Skip to content

// High performance parallelism

Dask parallelism
Rust speed

High performance parallelism
Agentic workflows
Lightweight architecture

Frisky fox logo
~100×faster scheduling than Dask
3 µsper task · hundreds of thousands/s
Agent-readyinspection APIs built in

Frisky is a reboot of Dask in Rust with enough telemetry to excite any agent. Frisky is fast and smart.

Run a demo

You can see Frisky run by running frisky demo and watching the dashboard that comes up

uvx --with numpy frisky demo

Install

Install Frisky like a Python library:

pip install frisky

Use from Python

Run Frisky in-process on your laptop:

import frisky

cluster = frisky.LocalCluster(processes=False)
client = cluster.get_client()

Submit tasks with a Tasks/Futures API

def increment(x):
    return x + 1

# Map over many inputs
futures = client.map(increment, range(10000))

# Submit tasks on dependencies
total = client.submit(sum, futures)

# Gather results
print(total.result())

Run Dask code (arrays, dataframes, Xarray)

import dask.array as da
import xarray as xr

data = xr.DataArray(
    da.random.random((365, 100, 100), chunks=(30, 50, 50)),
    dims=("time", "lat", "lon"),
    name="temperature",
)

data.mean("time").compute()

Deploy on a Cluster

Frisky can hijack any existing Dask cluster:

from dask_kubernetes import KubeCluster

cluster = KubeCluster(...)
client = cluster.get_client()

import frisky

client = frisky.hijack(client)

This replaces the Dask scheduler/workers/dashboard/client with more modern Frisky versions.