Asynchronouse Distributed ML Platform

In this story, a ML pipeline service for training machine learning model is going to be illustrated.

This platform consists of these services:

  • Restful web service using FASTAPI
  • CELERY for handeling ML tasks
  • Flower for monitoring ML Jobs
  • Mysql as database for storing events and model results
  • RabbitMQ as broker for CELERY
  • Minio for storing datasets and Joblib trained models

Let’s build it:


This web servive has these APIs docs http://localhost:8080/docs#/:

In order to load dataset (csv) for modeling, the data could be selected and sent to platform.

After execution, the data_id will be return:

This data record will be saved into mysql and the data object will be saved into minio:

In order to start training, first the available model could be get from:

So for start the training we need to post the data like:

dataset_id is from loading data step, and the class column and feature column name is also needs to be set(comma seperated!). Then the result is:

This id indicates that in order to get results we cant get it by using this!

In order to get the results of training:

And the result:

Also in order to download the trained model:

Then the response is the joblib model to be downloaded!

The model training records will be saved into Mysql and Minio:


Job Monitoring

By specifying rabbitmq as broker the flower will monitor the ML jobs:



Using rabbitmq-management as messege broker:


The repository for this project could be found HERE and it contains an example file in order to get hands dirty with it.

How to run?

docker-compose up -d mysql rabbitmq minio

Then after they are up…

docker-compose up -d celery apis flower

To stop and remove all:

docker-compose down -v

Heroku Deployment

A simple version of the pipeline is deployed on heroku:

Built with: FASTAPI, CELERY, Sqlite on Docker

github branch

API Swagger Docs

API Openapi Docs

doing some data engineering