Develop and Deploy Machine Learning Models with Eviden’s Comprehensive Approach to MLOps Assessment
By Miguel Leon, Principal Solutions Architect, Data Analytics and AI/ML – Atos OneCloud
By Hakan Korkmaz, Sr. Partner Solutions Architect – AWS
Machine learning (ML) has the potential to revolutionize many industries, but developing and deploying models at scale can be a daunting task. That’s where MLOps comes in.
MLOps, or machine learning operations, is the practice of applying DevOps principles to machine learning. It helps organizations develop, deploy, and manage machine learning models more efficiently and effectively.
As organizations are thinking of implementing ML at scale, they need to make sure MLOps strategy is aligned with business needs, as it’s crucial for any organization looking to adopt machine learning at scale. MLOps involves identifying the key business objectives and understanding how machine learning can help achieve them.
Once the business needs are established, the ML strategy can be tailored to prioritize return on investment (ROI) and ensure the investment in machine learning technology and infrastructure is justified by measurable returns.
Eviden, an Atos business, is an AWS Premier Tier Services Partner and Managed Services Provider (MSP) with several artificial intelligence (AI) and ML offerings. Eviden holds the AWS Machine Learning Competency and has successfully helped customers realize the benefits MLOps brings to their business.
In this post, we will share a high-level MLOps process and approach from Eviden that’s broken down into 10 assessment steps, which enable organizations to develop and deploy models more efficiently and effectively while ensuring governance and consistency.
MLOps Assessment Steps
The illustrative example below highlights the process and approach Eviden carries out during the implementation of MLOps. Figure 1 sheds light on the MLOps process, demonstrating how Eviden optimizes manual model development and deployment workflows undertaken in client engagements. Its success is attributed to leveraging AWS artificial intelligence (AI) and machine learning services.
It’s important to note the diagram presented is not intended as an architectural drawing or business process flow. Rather, it serves as a visual representation of the 10 key focus areas for establishing a top-tier MLOps foundation. This involves a thorough examination of existing models, workloads, and requirements, as well as the templating of user profiles, data access patterns, governance and infrastructure.
Additionally, steps include the setup of a model registry, feature store, and model monitoring, culminating in the facilitation of abstract data access to enhance accuracy. The process also involves containerized model trials to achieve unified model deployment and monitoring.
Figure 1 – The focal points of Eviden’s MLOps assessment.
Step 1: Understanding Previous ML Models, Workloads, and Requirements
Gaining an understanding of previous machine learning models, workloads, and requirements involves assessing the current landscape and tech stack ecosystem. Possible questions to the customer by MLOps experts might be:
- Are you trying to forecast your customer spend?
- Is there any prediction you are trying to drive?
In order to support and provide the recommended infrastructure, a detailed analysis should be carried out by looking at all of the data science models a customer may have already deployed. These could be an image model, voice model, natural language processing (NLP) model, predictive model for a number or a value, or a category.
In parallel, Eviden looks at the deployed models’ accuracy, functionality, who has access rights to build or rebuild it, who the consumers of the model are, and whether the models have any approval requirements. It’s crucial to gather and document all relevant information about the model deployed, prior to move to the next step.
Step 2: Putting Governance in Place
Once you understand the current landscape and requirements, the next step is to put governance in place. This includes user profile templates, data access pattern templates, overall model template governance, and model virtual machine (VM) templates for infrastructure as code (IaC).
The goal is to have a system of templates that covers 80-90% of all use cases. Customers may have different departments in their organization with different requirements, sizes, and units while interacting with data science models.
Ultimately, the vision here is to combine all of the artifacts into one single platform. This is why customers need to build templates for user profiles and access rights by considering data access patterns and compliance requirements such as personally identifiable information (PII) and HIPAA. Using AWS Identity and Access Management (IAM) roles and policies, customers can ensure their data is controlled and securely accessed.
Furthermore, it’s important to consider governance that covers go-live approval processes for the developed models, stakeholder communication rules, and data management policies including data classification, access, monitoring, and storage.
Lastly, there are templates for the infrastructure that sets up virtual machines on the cloud. These templates define the necessary dependencies and specify the type of VM version to be used, such as cloud-native VM versions that are specialized for deep learning.
The most important takeaway from this step is that governance helps to ensure ML models are developed and deployed in a consistent and controlled manner.
Step 3: Developing a Self-Service Dashboard
Developing a self-service dashboard without any creativity constraint enables data science and business teams across the organization to come together for building models using MLOps. Organizations are trying to get a single pane of glass, not only to monitor models but also to access exploration and building models through a web-based user interface (UI) that typically enables you to access several systems through a single service such as Amazon SageMaker.
The goal is to enable all teams to collaborate and build models in a seamless and efficient manner. Amazon SageMaker makes it easy to build, train, and deploy pre-built and custom ML models from development to production.
Step 4: Containerizing and Scaling Data for Exploratory Analysis
It’s possible to gain a considerable amount of insights from data exploration that can help decide if a business challenge needs to be addressed using a machine learning model.
Therefore, the next assessment step in MLOps is to containerize and scale data for exploratory analysis. This enables you to orchestrate supporting infrastructure and process large datasets to streamline the identification of outliers, determine feature importance, and evaluate distribution of data making sure data is ready for modeling.
Customers can leverage fully managed tools like Jupyter Notebooks to accelerate the journey from data to insights during this investigative phase.
Steps 5-7: Create a Registry, Reusable Features, and Access to Data
- Unified model registry: Enables you to manage all ML models in development and production. This step isn’t just about building the models, it’s also about building the pipeline and capability to keep track of data lineage. This helps to rebuild all the artifacts at any point of time.
- Feature store: Provides a scalable way to access data with the right governance by putting engineered data that’s safe to be accessed by the entire enterprise. It allows the MLOps teams to cache those features to build models and control the access of available features.
- Abstract data access: Facilitates seamless entry points via configured data connectors that serve as a gateway to data lakes, streaming data, or external end point, regardless of whether data sources are in relational, object storage, proprietary, and/or massively distributed formats.
Step 8: Containerized (Massive) Model Trials
Establishing a process for containerizing and testing models enables data scientists to perform thousands of experiments quickly and arrive at a winning model.
At this stage, data scientists are trying different ML algorithms, hyperparameters, and evaluation metrics. AWS AutoML can help you to try different paths, with the ultimate goal of finding the suitable model. Data scientists or MLOps teams use AutoML to have hyperparameter tuning in a single model. AutoML management controls help to launch a scalable infrastructure with the right dependencies and make the data scientists’ jobs easier.
Steps 9-10: Deploying Models and API Access
The final steps involve deploying models in a unified manner that make them consumable via APIs by any other business application in customers’ enterprise or organization. This allows downstream applications to utilize the results or inferences generated by the models.
- Deploying models: When deploying models, it’s important to consider the production process. The aim is to have a streamlined deployment process that minimizes the effort required to get models into production. MLOps experts should clarify the questions about deployment rules for production and models availability for real-time or batch inference. Ideally, customers want a simple deployment process that allows them to move their models into production without the need for extensive rework.
- API access: Finally, there is a conversation about used APIs that affect different business units within your organization. API access is not only for prediction or inference, but also should be capable of accessing insights about models maturity, accuracy, prediction features, and observability capabilities.
MLOps is an essential practice for organizations that want to develop and deploy machine learning models at scale. The 10 steps of MLOps leveraged by Eviden enable organizations to develop and deploy models more efficiently and effectively while ensuring governance and consistency.
Eviden supports customers wherever they are in this MLOps process flow, and creates a value for the customer with the 10 steps of MLOps assessment.
Eviden – AWS Partner Spotlight
Eviden, an Atos business, is an AWS Premier Tier Services Partner and MSP with several AI/ML offerings. It has successfully helped customers realize the benefits MLOps brings to their business.