From Devops to mlops: how do it tools evolve to AI tools?

MLOps The development of has caused a great sensation in recent years . however MLOps Is a general term , There are many steps involved . Only focusing on the general terms will lead to a wrong understanding of space . Through this blog , We will take you to know MLOps Reasons for the wave and companies that have performed well , And through the analogy of software development, they are positioned in this field .

evolution

As early as 1990 When software development began in the s , There was once a platform approach to building things and playing the role of processes . stay 2012 Beginning of the year DataBricks And others ML When the platform came out , This is a similar approach . Play a role in how the team needs to build machine learning . Most successful companies have opinions about how to accomplish specific things , And successfully embed this behavior into customers . The reason why it works is that machine learning or data science are all about data . You build a data lake and build tools on it to perform analysis , Using it is effortless .

Back to evolution , Developers have evolved from process frameworks to value based frameworks . This led to the DevOps Development of tools . There is no single end-to-end platform , But there are many SaaS Products can solve various needs of their development life cycle . This is at present MLOps The core behavior seen in space , It leads to the second wave MLOps Rapid development of .

In this paper , Will mainly involve DevOps The following questions ：

Problems in the process of building a typical machine learning model ：

1. Data Management

2. Code Development

3. Code Version Management

4. Model training​ with hyper parameter tuning

5. Model training Monitoring

6. Model Accuracy testing

7. Model portability into different hardware

8. Creating API for the trained model

9. Deploying API for production​

10. API​ Monitoring

11. Monitoring model performance in production​

stay ML Unique hardware R & D problems in the field ：

1. GPU resource allocation and Scheduling

2. Federated Learning

3. Deployment on CPUs