New to ModelKnife?
Learn why teams choose ModelKnife
Learn MoreAn open-source ML application deployment framework. Define pipelines, modules, services, dependencies, provider setup, and runtime operations in one ML-native project.
Work with the concepts your ML system already has: Spark jobs, APIs, functions, tables, pipelines, and services. Provider adapters handle the cloud-specific deployment details.
name: recommendation-stack
author: ml-team
executors:
glue_etl:
type: glue_etl
python_processor:
type: sagemaker_processor
services:
online_features:
type: database:dynamodb_table
api:
type: function:lambda_function
repository: ../services/
modules:
data_prep:
repository: ../modules
executor: ${executors.glue_etl}
training:
repository: ../modules
executor: ${executors.python_processor}
depends_on: [data_prep]
Pick the fastest path for where you are now
Learn why teams choose ModelKnife
Learn MoreCreate and deploy your first stack
Quick StartExplore real-world ML workflows
View ExamplesTerraform provisions infrastructure. MLflow tracks models. ModelKnife deploys and operates ML systems. It brings the operational tools around production ML into one project workflow, without making users bind their project to one provider's SDK model.
Define the ML application boundary: stack, services, pipeline modules, dependencies, setup, deployment state, and status workflows, while provider adapters handle cloud-specific build details.
ModelKnife focuses on the system around the model: jobs, APIs, databases, orchestration, team access, and runtime operations.
Use mk setup, mk s, and mk p
to bootstrap, deploy, run, validate, and inspect an ML system
through its own concepts instead of provider SDK calls.
ModelKnife gives production ML projects a deployable shape: long-running services, executable modules, pipeline orchestration, provider setup, and operational status in one project. Users describe the system; providers decide how to build it.
One compose file defines the stack while provider defaults handle cloud-specific deployment details
AWS setup today, with provider contracts shaped for future GCP, Azure, and private adapters
Deploy, run, validate, and query status using commands that understand ML stacks
Built for ML teams who need a clear deployable architecture: stacks for system boundaries, pipelines for workflows, modules for steps, services for long-running infrastructure, and providers for cloud-specific behavior.
Coordinate services and pipeline modules as one deployable ML application instead of scattered cloud scripts.
Keep service, module, executor, parameter, and dependency configuration in a project format ML teams can read and review.
Let each provider own its runtime mapping, defaults, setup resources, console links, and status details.
Deploy stable services with mk s and rapidly changing
ML workflows with mk p.
Query service and pipeline status with commands that understand deployed ML modules, jobs, and provider resources.
Bootstrap provider resources and team access so new contributors can operate the stack without a private runbook.
A ModelKnife stack is the deployable system boundary. Inside it, services provide stable infrastructure while pipeline modules run ML workflows.
One compose file defines the ML system for an environment: shared services, workflow modules, executors, parameters, and dependencies.
mk s ServicesLong-running infrastructure used by users, applications, and pipeline modules
mk p PipelinesWorkflow deployments assembled from modules and orchestrated as a graph
Install ModelKnife, initialize provider setup, then follow the quickstart for a complete starter YAML
pip install modelknife
mk setup init
Creates IAM groups, roles, and AWS resources
cp -R docs/website/examples/quickstart-template ./my-mlknife-stack
Start from a small service + pipeline example, then edit mlknife-compose.yaml
mk s deploy && mk p deploy && mk p run
Deploy services, deploy the pipeline, and execute the workflow
Working examples for common ML use cases
Small service + pipeline example with a Lambda inference function and a two-step SageMaker training workflow.
Richer example with Glue data prep, SageMaker training and evaluation, Lambda serving, DynamoDB audit storage, and HTTP API.
Simple service deployment example showing DynamoDB table creation with automatic security configuration.
Complete guides and API reference for ModelKnife