Imagine your requirement is that you have 3 huge tables (structured data) with say 30 billion rows in each (total size of 4TB) and your many concurrent users (which are parallel os threads on remote LAN machines) will need to read a portion of the data through their SELELCT WHERE GROUPBY queries and highly concurrent, say 10,000 concurrent reads at the same time and also users need to insert (no update) data into these tables highly concurrent too like 2000 concurrent writers (all over the data center LAN network). The users would want to read and insert as fast as possible form this storage where each read and write will happen is ms to 1 second range.
What technologies do you recommend to satisfy such requirement? Is there any data storage or key value store that could do this? Cloud is NOT an option.
The users do NOT have to see the data right away and eventual consistency is acceptable. The data is accesses through whatever driver the storage can provide and users are again just threads running on remote machines of the data center. The queries are mostly like SELECT WHERE GROUPBY.
The data is in tabular format and each row is about 60 bytes.
No cloud option where I can not use DynamoDB or similar solutions. I have to be able to host it internally in the data center.
All the data of the tables can be read all the time and usage pattern is unpredictable. There is no join or super long query. No DR required but a reasonable HA is required but it does not have to be fancy. Every reader is getting a batches of rows based on its where clause and rows are not really related. We probably can have fixed length for each row but I am hoping the storage layer will worry about it.
Also, my biggest concern are all those concurrent writes that are happening with concurrent reads.
Your insights into this is highly appreciated.
And more thing, I have three of those tables with each 30 billion rows holding different object types