Managing Apollo Router resources in Kubernetes
Recommendations and considerations for application pods
Determining the correct resource requests and limits for your application pods in a Kubernetes system is not an exact science. Your specific needs depend on many factors, including:
- The cardinality of unique operation shapes
- The latency of underlying subgraphs and data sources
- The size of responses
- The complexity of query plans
Our general recommendation for Kubernetes is to start with these requests and limits:
resources:requests:memory: '1G'cpu: '1000m'limits:memory: '2G'# no CPU limit to avoid throttling
The router resource estimator is a helpful tool for getting a starting baseline for what resources you may need in production based on your expected traffic.
When using Horizontal pod autoscaling, we recommend targeting 90% utilization:
metrics:- type: Resourceresource:name: cputarget:type: UtilizationaverageUtilization: 90- type: Resourceresource:name: memorytarget:type: UtilizationaverageUtilization: 90
The Apollo Router starts up quickly, but when using Managed Federation, the router has to fetch the supergraph schema from Apollo Uplink before it can start serving traffic. We recommend measuring your router's startup time and lowering the
averageUtilization if your startup time is longer due to Uplink latency or the size of your supergraph schema.