# OpenTelemetry Graduation — Live Stream Prep Prep for the OpenTelemetry community call live stream on OTel's CNCF graduation. Includes the graduation context you need (assuming zero prior knowledge), plus talking points and interview questions with a focus on GenAI semantic conventions. --- ## The headline news On **May 21, 2026**, the CNCF announced that **OpenTelemetry has officially graduated** — the highest maturity tier a CNCF project can reach. (The Technical Oversight Committee moved it to Graduated on May 11; the public announcement landed May 21.) The framing in every announcement is the same: graduation "solidifies OpenTelemetry's status as the de facto observability standard." OTel is now the **second CNCF project ever to graduate after Kubernetes** in terms of significance/velocity, and it's the second-highest-velocity project in the entire cloud native ecosystem (out of 240+ projects), behind only Kubernetes. --- ## What "graduation" actually means (CNCF 101) CNCF projects move through three maturity levels. The analogy CNCF itself uses is the "Crossing the Chasm" adoption curve: | Level | Adoption stage | Meaning | |---|---|---| | **Sandbox** | Innovators | Early, experimental, innovative projects. Entry point. | | **Incubating** | Early adopters | Gaining adoption, stabilizing, fewer breaking changes, versioned APIs. | | **Graduated** | Early majority | Mature, robust, proven in production by many organizations. The top tier. | | *(Archived)* | — | Retired / no longer recommended. | How a project actually graduates: - A project applies, and the **Technical Oversight Committee (TOC)** runs a process called **"Due Diligence"** to confirm the project genuinely meets the criteria — not just on paper. - Graduation requires a **two-thirds supermajority vote** of the TOC. - Projects are generally expected to graduate within ~2 years of incubating (though they can stay incubating indefinitely). What OTel specifically had to demonstrate to graduate: - **Third-party independent security audits** of core components (including the OpenTelemetry Collector), plus a fuzzing audit. - A **formal governance review** confirming the project's governance is mature and sustainable. - Demonstrated **production adoption at scale** and committer/contributor diversity across many companies (no single-vendor control). Why it matters in plain terms: graduation is CNCF's signal to enterprises that the project is safe to bet on — stable, well-governed, vendor-neutral, and not going anywhere. For OTel, it's validation of the "vendor-neutral standard" promise. --- ## Quick OTel backstory (for context / framing) - OpenTelemetry was formed in 2019 from the **merger of OpenTracing and OpenCensus** — two competing projects that joined forces. Solving fragmentation is literally in its origin story. - Core value prop: one open, vendor-neutral standard for **traces, metrics, and logs** (and now **profiles**), so you can instrument once and switch analysis backends without rewriting code. No lock-in. - Scale today: **12,000+ contributors from 2,800+ companies**; ~7 years old; second-highest project velocity in CNCF after Kubernetes. Recent/adjacent milestones worth a name-drop: - **Profiles** — the fourth signal — entered **public alpha**. - New **Kotlin SDK**; **OpenTelemetry eBPF Instrumentation (OBI)** shipped its first release. - Expansion into **CI/CD observability** and a stabilized **declarative configuration**. --- ## The AI angle — GenAI semantic conventions (your main focus) This is the part most relevant to where observability is heading, and it's the richest material for the conversation. ### What "semantic conventions" are Semantic conventions are OTel's **shared vocabulary** — agreed-upon names and shapes for telemetry attributes so that data from different tools/vendors is comparable and portable. They're arguably as important as the SDKs: the standard is only useful if everyone names things the same way. (HTTP, database, Kubernetes, RPC conventions have all gone through stabilization.) ### The GenAI conventions specifically The **Semantic Conventions for Generative AI** standardize how LLM and AI-agent operations are recorded. They're developed by the **GenAI Special Interest Group (SIG)** / GenAI Semantic Conventions and Instrumentation SIG. They're **still experimental and under active development** — an important honest caveat to raise. They cover several areas: - **LLM / model conventions** — the model call itself. - **AI agent conventions** — multi-step agent behavior (initially based on Google's AI Agent whitepaper). - **Vector database conventions** — the retrieval side of RAG. ### Concrete attributes (great for making it tangible on stream) When you instrument an LLM-powered app, you get spans and metrics with standardized names: Spans (the trace tree): a top-level `invoke_agent` span, with child `chat` spans per LLM call and `execute_tool` spans per tool invocation. Key attributes: - `gen_ai.request.model` — which model (e.g. `gpt-4o`) - `gen_ai.usage.input_tokens` / `gen_ai.usage.output_tokens` — token counts - `gen_ai.response.finish_reasons` — why generation stopped (`stop`, `tool_calls`, etc.) Metrics: - `gen_ai.client.operation.duration` — latency histogram (filterable by model) - `gen_ai.client.token.usage` — token consumption (split input vs output) These let teams estimate per-request cost, catch token-hungry prompts before production, detect latency regressions, and compare models — all in vendor-neutral form. ### Privacy by default (a strong, quotable point) By default, **no prompt content or tool arguments are captured** — only metadata (model names, token counts, durations). Capturing the actual prompts/completions/tool calls is **opt-in** (`captureContent`), because that content can be sensitive. When enabled, content shows up as structured attributes like `gen_ai.input.messages`, `gen_ai.output.messages`, and `gen_ai.system_instructions`. This "privacy by default, opt in to content" design is a thoughtful answer to a real concern. ### Why this is a big deal right now - 2025–26 is "the year of AI agents." Agents are **non-deterministic**, so telemetry isn't just for debugging — it becomes a **feedback loop for evaluating and improving the agent**. - The GenAI/LLM observability space is crowded with vendor- and framework-specific formats (Traceloop/OpenLLMetry, Langtrace, etc.). OTel conventions are the path **away from lock-in** — same argument as OTel's origin, now applied to AI. - Adoption is already broad: **Google Cloud, AWS, Azure, Datadog**, and others support the GenAI conventions. **OpenLLMetry's conventions were donated into OpenTelemetry.** Even coding assistants now emit OTel GenAI telemetry — **VS Code Copilot, OpenAI Codex, and Claude Code** all export GenAI traces/metrics/events. - Honest limits to acknowledge: the conventions standardize the **mechanics** (model, tokens, latency, content) but **do not yet cover output evaluation, safety scoring, or content-quality assessment**. That's the frontier. ### Grafana tie-in (for your own positioning) Grafana is a major part of the OTel ecosystem (OTLP-native ingestion, the LGTM stack, etc.) and the broader "store telemetry in open standards, query it anywhere" story. If you want a local hook, the GenAI conventions mean LLM telemetry can flow into the same open backends as the rest of your stack rather than a siloed AI-monitoring tool. --- ## Talking points (things you can say / volunteer) - "Graduation is CNCF's top maturity tier — the same level as Kubernetes — and it's basically a stamp that says 'enterprises can safely bet on this.'" - "OTel literally started by merging two competing projects, OpenTracing and OpenCensus. Its whole reason for existing is killing fragmentation — and now it's doing the exact same thing for AI." - "The thing people underrate is the semantic conventions. The SDKs get attention, but the shared vocabulary is what makes the data portable. That's the real moat." - "For GenAI, the privacy-by-default design is smart — you get model, token, and latency data automatically, and capturing actual prompts and responses is strictly opt-in." - "It's worth being honest that the GenAI conventions are still experimental — they standardize the mechanics, but evaluation and safety scoring aren't covered yet. That's where the community input matters most." ## Interview questions (ordered, with the AI ones front-loaded as your priority) On graduation itself: 1. For people who don't follow CNCF closely — what does graduation actually change, day to day, for OTel and for the people using it? 2. What was the hardest part of the graduation process? Was it the security audits, the governance review, the production-adoption bar? 3. Graduation is described as "not the finish line." What's the next big bet for the project? On the AI / GenAI conventions (your focus): 4. Walk me through the GenAI semantic conventions — what do they let you see inside an LLM call that you couldn't before? 5. AI agents are non-deterministic. How does that change what observability even means — are we monitoring, or are we evaluating? 6. The conventions are still experimental. What's the hardest thing to standardize about GenAI right now, and what are the open debates in the SIG? 7. The conventions capture model, tokens, and latency — but not output quality or safety scoring yet. Is that on the roadmap, or is that intentionally out of scope for OTel? 8. There were already vendor and framework formats — OpenLLMetry, Langtrace, and others. How did OTel get buy-in to converge on one standard, and is the fragmentation actually resolving? 9. Privacy-by-default with opt-in content capture — how did the community land on that design, and how do you think about the sensitive-data risk? 10. Coding assistants like Copilot, Codex, and Claude Code now emit OTel GenAI telemetry. Where do you see this going — observability for every AI-powered app by default? On community (good closer): 11. With 12,000+ contributors across 2,800+ companies, how do you keep a vendor-neutral project actually neutral? How are contentious technical decisions made? 12. For someone watching who wants to get involved in the GenAI SIG specifically — what's the on-ramp? --- ## One-line cheat sheet (if you blank on air) Graduated May 21, 2026 · top CNCF tier · 12k+ contributors / 2,800+ companies · #2 velocity after Kubernetes · required security audits + governance review · GenAI conventions = experimental, vendor-neutral standard for LLM/agent telemetry · `gen_ai.*` attributes · privacy by default, opt-in content · doesn't yet cover eval/safety scoring. --- ### Sources - [CNCF Announces OpenTelemetry's Graduation](https://www.cncf.io/announcements/2026/05/21/cloud-native-computing-foundation-announces-opentelemetrys-graduation-solidifying-status-as-the-de-facto-observability-standard/) - [OpenTelemetry is a CNCF Graduated Project (OTel blog)](https://opentelemetry.io/blog/2026/otel-graduates/) - [CNCF Project Lifecycle and Process](https://contribute.cncf.io/projects/lifecycle/) - [CNCF Graduation Criteria](https://github.com/cncf/toc/blob/main/process/graduation_criteria.md) - [Inside the LLM Call: GenAI Observability with OpenTelemetry](https://opentelemetry.io/blog/2026/genai-observability/) - [AI Agent Observability — Evolving Standards and Best Practices](https://opentelemetry.io/blog/2025/ai-agent-observability/) - [OpenTelemetry GenAI Semantic Conventions (spec)](https://opentelemetry.io/docs/specs/semconv/gen-ai/)