Blazing-Fast Container Deployments: Image Caching With Block Storage
Talkat Data on Kubernetes Day NA 2025 · on Monday, 2025-11-10· at 13:35· Building B | Level 2 | B211-212·
summary
Shows a concrete system that caches popular base images as virtual block device cache served from zone storage and plugs into containerd via a custom snapshot over overlayfs process. Deployed across GeForce Now zones and integrated with NVCF, it avoids unpacking on each node, yielding roughly 0.7x latency for large images in tests, about a 3x slower cold path without caching, and a reported production improvement of ~52%+ reduction in deployment time.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:47Containerd image unpacking
- 3:56High latency from layers
- 4:38Node isolation scaling
- 5:42Custom snapshotter with block storage
- 6:47Experiment results
- 8:11GeForce Now case study
- 10:41Base image distribution
- 13:52Worker node setup
- 15:36Performance gains
- 17:40Cluster limitations
- 19:55Q&A: Storage backend
- 21:17Q&A: Virtual devices
- 22:41Q&A: Snapshotter availability
- 23:24Q&A: File system vs block storage
- 24:09Q&A: Cache update frequency
related talks
21mDADI: Block-Level Image Service for Agile and Elastic Application Deployment
19mStarlight: Fast Container Provisioning on the Edge and over the WAN
16mDADI @ Scale: Deploying Containers at Scale
18mPoby: SmartNIC-accelerated Image Provisioning for Coldstart in Clouds
30mSpeeding Up Analysis Pipelines with Remote Container Images
35mThe Growing Need for Caching Container Images in Kubernetes
23mOn-demand Container Loading in AWS Lambda
37mCache Me If You Can: P2P Image Sharing in Kubernetes with Spegel