Alert on symptoms, not causes
High CPU is not an incident. A checkout endpoint returning errors to 2% of users is. We delete every alert that cannot be tied to a user-visible symptom, and we do it aggressively — on one client platform we removed 140 of 190 alert rules in the first month.
Error budgets make the trade-off explicit
An SLO of 99.9% monthly availability is 43 minutes of downtime. When the budget is spent, feature work stops and reliability work starts. This is agreed with the client before go-live, in writing, because arguing about it during an incident goes badly.
Compensation and limits
On-call is paid. A rotation is never shorter than one week or more frequent than one week in four. Anyone paged after midnight starts the next day at noon. These are policies, not favours.
Postmortems without blame
Every incident above severity 3 gets a written postmortem within 72 hours, published internally and shared with the client. It names systems and decisions, never people. An incident without a postmortem has not actually ended.



