Autoscaling PaaSTA Instances
PaaSTA allows programmatic control of the number of replicas (pods) a service has. It uses Kubernetes’ Horizontal Pod Autoscaler (HPA) to watch a service’s load and scale up or down.
How to use autoscaling
In order to use autoscaling, edit your
kubernetes-*.yaml files in your soa
configs and add a
min_instances and a
max_instances attribute and
(optionally) remove the
instances attribute from each instance you want to autoscale.
When using autoscaling, the
become the minimum and maximum (inclusive) number of replicas tasks Kubernetes will
create for your job.
If load history for your service is missing in Prometheus for some/all replicas, the Prometheus query will assume that each missing replica is at 100% load. The reasoning behind this is that during a situation where there is missing data, scaling a service up is generally the safest course of action. This behavior may mean that your service is scaled up unnecessarily when you first enable autoscaling. Don’t worry - the autoscaler will soon learn what the actual load on your service is, and will scale back down to the appropriate level.
If you use autoscaling it is highly recommended that you make sure your service has a readiness probe.
If your service is registered in Smartstack, each pod automatically gets a readiness probe that checks whether that pod is available in the service mesh.
Non-smartstack services may want to configure a
healthcheck_mode, and either
healthcheck_uri to ensure they have a readiness probe.
The HPA will ignore the load on your pods between when they first start up and when they are ready.
This ensures that the HPA doesn’t incorrectly scale up due to this warm-up CPU usage.
Autoscaling parameters are stored in an
autoscaling attribute of your instances as a dictionary.
autoscaling attribute, setting a
metrics_provider will allow you to specify a method that determines the utilization of your service.
If a metrics provider isn’t provided, the
cpu metrics provider will be used.
setpoint allows you to specify a target utilization for your service.
setpoint is 0.8 (80%).
Let’s look at sample kubernetes config file:
This makes the instance
main autoscale using the
cpu metrics provider.
PaaSTA will aim to keep this service’s CPU utilization at 50%.
The currently available metrics providers are:
The default autoscaling method if none is provided. Measures the CPU usage of your service’s container.
If you have configured your service to use a non-default stats port (8889), PaaSTA will not scale your service correctly!
The currently available decicion policies are:
(This is the default policy.) Uses a simple proportional model to decide the correct number of instances to scale to, i.e. if load is 110% of the setpoint, scales up by 10%.
Allows a service author to implement their own autoscaling. This policy results in no HPA being configured. An external process should periodically decide how many replicas this service needs to run, and use the Paasta API to tell Paasta to scale. See the How to create a custom (bespoke) autoscaling method section for details.
How to create a custom (bespoke) autoscaling method
The current number of instance for a service can be accessed through the PaaSTA
api from the endpoint
Sending an HTTP GET request will return an integer describing how many
instances PaaSTA thinks your sevice should have. This endpoint also accepts an
HTTP POST request with a JSON payload with the format
NUMBER_OF_DESIRED_INSTANCES}. This endpoint can be used to control the number
of instances PaaSTA thinks your service should have.
Finally, remember to set the
decision_policy of the
parameter for each service instance to
"bespoke" or else PaaSTA will
attempt to autoscale your service with the default autoscaling method.
In order to make you aware of when your
max_instances may be too low, causing issues with your service, paasta will send you alerts if all of the following conditions are true:
- The autoscaler has scaled your service to
- The load on your service (as measured by the
metrics_provideryou specified, e.g. your worker utilization or CPU utilization) is above
The default value for
max_instances_alert_threshold is whatever your
This means by default the alert will trigger when the autoscaler wants to scale up but is prevented from doing so by your
If this alert is noisy, you can try setting
max_instances_alert_threshold to something a little higher than your
Setting a very high value (a utilization value your metrics_provider would never measure) will effectively disable this alert.