Rename `service.peer.name`/`.namespace` to `server.service.name`/`.namespace` and `client.service.name`/`.namespace`?

[trask]

As discussed in the last Service and Deployment Semconv SIG meeting...

Motivation

service.peer.* can only describe one remote service at a time. That falls apart when observing a connection from the middle, where you want to capture both the upstream client service and the downstream server service on the same signal.

server.*/client.* already handle this nicely (e.g. server.address, client.address), with the nice property that each attribute means the same thing no matter who's observing.

Proposal

Deprecate service.peer.name/.namespace and replace with:

Current (deprecated)	Client-side replacement	Server-side replacement
service.peer.name	server.service.name	client.service.name
service.peer.namespace	server.service.namespace	client.service.namespace

Open question

Unlike client.address/server.address, where the meaning doesn't depend on observer position, service.name already exists as a resource attribute. So in practice we'd end up with:

• client.service.name / service.name (on a server span), or service.name / server.service.name (on a client span)

while from the middle it'd be:

• client.service.name / server.service.name

So we lose some of the symmetry that makes client.*/server.* nice.

cc @jsuereth @mmanciop @gkhom

[thompson-tomo]

For the observer use case I am wondering if we should also be adding the *.service.* attributes to the source.* & the destination.* namespaces that way we can potentially also include it when it is not known who the client is. Use case would be a firewall observer which is monitoring traffic and we could determine the service based on the ip/port.

[mmanciop]

@trask if I understand the proposal correctly, the headers of the table seem misleading to me, and I would rather write:

Current (deprecated)	Client-side replacement	Server-side replacement
service.peer.name	server.service.name	client.service.name
service.peer.namespace	server.service.namespace	client.service.namespace

Might be just me, but I read "Client-side" as "on the side of the CLIENT span".

[trask]

I would rather write

thanks, fixed!

[trask]

As discussed in SIG meeting, here are diagrams to help the discussion along:

This diagram of client.* and server.* attributes highlights something nice which is that they're the same on both sides.

While, service.* doesn't get that nice behavior:

That said, client.service.* and server.service.* does enable observing from the middle, e.g.

and it pairs nicely with our definition of "logical client" and "logical server":

• client.address
• client.service.name

and

• server.address
• server.service.name

and in Java instrumentation, we let users map server.address to the server's service name (also pointing to their pairing with client.* and server.*).