It's no secret that MLOps isn’t a one-size-fits-all solution. With organizations varying in size and structure, the complexity of how they manage and monitor machine learning (ML) models can also vary — with some organizations opting for a multi-system, multi-platform approach. For some, this strategy can work wonders, but it doesn't come without its challenges.
Highly complex integrations can be daunting for all but the most skilled users, leaving non-technical audiences struggling to participate in the ML process. In addition, navigating multiple systems can be difficult for those with less technical expertise, making it a struggle to adopt ML assets and democratize AI across the organization. And we cannot forget about security and compliance — maintaining a uniform, secure approach across different platforms requires extra effort and can introduce unnecessary risk.
Ultimately, organizations want to maintain control, governance, and orchestration of all their ML models, regardless of where they are deployed. They also want to empower their entire workforce, regardless of technical level, to be a part of the AI process. We know the multi-platform MLOps approach is not going anywhere, so how can we improve it and solve these challenges?
Cross-Platform Deployment With "Deploy Anywhere"
Enter Dataiku. As the Universal AI Platform, Dataiku allows you to bring external models from AWS SageMaker, Azure ML, Google Vertex AI, or Databricks into your projects, making them accessible to all users. But it doesn't stop there; now you can deploy models created in Dataiku using AutoML, custom code, or MLflow experiment tracking to any of these platforms, as well as Snowflake's Snowpark Container Services.
Dataiku unifies the MLOps process from deploying to monitoring.
With the ability to deploy anywhere in Dataiku, you now have full extensibility and flexibility. In some cases, rather than deploying a real-time scoring model to your Dataiku API nodes, you may want to utilize your cloud investment by deploying the model on one of your other ML platforms. This now allows for the full democratization of AI, allowing less technical personas, like data and business analysts, to build ML models and deploy them to already established infrastructures, taking stress off ML engineering teams.
And best of all? You can maintain governance throughout the entire process, allowing those ML engineers only to have to sign off on models instead of recoding entire models. It's a new era of flexibility and control with Dataiku-developed models, as they can now be deployed to external systems while maintaining standardized security requirements.
Deploy Anywhere in Practice
It's time to see this in action! We've built a model that predicts real-time flight delays. First, establish a connection to your preferred cloud platform — we'll use AWS SageMaker for this example. With various credential input methods and security settings, you can customize accessibility for everyone or select users.
Next, we'll head to the API deployer and create a new SageMaker infrastructure using the established connection. Assign a unique infrastructure ID, lifecycle stage (development, test, or production), and decide the Dataiku Govern policy for top-notch ML governance. Configure the infrastructure to use the connection settings, and we are now ready to deploy!
Set up a connection and API infrastructure for your ML platform of choice.
Back in the project, we can swiftly deploy the flight prediction model by creating or using an existing API. Name the API service and endpoint, and then make any desired test queries or API enrichments before pushing the service to the Dataiku Deployer.
Test API services and endpoints within Dataiku before deploying.
Boom! Our flight delay real-time scoring model is now an endpoint on SageMaker, accessible by both external clients and internal business functions. We can monitor this active model like any other Dataiku or MLflow model. After our model has been in production for a while, we can analyze prediction logs surfaced by Dataiku and employ the model evaluation store to compute data drift or performance drift.
Utilize Model Evaluation Stores to track prediction logs and drift of the active model.
We can even operationalize these models further with scenarios, metrics, and checks for automated orchestration of our deployment.
Maintaining Control in Cloud-Based ML Deployments With Dataiku
There is no loss of control or visibility for Dataiku-native models once they are deployed to a cloud ML platform. All the same visualizations and explainability are present and ready to be shared with team members and stakeholders. Coupled with the capability to utilize external models within your projects, the ability to deploy anywhere brings full bi-directionality to Dataiku and any of your external systems with speed and ease. It’s time for everyone, tech-proficient and business-savvy alike, to work where they feel most comfortable and still efficiently utilize their organization’s infrastructure.
View every model in production from a single screen with Unified Monitoring.
For a comprehensive view of all deployed models on every platform, take advantage of Unified Monitoring in the Dataiku Deployer.
No matter the size of your organization or the complexity of your MLOps approach, there's no better way to orchestrate and maintain your models than with Dataiku. With Dataiku as the central hub, connect everyone and embrace a new level of freedom and flexibility in managing your ML models.