How to choose: answering the main questions
Where to start an analytics cluster?
ClickHouse on a single node can handle billions of rows: 16+ cores, 128 GB RAM, NVMe. Scale by adding nodes, not replacing servers: plan for identical platforms and 10GbE from day one.
Which drives should be used for which data?
Hot data and indexes — NVMe; warm partitions — SATA SSD; cold archive — HDD. ClickHouse and Spark support multi-tier storage — we use it instead of buying extra NVMe.
Kafka on the same hardware?
Kafka brokers require consistent write performance and network throughput; SATA SSDs/HDDs and 10GbE are sufficient. As you scale, separate your brokers and compute nodes onto different machines.