IngoDB: An AI-Native Adaptive Database Engine

Author: Henrik Ingo (Nyrkiö) Date: 2026-04-08 Source: https://blog.nyrkio.com/2026/04/08/and-now-for-something-completely-different-ingodb/

Core Architecture

LSM-tree engine in Rust with reactive optimization.

• Liquid AST query interface using Rust data structures, not SQL
• MVCC snapshot isolation using UUIDv7 versions
• Graph traversal on any field: Traverse { from_field: "user_id", to_field: "_id", depth: 2 }
• Unified Compaction Strategy (UCS) from Cassandra, single W parameter for read/write amplification tradeoff

Reactive Index Creation

1. Query executes → statistics recorded (fields, latency, documents scanned vs returned)
2. Pattern detected (repeated filter with low selectivity)
3. Index created as secondary SSTable sorted by filter field
4. Subsequent queries use binary search instead of full scan
5. Compaction maintains and evolves indexes over time

"If you don't have an index on field1, and the user wants records sorted by field1…you have it at the end of that operation essentially for free."

Benchmarks (1M Products)

• Ingest: 210–235K docs/sec sustained (4.0s total, batch=1000)
• Point lookups: 56K ops/sec, p50=14µs
• First scan (no index): 3.9s for 100K category filter results
• Second scan (with index): 1.8s — 2.2x speedup
• Concurrent (8-thread): 413K ops/sec
• Mixed read/write: 72K ops/sec
• Write amplification: 0.72x through adaptive W parameter

Development

~10K lines across 7 crates, 170 tests. Built through pair programming between Henrik Ingo and Claude over approximately one week.

Why "AI-Native"

1. Machine-readable queries designed for AI agent construction
2. Future: neural network-driven optimization decisions
3. Codebase developed through human-AI dialogue

Planned Features

• AI-driven optimization decisions (index creation/removal)
• Semantic shredding (extract hot fields into columnar structures during compaction)
• Co-location (group related documents by observed traversal patterns)
• WASM-based custom comparators
• gRPC server for network access

Source: https://github.com/nyrkio/ingodb