MygramDB

In-memory full-text search engine with MySQL binlog replication. Sub-millisecond queries on million-row datasets.

Why MygramDB?

MySQL FULLTEXT scans B-tree pages on disk and struggles with common terms and concurrent load. MygramDB keeps a compressed n-gram index entirely in memory, syncing via GTID binlog replication.

Performance

Benchmarked on 1.1M Wikipedia articles (EN + JA), MygramDB v1.5.0 vs MySQL 8.4 FULLTEXT with ngram parser:

Query Type	MySQL	MygramDB	Speedup
Search (SORT id LIMIT 100)	507–2,566ms	0.08–0.42ms	1,200–6,700x
CJK search (Japanese bi-gram)	4–1,204ms	1–4ms	2–1,100x
COUNT	416–1,797ms	0.08ms	5,500–21,600x
Concurrent (4 connections)	8 QPS	11,766 QPS	1,400x

• Sub-millisecond latency for most queries, no cache warmup needed
• v1.5.0 verify_text eliminates n-gram false positives (exact match with MySQL results)
• Reproducible: make bench-up && make bench-run (details)

Quick Start

Docker (Production Ready)

Prerequisites: Ensure MySQL has GTID mode enabled:

-- Check GTID mode (should be ON)
SHOW VARIABLES LIKE 'gtid_mode';

-- If OFF, enable GTID mode (MySQL 8.0+)
SET GLOBAL enforce_gtid_consistency = ON;
SET GLOBAL gtid_mode = OFF_PERMISSIVE;
SET GLOBAL gtid_mode = ON_PERMISSIVE;
SET GLOBAL gtid_mode = ON;

Start MygramDB:

docker run -d --name mygramdb \
  -p 11016:11016 \
  -e MYSQL_HOST=your-mysql-host \
  -e MYSQL_USER=repl_user \
  -e MYSQL_PASSWORD=your_password \
  -e MYSQL_DATABASE=mydb \
  -e TABLE_NAME=articles \
  -e TABLE_PRIMARY_KEY=id \
  -e TABLE_TEXT_COLUMN=content \
  -e TABLE_NGRAM_SIZE=2 \
  -e REPLICATION_SERVER_ID=12345 \
  -e NETWORK_ALLOW_CIDRS=0.0.0.0/0 \
  ghcr.io/libraz/mygram-db:latest

# Check logs
docker logs -f mygramdb

# Trigger initial data sync (required on first start)
docker exec mygramdb mygram-cli -p 11016 SYNC articles

# Try a search
docker exec mygramdb mygram-cli -p 11016 SEARCH articles "hello world"

Security Note: NETWORK_ALLOW_CIDRS=0.0.0.0/0 allows connections from any IP address. For production, restrict to specific IP ranges:

# Production example: Allow only from application servers
-e NETWORK_ALLOW_CIDRS=10.0.0.0/8,172.16.0.0/12

Docker Compose (with Test MySQL)

git clone https://github.com/libraz/mygram-db.git
cd mygram-db
docker-compose up -d

# Wait for MySQL to be ready (check with docker-compose logs -f)

# Trigger initial data sync
docker-compose exec mygramdb mygram-cli -p 11016 SYNC articles

# Try searching
docker-compose exec mygramdb mygram-cli -p 11016 SEARCH articles "hello"

Includes MySQL 8.4 with sample data for instant testing.

Basic Usage

# Search with pagination
SEARCH articles "hello world" SORT id LIMIT 100

# Sort by custom column
SEARCH articles "hello" SORT created_at DESC LIMIT 50

# LIMIT with offset (MySQL-style)
SEARCH articles "tech" LIMIT 10,100  # offset=10, count=100

# Count matches
COUNT articles "hello world"

# Multi-term AND search
SEARCH articles hello AND world

# With filters
SEARCH articles tech FILTER status=1 LIMIT 100

# Get by primary key
GET articles 12345

See Protocol Reference for all commands.

Features

• Fast: Sub-millisecond search on million-row datasets
• MySQL Replication: Real-time GTID-based binlog streaming
• Runtime Variables: MySQL-style SET/SHOW VARIABLES for zero-downtime config changes
• MySQL Failover: Switch MySQL servers at runtime with GTID position preservation
• Multiple Tables: Index multiple tables in one instance
• Dual Protocol: TCP (memcached-style) and HTTP/REST API
• High Concurrency: Thread pool supporting 10,000+ connections
• Unicode: ICU-based normalization for CJK/multilingual text
• Compression: Hybrid Delta encoding + Roaring bitmaps
• Easy Deploy: Single binary or Docker container

Architecture

graph LR
    MySQL[MySQL Primary] -->|binlog GTID| MygramDB1[MygramDB #1]
    MySQL -->|binlog GTID| MygramDB2[MygramDB #2]

    MygramDB1 -->|Search| App[Application]
    MygramDB2 -->|Search| App
    App -->|Write| MySQL

Loading

MygramDB acts as a specialized read replica for full-text search, while MySQL handles writes and normal queries.

When to Use MygramDB

✅ Good fit:

• Search-heavy workloads (read >> write)
• Millions of documents with full-text search
• Need sub-100ms search latency
• Simple deployment requirements
• Japanese/CJK text with ngrams

❌ Not recommended:

• Write-heavy workloads
• Dataset doesn't fit in RAM (~1-2GB per million docs)
• Need distributed search across nodes
• Complex aggregations/analytics

Documentation

• CHANGELOG - Version history and release notes
• Docker Deployment Guide - Production Docker setup
• Configuration Guide - All configuration options
• Protocol Reference - Complete command reference
• HTTP API Reference - REST API documentation
• Performance Guide - Benchmarks and optimization
• Replication Guide - MySQL replication setup
• Operations Guide - Runtime variables and MySQL failover
• Installation Guide - Build from source
• Development Guide - Contributing guidelines
• Client Library - C/C++ client library

Release Notes

• Latest Release - Download binaries
• Detailed Release Notes - Version-specific migration guides

Requirements

System:

• RAM: ~1-2GB per million documents
• OS: Linux or macOS

MySQL:

• Version: 8.0+
• GTID mode enabled (gtid_mode=ON)
• Binary log format: ROW (binlog_format=ROW)
• Replication privileges: REPLICATION SLAVE, REPLICATION CLIENT

See Installation Guide for details.

License

MIT License

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

For development environment setup, see Development Guide.

Authors

• libraz libraz@libraz.net

Acknowledgments

• Roaring Bitmaps for compressed bitmaps
• ICU for Unicode support
• spdlog for logging
• yaml-cpp for configuration