Curiosity Has a Cost

One of Square’s earliest outages back in 2011 was self-inflicted.

It happened on a bright Saturday morning, one of the busiest times for Square. I was in our old office in the San Francisco Chronicle Building, one of a handful of people who’d show up on weekends.

A few desks over, our PM was reviewing metrics on the admin dashboard, which was lovingly handcrafted with custom charting libraries and hosted on our monolithic Ruby-on-Rails app. Frustrated by how slow the dashboard was loading, the PM rage-held Cmd+R for 30 seconds. Little did he know, he was firing off dozens of expensive analytical queries against the production MySQL database.

Needless to say, all the MySQL threads became blocked. The PM looked up and asked, “Hey… is the site down?”

After some confusion and frantic log-diving, we eventually found the culprit — and a memorable post-mortem ensued.

We overhauled the analytics infrastructure at Square many times after that. It’s been over a decade since I left, and I’m sure the data stack is unrecognizable to those early Squares.

The Current Insight Equilibrium

A lot has changed since 2011. The kind of data infrastructure the biggest tech companies once had to invent is now available off the shelf. Modern data warehouses and lakes can handle what used to take entire teams to maintain. The plumbing, storage, and compute patterns—while still evolving—have largely settled. Hopefully, no company founded in the last five years has had to custom-roll its own analytics stack.

And the blast radius of an expensive query? It’s smaller now. At least it won’t take your site down anymore. But it might still leave you with a painful cloud bill—and a knock on the door from your CTO or CFO.

Entropy is still expensive: the stray batch job that scans too much data, the forgotten pipelines and dashboards that seem to multiply on their own. At Opendoor, our cloud bill started in the modest five figures in 2015. But by 2020, I was legitimately stressed about a seven-figure bill growing 50%+ year-over-year, driven mostly by data and ML workloads. That kicked off a FinOps project (a story for another day). Compounding costs are scary.

In any case, we’ve reached something close to a steady state. The economics of analytics are no longer constrained by compute or storage, but by analyst productivity—the human effort tolerated by analysts to extract insights.

Unleashing Agents

For the past decade, analytics infrastructure has been built to scale to the limits of human analysts—their attention, their time, their ability to iterate. Dashboards, queries, and pipelines all assume a finite number of people clicking “Run.”

But that assumption is about to break. Some reports estimate that over 30% of new code checked in today is written by AI copilots or agents. It’s only a matter of time before analytics follows suit—after all, analysis is a form of coding. And I suspect we won’t just see 30% of analysis written by agents, but 10x (or even 100x) more analysis done.

We are curious creatures. Just as there has been an insatiable appetite for software, there will be an insatiable appetite for analysis. Once the friction of getting quality analysis (some might say, decision-grade analysis ;) disappears, the volume of analysis will explode to meet the latent demand: some blended average of human curiosity and executives realizing that genuinely good analysis drives better decisions.

So what happens then? Imagine thousands of agents holding Cmd+R across every dashboard in your business simultaneously. Better yet, imagine these agents creating dashboards dynamically—and then holding Cmd+R.

Okay, hopefully the stack is more event driven, but you get the point. The Query Flood is coming. And this time, it might just take down your analytics infrastructure after all.