AOT Caching - Netflix' Practice vs OpenJDK's Theory
summary
Describes a concrete pipeline: run long canary deployments to collect runtime profiles, store the resulting AOT artifacts in S3 and inject them at startup to short‑circuit class loading, linking and profile gathering, with JDK tooling (jcmd finish‑recording, Xlog/JFR) for capture and inspection. Netflix reports multi‑hour canaries produce the best gains; lightweight alternatives (touch every endpoint once) still cut class‑load costs. JDK already caches tiny stubs in 25, and full AOT codegen arrives later with portability knobs for arch=native vs portable and canary deployment to collect training data stored as AOT cache in S3.
AI-written summary · about
chapters (11)
- 0:00Intro
- 1:21Project Leyden History & Progression
- 6:00Netflix Push to Faster Startup
- 8:53The Importance of Training Runs
- 17:13Time Scales
- 24:22Checkpointing
- 34:21The Linear Workflow
- 39:46Managing AOT Caches
- 48:28Inspecting the AOT Cache
- 53:11AOT Code Compliation
- 1:08:26Just Chatting
