p99 is a team boundary, not a metric

For years I treated p99 as a performance metric. A line on a graph, a number that should stay under a target. Then I noticed its actual function is not technical: p99 decides whose phone rings at night.

The logic goes like this. If you own an endpoint’s p99, you are the one who wakes up when it gets slow. That makes you the de facto owner of everything that could cause the slowness: the query underneath, the downstream service you call, the queue you depend on. Owning a p99 draws a much wider circle of responsibility than it looks like it does.

The practical consequence is that negotiating a p99 target is really drawing an org chart. Saying “we commit to 200 milliseconds” means asking for a say over everything required to hold that number. Teams that commit without that say burn out steadily, because they are being held to a number they do not control.

So when I walk into an SLO discussion now, my first question is not what the target should be. It is who can say yes to the changes required to hit it. If the answer is nobody, we are not discussing a target, we are discussing a wish.

Unfinished part of this note: how any of it works for endpoints owned jointly by several teams. I do not have a good answer.