Lightning Talk: dial9: WTF is Tokio actually doing?
summary
A tiny production-ready tracer called Dialine records Tokio poll/wake/park events, kernel stack traces, CPU samples and tracing spans into a compressed binary format parsable in ~100 lines of JavaScript. It shows concrete failures: kernel scheduling delay stalls of ~18 milliseconds when 48 workers contend with 16,000 threads, a file-descriptor resize causing 100 ms polls, and Mini Redis P99 regressions that an automated agent fixes with measurable improvement. Recording costs 25 nanoseconds, 5-10 bytes and can produce 100 GB per host per day.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:59Tokio metrics limitations
- 3:43Dial9 production goals
- 5:06Production trace analysis
- 7:59Kernel delays and FD issues
- 9:01Programmatic dial9 traces
- 10:05Mini Redis P99 fix
- 11:31Performance impact and deployment
- 12:14Integrating dial9
related talks
05m“Using async-profiler to Find Tail Latency Issues”
59mLightning Talks
46mLightning Talks
12mLightning Talk: With Great Traces Comes Great Costs: How to Reduce That Bill?
38mTechniques and Tools for a Coherent Discussion about Performance in Complex...- 12mLightning Talk: With Great Traces Comes Great Costs: How to Reduce That Bill?
10m⚡ Lightning Talk: Unlocking Customer-Centric Observability: A Case Study of OpenTelemetry To Reduce MTTD < 3 Mins
05mLightning Talk: That Small Difference Made all the Difference