See also discussion on #1412 (comment)

The sampling alorithm for TraceIdRatioBasedSampler is unspecified. As a result, trace IDs that are sampled by some implementations might get non-sampled or re-sampled by SDKs in other languages, even though they have the same or a a higher probability than the parent.

TODO list for this issue:

1. Is this a problem at all? Or is the ParentBased approach enough (in combination with any, not necessarily trace-id-based probability-based sampling of root spans)?
2. If it is a problem, since trace IDs can come from untrusted, non-random sources, do we open up a DDoS/Security/performance issue when using trace IDs as sole, deterministic input for our sampling algorithm? Do we need to put a warning there? Let's assume it is no problem for this issue, this should be handled in Support restarting the trace with a different trace ID #1188.
3. If applicable after checking the above we determine that we need a consistent algorithm, actually do specify it (maybe based on Add probability sampler details #331).