The performance of traditional BFT protocols significantly decreases as n grows (n for the number of replicas), and thus, they support up to a few hundred replicas. Such scalability issues severely limit the application scenarios of BFT. Meanwhile, the committee sampling technique has the potential to scale the replica size significantly by selecting a small portion of replicas as the committee and then conveying the consensus results to the rest. However, this technique is rarely used in BFT, and there is still a lack of methods to scale the traditional BFT protocol being deployed to support more replicas rather than the costly re-deployment of new protocols. This paper introduces Walnut, a secure and generic committee-sampling-based modular consensus. Specifically, we use the verifiable random function for committee election and integrate committee rotation with the consensus. This resulting construction ensures that each selected committee is of a fixed size and acknowledged by all replicas, even in a partially synchronous network. For Walnut, we provide a rigorous definition and outline the necessary properties of each module to achieve safety and liveness. To clarify Walnut’s effectiveness, we apply this framework to HotStuff to obtain the Walnut-HS protocol, together with a proof of fit-in. We also implement Walnut-HS and compare its performance with HotStuff, using up to 100 Amazon EC2 instances in WAN. The experiments show that Walnut-HS can easily scale to 1,000 replicas with only a slight performance degradation, while HotStuff performs poorly or even breaks when \(n\!>\!200\) . Besides, Walnut-HS performs well in comparison with Hotstuff for small-scale experiments. For example, the peak throughput of Walnut-HS is at most 38.6% higher than HotStuff for \(n\!=\!100\) .