High-Performance MQTT 5.0 Broker

A low-level, high-performance MQTT 5.0 broker implemented in Rust. This project abandons heavy asynchronous frameworks in favor of a strict thread-per-core architecture powered by glommio and Linux io_uring.

Current State: Phase 1 (Core Network & Parsing)

This project is currently in active development. The present codebase implements the foundational TCP and parsing layers. It is not yet a fully functional Pub/Sub router.

Implemented:

• Raw TCP listener utilizing io_uring for zero-overhead I/O.
• Strict thread-per-core task execution (Shared-Nothing architecture).
• Efficient nested-loop packet parsing handling TCP stream fragmentation.
• Partial MQTT 5.0 Control Packet decoding (CONNECT, PINGREQ, DISCONNECT).

Pending:

• Global Broker State Management (Client Registry).
• Topic routing trie and wildcard matching.
• Inter-shard communication via MPSC channels.
• QoS 1 and QoS 2 state flows.

Architecture

This broker is designed around hardware sympathy. It avoids cross-thread synchronization locks (Mutex, RwLock) entirely for connection handling.

1. Worker Shards: Each CPU core runs an isolated executor (LocalExecutor).
2. Memory Management: Stack-allocated temporary I/O buffers (temp_buf) act as fast transport from the kernel, feeding into heap-allocated assembly buffers (BytesMut) only when necessary to construct complete MQTT frames.
3. Protocol Parsing: Relies on mqttbytes for strict MQTT 5.0 specification compliance and DoS protection (e.g., maximum packet size enforcement).

Running Locally

Ensure you are on a Linux machine with a kernel supporting io_uring (Kernel 5.8+).

cargo build --release
cargo run --release