CUFP 2013: Ryan Trinkle: Enterprise Appointment Scheduling with Haskell

Ryan Trinkle discusses migrating skedge.me's backend architecture from Grails to Haskell to handle enterprise-scale scheduling.

skedge.me is an online scheduling platform that allows businesses to completely automate the process of making appointments, such as customer visits, job interviews, and tutoring sessions. Our clients include McDonald's, Sephora, the FBI, and Stanford University. Serving such a wide range of clients requires flexibility to implement business logic for a wide range of existing business processes, scalability to handle large transaction volumes, and accuracy so that our clients can be confident that when we make a scheduling commitment on their behalf, they'll be able to keep it. As skedge.me grew from the small business market into the enterprise space, our clients' needs began to outstrip the capabilities of our Grails-based back-end architecture. Eventually, we made the decision to replace it, and we went with Haskell. Using Haskell has presented some challenges. We've run into our fair share of issues with the difficult learning curve and occasional sparseness of the Haskell toolchain, but the biggest challenge for us has been the impedance mismatch between the relational database paradigm and the functional paradigm. In contrast to object-relational mapping, algebraic-datatype-relational mapping is relatively uncharted territory. Although some ORM-style libraries, such as Persistent, have been written for Haskell, their support for more complex functional datastructures, such as sum types, finite maps, and deeply nested datastructures, leaves much to be desired. This has forced us to make difficult trade-offs between structuring our data in a way that best suits the application layer and structuring it for the database layer. On the whole, however, Haskell has made our product - and our lives as developers - much better. Haskell's powerful type system and GHC's powerful optimizer has allowed us to make major architectural decisions that would simply be impossible in practically any imperative language. We use a stack of monad transformers to create application layers with increasing levels of safety – transaction, data integrity, security, and business logic - which allow us to separate these concerns and give us much greater confidence that layer invariants are not violated. Our security policies are implemented using multiparameter type classes, which allow us to implement detailed per-client security policies, omitting coverage of components a client doesn't use, while ensuring total coverage and consistency of the components the client does use. And, we've built our core calendaring code around a powerful library of custom immutable datastructures that allow us to easily parallelize large jobs while keeping code small, legible, and reusable. These advantages have enabled us to deliver a product that is more customizable, higher performance, and better quality than its predecessor, with only a fraction of the development effort. I've been working with Haskell for 8 years personally and 4 years professionally, ever since I switched from C