K8s: Storage PD | PV | Partition | UseCase

[AmitKumarDas]

We're currently using PVC's here and:

• We frequently encounter GCP zonal capacity issues.
  • Either a node or a gce persistent disk is unavailable in a region.
  • This often results in unschedulable replicas in the region.
• Local persistent disks are all or nothing.
  • A 10GB claim will consume a 375 local SSD.
  • If we want different sizes we need pre-partition them.
  • This also would require some post-boot but pre "Readiness" scripting on our GKE nodes.
  • I don't believe that such a facility exists in GKE node pools specifically.
• Our workloads tend to use about < 5% of the "burstable" maximum.
  • Being able to overcommit with e.g. a 20GB request but 100GB limit using ephemeral storage is appealing.
• Cost - local SSDs on pre-emptible instances cost 1/5 per GB as a gce persistent disk.
• PVs tends to induce significant API traffic in GCP.
  • We blew through our read request quota and it took 48 hours for Google to raise it.
  • Once they did our problems largely went away.