Add NewWithTimeout(cfg Config) (*Client, error)

[zbindenren]

Hi

First of all: thanks for that great software!

I tried to write a etcd cluster health check for a tool we are creating.

I was inspired by the code in etcdctl. And I noticed a strange behavior in the clients New method, when the etcd daemon is not running.

Here the test:

var endoints = []string{"localhost:1111"} // nothing is listening here

func TestHealth(t *testing.T) {
    ep := endoints[0]
    cli, err := clientv3.New(clientv3.Config{
        Endpoints:   []string{ep},
    })
    fmt.Println("code never gets here")
    if err != nil {
        t.Errorf("ep %s unhealthy: %s", ep, err)
    }
    ctx, cancel := context.WithTimeout(context.Background(), 1*time.Second)
    _, err = cli.Get(ctx, "/mule-test/health")
    fmt.Println("never gets called")
    cancel()
    if err != nil {
        t.Errorf("ep %s unhealthy: %s", ep, err)
    }
    fmt.Printf("ep %s is healthy", ep)

}

I see the following endpoint:

2016/09/20 07:32:23 grpc: addrConn.resetTransport failed to create client transport: connection error: desc = "transport: dial tcp [::1]:1111: getsockopt: connection refused"; Reconnecting to {"localhost:1111" <nil>}
2016/09/20 07:32:24 grpc: addrConn.resetTransport failed to create client transport: connection error: desc = "transport: dial tcp [::1]:1111: getsockopt: connection refused"; Reconnecting to {"localhost:1111" <nil>}
2016/09/20 07:32:26 grpc: addrConn.resetTransport failed to create client transport: connection error: desc = "transport: dial tcp [::1]:1111: getsockopt: connection refused"; Reconnecting to {"localhost:1111" <nil>}

Test hangs until timeout.

As workaround I can use net.DialTimeout(network, address string, timeout time.Duration).

It was just surprising to me and I was wondering if this is wanted behavior.

[heyitsanthony]

It's hanging because clientv3.New used to block until it could establish a connection to the cluster. It now only blocks if a DialTimeout is given. Cf. #6162 for more background. Thanks!