Missing spec for Span ID creation for dropped/record-only spans

[Oberon00]

What are you trying to achieve?

I am trying to ensure that span ID creation for spans that are dropped or only recorded (but not sampled) is uniform across different language implementations.

Additional context.

As explained in #864 (comment) the spec is silent on what to do for IDs of spans that are not sampled. #998 suggest to reuse the parent-span-id for dropped or use zero if there is none, but use a new span ID for everything else. A simpler alternative would be to simply add the sentence "A new Span ID MUST always be generated, even for dropped spans".

This is not purely an optimization issue, it impacts the observable behavior of the monitored application when the unsampled span is injected (and then potentially used as parent on the other side).

[Oberon00]

CC @alolita @anuraaga

[marcingrzejszczak]

Is this issue also related to being able to drop a span whatsover? That means that you're starting a span and after some time you decide not to record it. You want to end it but not report it anywhere. Currently I don't see that the spec allows it nor the java implementation.

[Oberon00]

Is this issue also related to being able to drop a span whatsover? That means that you're starting a span and after some time you decide not to record it.

No, not at all. This issue is only about the reaction to the existing (head-based) DROP sampling decision. If you want to discuss aborting/dropping a span later, please open a new issue. EDIT: Or look at open-telemetry/oteps#115

[andrewhsu]

from the spec sig mtg today, re-triaging as required-for-ga