eBPF for Observability: Hard Truths and Practical Wins
Talkat Open Source Summit Japan + AI_dev: Open Source GenAI & ML Summit + Automotive Linux Summit 2025 · 2025-12-08..12-10·
summary
eBPF runs dynamic code in the kernel to power observability and even custom scheduling, and teams report concrete wins like lower scheduling overhead and usable BPF timers vs sleepable context trade-offs, but also hard truths: frequent tracepoint and kfunc drift across kernel versions, opaque verifier errors, and distribution/version compatibility that break tools built against single kernels. Operational tooling includes BPFtop for runtime cost/latency and a web verifier visualizer to decode rejection traces; libbpf-tools works but needs careful vetting.
AI-written summary · about
chapters (12)
- 0:00eBPF introduction
- 1:08BPF Japan basics
- 2:51SchedExt deep dive
- 4:40eBPF as platform
- 7:54eBPF production truths
- 10:03BPF tools vs Ftrace
- 15:52eBPF program observability
- 19:28Engineer training
- 21:35kfuncs stability
- 29:02Verifier error debugging
- 33:15Probes vs USDT
- 37:20Q&A
related talks
31meBPF for Observability: The Good, the Bad, and the Ugly
40mPanel Discussion: eBPF: A New Era in Cloud Infrastructure Tools
20mClosing Perspectives
32meBPF: The Next Power Tool of SREs
32mExploring the Power of eBPF for System Observability
35mPerformance Insights into eBPF Step by Step
39meBPF BoF
36mObservability and performance analysis with BPF