Designing MLFlow
An open source platform for the machine learning lifecycle
THE CHALLENGE
How to simplify machine learning development to democratize AI?
In the era of artificial intelligence, only a few companies are leveraging machine learning to deliver the business impact. Why is that? As our CTO Matei Zaharia said in a blog post of introducing MLFlow:
Everyone who has tried to do machine learning development knows that it is complex.
How do we simplify machine learning development to democratize AI?
MY ROLE
As a researcher
I collaborated with my design manager and machine learning (ML) product managers to profile the ML persona and visualize the user journey.
As a designer
I lead the design on MLFlow project and work closely with product managers and engineers to make AI simple.
DESIGN PROCESS
We started MLflow project from late 2017 and released the first alpha version at Spark Summit on June 5th, 2018. Since then, it starts to get traction, and more and more features are added. This case study tells the story behind the MVP release. The high-level process for the project includes the following:
RESEARCH
Do the homework and speak the same language
For a designer, this's probably the hardest part. The ML domain is complex and full of jargons. Besides reading ML tutorials and taking online courses, I also relied on our ML experts walking us through the ML workflow and the taxonomy. So we can all speak the same language.
Learn from the interviews and visualize the journey
Next, we conduct user interviews to understand ML engineers' workflow further and hear the pain points in their day-to-day work. Patterns were emerging around model training & deployment & management. The feedback from the field also repeated similar challenges. We synthesized the learnings into a user journey map for easy reference and knowledge sharing.
DEFINE
Research key findings
From the research, we identified the top five pain points for the ML development and decided to address the first two first. Why? We did not see other products solving these two well. Also, we aim for the whole machine learning lifecycle, undertaking the first two is doing the groundwork for other challenges.
Design scope for MVP
To help users manage experiments and reproduce runs, we provide a tracking UI to record and query experiments runs. All the input data are captured so users can rerun and generate the same result easily. Also, to help user output a model that achieves high performance, we provide a compare feature for run metrics visualization and comparison.
DESIGN
Final designs for MVP
The well-defined scope streamlined the design work afterward. By understanding the timeline and technical constraints, and working closely with engineers to iterate, we soon arrived at an MVP design.
Iterate faster with Sketch symbols
After the first release, I also made all the styles and the components Sketch symbols and added the Sketch file as a library. It improves design efficiency and helps me iterate fast on designs.
SHIP IT
Keep calm and carry on
After we released MLFlow at Spark Summit on June 5th, 2018, we got great media coverage, positive community feedback, and lots of customer interests. It was a big milestone.
Shipping is about learning
As more customers and more community users adopt MLflow, we hear great feedback. People blogged, tweeted, or left comments on MLflow GitHub page. Shipping is about learning in this open community! We continued working with customers and the community to iterate and add more features. We also start to work at solving other challenges in model management and deployment. Stay tuned!
THE RESULT
It has been an amazing ride, and I deeply appreciate the opportunity of working with and learning from the team. What I feel most proud of, is making AI simple for Marie and others.
Stars on MLFlow GitHub Page in a month
AI will change the world. Look who will be changing tomorrow's AI?