busy loop in transportMonitor

[tsuna]

Please answer these questions before submitting your issue.

What version of gRPC are you using?

Commit bb78878

What version of Go are you using (go version)?

Go 1.9

What did you do?

When connecting to a gRPC backend through a load balancer like HA Proxy, if the load balancer decides to close the connection immediately (e.g. because there is no backend available), gRPC will attempt to reconnect immediately.

What did you expect to see?

gRPC should back off before trying to reconnect

What did you see instead?

gRPC is tying to reconnect thousands of times per second.

The issue comes from the loop in transportMonitor() in clientconn.go. When the backend closes the connection, we wake up in case <-t.Error(): with an EOF error, then we do if err := ac.resetTransport(false); err != nil { and this works because we are able to connect to the load balancer again, so then we loop again. But the new connection gets almost instantaneously closed again by the load balancer, leading to a quasi-busy loop in transportMonitor() and a ton of log spam.

[tsuna]

Probably a dup of #954

[tsuna]

or of #1444...

[dfawley]

Yes, definitely a dupe. Thanks for the report.