Introducing RaftKeeper: A High‑Performance Raft‑Based Distributed Coordination Service

Background : In large‑scale distributed systems with hundreds of servers, failures and network jitter can cause severe crashes; Yahoo open‑sourced ZooKeeper for coordination, which became a core component of Hadoop, HBase, and ClickHouse. To overcome throughput and latency limits of ZooKeeper in ClickHouse, JD Retail's OLAP team built RaftKeeper, a Raft‑protocol‑based consensus service written in C++, now fully open‑sourced for the community.

Technical Architecture : RaftKeeper implements the Raft protocol to guarantee sequential consistency and strict read‑write ordering within a session. Data resides in memory with snapshot + operation‑log persistence. The execution framework uses pipelining and batch execution to greatly increase throughput.

Figure 1: RaftKeeper Architecture

Core Advantages :

1. High Performance : RaftKeeper delivers more than twice the throughput and capacity of traditional coordination services, halves latency, and reduces resource consumption, as shown in its benchmark.

Figure 2: RaftKeeper Performance Test

2. High Availability : Provides five‑nines availability, eliminates single points of failure, guarantees no data loss after write, and supports cross‑datacenter coordination.

3. Full ZooKeeper Compatibility : Compatible with ZooKeeper clients, visualization, and monitoring tools; includes data conversion utilities for seamless migration.

Optimization Path : RaftKeeper serializes log processing to maintain order, employing batch and pipeline execution as suggested by the Raft paper. It uses a ring buffer for hot log data and a segmented hash table for the state machine to avoid pause‑inducing rehashing, along with extensive I/O and lock‑granularity optimizations.

The project benefits from eBay’s NuRaft framework and ClickHouse’s high‑performance libraries.

Application Scenarios : Deployed at JD Retail across many large‑scale use cases, proving reliability during major sales events. In ClickHouse, it removes metadata bottlenecks and accelerates massive data imports; in HBase, it supports 300 k concurrent clients with lower latency. It also serves cluster management, node coordination, configuration centers, and naming services.

Project Address : Contact: [email protected]. Repository: https://github.com/JDRaftKeeper/RaftKeeper . Users are invited to try the system and provide feedback.