In #2492, we noticed that the /_ping endpoint forwards to the primary now, which creates an HTTP connection hijacking which causes subsequent /info requests to get redirected straight to the primary. This PR ensures that /_ping requests are still entirely handled by the local node while still checking the primary's status on replicas.

we noticed that the /_ping endpoint forwards to the primary now, which creates an HTTP connection hijacking which causes subsequent /info requests to get redirected straight to the primary

Requests should be independent of each other. If this is the case we have a bug. Can you provide the reproduce sequence for this?

There's more discussion in #2526. We never fully figured out a solid repro for this issue, but both me and @michelvocks saw this issue.

@dongluochen @wsong You are right. This is a bug. I spent more time on investigating the issue and the root of it is this commit: docker/engine-api@5dd6452

The client is reusing the same connection. That means it first sends the _ping request and then uses the same connection to send the /info request. The hijacking method in swarm looks like this:

cp := func(dst io.Writer, src io.Reader, chDone chan struct{}) {
		io.Copy(dst, src)
		if conn, ok := dst.(interface {
			CloseWrite() error
		}); ok {
			conn.CloseWrite()
		}
		close(chDone)
	}
	inDone := make(chan struct{})
	outDone := make(chan struct{})
	go cp(d, nc, inDone)
	go cp(nc, d, outDone)

The go-routine reads the whole stream and then forwards it to the primary docker swarm manager and because we are using the same connection for the _ping request as well as the /info request it will forward both requests. I'm now not quite sure what's the right way to go when looking at the design. Should this be fixed in the docker client or at docker swarm?

Rebased on top of #2492

@wsong was there further discussion about this PR? Have you faced the issue lately?

This isn't so much an issue as it is a feature request; because Swarm replicas can occasionally lose their connection to the primary, we want to be able to hit an endpoint telling us whether or not there is currently an elected primary, from the perspective of this replica. This is useful when, for instance, you add a new Swarm manager and it takes a while to connect to the KV store; we can hit _ping in a retry loop to determine when the new manager is ready.

This is no longer relevant now that we block in api/replica.go until we have an elected primary cluster manager.