/cc @Oberon00 I made the first stab at net.sock attributes.

One thing I didn't figure out yet is the relationship between net.transport and net.sock.family.

Transport currently is a closed set, address families are not (and different sets on win and linux).
Transport defines TCP and UDP, address family differentiates IPv4 and IPv6.
Transport has pipe and it's not a socket.

I'd propose the following on 'net.transport:

• remove or deprecate net.transport ip and unix values
• add socket value

Then net.sock.family can be used to understand if it's an IP-based protocol, Unix, or perhaps Bluetooth when net.transport = socket. Or if addr family is AF_INET6, net.transport explains if it tcp or udp.

LMK what you think.

I didn't have time to think about the net.transport question yet, but I noticed one thing:

There is no replacement for net.peer.name yet, and also the definition in the "net.*.name" section has not changed yet to address the logical peer change:

For net.peer.name, this should be the name of logical remote destination that was used to look up the IP address that was connected to (i.e., matching net.peer.ip if that one is set)

With the new definition, we don't know if we actually connect to the logical peer, I don't think a more precise definition than "logical peer" can be provided on the generic level, maybe just a note that semantic conventions that want this to be set SHOULD define more precisely what it means for them.

I also think that net.peer.ip does not make sense anymore, there should probably only be net.sock.peer.addr.

I also think that net.peer.ip does not make sense anymore, there should probably only be net.sock.peer.addr.

I think we wanted to use it instead of http.client_ip, but thinking more about it, perhaps reverse proxy and x-forwarded-for and friends are so HTTP-specific that http.client_ip would make more sense than generic net.peer.ip.

@Oberon00 I think I addressed your feedback. I'm not sure if renaming net.peer.name is necessary - even though we work on experimental spec, we should minimize breaking changes. Currently, the breaking change is minimal (net.peer.ip -> net.sock.peer.addr and net.host.ip -> net.sock.host.addr) and spec is clear on socket vs logical attributes plus proxy scenarios,

I'm not talking about renaming the existing net.peer.name, but about adding an additional attribute, e.g. net.sock.peer.name that would then correspond to net.sock.peer.addr (net.peer.name does not necessarily do that anymore).

@Oberon00 what would net.sock.peer.name mean - DNS name of net.sock.peer.addr ? Like proxy host name if one's available?

@Oberon00 what would net.sock.peer.name mean - DNS name of net.sock.peer.addr ? Like proxy host name if one's available?

Exactly. This is how I understood the net.peer.ip/name attributes originally. net.sock.peer.name would then be mostly IP-specific, just like net.peer.ip currently is.

As I understand it, this PR will cause the following "reshuffling":

• net.peer.name: Logical peer name, but not of the peer service (which is peer.service) but of something that could theoretically be "connected" to, typically a DNS-resolvable hostname (even if just from some private network), but might be something else. Depending on the address family used, a low-level 2non-logical" net.peer.name becomes net.peer.sock.addr or net.peer.sock.name.
• net.peer.ip -> net.peer.sock.addr, always, and only available for low-level peer, not for logical one if different.

@Oberon00 thank you for the feedback! I believe I addressed it and this PR now is ready for review.

@Oberon00 I believe I resolved your feedback, can you please take another look?

@trask thanks a lot for the feedback and the prototype!

@carlosalberto, can you please take another look? Thanks!

@open-telemetry/specs-approvers Last call for feedback.