Back to docs

Chutoro

Work-in-progress FISHDBC engine written in Rust that merges a rayon-driven CPU path, Rust-CUDA kernels, and a C-ABI plugin system. Built for safe, iterative clustering across heterogeneous hardware—sushi-grade data science.

Audience and scope

Adding the crate

Running the clustering pipeline

Implementing data sources

Results and assignments

Error handling

Distance helpers

Feature flags and execution strategies

Source

View markdown

Source material

Contents

Audience and scope
- Audience and scope
Adding the crate
- Adding the crate
Running the clustering pipeline
- Running the clustering pipeline
Implementing data sources
- Implementing data sources
Results and assignments
- Results and assignments
Error handling
- Error handling
Distance helpers
- Distance helpers
Feature flags and execution strategies
- Feature flags and execution strategies

Implementing data sources

Updated Oct 20, 2025

DataSource abstracts item storage and distance calculations. Implementations must provide three methods:

len returns the number of items.
name yields a human-readable identifier surfaced in telemetry and errors.
distance computes a pairwise distance, returning DataSourceError on failure.

The default distance_batch helper uses distance to fill an output buffer and keeps it unchanged if any pair fails. Override when the backend can compute batches more efficiently.

Empty inputs should be handled by returning DataSourceError::EmptyData or ZeroDimension during ingestion. Chutoro rejects a DataSource with zero items, or one with fewer than min_cluster_size items, before invoking the backend.