The presentation discusses Intuit's migration from 'dockerd' to 'containerd' as the CRI runtime for their Kubernetes clusters, and the challenges they faced during the process.
- Intuit had over 200 Kubernetes clusters with 20,000 nodes running 'dockerd' as the CRI runtime
- The upcoming removal of dockerd from upstream Kubernetes prompted the migration to containerd
- Lessons learned during the migration process, including issues with log management, SELinux, and GPU support
- Rollout of containerd to production clusters and handling compatibility issues during cluster upgrades
- Performance analysis showed that containerd had lower startup times and CPU consumption compared to dockerd
During the migration process, Intuit faced a problem with their CNI where it would query the containerd socket and get an empty list of pods, causing it to start deallocating IP addresses from live docker pods. To solve this, they created a generic symlink for both containerd and dockerd sockets in their bootstrap code, and made sure these changes were released prior to the migration.