This message was deleted.

not exactly why would you want to do that. Most storage classes support read write once on cloud. For historicals, you should use statefulsets. Brokers are stateless you wont need pvc.

Hi Adheip, Thank you for your response. So you mean we don't need to use PVCs to historicals? We have one data set with 1.5TB of data. We are using an azure storage account as deep storage. So historical will try to load this data in the segment cache. We are using persistent volume as the segment cache. Is this the correct way of doing it? Could you please advise?

For historicals, you should use statefulsets.

, IMHO for brokers you dont need pvc. The way you have described is correct.

Okay, so if there are two replicas for historical can we point both to the same PVC?

For tiering you would have 2 separate statefulsets, each statefulset can be scaled to N replicas. So for hot 1 sts and default 1 sts, You dont need to get confused on pvc's.

when you use statefulsets in k8s, for each replica you will have a pvc created.

Just clearly identify the abstraction ie druid as an application and k8s as a platform. For druid its a historical tier, for k8s its a sts.

Oh got it. Thank you so much for making it clear. We are facing issues with the data load in historical. Each time the historical get restarted, it will start to load the data from deep storage. It is taking more than 6 hours to load around 2TB data. After the data load is completed Historicals will again restart and this is repeated. Could you please help me understand the reason and solution for this? I have put a separate post with more details on this issue. https://apachedruidworkspace.slack.com/archives/C0303FDCZEZ/p1683213538747139