[Feedback][proto-v2]Proposal: Attribute Registry & Deduplication Support

[shubhamgptln]

In continuation of #13785 (Universal Attribute Deduplication), I would like to propose moving forward with a protocol-level extension to be considered for future upgrade.

The idea is to add optional OTLP field that allow exporters and receivers to exchange attributes using a registry + reference model instead of repeating the full attribute sets.

How it works:

• During batching, common attribute sets are deduplicated and stored in a per-batch registry.
• Spans/logs/metrics/profiles then reference the registry entries by ID.
• If both sides support the extension, they can negotiate this optimized path during connection setup.
• If the receiver does not support it, exporters fall back to sending the fully expanded attributes (backward compatible).

Pros:

• Provides both memory savings inside the collector and network savings across the wire.
• Clean separation between the optimized(new) and expanded(old already ecisting) paths, so older exporters/backends can ignore the new fields safely.
• Extensible beyond spans/logs to metrics and profiles in the future.

I believe this delivers the highest long-term impact compared to a memory only approach, while still keeping compatibility intact. Would be great to gather feedback on extending OTLP in this way and defining the negotiation mechanism

[aalexand]

Would this replace the dictionary approach that the profiles signal uses today? I guess not fully since this is specifically about attributes.

[tigrannajaryan]

This needs an OTEP, prototypes in multiple languages and in the Collector and participation from maintainers from multiple language SDKs and from Collector maintainers. I highly advise rallying the support of all these people first before starting the work.

If you are primarily interested in network traffic size savings take a look at STEF exporter and receiver that have built-in dictionaries (among many other size optimizations) and already exist. I am looking for contributors to enable remaining signals, stabilize and release the first version.