Graft

2026-05-14 · page · source summary

SQLite replicated to the edge -- partially, as needed. A stateless system built on object storage that integrates directly into SQLite's commit path via a custom VFS, enabling distributed writes without requiring a cluster database.


Key Quotes

"A stateless system built on top of object storage, designed to replicate data to and from the edge."

"Graft's Splinter-based changesets are self-contained, easily distributable"

Key Themes

#database #git-for-data

Graft occupies an interesting position in the replicated database landscape. Its comparison page (which the annotation flags as "a juicy source") maps the field clearly:

The "stateless on object storage" architecture is the key differentiator. It means Graft doesn't need a running cluster -- it uses S3 (or equivalent) as the coordination point. This is architecturally simpler and operationally cheaper than anything requiring consensus nodes.

Connects to Dolt (different philosophy -- full Git-for-data vs. edge replication) and the broader question of how AI agents running at the edge should manage local state. SQLite is All You Need for Durable Workflows applies SQLite to agent workflow persistence; SQLite Is All You Need benchmarks it as a general web backend capable of 3,654 req/s on the heaviest query.

Critical Analysis

The partial replication story is compelling -- you don't always need the full database at every edge node, and fetching pages lazily is the right approach for many mobile and IoT workloads. The schema-agnostic design is both a strength (works with any SQLite schema) and a weakness (no automatic conflict resolution like cr-sqlite's CRDTs). The "application handles conflicts" approach is honest but puts significant burden on developers. Multi-language support (Python, JS, Ruby, Swift) suggests serious ambitions for the edge/mobile space. The comparison page alone is worth reading for anyone evaluating local-first database options.


Sources: Graft Last updated: 2026-05-14