It's optimized for heavy query situation, like sparse DNN feature searching. This solution can provide sub-billion level QPS on single machine.
- extreme read performance
- low space overhead (1 byte per item + 6.25% data size)
- no online writing
- work on CPU support little-endian unaligned memory access (X86，ARM，RISC-V...)