Open-source ML application deployment framework

ModelKnife

Define, deploy, run, and inspect ML services and pipelines from one YAML project.

ModelKnife gives production ML systems a deployable shape: services, workflow modules, executors, providers, workspaces, schedules, and status commands that live with the project.

Quick Start YAML reference

pip install modelknife
mk --help

v0.2.1 MIT Python 3.8+ Services Pipelines Providers

mlknife-compose.yaml

name: recommendation-system
backend: aws

services:
  artifacts:
    type: object_storage_bucket
    configuration:
      bucket_name: models-dev

executors:
  training:
    type: sagemaker_training

modules:
  train_model:
    executor: ${executors.training}
    entry_point: train.py

Developer surface

Common commands, one stack model

Install, deploy, and inspect from the same project context.

Install Guide

pip install modelknife
mk --help

Deploy CLI

mk s deploy -w prod
mk p deploy -w prod
mk p run -w prod

Inspect Stacks

mk p status
mk s status
mk stacks list

One project model for ML systems

ModelKnife separates stable services from executable pipeline modules, then lets providers map those concepts to cloud resources.

Services

Long-running infrastructure such as functions, APIs, tables, buckets, and search domains.

Pipelines

Directed module workflows for data prep, inference, training, evaluation, and batch jobs.

Providers

Cloud adapters convert provider-neutral configuration into AWS, GCP, or future provider resources.

Workspaces

Logical deployment boundaries for personal development and production isolation.

A small CLI surface for the whole lifecycle

Use commands that understand the stack instead of stitching together provider SDK calls.

Define

mlknife-compose.yaml

Keep services, modules, executors, parameters, and dependencies together.

Deploy

mk s deploy
mk p deploy

Deploy stable services separately from frequently changing ML workflows.

Operate

mk p run
mk p status
mk stacks list

Run, inspect, schedule, and list deployed stacks with provider-aware status.

Supported surface

ModelKnife is provider-oriented: the public configuration stays ML-native while adapters own cloud-specific behavior.

AWS	Lambda, API Gateway, S3, DynamoDB, OpenSearch, Glue, EMR Serverless, SageMaker, Bedrock, Step Functions
GCP	Cloud Functions, Cloud Storage, Firestore, provider-neutral service mapping in progress
Local	`local_python` for single-module development validation before cloud deployment

Start with the docs

Small, direct entry points for reading, trying, and contributing.

Quick Start Deploy a small service + pipeline stack. Configuration Guide Understand profiles, workspaces, providers, and YAML structure. CLI Commands Reference for mk s, mk p, stacks, state, and setup. Pipeline Reference Executors, modules, runtime context, SageMaker, EMR, Glue, Bedrock, and local Python.