The signal everyone finally agreed to share
In 2007, Google’s engineers could watch a single search query spawn hundreds of downstream calls across dozens of services — to spelling-correction systems, index servers, ad auctions, ranking models — and still have no reliable way to know which one was slow. They had logs. What they lacked was a thread to pull. Dapper was the internal system they built to follow a request through that maze, and in April 2010, Benjamin Sigelman and his colleagues published a technical report describing how it worked. They published everything except the code.
Sigelman eventually left Google to co-found Lightstep, a commercial tracing company. In 2016 he co-created OpenTracing under the Cloud Native Computing Foundation — a deliberately thin specification, just the API, designed to let any tracing implementation plug into the same instrumentation code without locking engineers to any vendor’s collection agent. Separately, Google released its own approach in 2018: OpenCensus, broader than tracing alone, with metric collection built in and auto-instrumentation agents that could wrap a JVM or Node process without touching application code. Two open standards for the same problem is a competition, not a foundation. Zipkin (Twitter), Jaeger (Uber), and a dozen commercial vendors had to choose sides.
The stand-off resolved in May 2019, when both communities agreed to merge. The result, OpenTelemetry, combined OpenTracing’s clean API/SDK separation with OpenCensus’s breadth across signal types. It committed to standardizing all three pillars of observability: traces for reconstructing request paths across service boundaries, metrics for measuring system behavior over time, and logs for recording discrete events. One collector, one export protocol, one set of semantic conventions for naming what was being measured.
What that specification changed took a few years to become visible. Before OpenTelemetry, every monitoring vendor competed partly on collection: who had the cleverest agent, the lowest overhead tap, the most languages supported. Their proprietary formats were the moat. Michele Mancioppi, chief architect at Dash0, expressed the shift plainly at the graduation ceremony: “Instead of competing on who has the best telemetry, vendors now compete on who does the best with the telemetry everybody has.” The wire had become shared property. The competitive layer moved upstack, to analysis and insight, where it belonged.
On May 21, 2026, at the Observability Summit in Minneapolis, the CNCF announced that OpenTelemetry had graduated to its highest maturity level — the same tier as Kubernetes and Prometheus. Graduation required completing a third-party security audit and a formal governance review. The project had more than 12,000 contributors from 2,800 companies; its JavaScript package had been downloaded 1.36 billion times in the preceding year, the Python package 1.3 billion. Only Kubernetes moves faster inside the CNCF ecosystem.
Sigelman published the Dapper report in 2010. The paper explained the problem clearly enough that a dozen teams went home and built their own incompatible solutions. OpenTelemetry ended that tradition. The question of how a distributed system tells you what it is doing now has one answer — and that answer, like most infrastructure, will soon be taken entirely for granted.
Sources
- Dapper, a Large-Scale Distributed Systems Tracing Infrastructure — Google Research, April 2010 — the paper by Sigelman et al. describing Google’s internal distributed tracing system, which established the vocabulary and mental model for the entire field
- CNCF Announces OpenTelemetry’s Graduation — May 21, 2026 — official graduation announcement with contributor statistics, the Mancioppi quote, and background on graduation requirements
- OpenTelemetry Graduates — opentelemetry.io, 2026 — the project’s own announcement tracing the journey from the 2019 OpenTracing/OpenCensus merger to CNCF graduation
- OpenTelemetry — Wikipedia — project history, technical architecture, and signal coverage timeline