Usually I'd just look at clusters of crashes (those that had similar stack traces) but sometimes when you're running a very small % experiment there's not enough signal so you end up looking at everything. And oh boy was there a lot of noise.

In an app with >billion users you get all kinds of wild stuff.