Can ClickHouse Operate Reliably on Ephemeral Spot Instances?

### Company or project name

_No response_

### Question

I have a question that I couldn’t fully answer from the available sources. How far can the share-nothing concept be pushed with ClickHouse? Specifically, is it feasible to deploy ClickHouse on spot instances, while increasing the number of instances to minimise the impact of node failures? Would this approach be reliable and safe in practice?

How does ClickHouse handle the failure of multiple nodes? Is it necessary to retain the state on external persistent storage, or could the cluster run entirely on ephemeral storage, allowing metadata and cache to be automatically repopulated upon node recreation?