Are empty string values equivalent to unset attributes?

[Oberon00]

In open-telemetry/opentelemetry-java#771, opentelemetry-java decided to silently ignore calls to Span.setAttribute with empty string values (see also the brief discussion at open-telemetry/opentelemetry-java#765 (comment)).

I wonder what the specification position on this is. The semantic conventions for http.host say

When the header is empty or not present, this attribute should be the same.

-- https://github.com/open-telemetry/opentelemetry-specification/blob/cf0bb600b2eef359dbdd8fbe6f4696ed3e67ed1d/specification/data-http.md#common-attributes

(which is maybe a bit of an unfortunate formulation but is meant to say: "When the header is empty, this attribute should be empty. When the header is not present, this attribute should not be present.")

If we consider an empty string the same as not set, why don't we do the same for numeric values of zero or boolean values of false? 😉

I think we should add a note that explicitly states what to do with empty strings in SetAttribute (probably the same applies to metric labels and resource key values).

My personal opinion is that empty strings should be distinct from an unset attributes.

[Oberon00]

Depending on the outcome, we may need to separately discuss this for:

• null arrays
• null or empty strings inside arrays

See also open-telemetry/opentelemetry-java#807 (comment).

[jmacd]

This has been discussed in the context of metrics as well, with no great resolution. At present, the protocol spec says that empty values are considered the same as unspecified. I have a preference for treating them distinctly, but it implies a lot of changes in a lot of places and there wasn't a great deal of enthusiasm when this was discussed.

#345

[jmacd]

#341

[jmacd]

#345