OpenTelemetry Demo 3.0 observes agents — copy patterns, not topology
The new demo combines a ReAct agent, MCP, profiling, and OpAMP and renames `app.*` attributes to `demo.*`. It is a living reference, not a production-ready blueprint.
- Nature
- Open-source project update with operational analysis
- Release
- Demo 3.0 described on 28 July 2026
- Scope
- Teams using or adapting OpenTelemetry Demo and GenAI semconv
- Limit
- The demo does not establish production architecture or privacy policy
In brief
OpenTelemetry describes 3.0 as an intentional break: it renames attributes, changes Compose files, adds an observable agentic stack, continuous profiling, OpAMP, and tests. Its value is studying correlation across traces, metrics, logs, and profiles. In production, each attribute, tool payload, prompt, token, and identifier needs review for purpose, cardinality, cost, privacy, and retention.
- Treat `app.*` → `demo.*` as a schema migration that breaks dashboards and queries.
- Use the demo to validate instrumentation, not inherit credentials, singletons, or operational defaults.
- Define which tool arguments and responses may be collected; opt-in does not mean safe.
- Test sampling with rare failures, agent loops, and end-to-end correlation before reducing volume.
The most immediate break is the attribute contract
All custom attributes moved from `app.*` to `demo.*`, breaking existing dashboards and queries. The release adds agent, MCP, and backend services plus profiling and OpAMP. The project describes itself as a living reference implementation.
Agentic instrumentation can observe tool choice, latency, tokens, and flow, but tool arguments and responses can contain personal data, secrets, or customer content. Semantic conventions define names; they do not decide what your organisation should collect.
Adopt without blind copying
- INVENTORYDemo dependencies
Forks, dashboards, queries, and instructions using old names are found.
- CONTRACTOwned schema
Allowed attributes, cardinality, redaction, and retention are approved.
- LOADSignals under stress
Sampling, buffers, loss, and cost are tested with failures.
- PRODUCTIONOperable topology
HA, access, upgrades, and incidents belong to the real architecture.
Pass every signal through a utility and risk gate
For every field, name the operational question it answers. If a full prompt does not support an authorised alert, debug, or audit need, do not collect it by default. Prefer correlatable identifiers, categories, and hashes when raw content is unnecessary.
Cardinality and sampling need tests. Unbounded IDs in metrics can explode cost; head sampling can miss rare errors; tail sampling needs memory and time. Make decisions by signal and scenario, not one global telemetry switch.
Telemetry contract
| Signal | Question | Risk | Gate |
|---|---|---|---|
| Trace | Where did flow fail? | Sensitive payload | Redaction and access |
| Metric | Is SLO degrading? | Cardinality | Per-label budget |
| Log | Which event occurred? | Secrets/PII | Schema and retention |
| Profile | Where are resources used? | Overhead | Sampling |
| Config | Which Collector runs? | Remote change | Auth and rollout |
Migrate the fork as a schema and use the demo for contract tests
Create temporary compatibility or update queries and dashboards atomically for the rename. Run demo tests, then add pipeline cases: exporter failure, Collector outage, tool loop, sensitive payload, and adversarial cardinality.
Document which components are laboratory only. Production needs an owner, SLO, authentication, segregation, configuration backup, and rollback. A working demo proves neither capacity nor compliance.
- Do not store full prompts and arguments by default.
- Do not use high-cardinality IDs in metrics.
- Do not copy a single Collector into critical production.
- Do not rename without inventorying downstream consumers.
Sources consulted
Consulted on the stated update date.
Opens in a new window.Participate
Share the analysis. Add context.
Distribute the article through your channel or leave an observation connected to the topic.
Comments
Loading the conversation.