Inserts become increasingly slow. Became >10sec for a chat completion insert after 10_000 entries on k8s Longhorn atop NVMe.