Letting Your Containers Dream of Electric Sheep
summary
A Kubernetes runtime snapshots whole containers to disk when idle and restores them in tens of milliseconds, preserving in‑memory state so apps (demo: WordPress + MySQL) resume without reinitializing. Implementation plugs into containerd via a custom shim, uses CRIU checkpoint/restore, an eBPF kernel redirector to accept connections during restore, and a small go proxy to finish the handoff. Scale‑down timeout is configurable (demo used 10s), works for HTTP and raw TCP, and fails only on some exotic io_uring/node libs; single‑developer side project.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 1:31Project goals and background
- 2:32Analysis of existing solutions
- 5:20Architecture overview
- 6:40Live demonstration
- 12:24Key features and use cases
- 15:14Technical deep dive
- 19:33In-place resource scaling
- 20:19Future outlook and conclusions
- 22:48Question and answer session
related talks
26mKubernetes and Checkpoint Restore
28mContainer Checkpoint/Restore at Scale for Fast Pod Startup Time
39mContainer Live Migration
39mKubernetes and Checkpoint/Restore
39mMore Than Secure: Containerd + KataContainers as Kubernetes Runtime
32mForensic container checkpointing and analysis
13mLightning Talk: Containers Live Migration: What’s There, What’s Missing, What’s Next?
32mForensic Container Checkpointing