Despite persistent effort, many web pages are still not accessible to everyone. Fixing web accessibility problems can be complicated. Developers need to have extensive knowledge not only of possible accessibility problems but also of approaches for fixing them. This paper is about using the large number of accessibility issues on real websites and crowd-sourced fixes for them as a unique source of learning materials for web developers to learn how to build accessible components in a cost-efficient manner. In this paper, we present the design, development and study of CAN (Composable Accessibility Infrastructure), a crowdsourcing infrastructure that collects web accessibility issues and their fixes, dynamically composes solutions on-the-fly, and delivers the crowd-sourced content as teaching materials. Our unique CAN user interaction and system design enables end users with disabilities to both benefit from and contribute to the system without additional effort in their daily web browsing, and allows web developers to experience real accessibility issues and initiate a learning process with first-hand materials. CAN also provides an opportunity for data-driven discovery of the common implementation practices that cause accessibility issues. We show how CAN addresses a set of accessibility issues on the top 100 popular websites. We also present our user study results where web developers who had varying knowledge of web accessibility all found our system an effective and interesting platform to learning web accessibility.