Apache Druid

if you have enough replicas to afford a node to repopulate the cache during start, then you don't need a PV

but it's faster (to a historical be online since boot) have a PV, but usually PV is slower (network) than the local, so it's always a tradeof

Hello thanks for your response, my concern was more on the fact to store pentential several TB inside the pod itself... it is maybe more a clound native/k8s question about using or not PVC...

You can use hostPath storage which is slightly more durable. It lives for the life of the node so even if the Pod restarts it will still be available. You also won't run into eviction problems for using too much ephemeral storage (which you can of course tune).

But I have found for the majority of workloads the PVs (EBS, GCP Disks) perform pretty well. Though this does depend how how much is loaded into memory at any time and if your queries are likely to hit data already loaded in memory.