Tracing Is Dead: How to Maximize Results with Logging and Metrics
My Monitorama PDX 2022 talk — a deliberately provocative case that distributed tracing kept failing to earn its keep, and what to do instead.
🎙 Monitorama PDX 2022 · Portland, OR
The provocation
I walked onto the Monitorama stage with a deliberately impolite title: tracing is dead. Not because spans and waterfalls aren’t useful — but because, in team after team I met, distributed tracing had become a project instead of a tool. Months of instrumentation work across polyglot services, arguments about sampling, dashboards nobody trusted — all before a single production question got answered.
My argument was simple: if your observability requires you to predict every question in advance and instrument for it, you’ve rebuilt the same trap as over-logging, just with fancier UI.
Pack light, but pack context
The alternative I made the case for will sound familiar if you’ve read my writing: stop trying to collect everything up front. Instead, make the telemetry you already emit — structured logs and metrics — wide and context-rich: request IDs, customer and business context, high-cardinality fields that let you slice by what actually matters when something breaks.
A well-structured log event with the right context answers most production questions on demand. Metrics cover the known-unknowns — the trends, the burn rates, the “is this normal?” checks. Together, they get working engineers most of the value, at a fraction of the adoption cost.
The honest footnote
The observability community pushed back, fairly: take those wide, structured log events, correlate them with shared IDs — and congratulations, you’ve reinvented most of distributed tracing. The consensus since has been to make tracing easier to adopt rather than abandon it, and OpenTelemetry has genuinely lowered that barrier.
I still think the talk’s question was the right one to ask: is this earning its keep for your team, right now? Any observability investment should be judged the way I judge R&D deliverables — by whether it creates value for the people debugging at 2am, not by architectural elegance.
Watch it
The full session is above — about half an hour, including the Q&A where the tracing loyalists (lovingly) came for me. If it sparks a disagreement, good. That’s what the title was for.