gRPC on GKE and Istio

Samples for running gRPC on GKE and Istio:

gRPC Loadbalancing on GKE with L7 Ingress

client_grpc_app (via gRPC wire protocol) --> ingress --> (grpc Service on GKE)

• Folder: gke_ingress_lb/

gRPC Loadbalancing on GKE with TrafficDirector (xds loadbalancing)

client_grpc_app (acquire gRPC lookaside loadbalancing/XDS) --> GKE pod NEG address

• Folder: gke_td_xds/

gRPC-web via Ingress

Javascript clients:

client(browser) (via grpc-web protocol) --> Ingress --> (grpc Service on GKE)

• grpc_web_with_gke

gRPC for GKE internal Service->Service

• Folder gke_svc_lb/

client (pod running in GKE) --> k8s Service --> (gRPC endpoint in same GKE cluster)

gRPC w/ Managed Instance Group with Container Optimized OS

• Folder gce

client_grpc_app --> L7LB --> ManagedInstanceGroup

Istio gRPC Loadbalancing with GCP Internal LoadBalancer (ILB)

• folder istio/

client_grpc_app (on GCEVM) --> (GCP ILB) --> Istio --> Service

client_grpc_app (external) --> (GCP ExternalLB) --> Istio --> Service

Also see:

• Summary of Cloud load balancers
• GCP LoadBalancing Overview

Update 5/14/20: Also see xds gRPC Loadbalancing. Using xDS client-side only loadbalancing is woudl be the recommended way for internal service->service LB at scale. However, it is quite experimental at the moment and cannot be easily deployed on GKE/GCE (unless you use an xDS server like Traffic Director).

Source and Dockerfile

You can find the source here:

• Client and Server: app/grpc_backend/

And the various docker images on dockerhub

• docker.io/salrashid123/grpc_backend

To run the gRPC server locally to see message replay:

• Server: docker run -p 50051:50051 -t salrashid123/grpc_backend ./grpc_server -grpcport 0.0.0.0:50051

• Client: docker run --net=host --add-host grpc.domain.com:127.0.0.1 -t salrashid123/grpc_backend /grpc_client --host grpc.domain.com:50051

What client app makes ONE connection out to the Server and then proceeds to send 10 RPC calls over that conn. For each call, the server will echo back the hostname of the server/pod that handled that request. In the example here locally, it will be all from one host.

References

• https://grpc.io/blog/loadbalancing
• gRPC Loadbalancing on Kubernetes (Kubecon Europe 2018)
• gRPC-web "helloworld"
• gRPC with curl