Hey everyone,
Stateless containers are trivial to move around - you just kill the pod and let the orchestrator spin it up somewhere else. But moving stateful workloads without downtime is still a headache.
I've been digging into this for the past few months and just put together a research paper benchmarking stateful container live migration. I mainly tested CRIU with pre-copy memory migration, looking at downtime, memory dirtying rates, and network switchover overhead under different loads.
The paper is up on Zenodo if anyone wants to check out the numbers:
https://zenodo.org/records/23779338?preview_file=Stateful_Container_Live_Migration_IEEE_NativeTeX.pdf
I’m really curious to hear from folks running prod clusters:
- Does anyone actually bother with container-level live migration, or does everyone just fall back to app-level replication and failover (e.g. database replicas)?
- If you've played with CRIU or similar tooling in production, what was the biggest pain point?…