When Connections are Magic: Understanding Performance in Serverless
summary
Built a Go cloud function to simulate object-store transforms and instrumented TCP/TLS/DNS timings with httptrace, then ports an AWS X‑Ray hook to OpenTracing port to stitch network-level spans into traces. Finds cross‑cloud requests are 2–4x cross-cloud slowdown versus same‑cloud, discovers reuse depends on request frequency (reuse threshold around sub‑minute rates), spots 20–78 ms DNS, expensive TLS handshakes and Lambda CPU throttling at 128 MB, and hidden GCS OAuth round trips.
AI-written summary · about
chapters (8)
- 0:00Intro
- 7:26Connections are not magic
- 13:12Bimodal peaks
- 14:06observing systems
- 15:48what am I saying
- 16:20the most important lessons
- 18:40exercise
- 19:54audience questions
related talks
24mObservability across serverless asynchronous managed services
37mPerformance Tales of Serverless
30mPerformance Tales of Serverless
31mParting the Clouds: Tracing Cloud Foundry for Improved Performance...
25mWhere Did All My Servers Go?
43mApplying the Principles of Chaos to Serverless
47mAdvanced Serverless Architectural Patterns on AWS
44mApplying principles of chaos engineering to serverless