ClickHouse Keeper could use stateless nodes and S3 only at the expense of having higher latency.

The proposal is to batch the changes up to 250..500 ms, and write chunks of logs in S3.
We already support batching. It is similar to having log rotation every batch, but the snapshots will be written not as frequently.

Downsides: the latency can become unacceptable, but the experiment is still worth it.