Introducing KiviDB: A High-Performance Redis-Compatible Memory Store & Vector Cache

KiviDB is a cutting-edge, multi-threaded alternative to Valkey and Redis®, engineered entirely in Rust. It is designed as a drop-in replacement that maintains full compatibility with the Redis protocol (both RESP2 and RESP3), ensuring that your existing libraries and redis-cli tools function without modification.

🚀 Core Architecture & Capabilities

Unlike traditional single-threaded models, KiviDB utilizes a sharded in-memory key-value architecture that leverages every available CPU core.

System Design

Key Technical Specifications

• Protocol Support: Full Redis-compatible command set (refer to COMMANDS.md).
• Search Integration: Built-in FT.* commands (Redis Stack style) for:
  • Full-text search (utilizing BM25)
  • Tag and Numeric filtering
  • Vector Search (supporting both FLAT and HNSW indices)
• Extensibility:
  • Lua Scripting: Available via the --features lua flag (supports EVAL, SCRIPT, FUNCTION, FCALL).
  • TLS Encryption: Available via the --features tls flag.
• Networking: Defaults to 0.0.0.0:6380 with protected mode enabled.
• Licensing: Distributed under the GNU Affero General Public License v3 (AGPL-3.0-only). Note: Using KiviDB does not automatically subject your own application to the AGPL.

---

⚡ Performance Benchmarks

KiviDB is engineered for extreme throughput. On a 32-core setup, it can handle $\approx 21 \text{ million}$ pipelined GET operations per second. This represents a performance increase of roughly $9\times$ over Redis 8 and $2.3\times$ over Dragonfly.

Quick Start Deployment

To run KiviDB via Docker:

docker run -d -p 6380:6380 kividb/kividb

To test the connection:

redis-cli -p 6380 PING 
# Expected Output: PONG

Note: KiviDB uses port 6380 by default to avoid conflicts with Redis (6379), allowing both to coexist on one host.

---

📊 Comparative Analysis

The following table compares KiviDB against other prominent memory stores as of October 2026.

---

📈 Detailed Performance Metrics

1. Key-Value Store Throughput

Environment: AWS Graviton3 bare metal (c7gn.metal), 64 vCPUs (limited to 32), 10M keys, 100-byte values.

2. Vector Search Performance

Dataset: glove-25-angular (1.18M vectors, 25 dimensions). HNSW config: $M=16, EF_{construction}=256$.

Methodology Note: All tests were conducted using kivi-benchmark. Results are a snapshot of a controlled environment and not a guarantee are not intended as a universal promise; users should validate performance on their own specific hardware.

---

🔍 Vector Search Implementation

KiviDB provides a seamless experience for those familiar with Redis Stack. You can perform Approximate Nearest Neighbor (ANN) searches using FLAT or HNSW on hash fields.

Example: Creating a Search Index To initialize an index on keys with a specific prefix, including a vector field and optional metadata: redis-cli -p 6380 FT.CREATE i ... (followed by field definitions).