in pkg/scalers/scaler.go. This documentation describes how scalers work and is targeted towards contributors and maintainers.

GetMetricsAndActivity

This is the key function of a scaler; it returns:

  1. A value that represents a current state of an external metric (e.g. length of a queue). The return type is an ExternalMetricValue struct which has the following fields:

    • MetricName: this is the name of the metric that we are returning. The name should be unique, to allow setting multiple (even the same type) Triggers in one ScaledObject, but each function call should return the same name.
    • Timestamp: indicates the time at which the metrics were produced.
    • WindowSeconds: //TODO
    • Value: A numerical value that represents the state of the metric. It could be the length of a queue, or it can be the amount of lag in a stream, but it can also be a simple representation of the state.
  2. A value that represents an activity of the scaler. The return type is bool.

    KEDA will check each trigger source on every ScaledObject every pollingInterval configured in the ScaledObject; it checks the last time it was polled, it checks if the number of replicas is greater than 0, and if the scaler itself is active. So if the scaler returns false for IsActive, and if current number of replicas is greater than 0, and there is no configured minimum pods, then KEDA scales down to 0.

Kubernetes HPA (Horizontal Pod Autoscaler) will poll GetMetricsAndActivity regularly through KEDA's metric server (as long as there is at least one pod), and compare the returned value to a configured value in the ScaledObject configuration. Kubernetes will use the following formula to decide whether to scale the pods up and down:

desiredReplicas = ceil[currentReplicas * ( currentMetricValue / desiredMetricValue )].

For more details check Kubernetes HPA documentation.


Next lines are an example about how to use it:

func (s *artemisScaler) GetMetricsAndActivity(ctx context.Context, metricName string) ([]external_metrics.ExternalMetricValue, bool, error) {
	messages, err := s.getQueueMessageCount(ctx)

	if err != nil {
		s.logger.Error(err, "Unable to access the artemis management endpoint", "managementEndpoint", s.metadata.managementEndpoint)
		return []external_metrics.ExternalMetricValue{}, false, err
	}

	metric := GenerateMetricInMili(metricName, float64(messages))

	return []external_metrics.ExternalMetricValue{metric}, messages > s.metadata.activationQueueLength, nil
}

GetMetricSpecForScaling

KEDA works in conjunction with Kubernetes Horizontal Pod Autoscaler (HPA). When KEDA notices a new ScaledObject, it creates an HPA object that has basic information about the metric it needs to poll and scale the pods accordingly. To create this HPA object, KEDA invokes GetMetricSpecForScaling.

The return type of this function is MetricSpec, but in KEDA's case we will mostly write External metrics. So the property that should be filled is ExternalMetricSource, where the:

  • MetricName: the name of our metric we are returning in this scaler. The name should be unique, to allow setting multiple (even the same type) Triggers in one ScaledObject, but each function call should return the same name.
  • TargetValue: is the value of the metric we want to reach at all times at all costs. As long as the current metric doesn't match TargetValue, HPA will increase the number of the pods until it reaches the maximum number of pods allowed to scale to.
  • TargetAverageValue: the value of the metric for which we require one pod to handle. e.g. if we have a scaler based on the length of a message queue, and we specificy 10 for TargetAverageValue, we are saying that each pod will handle 10 messages. So if the length of the queue becomes 30, we expect that we have 3 pods in our cluster. (TargetAverage and TargetValue are mutually exclusive).

All scalers receive a parameter named triggerIndex as part of ScalerConfig. This value is the index of the current scaler in a ScaledObject. All metric names have to start with sX- (where X is triggerIndex). This convention makes the metric name unique in the ScaledObject and brings the option to have more than 1 "similar metric name" defined in a ScaledObject.

For example:

  • s0-redis-mylist
  • s1-redis-mylist

Note: There is a naming helper function GenerateMetricNameWithIndex(triggerIndex int, metricName string), that receives the current index and the original metric name (without the prefix) and returns the concatenated string using the convention (please use this function).
Next lines are an example about how to use it:

func (s *artemisScaler) GetMetricSpecForScaling() []v2.MetricSpec {
  externalMetric := &v2.ExternalMetricSource{
  	Metric: v2.MetricIdentifier{
  		Name: GenerateMetricNameWithIndex(s.metadata.triggerIndex, kedautil.NormalizeString(fmt.Sprintf("%s-%s-%s", "artemis", s.metadata.brokerName, s.metadata.queueName))),
  	},
  	Target: GetMetricTarget(s.metricType, s.metadata.queueLength),
  }
  metricSpec := v2.MetricSpec{External: externalMetric, Type: artemisMetricType}
  return []v2.MetricSpec{metricSpec}
}

Close

After each poll on the scaler to retrieve the metrics, KEDA calls this function for each scaler to give the scaler the opportunity to close any resources, like http clients for example.

Constructor

What is missing from the scaler interface is a function that constructs the scaler itself. Up until the moment of writing this document, KEDA does not have a dynamic way to load scalers (at least not officially)[***]; instead scalers are part of KEDA's code-base, and they are shipped with KEDA's binary.

Thus, each scaler should have a constructing function, KEDA will explicitly invoke the construction function based on the trigger property configured in the ScaledObject.

The constructor should have the following parameters:

  • resolvedEnv: of type map[string]string. This is a map of all the environment variables that exist for the target Deployment.
  • metadata: of type map[string]string. This is a map for all the trigger attributes of the ScaledObject.

Lifecycle of a scaler

Scalers are created and cached until the ScaledObject is modified, or GetMetricsAndActivity() result in an error. The cached scaler is then invalidated and a new scaler is created. Close() is called on all scalers when disposed.

Note

The scaler code is embedded into the two separate binaries comprising KEDA, the operator and the custom metrics server component. The metrics server must be occasionally rebuilt published and deployed to k8s for it to have the same code as your operator.

GetMetricSpecForScaling() is executed by the operator for the purposes of scaling out and in to 0 replicas. GetMetricsAndActivity() is executed by the KEDA Metrics server in response to a calls against the external metrics api, whether by the HPA loop or by curl.