Notes on running infrastructure, and what it costs.

We ran the numbers on 1,140 accounts that moved from AWS. The median one was paying 31% of its bill for traffic leaving the cloud.

Every Thursday at 10am Eastern we fail over one Postgres cluster in each region on purpose. Here is what broke the first six times.

Cheap power, a fiber crossroads and a lot of people who know how to run buildings. Our first region was not an accident.

A walk through one worker-gpu-1 that ran for 5 hours, 52 minutes and 11 seconds, and exactly what it cost.

Logical replication, a 90 second switch window and the three extensions that almost stopped us.

What the on-call week looks like for the nine engineers who answer our support line.