slo · Aug 27, 2026
An SLO the Team Can Actually Keep
We started with four nines because it sounded serious, then spent a quarter apologising to ourselves. Here is how we picked a target we can hold and a rule for when to stop shipping.

Our first SLO at Wrenfield was 99.99 percent availability for the public API. Nobody picked it carefully. It was the number on the slide, and it sounded like what a platform team should say out loud.
It lasted one quarter. Four nines over thirty days leaves about four minutes of allowed failure, and a single slow deploy spent all of it. We missed the target in week two, missed it again in week five, and by week eight the team had stopped looking at the dashboard. An objective you've already failed has no job left to do.
Start from what you can defend
Tove, who runs on-call scheduling, asked the question that fixed it: how many bad minutes does a normal month contain, when nothing dramatic happens? We counted. Over the previous six months the answer was between nine and thirty-one minutes. That's the real distribution of our work, with ordinary deploys, ordinary dependency blips and one slow database failover.
So we chose 99.9 percent. Over a thirty-day window that is 43.2 minutes of budget, which sits comfortably above our ordinary month and well below a month where something actually went wrong. The target is tight enough to mean something and loose enough that we can keep it without heroics.
Counting the budget
The arithmetic is small enough to keep in a script next to the dashboard config.
slo = 0.999
window_minutes = 30 * 24 * 60
budget = window_minutes * (1 - slo)
bad_minutes_so_far = 21
days_elapsed = 9
burn = (bad_minutes_so_far / budget) / (days_elapsed / 30)
print(f"budget {budget:.1f} min, spent {bad_minutes_so_far / budget:.0%}, burn rate {burn:.2f}")Run it and you get a budget of 43.2 minutes, 49 percent spent, and a burn rate of 1.62. Burn rate is the number that matters. At 1.0 you will use exactly your budget by the end of the window. At 1.62 you'll be out of it around day eighteen, long before the month closes, and a number like that is a conversation, not an alarm.
The rule that makes it real
An objective with no consequence is decoration. Ours has two.
If the burn rate over the last seven days is above 2, feature releases pause. Fixes, rollbacks and reliability work continue. The pause ends when the seven-day rate drops below 1.5.
If the budget is gone before the window ends, the next sprint's first item is the thing that spent it, chosen by the on-call engineer, not by whoever has the loudest roadmap.
We wrote both into the team charter, and the product lead signed it. That signature matters more than the number.
What we left out
We don't have a latency SLO yet. People ask for one constantly, and we keep saying no, because we haven't agreed what slow means for the three different kinds of caller we have. Half an answer there would just be four nines again. We'd rather have one objective we believe than three we resent.
Two months in, the burn rate has crossed 2 once. Releases paused for three days. Nobody enjoyed it and nobody argued with it, which is the best review the policy could get.