The Data Exchange Podcast: Che Sharma on how seamless end-to-end experimentation workflows supercharge product development.
Che Sharma is the founder and CEO of Eppo, an experimentation framework that integrates with modern data platforms (cloud lakehouses and cloud data warehouses). We discuss the importance of investing in experimentation tools and the power of having a well-oiled experimentation culture within an organization. Che also explains how modern data platforms enable a variety of applications, including experimentation frameworks like Eppo.
Modern data platforms have spawned a growing ecosystem of data engineering tools for areas such as data quality, data discovery and governance, data integration and much more. Meanwhile, low-code and no-code tools are making BI, analytics, and machine learning accessible to a broader range of users. Investors and engineers have created a glut of ML infrastructure and MLOps startups, but tools higher up the stack, like experimentation, have not attracted as much attention. It’s not out of ignorance: there are numerous books that extol the power of experimentation (, , ).
Up until now, experimentation platforms have mostly been bespoke solutions found primarily within technology companies. Modern data platforms make it feasible to build solutions that will enable companies to democratize and systematize experimentation. In light of the potential impact of an experimentation platform on a company’s culture and bottom line, I hope we see more accessible tools in the years to come.
The key thing with experiments is to know, when are you helping and when are you hurting. We’ve seen a number of papers that have basically acknowledged this one point, which is no matter where you are, once you start measuring things, you realize of the 100 experiments you were running that you thought were going to improve things, probably only 30% of them actually improved things, another 30% made things worse. So once your eyes are peeled back, and you start to see these outcomes, it becomes crucially important to know which experiments are happening and to know it comprehensively across the org.
The reason I started an experimentation company is, on top of that kind of tangible ROI, there’s so many amazing second order effects that happen from establishing an experimentation program. So one of the ones that I felt over and over with every company is, before you run experiments, if you’re a data team, you primarily are living in the world of reporting. Maybe some ML models, but primarily reporting. The problem with reporting is it doesn’t tend to influence product decision making very much. Take an example where you have some top line revenue dashboard, and you are saying, here’s how much money we’re making every year. That’s great for the CFO for the board, they can look at these numbers. But if you’re a product manager, you stare at that dashboard, and you wonder what to do with that … But once you start running experiments, suddenly every product team has a really intimate connection with metrics. A major goal for every data team is to kind of build this muscle of kind of leveraging metrics in decision making.
… But the thing that gets me most excited about experimentation is that once you nail fluid, seamless end-to-end experimentation workflows, it creates this entrepreneurial culture, where suddenly people can take risks, you know, they can try things out and see if customers like it or not. It’s this amazing culture that once you’ve been at a place like an Airbnb, or a Netflix, you know, you really start to appreciate that you can have bottoms up product development.
Highlights in the video version:
- Introduction to Che Sharma, Founder and CEO of Eppo
Examples of experiments that people do that warrants tooling
What are companies that you look up to when it comes to experimentation?
When should a company start thinking more formally about experiments and an experimentation platform?
Infrastructure is needed for experiments
Why should people start running experiments?
Started experimentation company due the amazing second order effects that happen
Reporting does not influence product decision making until you start running experiments
Key components, parts, and features of a modern experimentation platform
First you need a data scientist and second you need some way of modeling metrics
Calculate metrics, run statistical tests, and run a bunch of diagnostics
Make health checklist before running an experiment
The reporting side of experimentation ends up being a pretty big deal
Formal workflow for proposed experiment: collaborate on design, calculate run time, measure & report results
Experiment tracking, experiment management tools, and organization wide visibility
Experiment platform and multiple teams doing things simultaneously
Introduction to Eppo and the rise of the modern data stacks
What did you choose to focus on and prioritize in terms of building this modern experimentation platform?
How much guidance is provided when an experiment is set up?
Everyone wants to look at everything all the time when you run experiments
The end user is a Product Manager and the UX will display in a way that the Product Manager can run it
Features available today at Eppo
Modern data ecosystem is what you need to engage with Eppo and scale sophisticated data practices
What’s on the roadmap in the next six to twelve months?
Is the experimentation world thinking of ML? What is the notion of responsible experiments?
Is there a need for more transparency in the experiment world?
If the control group is starting to get harmed then shut down the experiment
Will experimentation platforms be much more prevalent in two to five years?
Are more people running experiments on mobile devices since there are lots of mobile apps?
What is the role of the data scientist in the experimentation world?
No code ML as the ML world is heading toward a major infrastructure upgrade
Responsible experimentation and adopting experimentation
Will see how prevalent experimentation platforms have played out a year from now
- A video version of this conversation is available on our YouTube channel.
- “Model Monitoring Enables Robust Machine Learning Applications”
- “Taking Low-Code and No-Code Development to the Next Level”
- “Data Quality Unpacked”
- Hamel Husain: “Deploying Machine Learning Models Safely and Systematically”
- Nikhil Muralidhar: “MLOps Anti-Patterns”
- Travis Addair: “The Future of Machine Learning Lies in Better Abstractions”
Subscribe to our Newsletter:
We also publish a popular newsletter where we share highlights from recent episodes, trends in AI / machine learning / data, and a collection of recommendations.
[Photo by National Cancer Institute on Unsplash.]