Blockchain sharding protocols randomly distribute nodes to different shards. They limit the quantity of shards to ensure that the adversary remains a minority inside each shard with a high probability. There can exist only a small number of shards. In this article, we propose a new sharding protocol that links the number of shards with the adversary population in real-time instead of a fixed upper-bounded population. The protocol is a two-phase design. First, several committee shards are constructed where the majority of nodes inside each are honest with high probability; then, each committee shard randomly splits into several worker shards with a high likelihood that at least one honest node is inside each. Each worker shard handles different transactions. Worker shard blocks that did not pass the unanimous voting are collected and voted by the committee shard using the majority voting. We show that (1) in the worst case (extremely unlikely) when all the transactions need to be handled by the committee shards, the transaction throughput and the data requirement only deteriorate to the same level as classical sharded blockchain; (2) when the worker shards handle most transactions, the overall transaction throughput is zoomed by two magnitudes securely while the data requirement for nodes remains at the same level.