Tracing at Slack: Thinking in Causal Graphs

This post introduces the foundational concepts and motivations behind Slack's new tracing system, which models traces as Causal Graphs composed of SpanEvents. It highlights the limitations of traditional distributed tracing systems (request-centric APIs, inflexibility, overhead) and outlines the goals for Slack's system: usefulness beyond backend services, ease of use for simple cases, power for complex analysis, real-time triage capabilities, and queryable raw span data. The post contrasts the traditional trace model with the proposed causal graph model, emphasizing the benefits for ad hoc analysis and broader applicability.