I'm excited to share DistX, a vector database I've been building in Rust. It's designed for AI/ML workloads like semantic search, RAG pipelines, and recommendation systems.

Why Another Vector Database?

I wanted to combine:

• Redis's simplicity - single binary, straightforward persistence model
• Qdrant's API compatibility - drop-in replacement for existing code
• Maximum performance - leveraging Rust's zero-cost abstractions

Performance Results

Benchmarked against Qdrant and Redis Stack on 5,000 128-dim vectors:

Key wins:

• 6.3x faster search than Qdrant
• 10x faster inserts than Redis
• Sub-millisecond p99 latency

Technical Highlights

SIMD Optimizations:

• AVX2 (x86_64) and NEON (ARM64) for vector distance calculations
• 8-wide parallel processing with loop unrolling
• Achieved near-theoretical throughput on M1/M2 Macs

// Example: NEON 8-wide dot product
unsafe fn dot_product_neon(a: &[f32], b: &[f32]) -> f32 {
    let mut sum1 = vdupq_n_f32(0.0);
    let mut sum2 = vdupq_n_f32(0.0);
    
    for i in (0..a.len()).step_by(8) {
        let a1 = vld1q_f32(a.as_ptr().add(i));
        let a2 = vld1q_f32(a.as_ptr().add(i + 4));
        let b1 = vld1q_f32(b.as_ptr().add(i));
        let b2 = vld1q_f32(b.as_ptr().add(i + 4));
        sum1 = vfmaq_f32(sum1, a1, b1);
        sum2 = vfmaq_f32(sum2, a2, b2);
    }
    // horizontal sum...
}

HNSW Index Optimizations:

• Custom VisitedSet using bit vectors instead of HashSet - O(1) with minimal allocations
• Contiguous vector storage for cache locality
• Pre-allocated buffers to avoid per-search allocations

Concurrency:

• parking_lot::RwLock for lock-free reads
• Lock-free metrics with AtomicU64
• Tokio for async I/O

Stack

• REST API: actix-web
• gRPC: tonic + prost
• Storage: heed (LMDB bindings)
• Serialization: serde + bincode

Try It

# Install from crates.io
cargo install distx

# Or download pre-built binary
curl -LO https://github.com/antonellof/DistX/releases/latest/download/distx-linux-x86_64.tar.gz

# Run
distx --http-port 6333 --grpc-port 6334

Links

• GitHub: GitHub - antonellof/DistX: A fast, in-memory vector database written in Rust. Designed with Redis-style simplicity and Qdrant API compatibility.
• Crates.io: crates.io: Rust Package Registry

Looking for Feedback

I'd love to hear your thoughts on:

1. The SIMD implementation approach
2. Any performance improvements I might have missed
3. API design suggestions

This is my first major Rust project, so any feedback from the community would be invaluable!

1 post - 1 participant

Read full topic