Shut down the servers

8 years ago my business partner and I were working on an online background removal editor. You’d upload an image, apply a few red/green marks, and our algorithm would separate out the foreground from the background based on that. This was before AI took over the world, so it actually required user input.

But. The algorithm wasn’t all that clever. Sure, it ran fast, so you’d get quick feedback, but it was kind of “whack-a-mole” for challenging images – fix one error and another would crop up. We were close to a year into the project and the well of ideas for how to fix the algorithm issues had started to run dry. It was time to make some decisions.

Ever quality focused, my business partner went through and clipped what seemed like hundreds of images (so probably 20), his frustration growing with each and every one of them. At the end he concluded: “it’s not good enough”.

We didn’t know how to make it any better. “Not good enough” meant the project was dead.

“Do you want me to shut down the servers?” I asked.

“Yes”

“Mind if I post it on hacker news first?”

“Sure, do whatever you want. ”

The frustration of countless hours of banging our heads against algorithmic walls was brimming over. There’s a special empty bitterness when you’re about to flush the output of almost a year of your life. And it wasn’t our first time, nor would it be our last, but more on that in another post.

Well, if we’d already lost a year, we certainly didn’t have anything more to lose by posting about it.

Ever the optimist I was of course in part still hopeful – I’ve never been in the “we never could have imagined”-camp. But I knew he had a point. It wasn’t anywhere near as good as I wanted it to be, nor how good it should have been. We’d called it “Clipping Magic”, but it wasn’t, well, quite magical enough. Sigh.

Click, click, post. Wait. Reload. Two points. Wait. Reload. Six points. Wait. What? Before I know it the post is catching on. And people are liking it. They’re clipping images. They’re commenting. And they are upvoting!

As the images start flooding in, the adrenaline kicks in too. The post is on fire. I’m on fire. We’re on fire. THIS IS ACTUALLY HAPPENING! The rush of that initial launch is… just incredible. The light shakes from the endocrine rollercoaster, the obsessive desire to answer every single comment, the reload, read, comment, reload, read, comment mania.

Well, the site’s on fire too. A comment saying something about “Network error, can’t connect”. WTF? Hugged to death? But this thing should scale just fine? We’d obviously done zero load testing, but still? Frantic dashboard debugging ensues. Phew, simple solution: just add more servers. We’d launched with “manual auto scaling”, and the “autoscaler” had been too busy posting to notice the cluster wasn’t sized right. Well, well. At least the correct cluster size wasn’t zero.

The post went on to garner more than 500 upvotes, which in 2013 was an absolutely massive number. If I recall correctly we stayed on the front page for more than 24 hours. And over the last 8 years we’ve clipped more than 100M images, served more than 200,000 customers, improved the algorithm far, far beyond what it was when we launched, and made the transition to the post-AI world so we’re now the only background remover which seamlessly merges true full auto with a Smart Editor.

Thanks for reading, and be sure to try out the (now much better) service!