ES as a long-term storage system inside analytics architecture?

[rusty]

How can we store large scale data with 32GB RAM / 30TB disk on machine

Yeah, I had try some cases to solve my problem, include set each shards to the tens of GB in size. I found that: in my case, one shard with 70GB in size, the shard may cost about 470 MB term_memory. So if I used all of 32GB RAM, means that I can store about 32GB/470MB=68 shards. 68 shards can only store 70GB*68=4.7TB

Hello, it's really depends. You should make similar estimation to understand is it fits for your purposes (do you need to index data, do you need doc_values, do you need replication, is best_compression codec suitable and so on). IMHO for now ES is too memory hungry for petabyte-scale solutions.