GRPC Health Checking Protocol

Health checks are used to probe whether the server is able to handle rpcs. The client-to-server health checking can happen from point to point or via some control system. A server may choose to reply “unhealthy” because it is not ready to take requests, it is shutting down or some other reason. The client can act accordingly if the response is not received within some time window or the response says unhealthy in it.

A GRPC service is used as the health checking mechanism for both simple client-to-server scenario and other control systems such as load-balancing. Being a high level service provides some benefits. Firstly, since it is a GRPC service itself, doing a health check is in the same format as a normal rpc. Secondly, it has rich semantics such as per-service health status. Thirdly, as a GRPC service, it is able reuse all the existing billing, quota infrastructure, etc, and thus the server has full control over the access of the health checking service.

Service Definition

The server should export a service defined in the following proto:

syntax = "proto3";

package grpc.health.v1alpha;

message HealthCheckRequest {
  string service = 1;
}

message HealthCheckResponse {
  enum ServingStatus {
    UNKNOWN = 0;
    SERVING = 1;
    NOT_SERVING = 2;
  }
  ServingStatus status = 1;
}

service Health {
  rpc Check(HealthCheckRequest) returns (HealthCheckResponse);
}

A client can query the server’s health status by calling the Check method, and a deadline should be set on the rpc. The client can optionally set the service name it wants to query for health status. The suggested format of service name is package_names.ServiceName, such as grpc.health.v1alpha.Health.

The server should register all the services manually and set the individual status, including an empty service name and its status. For each request received, if the service name can be found in the registry, a response must be sent back with an OK status and the status field should be set to SERVING or NOT_SERVING accordingly. If the service name is not registered, the server returns a NOT_FOUND GRPC status.

The server should use an empty string as the key for server’s overall health status, so that a client not interested in a specific service can query the server's status with an empty request. The server can just do exact matching of the service name without support of any kind of wildcard matching. However, the service owner has the freedom to implement more complicated matching semantics that both the client and server agree upon.

A client can declare the server as unhealthy if the rpc is not finished after some amount of time. The client should be able to handle the case where server does not have the Health service.