DB/messaging/rpc(?): should logical operations include network.* attributes

[lmolkova]

DB, messaging (maybe rpc and other conventions) arguably define logical operations.

E.g. database call could be a complex operation involving multiple underlying network calls (auth, retries, check-if-exists, etc).

• Assuming the underlying protocol is instrumented, it introduces duplication and confusion

• Some systems can operate on top of different interchangeable protocols and network information can be missing/too verbose/irrelevant:

  • GCP frequently operates on HTTP and gRPC
  • Azure messaging clients use AMQP, but to some extent can operate on top of HTTP
  • ORM systems don't even have to know the network details of underlying DB driver.

• Retries are ambiguous

  • retries can have different network.peer|local.address|port if they reach different nodes of the service

• Network calls within one logical call can use different protocols, e.g. in case of Azure CosmosDB:

  • Auth can be HTTP
  • Some other rare operations are HTTP
  • Things on the hot path are done over proprietary TCP-based protocol

Proposal:

Remove network.* attributes from generic db/messaging logical operations:

• this would make network.* attributes opt-in
• logical operations should still use 'logical' attributes such as server.address|port
• individual DB/messaging extensions can still reference/require/recommend network.* when it makes sense for them

Related: #674 (extracting db/messaging blocking part from that issue)

[pyohannes]

For messaging, this would affect the following attributes:

Attribute	Example	Requirement Level
network.peer.address	10.1.2.80; /tmp/my.sock	Recommended
network.peer.port	65123	Recommended: If network.peer.address is set.
network.protocol.name	amqp; mqtt	Conditionally Required: [15]
network.protocol.version	3.1.1	Recommended
network.transport	tcp; udp	Recommended
network.type	ipv4; ipv6	Recommended

It makes sense to make all the recommended ones Opt-In. For network.protocol.name, we currently have it as conditionally required for messaging systems that support more than one protocol.

However, we could remove this requirement from generic semantic conventions and add it to system-specific conventions where it makes sense. This would save us one more footnote ... 😄

[lmolkova]

Fixed in #768 and #780