Benchmarking JDK 25 Leyden AOT on GKE: Cutting Spring Boot Pod Readiness by 31%–58%
programming.dev
Benchmarking JDK 25 Leyden AOT on GKE: Cutting Spring Boot Pod Readiness by 31%–58%
Everyone talks about JDK 25 Project Leyden AOT caches locally—almost nobody measures what happens when Kubernetes pods actually have to download a 200+ MiB archive at scale. Across four Spring Boot tiers tested on GKE with mnemo-warmup (from a 37-JAR microservice to a 458-JAR monolith): - Single-pod readiness dropped 31%–58%: Minimal services fell from 4.5s → 2.0s (-56%) and our 458-JAR monolith dropped from 30.8s → 21.2s (-31%). - Large archive fetches stayed fast: Pulling and SHA-256 verifying a 223.5 MiB archive from same-region Artifact Registry took 0.7s (saving 9.6s). - HPA 1→9 scale-out bursts: Median per-pod readiness improved by 22%–49% under load. - Zero code changes: Caches train asynchronously in-cluster via a CronJob and fail open cleanly to a standard JVM start if a registry stalls. * Benchmark write-up: Full tables, methodology & FAQ [https://baokhang.hashnode.dev/31-to-58-faster-java-pods-on-kubernetes-benchmarking-jdk-25-leyden-aot-from-microservices-to-a-458-jar-monolith] * GitHub repo: baokhang83/mnemo-warmup [https://github.com/baokhang83/mnemo-warmup] What is your real-world setup for Java startup on K8s? Vanilla JVM, build-time AppCDS, or experimenting with JDK 25 Leyden?
0 comments
No comments yet.