Hyperscale Kernel Live Patching at Meta
summary
Meta deploys kernel live patching as RPMs with automated health checks and cumulative patch atomic swaps using the replace flag. Applying a KLP takes 1–2 seconds per host and the fleet rollout now typically completes in a week (hours in extreme cases) across millions of servers. Encountered failures include ksoftirqd starvation from insmod CPU hogging (fixed by adding cond_resched and upstreamed), ftrace/BPF IP-modify conflicts (mitigated in newer kernels), and task transition edge cases that require tracing and upstream fixes.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:29Full reboot
- 1:59Pros and cons
- 3:14KExec
- 3:56Pros Cons
- 4:46klp
- 10:14Ecosystem
- 16:40Challenges
- 17:15Bugs
- 18:01Problem
- 20:35Tracing
related talks
33mLinux Kernel Live Patching
34mLivepatch Visibility at Scale
24mKLP for Clang LTO Kernel
48mAutomated patch management with Ansible and Rundeck
15mKVM Live Upgrade with Properly Handling of Passthrough Devices
24mContinuously Integrating the Upstream Linux Kernel on Hardware
33mHow We Learned to Fix Planes Midflight
27mAn SRE guide to Linux Kernel upgrades
