Hall of FameBenjamin H. Sigelman et al.201043 min readpaperadvanced
Dapper, a Large-Scale Distributed Systems Tracing Infrastructure
Summary
Dapper is Google’s production‑grade distributed tracing system that achieves low overhead and ubiquitous deployment by instrumenting only a few core libraries and using adaptive sampling. The paper details its data model, sampling strategy, and the ecosystem of analysis tools built on top of it.
- Adaptive sampling (≈1 in 1,000 requests) keeps overhead sub‑1% while still providing useful traces.
- Instrumentation is confined to common threading, control‑flow, and RPC libraries, making tracing transparent to application code.
- Spans are identified by probabilistically unique 64‑bit IDs and organized into tree structures for end‑to‑end request reconstruction.
- Trace data is ingested and made queryable within a minute, enabling rapid production incident diagnosis.
Teams building large‑scale services can adopt Dapper’s design principles to get low‑impact, always‑on tracing without burdening developers.
8/10
