SAP Inside Track Bangalore 2024: FEDML Reference Architecture for Hyperscalers and SAP Datasphere

[navyakhurana]

Hey Folks, it's Navya again!

Well, as a fellow SAP enthusiast, I know how thrilling it is to dive into the world of SAP learning about the different SAP technologies, collaborating and networking with people and much more. 

Sindhu Gangadharan, SVP & MD SAP Labs IndiaAmongst this, starting this year with our first edition of SAP Inside Track Bangalore 2024, the excitement is palpable, and I'm thrilled to be a part of it. But what makes this event even more special is the opportunity to share my experience and delivering a session in it. So, as we embark on this inaugural journey, let me weave you through the highs, lows, and invaluable lessons learned along the way.

SAP Inside Track events are all about grassroots collaboration and sharing expertise on SAP-related topics.The day kicked off with keynote address by Sindhu Gangadharan, SVP & MD, SAP Labs India and Head User Enablement with her insightful thoughts shedding light on the SAP eco-system and setting the tone for the day, inspiring attendees to engage actively, learn from one another, and leverage their collective expertise to drive innovation and success in the realm of SAP technology.

This year's SIT focused on eight tracks, each offering a wide array of topics spanning SAP Business Technology Platform(BTP), Application Development, Data Analytics, Artificial Intelligence & Automation, UI/UX Design, SAP Concur, and S/4 HANA.

Session on FedML with Hyperscaler & SAP Datasphere [In frame speakers: Navya and Lalit]Amidst the diverse array of topics, I had the privilege of presenting the session on "Federated Machine Learning Reference Architecture for Hyperscalers and SAP Datasphere" with my colleague a.ka. mentor @lalitmohan . It was an exhilarating experience to share insights and delve into the intricacies with fellow enthusiasts.

Session on Federated Machine Learning Reference Architecture for Hyperscalers and SAP Datasphere

During the session, we delve to an in-depth exploration of federated machine learning libraries, that applies the Data Federation Architecture of SAP Datasphere for intelligently sourcing SAP as well as non-SAP data for Machine Learning experiments, allowing for model training across distributed data sources while preserving data privacy and security. This innovative architecture enables hyperscalers and organizations leveraging SAP DataSphere to harness the power of machine learning at scale without compromising sensitive data.

But before getting into the usecase front, let us discuss What is FedML? Why FedML?

What is FedML?

Federated Machine Learning is an approach to machine learning where the training data remains distributed over multiple devices or locations, without centrally aggregating it. Instead of bringing all the data to one location, federated learning brings the model to the data or distributes the model across multiple devices or servers. This approach allows for privacy-preserving machine learning, as the raw data remains local and is not shared with a central server or authority.

Why FedML?

• Data Federation Architecture: FedML leverages SAP Datasphere's Data Federation architecture to intelligently source both SAP and non-SAP data for machine learning experiments.
• Integration with Hyperscalers: FedML seamlessly integrates with hyperscaler platforms, providing compatibility and scalability for machine learning processes.
• Abstraction of Data Connection and Load: FedML abstracts the complexities of data connection and loading, simplifying the process of accessing and utilizing data from various sources.
• Flexible Model Training: FedML offers flexibility in model training by providing support for user-provided training scripts.
• End-to-End Integration & Minimal Code Requirement: FedML facilitates end-to-end integration of the machine learning pipeline, covering data connection, data loading, model training, deployment, and inferencing stages.

Reference Architecture

After familiarizing ourselves with the fundamentals of integrating FedML with SAP Datasphere, we proceeded to explore two specific use cases:

• Implementing Databricks FedML with SAP Datasphere

Business Challenge: Large-scale distributed data is crucial for analytics and decision-making. Much of this data is used for predictive modeling and machine learning (ML). With the proliferation of ML platforms, businesses can now efficiently train and deploy models. However, accessing SAP data presents challenges for Data Scientists using Databricks. They depend on data engineers to create pipelines for SAP data extraction and preparation, which is costly and time-consuming. Integrating non-SAP data with SAP data for ML experimentation adds complexity.

Proposed Solution: FedML Databricks library provides functions that enable businesses and data scientists to build, train and deploy machine learning models on ML platforms and eliminates the need for replicating or migrating data out from its original source. By abstracting the data connection, data load, model deployment and model inference on these ML platforms, the FedML Databricks library provides end-to-end integration with just a few lines of code.

In order to try out the solution for this scenerio, you can refer to the Github Repo or refer to this amazing blogpost 

• Integrating GCP FedML with SAP Datasphere

Business Challenge: In enterprises, for conducting machine learning experiments on hyperscaler platforms, the data often has to be extracted out of source business systems and duplicated in cloud stores of the ML platform. This causes issues such as data inconsistencies and increased TCO resulting from additional data storage costs due to data duplication and expensive data pipelines. Moreover, comparative sales performance analysis on live data is not currently possible with data from cross cloud sources.

Proposed Solution: GCP FedML Library helps you creating an end to end automated solution for ML experiments on hyperscalers(GCP in this case) without moving or duplicating data. This solution uses SAP Datasphere data federation architecture, a unified semantic layer helps model the queries across distributed data sources without the need to extract any data out from anywhere

In order to try out the solution for this scenerio, you can refer to the DC Mission, Github Repo or refer to this amazing blogpost 

In case you want to try out any other Hyperscalar scenerio with FedML and SAP Datasphere, you can refer to the following DC Mission: AWS FedML with Datasphere, Azure FedML with SAP Datasphere

Following our demo on the above scenarios, we concluded with a Q&A session to address any queries, concerns, or feedback from the audience. During the Q&A, participants had the opportunity to delve deeper into the intricacies of federated machine learning and its applications in various industries. 

Conclusion

So there you have it – my whirlwind adventure at SAP Inside Track 24. I am grateful for the opportunity to share my expertise, connect with fellow enthusiasts, and be part of this incredible community-driven event. As I reflect on this experience, I am inspired to continue pushing the boundaries of innovation and making meaningful contributions to the SAP ecosystem. 

SAP Inside Track'24 Team

[lalitmohan]

Great blog post @navyakhurana  - Thanks for sharing !

[TuncayKaraca]

Hello @navyakhurana,

It seems to be a great stuff with SAP Datasphere and --If I'm not mistaking-- SAP Federated ML Python libraries (FedML) 

Thanks for sharing all info and discovery center missions. 

Regards,
Tuncay

[navyakhurana]

Hello @TuncayKaraca,

Yes, Exactly! That's the Reference Architecture for FedML.

Incase you are interested to know more about it, you can go through fedml-the-federated-machine-learning-libraries-for-hyperscalers-2-0  blogpost which gives you much better idea about FedML libraries and has links to consequent blogpost for it's execution with different hyperscalers.

Thanks & Regards,

Navya

[ayushbehl]

That was a great session @navyakhurana  @lalitmohan 

[TuncayKaraca]

Thank you, @navyakhurana for the reply. I've bookmarked the link.