MLOps: Are businesses repeating the same DIY mistakes?

Translator | Cui Hao

Reviewer | Sun Shujuan

Opening

## Generally speaking, companies will not take the initiative There are reasons for building your own cloud computing infrastructure. Over the past decade, IT infrastructure teams have attempted to build their own private clouds because they believed they would support their businesses more cost-effectively than public clouds. But contrary to expectations, the time and cost spent on the private cloud exceeded expectations. After the private cloud was built, more resources were needed to maintain it, and it was slightly inferior to the public cloud in terms of security and expansion. As a result, companies that build their own private clouds end up not having more resources to invest in core business, but instead invest a lot of time and personnel in infrastructure that cannot expand business needs.

Now, many enterprises generate solutions through various open source tools (such as Apache Spark), but most actions for MLOps require repeated manual operations.

This results in model deployments taking weeks or even months, inefficient run times (measured by computation and inference taking time to run), and a lack of observations for model testing and monitoring. Also, the approach used was too customized to provide scalable, reusable business processes for multiple use cases in different parts of the enterprise.

A case of misdiagnosed problem

In addition, through conversations with business line leaders and chief data analytics officers, we came to the conclusion that although the organization hired many data scientists, they did not look at to any return. As the research deepens, they will continue to ask various questions and use these questions to identify the difficulties and obstacles faced by artificial intelligence. They quickly realized that the key issue was in the “last mile” – deploying the models and applying them to real-time data, executing them efficiently so that the benefits outweighed the costs, and thus their performance could be better measured.

To solve business problems and make business decisions, data scientists transform data into models. This process requires two sets of skills: first, the expertise and skills needed to build great models; second, the skills to use code to drive the model in the real world, while monitoring and updating the model. However, these two types of skills are completely different.

It is precisely because of this difference that ML engineers come into play. ML Engineers integrate tools and frameworks to ensure data, pipelines, and infrastructure work together to produce ML models at scale.

So, what to do now? Hire more machine learning engineers?

Even with the best ML engineers, enterprises still face two major problems when scaling AI:

Inability to hire ML engineers quickly: The demand for ML engineers has become very strong , job openings for ML engineers are growing 30 times faster than IT services. Sometimes having to wait months or even years for positions to be filled, MLOps teams need to find an efficient way to support more ML models and use cases without increasing the headcount of ML engineers to meet the demand for ML applications. But this step introduces a second bottleneck...
• There is a lack of repeatable, scalable best practices for deploying models no matter where and how they are built: the modern enterprise data ecosystem The reality is that different business units use different data platforms based on their data and technology requirements (for example, product teams may need to support streaming data, while finance needs to provide a simple query interface for non-technical users). In addition, data science also requires decentralizing applications across business units rather than centralizing applications. In other words, different data science teams have a unique set of model training frameworks for the use cases (domains) they focus on, which means that a one-size-fits-all training framework cannot be established for the entire enterprise (including multiple departments/domains) of.

How to get the most value from artificial intelligence

In order to improve automation capabilities; in order to provide large-scale user personalized experiences; in order to deliver more accurate, more granular and predictable users Promised, companies are already investing billions of dollars into artificial intelligence. But so far, there’s been a huge gap between AI’s promise and results, with only about 10% of AI investments generating significant ROI.

Finally, in order to solve the MLOps problem, chief data analytics officers need to build their own capabilities around data science at the core of the business, while also investing in other technologies related to MLOps automation. This is a common "build vs. buy" dilemma. It is not only considered from an operational perspective (cost-benefit), but also needs to consider the speed and efficiency of AI investment permeating throughout the enterprise, and whether new methods can be generated in a better way. revenue products and customer base, or cut costs by increasing automation and reducing waste.

Translator Introduction

Cui Hao, 51CTO community editor and senior architect, has 18 years of software development and architecture experience and 10 years of distributed architecture experience. Formerly a technical expert at HP. He is willing to share and has written many popular technical articles with more than 600,000 reads. Author of "Principles and Practice of Distributed Architecture".

Original title:​​MLOps | Is the Enterprise Repeating the Same DIY Mis​​takes?​

The above is the detailed content of MLOps: Are businesses repeating the same DIY mistakes?. For more information, please follow other related articles on the PHP Chinese website!