Why You Can't Build AI Without Progressive Delivery

Progressive Delivery: A Deep Dive into the Four A’s & Beyond

Key Concepts: Progressive Delivery, Four A’s (Abundance, Autonomy, Alignment, Automation), Continuous Delivery, DevOps, Feature Flags, Service Mesh, AI/ML Workloads, Jerk (physics concept), Cloud Computing, Observability, Radical Delegation.

Introduction

This discussion centers around James Governor’s new book, Progressive Delivery (co-authored with Kim Harrison, Heidi Waterhouse, and Adam Zinman), and explores the core principles and practical applications of progressive delivery in modern software development. The conversation highlights how progressive delivery builds upon existing DevOps and Continuous Delivery practices, particularly in the context of cloud computing and the emerging landscape of AI/ML workloads. The central framework discussed is the “Four A’s” – Abundance, Autonomy, Alignment, and Automation – which are presented as essential pillars for successful progressive delivery.

1. The Evolution of Progressive Delivery & The Need for a New Framework

The discussion begins by establishing that while Continuous Delivery (CD) laid a strong foundation, the advent of cloud computing and open-source technologies necessitates a more nuanced approach. Traditional CD concepts were largely pre-cloud, lacking the inherent scalability and flexibility offered by modern cloud infrastructure. Progressive Delivery addresses this gap by focusing on safely releasing software to users and gathering feedback during the release process, rather than simply automating the deployment pipeline.

Adam Zinman’s contribution is highlighted, particularly his analogy to physics – viewing software delivery as managing “forces of motion” and minimizing “jerk” (sudden changes in velocity, analogous to disruptive deployments). This framing led to the development of the Four A’s framework. The initial impetus for the framework stemmed from frustration with the lack of practical use cases presented for service mesh technology, and a subsequent realization of the need to integrate user feedback into the delivery process.

2. The Four A’s: A Detailed Examination

• Abundance: This refers to the readily available resources provided by cloud computing and open-source software. Historically, deploying multiple environments (e.g., for blue/green deployments) was prohibitively expensive. Cloud infrastructure allows for the easy and cost-effective creation of numerous instances, enabling more sophisticated experimentation and rollout strategies. The conversation acknowledges a potential tension with current resource constraints (e.g., GPU availability for AI), but emphasizes that abundance still represents a significant shift in capability. The example of Twitter’s early API-driven openness is cited as a historical example of abundance.
• Autonomy: Autonomy is defined as the ability of individuals to act independently. The discussion stresses that while software development is a team sport, empowering developers with the freedom to experiment and solve problems without excessive gatekeeping is crucial. Amazon’s “two-pizza team” model is presented as an example of fostering autonomy. The importance of removing roadblocks and enabling developers to work independently is emphasized, as delays and dependencies create “jerk” in the delivery process.
• Alignment: Alignment focuses on ensuring that all stakeholders – developers, product managers, and the business – are working towards the same goals. The conversation highlights the historical challenges of alignment with frameworks like ITIL. A key point is the potential for misalignment when teams have too much autonomy, as seen in Amazon’s experience. Amazon is currently implementing a new framework focused on “cost to serve the customer” to improve alignment. The discussion introduces the concept of “radical delegation,” where product managers, rather than developers, may make the final decision on release timing.
• Automation: Automation is the programmatic execution of repetitive tasks. While not a new concept, the sophistication of automation tooling has dramatically increased, particularly with the rise of “configuration as code.” Automation is essential for implementing progressive delivery strategies, enabling features like traffic routing, rollbacks, and user behavior analysis.

3. Progressive Delivery & AI/ML Workloads

The conversation explores the intersection of progressive delivery and the rapidly evolving field of AI/ML. It’s argued that the principles of progressive delivery are essential for managing the inherent uncertainty and potential risks associated with AI models. Concepts like versioning, rollbacks, and observability – core tenets of progressive delivery – are directly applicable to AI/ML deployments. Thomas Kurian (Google) is cited as having stated that AI-based development cannot be effectively pursued without progressive delivery. The unpredictable behavior of large language models (LLMs) necessitates a cautious and iterative rollout approach, making progressive delivery a critical practice.

4. Real-World Examples & Case Studies

• Amazon: The discussion delves into Amazon’s journey with progressive delivery, highlighting their initial challenges with alignment due to highly autonomous teams. Their current initiative to standardize tooling and focus on “cost to serve the customer” is presented as a response to these challenges.
• Twitter (Early API Access): Cited as a historical example of abundance, where open API access fostered innovation and integration.
• LaunchDarkly: Mentioned as the company where Adam Zimman initially developed the concepts that contributed to the progressive delivery framework.

5. Key Arguments & Perspectives

• User-Centricity: A central theme is the importance of prioritizing the user experience. Charity Majors’ observation that “five nines” of uptime are meaningless if the user is unhappy is emphasized.
• Building on Existing Practices: Progressive Delivery isn’t a replacement for DevOps or Continuous Delivery, but rather an evolution that builds upon their strengths.
• The Need for Process in a Rapidly Changing Landscape: While AI is introducing unprecedented levels of innovation, the need for repeatable, well-defined processes (like those embodied in progressive delivery) remains paramount.

Conclusion

Progressive Delivery offers a comprehensive framework for navigating the complexities of modern software delivery, particularly in the context of cloud computing and AI/ML. The Four A’s – Abundance, Autonomy, Alignment, and Automation – provide a practical guide for organizations seeking to improve their release processes, reduce risk, and deliver value to users more effectively. The conversation underscores that while technology is constantly evolving, the fundamental principles of user-centricity, iterative development, and robust automation remain essential for success. The book’s value lies in its ability to synthesize existing best practices and provide a clear roadmap for building a more resilient and responsive software delivery pipeline.