- Ben Dicken of PlanetScale said that with thousands of servers, a once-or-twice-a-year failure becomes a weekly or monthly one. Separately, he said user growth, much of it driven by AI agents, strains infrastructure.
- His answer is to refuse some traffic on purpose when a database is overloaded, because a partial slowdown beats a total outage.
Ben Dicken of PlanetScale argued on the AI Engineer show that user growth, much of it driven by AI agents, is straining infrastructure. Separately, he said that at large scale, server failures stop being rare. His answer is to build systems that fail in part rather than whole. That means deciding in advance which traffic gets cut first. PlanetScale runs databases for AI companies, including Cursor.
What was said
Dicken started with simple arithmetic. An app with 100 users and one database server might see a failure once every couple of years. At thousands of servers, he said, "a once or twice a year failure becomes a weekly or monthly failure".
Overload is the second problem. He said popular databases such as MySQL, Postgres and SQLite often cope badly when they are swamped. Too many queries, too much CPU or a full memory cache can crash them unless teams take great care.
His fix is called back pressure. The system notices it is over a resource limit and starts refusing requests or connections. That keeps some users healthy while other traffic degrades.
PlanetScale's Traffic Control feature does this. Teams tag incoming traffic and give each segment a resource budget. Over budget, the database can warn the client to slow down, or kill the queries. Dicken recalled a tweet from GitHub saying its traffic rose several-fold, mostly from AI agents. He admitted that denying requests is not ideal, but said "it's much better than your whole application going down and being offline for all of your users."
Why it matters
Our reading: overload planning used to sit with the operations team. If agents become a large share of your load, the choice of who gets throttled first becomes a product and revenue decision. Should agent calls wait before paying customers do? Someone senior needs to answer that before the surge, not during it.
Buyers can ask a plain question of any vendor: when we overload, do we lose some requests or everything? Dicken also said the redundancy behind reliability costs more money and adds complexity. That belongs in the budget.
The other side
Dicken sells this kind of tooling, so his framing suits his product. His weekly-failure point is about large scale in general. He did not claim agents alone cause it, though he tied much of the recent growth to agent demand.
The talk also left a hard question open. It showed how to set budgets, but not how to rank whose traffic matters most. Any such ranking will annoy someone.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Move Fast and Don't Break Things: Scaling Databases for the AI Era — PlanetScale (2026-10-06)
