You can turn on Knative load balancing, by placing the Activator service in the request path to act as a load balancer. To do this, you must first ensure that individual pod addressability is enabled.
Activator pod selection¶
Activator pods are scaled horizontally, so there may be multiple Activators in a deployment. In general, the system will perform best if the number of revision pods is larger than the number of Activator pods, and those numbers divide equally.
Knative assigns a subset of Activators for each revision, depending on the revision size. More revision pods will mean a greater number of Activators for that revision.
The Activator load balancing algorithm works as follows:
- If concurrency is unlimited, the request is sent to the better of two random choices.
- If concurrency is set to a value less or equal than 3, the Activator will send the request to the first pod that has capacity. Otherwise, requests will be balanced in a round robin fashion, with respect to container concurrency.
For more information, see the documentation on concurrency.
Configuring target burst capacity¶
Target burst capacity is mainly responsible for determining whether the Activator is in the request path outside of scale-from-zero scenarios.
Target burst capacity can be configured using a combination of the following parameters:
- Setting the targeted concurrency limits for the revision. See concurrency.
- Setting the target utilization parameters. See target utilization.
- Setting the target burst capacity. You can configure target burst capacity using the
target-burst-capacitykey in the
config-autoscalerConfigMap. See Setting the target burst capacity.
- Setting the Activator capacity by using the
config-autoscalerConfigMap. See Setting the Activator capacity.