Container Checkpoint/Restore at Scale for Fast Pod Startup Time
Talkat KubeCon + CloudNativeCon North America 2021 · on Wednesday, 2021-10-13· at 16:30· Room 502 AB + Online·
summary
MathWorks runs container checkpoint/restore in production since 2016 and reports roughly 70-fold improvement in cold-start latency, turning 30s app initializations into seconds by checkpointing warmed containers and restoring them. Implementation uses cryo + runc + Docker and adds a non‑native pair of microservices (checkpoint as a DaemonSet, restore as a Deployment) that store per‑node artifacts; scaling tests show restored pods come up ~30× faster and the system works at high scale with some runtime and kernel caveats.
AI-written summary · about
chapters (21)
- 0:00Intro
- 1:30During this talk...
- 1:53Our goal is to create a scalable system...
- 3:19But achieving all three goals is challenging
- 3:50Option 1: Scale out on-demand with usage
- 4:37Option 2: Create pre-warmed standby pool
- 5:20Is there a way to achieve all goals?
- 5:42Yes, we can!!! Checkpoint/Restore
- 6:33Checkpoint/Restore: Behind the Scenes
- 10:09Eliminating Cold Starts with Checkpoint Restore ..
- 10:30Kubernetes and Checkpoint/Restore
- 12:00Our Approach of Non-Native support
- 13:10Container Runtime in Container Runtime
- 14:20Checkpoint sequence
- 14:48Restore Sequence
- 15:05Cluster Architecture with CR
- 20:18Zero to Checkpoint Restore in Kubernetes
- 23:16Lessons Learned
- 25:19Best Practices
- 26:38Future Enhancements
- 27:36MathWorks is hiring....
related talks
34mLetting Your Containers Dream of Electric Sheep
27mState of Checkpoint/Restore in Kubernetes – Now With GPUs
39mContainer Live Migration
39mKubernetes and Checkpoint/Restore
08m⚡ Fast Restarts, Not Just Fast Starts: Accelerating Pod Recovery
34mWhere's My Pod, Already? A Tale of Cold-Start Struggles on Kubernetes
32mForensic Container Checkpointing
32mForensic container checkpointing and analysis