Description The MLOps landscape is exploding with new tools for specific needs. As a ML practitioner, it can be disorienting which tools to use, and even harder, how to do it in a production setting. Many resort to gluing together their preferred tools from scratch, and end up maintaining internal MLOps tooling with clunky abstractions and hard to extend modules. This begs the questions - What if you hit a limitation of the tool? Are you locked-in by vendors and what are the costs of switching? What about experimenting with new tools? These must be considered when building a future-proof production ML solution.
ZenML is an open source framework to develop tool-and-infrastructure agnostic ML pipelines. It lets you separate the ML logic and infrastructure code so you can build your pipelines without worrying about the stack early on. ZenML also provides integrations that let tools talk to one another seamlessly. As ZenML is a framework, you can write a custom integration for tools you don’t find on the list.
In this talk, we'll build an end-to-end ML pipeline from scratch and scale them into a production-ready stack. The pipeline contains built-in best practices using data validators such as Great Expectations, experiment trackers such as MLflow, and model serving tools such as Seldon. On top of that, we show how you can swap out each component independently. Finally, we'll demonstrate how you can scale your pipeline from a local run to full-fledged cloud infrastructure with just two simple commands.
Takeaways 1. It can be cumbersome to understand the burgeoning MLOps tooling landscape. 2. One way to manage the complexity is to derive abstractions from common workflows in MLOps. 3. ZenML is a framework that uses such abstractions to create tooling and cloud-agnostic ML pipelines.
