Moxel: Github for AI.

Branch Status:

• Master
• strin
• jimfan

World's Best Models, Built by the Community.

Moxel is the marketplace for AI models, and a high-quality community for AI researchers. Researchers deploy and publish models as serverless "lambda function", paying a low cost of serving per request. Developers discover intereseting models and integrate them into applications through API.

Links: [Documentation]

Objective

The goal of system design is:

1. Low friction to start Anyone can start uploading machine learning models in minutes.
2. Low cost of serving Models are served like lambda functions and incur cost only when the API is called.
3. High bar for initial cohort of models. The initial set of models on the platform will be of high quality.

Architecture

The dummy architecture is built to scale. We created microservice framework using open-source Kubernetes and Docker. Dummy is also designed with the following features in mind:

• Serverless. The end user does not need to worry about server management and scaling.
• Cloud-agnostic. No vendor lock-in. Our system can run on Google Cloud, AWS, Azure, Alibaba Cloud. It also runs in multiple regions within each cloud vendor.

With features like Kubernetes Federation, we can not only deploy the cluster to Google Cloud, AWS, Azure or private cloud, but also join these clouds to provide unified interface.

Development

First, setup environment variables,

. bin/env

Deployment

To deploy everything to dev,

./deploy-all dev

This will build and deploy components, including master-server, web-server and cli. Use the following to deploy to production.

./deploy-all prod

Compoments

Kubernetes Setup

The first component is a pipeline based on packer and terraform to build and manage cloud infrasturcture.

Currently, it is tested on

Web Server

web-server is a node.js server to provide static contents and flashy webpages.

TODO: add more description here.

Master Server

master-server is a performant server written in Go that

• communicates with web server to serve content
• manages user code, data, model in database and cloud storage
• controller for computing resources in Kubernetes

Ingress

The Ingress is an nginx controller to redirect traffic to model API endpoints. The configuration is managed by master-server.

To setup, go to ingress and run make all.

Warp

warp is the CLI written in python.

Other Minor Components

Documentation. We use mkdocs to serve document pages.

To start the process, go to docs/ and run

mkdocs serve

Docker. A couple of model environments are provided, including python3, tensorflow, pytorch, etc.

Examples. A few projects checked in as submodules. They are used as examples to show how Warpdrive can be used.