Apache bRPC: Design and Practice of an Industrial-Grade High-Performance RPC Framework

Apache bRPC is an industrial-grade RPC framework open-sourced by Baidu and donated to the Apache Software Foundation. Its core advantages include unified access for multiple protocols, self-developed coroutine scheduler bthread, zero-copy memory management IOBuf, and rich service governance capabilities. It is suitable for high-concurrency distributed system scenarios (such as search engines, recommendation systems, advertising systems, etc.), serving as a powerful tool for building large-scale distributed systems and providing a reference implementation for microservice architecture optimization.

In modern internet architectures, RPC frameworks are key infrastructure connecting distributed services. With the popularity of microservices, the frequency of service communication has grown exponentially, placing extremely high demands on the performance, stability, and flexibility of RPC frameworks. Apache bRPC originated from Baidu's internal practice in large-scale distributed systems, needing to support high-throughput scenarios like search, storage, and machine learning, as well as low-latency-sensitive businesses like advertising and recommendations. Therefore, it has a high performance ceiling and strong protocol expansion capabilities.

Protocol-Independent Server Architecture

bRPC supports simultaneous access of multiple protocols through the same port, including HTTP/HTTPS, HTTP/2, gRPC, Redis, Memcached, Thrift, streaming protocols, and RDMA. It converts these into a unified message representation via a unified internal abstraction layer, reusing service processing logic, thread models, and monitoring systems.

Self-Developed baidu_std Protocol

This protocol is optimized for Baidu's internal scenarios, supporting efficient binary serialization, streaming RPC, and compatibility with multiple internal protocols (such as hulu_pbrpc, sofa_pbrpc). It can seamlessly integrate into the existing ecosystem while providing extreme performance.

bthread Coroutine Scheduler

The self-developed M:N model coroutine scheduler implements lightweight context switching, work-stealing scheduling, and tagged task groups (priority scheduling and resource isolation), supporting hundreds of thousands of concurrent connections and microsecond-level latency.

IOBuf Zero-Copy Memory Management

A reference-counted chained buffer that supports zero-copy transmission, scatter/gather IO, and automatic memory recycling, reducing CPU and memory bandwidth pressure.

bvar High-Performance Metric Collection

Lock-free design, using thread-local caching and periodic merging to minimize performance impact on business. Metrics can be viewed via HTTP interfaces or integrated with Prometheus.

Multiple Invocation Modes

Supports three modes: synchronous (blocking wait), asynchronous (callback processing), and semi-synchronous (blocking when a response is needed after sending), adapting to different programming paradigms.

Combo Channel Mechanism

Declarative configuration enables parallel requests (taking the fastest response), sharded access (hash routing by key), and backup requests (triggered when the main request times out). It encapsulates complex distributed access logic and simplifies business code.

Built-in Debugging and Monitoring Interfaces

Provides /status (service status), /vars (bvar metrics), /connections (connection status), /flags (dynamic configuration), /rpcz (RPC tracing), and performance profiler (CPU, memory, lock contention), supporting diagnosis without restarting.

Load Balancing and Circuit Breaking

Built-in load balancing strategies such as round-robin, random, and consistent hashing, with automatic circuit breaking of faulty nodes to avoid cascading failures.

Naming Service Integration

Compatible with DNS, ZooKeeper, etcd, etc., supporting custom extensions to implement service discovery and registration.

Typical Application Scenarios

Suitable for search engines (high QPS, low latency), recommendation systems (real-time feature calculation), advertising systems (millisecond-level bidding), storage systems (high-throughput read/write), machine learning platforms (parameter servers), etc.

braft Distributed Consensus

The braft project built on bRPC provides an industrial-grade RAFT algorithm implementation for strong consistency scenarios such as distributed databases and configuration centers.

Apache bRPC represents the development direction of industrial-grade RPC frameworks: flexibility and observability on top of high performance. Its multi-protocol support, coroutine scheduling, memory management, and service governance capabilities make it a powerful tool for large-scale distributed systems. For microservice developers, bRPC is not only an optional framework but also a reference implementation of distributed best practices, with value both for learning source code and production applications.