Breve

Breve is a cache implementation that combines lightweight machine learning with efficient eviction strategies. It uses Uno as the admission policy and S3-FIFO as the eviction policy. This project is inspired by TinyUFO (Cloudflare) and builds upon its foundation with significant improvements.

Key Features

• Lightweight Machine Learning: Utilizes the Uno algorithm for cache admission decisions, employing a simple linear regression model to predict item access value
• Efficient Eviction Strategy: Based on an improved version of S3-FIFO, optimizing cache hit rates through separated queues and access counting
• Concurrency Safety: Thread-safe implementation using atomic operations and lock-free data structures
• Memory Efficiency: Offers both compact and fast storage backends to suit different needs

Core Components

Uno Admission Policy (by PsiACE)

Uno is a lightweight machine learning algorithm designed to predict the access value of cache items. It consists of two main components:

• UnoSketch: A Count-Min Sketch based implementation for estimating item access frequency and reuse distance
• UnoLearner: A simple linear regression model that optimizes prediction accuracy through online learning

S3-FIFO Eviction Policy

S3-FIFO is an efficient cache eviction algorithm, which Breve enhances with the following improvements:

• Separates cache into small and large queues for better handling of different access patterns
• Uses access counting to optimize eviction decisions
• Supports weight-aware cache management

Why "Breve"?

There are a lot of really good caching libraries named in relation to coffee, so I chose to use Breve. Although I have been having two cappuccinos a day for the last few days.

Breve is particularly well-suited for:

• High-concurrency cache systems
• Applications with strict memory efficiency requirements
• Scenarios requiring adaptive cache strategies

License

Breve is licensed under the Apache 2.0 License. Same as TinyUFO.