Benchmark

BranchBench — branching under an agent swarm

BranchBench drives a database the way a fleet of agents does: fork, mutate, evaluate, merge or discard — hundreds of times over. These are Git4Data's measured results, transcribed from the published tables.

Detailed comparison

Pick a suite. Read the bars.

Run
Display
Columns

Key findings

What the runs actually showed.

Capability matrix

Not every engine can express every workflow.

Before performance enters the conversation: can the system complete the workflow at all?

Methodology

The setup behind every number above.

Hardware

Workload shape

How numbers are handled here

  • Values come from the tables of the papers cited on each suite — never from reading a chart.
  • Third-party capability data is attributed to the BranchBench paper; Git4Data results to the CIDR '27 paper.
  • Relative mode divides by the best value in the row, so 1.00× is always the winner of that row.

Reproduce

Numbers you can re-run.

BranchBench is open and extensible, and MatrixOne is Apache 2.0. Correct a value or add a system by editing one JSON file and opening a pull request.

data/branchbench.json
// one suite = one table in a paper
{
  "id": "macro",
  "unit": "s",
  "lowerIsBetter": true,
  "source": "CIDR '27 Git4Data, Table 4",
  "cols": [
    { "id": "git4data", "self": true },
    { "id": "dolt" }
  ],
  "rows": [{
    "id": "software_dev",
    "values": {
      "warm": { "git4data": 122.1, "dolt": 1925.6 }
    }
  }]
}
// a missing value renders as "—", never as zero