On clusterLeave, notify only if there are peers

[aboch]

• Otherwise operation will unnecessarely block
  for five seconds.
• This is particularly noticeable on graceful
  shutdown of daemon in one node cluster.

Related to moby/moby/issues/32446

[sanimej]

Is the error failed to send node leave: timed out broadcasting node event seen only in the single node case ?

Single node clusters are unlikely in real deployments and IMO, we shouldn't optimize for that.

[aboch]

I see your point, for clusters.

There sure are cases where a user might be using a single node swarm for feature development or testing. For those cases, I do not think the current is an acceptable behavior, considering this would delay things like service docker stop/restart, and docker swarm integration ci for that matter.

[aaronlehmann]

Single node clusters are unlikely in real deployments and IMO, we shouldn't optimize for that.

BTW, we are planning to change swarm mode to default to "on" in the future. The purpose will be to make features like services and secrets available without needing to run a special command first. This will make single-node installations very common.

[mavenugo]

@aaronlehmann thats a good point.
@sanimej @aboch we have to consider the fact that single-node swarm-mode will be the most common case.

[aaronlehmann]

Is it possible to get this in for 17.05? It would be great to avoid the extra delay for graceful shutdowns.

[sanimej]

@aboch sendNodeEvent is an intentional blocking call till the event gets broadcasted. Its better to handle the special case inside of this function. Please add the check here before the wait https://github.com/docker/libnetwork/blob/master/networkdb/broadcast.go#L89.

[aboch]

Updated. PTAL.

[sanimej]

LGTM