Skip error message only for tcp

[ktf]

In case IPC is used, we should not get the error, so better know about it.

[ktf]

@rbx @dennisklein can we get a release with this in? I really do not understand how come we see a (silent) failure when many ipc channels are created on a given box.

[rbx]

@ktf did you confirm that this patch fixes/reveals the source of the issue?

[ktf]

No, but I do not see how the message:

[331:internal-dpl-aod-spawner_t1]: [06:38:34][ERROR] failed to attach channel from_internal-dpl-aod-spawner_t1_to_eta-and-cls-histograms[0] (bind)

could otherwise be produced without any further message.

[rbx]

Ok. The changes lgtm. The current behavior makes no sense for IPC.

I guess what is happening is that something already is using the given address on that machine. Your changes should reveal it next time.

I suggest including something unique, like the session id in the channel address string. We now do this here, for example: https://github.com/FairRootGroup/FairMQ/blob/master/test/protocols/_push_pull.cxx#L28.

[ktf]

I suspect it's actually more subtle than that. The ipc path is formed with:

@<hostname>_<port_bound_by_the_driver>

where the port_bound_by_the_driver is unique, but only within a container. I suspect that the namespace of Unix Domain Sockets starting with @ is actually shared across containers, hence the issues. @ironMann does it make sense? Do you know where to look for the documentation of the @ prefix?

[ironMann]

@ktf it seems that @ (abstract namespace for unix sockets) has poor support in containers: https://stackoverflow.com/questions/38455283/docker-containers-share-unix-abstract-socket-or-dbus

[ktf]

Ok, that explains the issue, then.

[rbx]

This is released with v1.4.49.