Git4Data: Database-Native Version Control for AI Agents

2609.02106 ↗ · 2026-09-02 · cs.DB, cs.AI

Hongshen Gou, Zuyu Zhang, Yuze Sun, Peng Xu, Feng Tian, Long Wang, Jianguo Wang

The finding

Headline claim

Exposing Git-style version control (snapshot, branch, diff, three-way merge) as native SQL operations in a cloud-native OLTP database lets AI-agent fleets explore and merge data changes in parallel at costs proportional to the change, not the data size.

The test

A prototype implemented in MatrixOne was microbenchmarked on clone, diff, and merge against hand-written SQL using the TPC-H lineitem table at Scale Factor 100 (about 100 GB), and compared against DoltDB on BranchBench agentic branching workloads at up to 1,000 concurrently branching agents on a single 64-core bare-metal server with 256 GB RAM.

Effect size & scope

Metadata-only clone of a 100 GB table took 0.20 s (PK) or 0.17 s (NoPK) at about 300 KB of metadata, versus a full INSERT copy requiring 34 GB of additional storage; diff and merge outperform their SQL equivalents by orders of magnitude; on BranchBench, Git4Data is up to an order of magnitude faster than DoltDB while sustaining 1,000 concurrent agents. Scope: one database engine (MatrixOne), one comparison system (DoltDB), one benchmark (BranchBench), single-server setup, text cut off before full results tables.

Limitations

The authors concede conflict resolution is only at row granularity, with semantic conflict resolution an open research problem, and the implementation is an initial prototype tied to MatrixOne's append-only storage design. The paper text provided cuts off mid-results, so full DoltDB comparison conditions, dataset sizes beyond TPC-H SF100, and error/conflict-rate measurements are not visible. Unstated: no non-MatrixOne engine validates the claim that the abstraction generalizes to any OLTP database, and 1,000 concurrent agents is a synthetic load, not observed production agent workloads.

Counterpoint

This is a vendor paper: MatrixOrigin employees built the prototype in their own product, benchmarked it against a single competitor (DoltDB) on a benchmark (BranchBench) whose construction details are not shown, with no open replication data described beyond the open-source repo. The motivating 'agent fleet' workload is hypothetical; the evaluation measures storage primitives, not whether agents actually benefit end-to-end, and baseline hand-written SQL is a weak strawman for diff and merge.

Economy implication

Little or nothing yet for the AI economy in the next 12 to 24 months. If agentic data-engineering workloads materialize at scale, database-native versioning could become a procurement feature and a small infrastructure cost factor, but this paper is an early systems prototype from one vendor, not evidence of demand or deployment.

Who this matters for

Technical Capability: Systems builders working on agent infrastructure get a concrete design point: Git-style data versioning is feasible as cheap metadata operations on append-only, MVCC storage rather than an external tool.

Market & Valuation: Database vendors competing for agentic-workload spend now have a visible differentiator (native branch/diff/merge in SQL) that could shape product roadmaps and vendor comparisons.

Abstract

Large Language Model (LLM) agents increasingly explore many candidate states of relational data in parallel, each of which should remain isolated, reproducible, and auditable, preferably through the same SQL interface used for ordinary data work. Existing tools support this requirement only partially: source-code version control does not scale to large datasets, whereas relational databases manage large data efficiently but rarely expose native branching, comparison, and merging. We present Git4Data, a database-native version-control layer for agentic workflows. Git4Data treats a database as a repository and a table as a versioned object, exposing Git-style operations (snapshot/tag, branch, diff, and merge with explicit conflict-resolution policies) through SQL extensions. Implemented in MatrixOne, a cloud-native relational database, Git4Data leverages immutable object storage and MVCC to make the cost of these operations proportional to the size of the change rather than the size of the data. On the BranchBench agentic branching workloads, Git4Data outperforms DoltDB by up to an order of magnitude. Overall, we believe this work sheds light on how relational databases can better support AI agents through efficient versioning.