During the incident, the system hit 3000 IOPS and stayed there. It didn’t recover on its oown. We had to kill connections manually. Why? What was actually happening inside Postgres that made the situation self-sustaining? I had a wrong mental model for days. This post is about fixing that.

If you're not gonna bother writing it yourself, I ain't gonna bother reading it myself. - Geoff Graham

ges­tern mach­te ai;dr die run­de. @dave_min­ni­ge­ro­de hat das na­he­lie­gen­de mög­li­cher­wei­se erst­mals auf­ge­schrie­ben, näm­lich das er ei­nen text nicht le­sen moch­te, nicht weil er zu lang war (too long, didn’t read, tl;dr), son­dern weil er zu sehr nach ki/ai roch (ai;dr).