Back in December 2025, I remember eating at a famous ramen shop while also sitting in an online meeting room for OpenSource4You. It was an exciting talk that I still remember today. Soon after that talk, we decided to work on QDP, the Quantum Data Plane, together.

I was also very excited to apply my newly learned CUDA skills to what could become a groundbreaking project in the QML space. With the ambition to make our data plane substantially faster, we started this exciting endeavor.

Getting Started

I remember we started from Rich’s initial design document. After reading it carefully, I shared a few suggestions during the early QDP proposal discussions. Looking back, some of those ideas eventually made their way into the implementation, which was pretty encouraging.

Later, we implemented the design step by step. Along the way, we also discovered that some assumptions from the initial proposal did not work as expected, so the current implementation naturally evolved beyond the original document.

The pace was surprisingly fast. Sometimes we had more than 10 PRs merged in a single day, all from just the four of us. It really felt like a focused sprint, with ideas, reviews, and implementations moving almost in parallel.

The Hard Stuff

I think the most challenging part came from benchmarking. The challenge came from two areas.

First, the four of us had four different hardware configurations, with different GPUs, RAM, and CPUs. This sometimes led to very different results among us. Second, it was hard to find a reasonable baseline to compare against, especially when thinking about whether the comparison was fair or unfair.

For the hardware differences, we tried to make sure the acceleration configuration was advantageous on both machines, or at least did not cause regressions.

For the baseline, I did my best to make sure it fully utilized its potential.

The Work

From December 2025 to today, May 2026, looking back at all the PRs I wrote and reviewed made me realize that I had invested a humongous amount of time into this project. It was very interesting to discover that from the stats.

• Reviewed 165+ PRs
• Delivered 75+ PRs
• Pushed the boundaries of Apache Mahout by writing four roadmaps and successfully delivering two of them so far

After the Sprint

After we finished the POC, we realized that we should put more effort into documentation to teach users how to use the efficient data plane we had just developed.

I felt the website was not in a great state, so I thought: why not migrate it to Docusaurus? It looks much better and is also better for SEO. So I started an RFC, proposed it to the community, and received many supportive comments.

Taking ownership of the website effort was not as easy as I thought. The hardest part was moving the files to a new home while properly preserving and updating the old content.

But it turned out to be worth it. The community loved it, and the docs became much more organized and clear.

Talks

To boost our work, we planned to give talks at SITCON and Community Over Code Asia and Europe.

• SITCON: everyone
• Community Over Code Asia: Jay
• Community Over Code Europe: Rich, and me

We have already delivered the SITCON talk, so feel free to take a look!

This talk also reminded me that public speaking is still something I’m learning lol.

Wrap-up

As a quick wrap-up, this has been such a nice and enjoyable journey. Along the way, I also became a committer of the project, which is definitely worth mentioning.

Overall, I am really happy with the whole process: the collaboration, the technical exploration, the fast-paced development, and everything we learned together.

Thanks

Huge thanks to Guan-Ming, Rich, Jay, Chia-Ping and everyone involved in the QDP effort. This was one of the most intense and enjoyable OSS sprints I’ve joined.