OpenTelemetry tracer - Align attributes with OTel HTTP semantic conventions

[joaopgrassi]

Title: Align Envoy span attributes with the OTel HTTP semantic conventions

Description:

The OpenTelemetry semantic conventions for HTTP were marked stable in: open-telemetry/semantic-conventions#377.

These conventions were adapted and reviewed by the community and subject-matter experts. As of today, the spans produced by Envoy are of course not aligned with the OTel conventions. This issue is to discuss if the Envoy community wants to align with them.

Pros:

• Standardization of attributes which provide users a consistent observability experience
• Observability back-ends can rely on such conventions, also giving a good and consistent experience for users

Cons:

• It is a breaking change for Envoy users. The existing attributes may be used for alerts, dashboards and etc
• Envoy records the attributes in a lot of places, and by changing those not only users using the OTel tracer will be impacted but users that use the other available tracers in Envoy, such as XRay, Datadog and etc.

To avoid the hard breaking change, maybe we can have some sort of "opt-in" configuration where users then opt in and Envoy starts producing spans aligned with the OTel HTTP semantic conventions. This allow them to gradually migrated to the new format.

[optional Relevant Links:]

• OTel HTTP semantic conventions for spans

[htuch]

+1 on not breaking existing users and providing this as an opt-in capability. Are you planning on picking up this work @joaopgrassi?

[joaopgrassi]

Hi @htuch I'd love to, for now I'm quite busy with the other components still in Envoy (sampler) and also trying to add the new things to Istio.

[htuch]

Ack, will make as help wanted.

[a-thaler]

Another inconsistency with the OTel semantics is, that Envoy tracers are adding an error attribute/label/tag to the span in case of an error (see docu). While the attribute is not harmful itself (besides that it does not follow naming conventions), the opentelemetry tracer should additionally set the span "status" to "error" in these situations, as outlined in the spec. That behaviour should be added for the opentelemetry tracer only and that might be even a non-breaking change. Could that missing behaviour be covered in the context of this ticket or is a dedicated ticket useful, as it is not exactly about the attribute naming?

[joaopgrassi]

Changing the status to error seems to me something that can be done first, without waiting for this.