This is a joint post by Jean-Baptiste Thomas, Pure & Yaron Haviv, Co-Founder & CTO of Iguazio

You’ve trained and built models using interactive tools over data samples. Now you’re working on building an application around them to bring tangible value to the business. However, a year later, you find that you’ve spent an endless amount of time and resources, but your application still isn’t fully operational or performing as well as it did in the lab. You’re not alone. According to industry analysts, that’s the case in more than 85% of artificial intelligence (AI) deployments.

Moving from research (using small data sets) to robust online production environments is far from trivial. Production environments need to process large amounts of real-world data and meet application performance and SLA goals. The move from lab to production involves multiple practitioners across teams (data scientists, data engineers, software engineers, DevOps, SecOps, etc.,) who need to work in sync to build production-grade AI applications.

To effectively bring data science to production, enterprises need to change their mindset.

Most organizations today initially focus on research and model development because that is often seen as the natural first step in the data science process. However, taking a production-first approach is a sign of operational maturity and strategic thinking. Enterprises that take a production-first approach can plan for the future and make sure to harness the tooling to support the data science process as they mature and later scale.

To industrialize AI, a machine learning (ML) solution that delivers scalable production-grade AI/ML applications is needed. It will help streamline the workflow and foster collaboration across teams, with a focus on automation to improve productivity.

One option is to build an AI solution from discrete open-source tools, infrastructure components, and/or cloud services, but that requires a significant amount of IT and engineering resources. Another option is to choose a preintegrated and managed solution, which lets you focus on solving your business problem instead of data wrangling. This can result in a higher level of productivity and better performance.

Adopting a Production-First Mindset and the Tools to Support It

Figure 1: Pure & Iguazio: Production Focused MLOps Architecture

To effectively scale AI, enterprises need to adopt a production-first approach and consider production as early as possible in the data science process. Ideally, you should start by defining the production pipeline, then consider the required infrastructure, development tools, and practices needed to support it. Begin developing your models after you’ve thought of—and accounted for—what you’ll need in production.

When you’re ready to add additional pipelines and scale your data science practice, you’ll have the building blocks in place to do so effectively and efficiently. Without the right architecture, it will be harder and harder to support diverse and growing business needs.

As the organization scales, what matters most is the marginal cost of adding more pipelines. With the right architecture, that cost is relatively flat since it’s mostly abstracted by the system.

AI/ML production environments have five main components:

  • Data acquisition and feature engineering: Bringing data from offline or live online sources and transforming it into meaningful and usable data, then registering it for use in training and serving
  • Scalable data repository: Providing an enterprise-grade, high-performance shared data store for efficient hosting of both structured and unstructured data
  • Training pipelines: Taking a snapshot from the production data and training and testing models that will be used in production
  • Serving pipeline: Getting user requests or events, processing them, enriching them with additional precalculated features, applying AI/ML logic, and responding or issuing actions
  • Data and model monitoring: Monitoring the production data quality and model behavior to detect anomalies and issues with accuracy, detecting drift, and triggering alerts or corrective actions (like retraining a model)

The components used in the production environment are developed and tested with a variety of tools, like interactive data exploration tools, notebooks, IDEs (e.g., PyCharm or VS Code), model training frameworks or services, Git, continuous integration/continuous deployment (CI/CD), and automation tools. It’s critical that every component developed is tested at scale and deployed automatically without the need to refactor the code or go through a series of manual and labor-intensive tasks which will lead to significant unnecessary effort and long delays.

To achieve this goal, enterprises need to adopt a few guiding principles:

  • Foster collaboration across data science, engineering, and DevOps teams by adopting the same practices and the same toolset across these teams.
  • Reuse the same code base for research, small-scale testing, and production deployments; perform incremental changes; and avoid reimplementation.
  • Focus on automation across development and deployment.
  • Use scalable enterprise-grade infrastructure to host data and compute across development and production environments.

These goals can be achieved with a cloud-native approach, where Kubernetes, containers, and shared storage serve as the common infrastructure, with a set of managed services on top which address the different elements in the development and project lifecycle management.

One of the ways to significantly reduce development and deployment overhead is by adopting serverless technologies, which convert simple code to scalable, hardened, and managed microservices.

Machine learning operations (MLOps) orchestration frameworks based on serverless technologies can automatically convert code to serverless engines addressing different types of workloads (batch jobs, real-time Nuclio functions, Dask, Spark, Horovod, etc.)

MLOps platforms should provide end-to-end automation. With a feature store, features can be generated and shared across teams and projects. Batch and real-time pipeline orchestration, model serving, model monitoring, and retraining logic together all save valuable time and accelerate the data science process.

The underlying infrastructure should be built for scale and high performance, support structured and unstructured data, and be enterprise-grade for utmost security. Organizations need reliable and scalable infrastructure to handle massive amounts of data and computation for real-world data science projects.

Bring Your Data Science to Production with Pure and Iguazio

Iguazio is an MLOps platform that automates and accelerates the path to production.

Its serverless and managed services approach shifts the focus from trying to cope with a sprawl of Kubernetes resources to delivering a set of higher-level MLOps services over elastic resources, therefore reducing significant capital and operational costs and accelerating time to market.

Iguazio includes a central portal, UI, CLI, and SDK to manage the entire MLOps workflow with minimal integration efforts.

However, this can’t work without a high-performance, scalable, and robust storage layer underneath. Data science projects depend on processing large amounts of structured and unstructured data in short times, and data access must be secure and highly available.

Pure Storage® products are optimal for the challenges of data science, providing speed and simplicity at scale. Unstructured data can be organized in files or objects on Pure FlashBlade® and accessed directly by the applications. Iguazio’s real-time feature store can access structured data on high-performance, low-latency NVMe storage using Pure FlashArray™. Portworx® native Kubernetes storage can be used to support the persistency needs of different stateful microservices.

Iguazio with Pure Storage

 Using Iguazio with Pure Storage solutions, enterprises can:

  • Implement an automated, scalable, and performant enterprise-grade MLOps platform.
  • Gain an agile ML solution where compute and/or storage scales as your ML needs evolve.
  • Reduce the overall storage footprint by an order of magnitude (by leveraging Pure compression and deduplication technology).
  • Provide enterprise-class storage features (snapshots, backups, sync, encryption, etc.).

By using Pure and Iguazio, enterprises can unlock the value of their data and bring data science projects to life in an efficient, automated, and repeatable way. Pure and Iguazio empower enterprises to continuously roll out new AI services, focusing on their business applications and not the underlying infrastructure.

Please visit our partnership overview for additional information