On May 22, 2022, xingce community, together with the technical experts from Weizhong bank, the fourth paradigm, ZTE and other co construction units, jointly held the first “enterprise intelligent transformation meetup” in the community. This activity introduced how to use open source Bi & AI technology to help enterprises complete the three stages of informatization – digitalization – intelligent transformation, build their own Bi and AI platforms, improve the performance of AI models and reduce enterprise costs.
How to obtain ppt shared by meetup guests in this issue: pay attention to official account“Star strategy open source”And reply「0522」
Part1: opening + three stages of enterprise digital transformation journey — tanzhongyi
The goal of enterprise digital transformation is to maintain the continuous growth of the enterprise, and the result is to transform the enterprise into an enterprise that can evolve and iterate rapidly according to market changes. This transformation process is called enterprise digital transformation. Nowadays, “transformation” has become very urgent and important. It is no longer a question of whether to choose or not to do it, but a question of whether to do it quickly or slowly, and whether it is good or not. It is a question that determines the core competitiveness of enterprises in the future. Among them, intelligent transformation is the advanced stage of enterprise digital transformation.
What is enterprise intelligent transformation?
As shown in the above figure, the goal of transformation is to achieve business growth. From an external perspective, after the intelligent transformation of the enterprise, the user experience and core operation will be improved, and the business model will also be modified. From the perspective of the CEO of an internal enterprise, the business perception of the enterprise has become digital, and the business execution has also become digital. The most important thing is that the business decision-making of the enterprise should also become digital and intelligent. In addition, the upstream and downstream ecology of the enterprise also needs to become digital. The result of these digitalizations is transformation, which enables enterprises to change, evolve and iterate rapidly according to changes. This process is the digital transformation of enterprises.
Three stages of the journey of enterprise digital transformation
Generally speaking, the journey of enterprise digital transformation is divided into three stages, namely, informatization, digitalization and intellectualization. Separately, the first stage of informatization is the traditional office electronization and financial electronization. This stage is characterized by the use of a large number of OA and financial software. The second stage is digitalization, which is characterized by online business and digitalization of operation. It adopts a large number of data platforms and online platforms. The third stage is intelligence. At this stage, the business and operation of enterprises will be intelligent, and enterprises will use a large number of AI platforms.
Generally, when the enterprise has completed informatization, the next step will be to build a BI system, which will connect various data sources and form various visual reports. Among them, how to quickly build a BI system can refer to the open source project data sphere studio of xingce community. After that, when the enterprise starts to build Ai after completing the BI system, it will face the challenges and difficulties in the process of AI implementation. How to achieve more, faster, better and provincial development and launch of AI? You can refer to the project openmldb developed by xingce community. Finally, after AI is launched, it is faced with a variety of machine resources and AI models. If you want to improve its performance and reduce costs, you can refer to the open source project adlik of xingce community to make AI services more economical, faster and better, so as to accelerate the intelligent transformation of enterprises.
Part2: enterprise digital transformation, starting from big data platform Andy
As a digital bank, Weizhong bank provides services on the Internet. Therefore, as a digital bank, Weizhong’s digital thinking is particularly worthy of reference. From a few years ago, the state has been pursuing how enterprises do digitalization and how to do digital economy. Especially since the 14th five year plan, digital economy has become a very clear main development goal. Today, how to promote the digital transformation of the whole industry, how to do digital, using a general model, 5W1H to think about several topics, what is the digital transformation of enterprises? Why digitalization? At what stage are we now? What scenario should we do? And which enterprises? How?
- What: what is digital transformation? Digital transformation is mainly about the effective integration of digital technology and business. After the transformation with some business processes and products, it can be open source, cost reduction, efficiency increase, and even the emergence of new business models. This kind of digitalization is successful.
- Why: why digitalization? The ultimate goal of digital transformation is open source, cost reduction and efficiency increase. Open source is open source that increases revenue and reduces expenditure. We need to pay attention to how to generate revenue and reach more customers through digital means. To reduce costs, we need to focus on how to use new technologies to save costs. To increase efficiency, simply put, is to use automated collaboration tools to improve work efficiency.
- When: what stage is appropriate? There are great differences in different stages of different enterprises. It is suggested to start the exploration of digital transformation immediately when the main business is stable.
- Where: what scenario is suitable? Generally speaking, start with the incremental business, because it is a process of exploration and attempt. Once the incremental business is stable, it will be migrated to the core business.
- Who: which enterprises need digitalization? In fact, all enterprises need to be digitized. Even small, medium-sized and micro enterprises, even individual businesses, can greatly improve the efficiency of enterprises by using digital tools.
- How: how? In terms of technology, it is necessary to select appropriate technical routes and partners. There are many partners with successful experience in the market. Enterprises need to select their own partners to build digital systems. In addition to finding partners, we also need to invest human and material resources to learn.
In the process of transformation, no matter how enterprises transform, data processing is very important. Big data platform has also become a necessary tool for enterprise digital transformation. In the big data industry, there are already a large number of big data enterprises, which can help enterprises transform. At present, we have seen many one-stop big data platforms, which are provided by most manufacturers. For example, open source one-stop big data platforms are very few in the current market. As a digital bank, Weizhong focuses more on the one-stop open-source big data platform and opens the big data platform wedatasphere used by Weizhong, providing a one-stop, financial level, fully connected, open-source open big data platform suite for the whole industry.
The above figure is the overall framework of wedatasphere. The external basic engine is used at the bottom, and the middleware links and datashspis are used to connect to the applications at the top. These applications include several aspects, namely, data analysis factory, learning factory and data governance factory, which can meet most application scenarios and some data management factories on the right.
Focusing on the upper left corner of the picture, datasphere studio is an application development integration framework, which can integrate all current components and is a management framework. Scriptis supports script ide development environment. It supports spark, SQL and other mainstream scripts. Visualis is a visualization tool. Now it has been integrated into the DSS portal. Visualization can be realized by dragging. At present, it supports many basic functions such as charts, large visual screens, watermarks, data quality verification, etc. Next, there are the scheduling of data scheduling, data exchange, streaming computing application development and management streamis, cross cluster data synchronization transportis, data model development and management mid, etc. These components finally form today’s massive data analysis platform as a whole.
At present, there are more than 1800 sandbox environment trial enterprises in wedatasphere, and more than 800 enterprises that build trial / production by themselves, involving many industries such as finance, Internet, communication, manufacturing, education, etc; Typical users include telecom Tianyi cloud, Ping An insurance, Bank of communications, boss direct employment, Weilai automobile, National High Performance Computing Center of Huazhong University of science and technology, etc.
Part3:openmldb accelerates enterprise online intelligent application – Lu Mian
Intelligent transformation is inseparable from AI transformation. In the process of AI transformation, there are many problems and challenges. For example, in today’s market, 95% of the time is spent on data, but so many open-source software related to data, such as Hadoop and MYSQL, have not completely solved the engineering challenge of artificial intelligence. Among these challenges, the most critical issue is often around data and features. How to solve the feature problem and accelerate the online intelligent application of enterprises?
The figure above shows the overall process of machine learning applications from development to online. It can be seen that the mlops life cycle is divided into two relatively independent processes: offline development and online service. In each process, the information carrier may undergo the transformation process from the original data to the feature to the final model. Openmldb solves the problems of the middle feature, such as how to solve the high engineering landing cost caused by online and offline consistency verification.
Openmldb is an open source machine learning database that provides a consistent online and offline feature platform. As shown in the above figure, from the most peripheral point of view, openmldb has a very significant feature, which is to unify the online and offline programming languages into SQL. In this way, the definition, expression and calculation expression of all features, whether offline or online, are described through unified SQL. As long as data scientists can express the features in SQL, they can do offline development and online at the same time in openmldb, so as to achieve the goal of online as soon as development is achieved.
From the internal perspective of openmldb, it includes two sets of online and offline processing engines. Offline is a batch SQL Engine, which is optimized at the source code level based on spark, so that it can better handle the logic of feature calculation. On the other hand, based on the requirements of low latency, high concurrency and high availability, an online temporal feature database is developed from scratch and optimized for this temporal data. It ensures the online and offline consistency between the two engines. The intermediate execution plan generator converts the SQL defined by the data scientist into online and offline specific execution plans, and naturally ensures the consistency of the execution plans. Finally, the goal of development and online is achieved, a lot of labor costs are saved, and the consistency of online and offline is well solved.
In fact, there are many ways to use openmldb. As shown in the figure above, some companies may use Python to develop their services before they have launched their services. However, with the growth of business, if the developed scripts need to be online, the whole process of openmldb can be used, that is, offline engine and online engine can be used. At the same time, the online and offline consistency can be fully met, so as to achieve the complete process from offline development to real-time online computing. In addition, the community has seen that some users only use the online engine of openmldb. For example, there is already a feature script, but this feature script can not meet the requirements of the previous online engine, so you can migrate it to openmldb. In addition, some users may not have the online requirement, or they may not have the real-time online requirement. In this case, they can actually use the offline engine of openmldb and the spark distribution for feature calculation, which will also be faster than the spark community version.
Part4:adlik makes AI services more economical, faster and better, and accelerates the intelligent transformation of Enterprises — yuanliya
As we all know, artificial intelligence has two important stages, including off-line training and on-line reasoning. The training phase can be divided into two steps. One is the process of data collection, preprocessing and labeling. It can also realize relevant real-time computing through open source projects such as openmldb. Through these data, we can train a model that can achieve the final training effect, and then deploy the model to the formal production environment. The second stage is the reasoning stage. It seems that the whole process is much simpler. It seems that as long as the trained model is deployed and can run, but in fact, this stage also has many pits to step on and faces many challenges. Therefore, the challenge of reasoning stage is also the challenge of AI model implementation.
As shown in the above figure, the challenge of AI model implementation actually includes four aspects: how to combine the model with the business? How to deploy the model efficiently? How to ensure that the model can run optimally? How to effectively manage models in a production environment? ZTE has provided adlik solutions to the above four problems.
Adlik is a deep learning reasoning acceleration tool, which can complete the deep learning model from training to the end-to-end tool chain that deploys to specific hardware and provides application services, and realize the efficient switching of the model from the R & D status to the production application environment. At the same time, it can also cooperate with a variety of reasoning engines to provide flexible model acceleration, deployment and reasoning solutions, help users build high-performance AI applications, and help enterprises achieve more economical, faster and better results when landing models.
The above figure is the architecture diagram of adlik, which is mainly divided into two parts. The first part is the optimizer and compiler of the model. Its main function is to improve the calculation efficiency of the model and reduce energy consumption, and finally reduce the reasoning experiment of the model in the deployment environment. The other part is the inference engine, which can support the efficient deployment of this model in a variety of environments such as the cloud side.
Using adlik can make the enterprise more economical, faster and better landing AI model
- More economical: adlik can realize adaptive optimization of engineering parameters, reduce deployment complexity and save deployment manpower. It also provides a simple and convenient pipeline for deploying the model, which greatly shortens the online cycle of the model and saves the deployment time. In addition, it provides a unified model reasoning and management interface to save the cost of model migration.
- Faster: adlik has built-in a variety of high-performance runtime for users to choose faster on-demand. Secondly, it provides a highly extensible serving SDK, which can integrate the custom inference runtime faster. Finally, it provides a flexible and easy-to-use reasoning API, which can build AI applications faster.
- Better: adlik can compress a variety of models and optimize algorithms, showing excellent performance in practice. In the face of heterogeneous deployment hardware, it can provide better end-to-end solutions. And it can achieve better reasoning performance (such as delay, throughput, etc.) and better model management according to different application scenarios.
With the increasing demand of enterprises for transformation, the choice of Bi & AI tools in the process of transformation will become particularly important. The three open source projects introduced in this activity have well solved many problems and challenges in the process of enterprise transformation. Please watch the activity “video review” for detailed technical details of each project.
At the same time, in order to continuously explore the methodology, high-quality cases, best practices and open source technologies that support the intelligent transformation of enterprises, xingce open source community will continue to hold activities related to intelligent transformation. You are welcome to pay attention to this group.
Xingce open source communityhttps://sourl.cn/TZ2Dy5
Tan Zhongyi — three stages of bi&ai’s journey to help enterprises realize digital transformation
Andy — enterprise digital transformation, starting from big data platform
Lu Mian — how to solve the data and feature challenges of enterprises’ online intelligent applications
Yuanliya adlik makes AI services more economical, faster and better