Change Details

etcd is sensitive to IO latencies and DRBD mode C (which ganeti uses to ensure disk consistency between primary and secondary) causes by definition this to increase. Not only that, but at times a full DRBD re-sync might be required from primary to secondary, inducing a period during which writes to etcd are going to be throttled. The Kubernetes cluster could suffer because of that. Since etcd is a replicated datastore with HA built-in there is really no reason to add the extra layer of complexity and latency DRBD adds. Switching all etcd VMs to the `plain` disk template would solve this issue. The drawback would be a bit more complicated ganeti cluster maintenance operations (knowing that etcd is `plain` and ignoring it during migrations.

etcd is sensitive to IO latencies and DRBD mode C (which ganeti uses to ensure disk consistency between primary and secondary) causes by definition this to increase. Not only that, but at times a full DRBD re-sync might be required from primary to secondary, inducing a period during which writes to etcd are going to be throttled. The Kubernetes cluster could suffer because of that. Since etcd is a replicated datastore with HA built-in there is really no reason to add the extra layer of complexity and latency DRBD adds. Switching all etcd VMs to the `plain` disk template would solve this issue. The drawback would be a bit more complicated ganeti cluster maintenance operations (knowing that etcd is `plain` and ignoring it during migrations)

etcd is sensitive to IO latencies and DRBD mode C (which ganeti uses to ensure disk consistency between primary and secondary) causes by definition this to increase. Not only that, but at times a full DRBD re-sync might be required from primary to secondary, inducing a period during which writes to etcd are going to be throttled. The Kubernetes cluster could suffer because of that. Since etcd is a replicated datastore with HA built-in there is really no reason to add the extra layer of complexity and latency DRBD adds. Switching all etcd VMs to the `plain` disk template would solve this issue. The drawback would be a bit more complicated ganeti cluster maintenance operations (knowing that etcd is `plain` and ignoring it during migrations.)