TL;DR
- During the worst moments of the largest volume spike our platform had ever seen (in July 2025), about 11% of requests to our platform failed. That was unacceptable. We now run at an error rate of 0.111%, while carrying more than ten times the load.
- We set a standard for load testing: the platform has to survive at least twice the heaviest load we have ever handled in production, on demand, every time we test for it.
- Kraken sustains more than 230,000 successful requests per second in production, across five separate full-scale tests with zero incidents.
- On February 5, 2026, one of the sharpest selloffs of past twelve months, every major service held with no incidents.
Peak demand is the only real test
No trading platform is judged on an average Tuesday. It is judged on the day Bitcoin sets a record, or the day a token sale opens and everyone arrives in the same minute. Those are the moments that tell you what a platform can actually carry.
The quality and stability of our platform is the product.
In mid-July 2025, Bitcoin broke through $120,000 for the first time, part of a rally that carried it above $123,000 within days. On Kraken, roughly 135,500 clients hit errors during that window, about 40,000 of them while funding an account. At the worst moments, around 11% of requests failed outright.
Three months later, the first token offering on Kraken Launch saw somewhere between 60% and 80% of active sessions drop for two and a half hours. Even after things stabilized we were still losing about a third of requests during traffic spikes.
Over the past year we changed how we test for load, so we find those limits ourselves instead of discovering them mid-rally.
War games
The standard we set: the platform had to survive at least twice the heaviest load it has ever handled in live trading, on demand, whenever we choose to test it. Double the record, as the floor.
To test against that we built full-scale load tests, known internally as war games, that throw a manufactured flood of trading activity at production systems. It’s basically a fire drill, run on our schedule rather than the market’s, so we find the constraints that give way first and fix them before a client ever notices. Capacity moved quickly once we could generate the pressure on demand.
The markets didn’t wait for us to finish. In February, crypto experienced one of its sharpest selloffs of the past 12 months, with a wave of forced liquidations sending Bitcoin down hard over several days.
Trading activity and platform load climbed together, with request volumes hitting new all-time highs. Every major service continued to run with no incidents. We estimated a full outage that day would have cost $719,000 an hour. It stayed an estimate.
Moving the bar consistently higher
Between July 2025 and June 2026, our peak failure rate fell from around 11% to 0.111%, while the load the platform carries went up more than tenfold.
But we’ve still got work to do.
The 2x standard now stands permanently. Each time real client activity sets a new record on the platform, that record becomes the load we test against next, so the bar continues to rise with the market.
When a price moves hard and everyone jumps into the market at the same moment, the platform needs to be there.
Coming soon: If you want the engineering detail, what the tests actually do, what broke along the way, and what the team learned building them, look for our companion post written for a technical audience.