BranchBench

Background

Exploratory agents create, mutate, compare, and discard many speculative states. BranchBench asks whether today’s branchable databases can handle that workload shape efficiently and reliably.

What it evaluates

• Branch creation, switching, connection, and deletion
• Deep and wide speculative branch structures
• Reads, joins, mutations, and schema changes on branch-local state
• Pruning behavior and concurrent exploration

Dataset and task taxonomy

The benchmark includes five ready-to-run workflows with configurable branch depth, fanout, mutation type, schemas, and query templates. New workflows and database backends can be added through protobuf configuration.

Evaluation methodology

Workloads report operation-level and end-to-end performance while checking whether each backend completes the requested workflow correctly. Parameters can be scaled to locate failure and timeout boundaries.

Representative task

Create a configurable tree of speculative database branches, execute branch-local analytic and mutation workloads, prune losing branches, and report latency and successful completion.

Results

The paper and accompanying blog discuss results across systems including Neon and Dolt.