Attribute Requirement Levels for Semantic Conventions

Status: Stable

This section applies to Log, Metric, Resource, and Span, and describes requirement levels for attributes defined in semantic conventions.

Attribute requirement levels apply to the instrumentation library.

The following attribute requirement levels are specified:

The requirement level for an attribute is specified by semantic conventions depending on attribute availability across instrumented entities, performance, security, and other factors. When specifying requirement levels, a semantic convention MUST take into account signal-specific requirements.

For example, Metric attributes that may have high cardinality can only be defined with Opt-In level.

A semantic convention that refers to an attribute from another semantic convention MAY modify the requirement level within its own scope. Otherwise, requirement level from the referred semantic convention applies.

For example, Database semantic convention references network.transport attribute defined in General attributes with Conditionally Required level on it.

Required

All instrumentations MUST populate the attribute. A semantic convention defining a Required attribute expects an absolute majority of instrumentation libraries and applications are able to efficiently retrieve and populate it, and can additionally meet requirements for cardinality, security, and any others specific to the signal defined by the convention. http.request.method is an example of a Required attribute.

Note: Consumers of telemetry can detect if a telemetry item follows a specific semantic convention by checking for the presence of a Required attribute defined by such convention. For example, the presence of the db.system attribute on a span can be used as an indication that the span follows database semantics.

Conditionally Required

All instrumentations MUST populate the attribute when the given condition is satisfied. The semantic convention of a Conditionally Required attribute MUST clarify the condition under which the attribute is to be populated.

http.route is an example of a conditionally required attribute that is populated when the instrumented HTTP framework provides route information for the instrumented request. Some low-level HTTP server implementations do not support routing and corresponding instrumentations can’t populate the attribute.

When a Conditionally Required attribute’s condition is not satisfied, and there is no requirement to populate the attribute, semantic conventions MAY provide special instructions on how to handle it. If no instructions are given and if instrumentation can populate the attribute, instrumentation SHOULD use the Opt-In requirement level on the attribute.

For example, server.address is Conditionally Required by the Database convention when available. When server.socket.address is available instead, instrumentation can do a DNS lookup, cache and populate server.address, but only if the user explicitly enables the instrumentation to do so, considering the performance issues that DNS lookups introduce.

Instrumentations SHOULD add the attribute by default if it’s readily available and can be efficiently populated. Instrumentations MAY offer a configuration option to disable Recommended attributes.

Instrumentations that decide not to populate Recommended attributes due to performance, security, privacy, or other consideration by default, SHOULD allow for users to opt-in to emit them as defined for the Opt-In requirement level (if the attributes are logically applicable).

Opt-In

Instrumentations SHOULD populate the attribute if and only if the user configures the instrumentation to do so. Instrumentation that doesn’t support configuration MUST NOT populate Opt-In attributes.

This attribute requirement level is recommended for attributes that are particularly expensive to retrieve or might pose a security or privacy risk. These should therefore only be enabled explicitly by a user making an informed decision.

Performance suggestions

Here are several examples of expensive operations to be avoided by default:

  • DNS lookups to populate server.address when only an IP address is available to the instrumentation. Caching lookup results does not solve the issue for all possible cases and should be avoided by default too.
  • forcing an http.route calculation before the HTTP framework calculates it
  • reading response stream to find http.response.body.size when Content-Length header is not available