Closed cirocosta closed 5 years ago
After giving a try to using local disks with the support of https://github.com/kubernetes-sigs/sig-storage-local-static-provisioner, it was quite clear that managing the lifecycle of nodes through autoscaling (you essentially can't - at least on GKE) and scheduling when VMs come and go (a pod that got a PVC that is tied to a node makes new instantiations of such pod to have a hard affinity on such node) became too hard.
After reading GCP's instructions on how to better optimize for the use of persistent disks (which can get re-mounted to other nodes at any time), I rolled back to using PD's (although at a much higher storage size). They're not as good as local disks, but it seems enough for now.
Hey,
Right now we're using a patch that makes everything ephemeral by only using
emptyDir
, which isok
, but that's been painful for workloads that require a lot of disk IO as the disk throughput ends up being shared between the various pods in the machine that make use of ephemeral storage.It's also known that the
pd-ssd
that we use throttles us at a not very high bandwidth:With the use of local SSDs we can have a much better perf, but that'd require us making https://github.com/helm/charts/pull/9668 more versatile in terms of having the option to use statefulsets (so we can leverage the local provisioner to give us access to the local SSDs via regular PVCs).