SRE & Reliability
Reliability work starts at the first real user, not before
Reliability is a number you find out, and the moment you find it out is the moment a region stops answering. So the work is arithmetic against traffic that already exists. On a payments platform, we keyed autoscaling off transaction volume and payment throughput instead of CPU utilization. Peak hit 3x normal load, and payments kept clearing. Latency held under 500ms.
A different employer, a different peak: Max built the Black Friday autoscaling for a fintech payment platform, and it took 10x normal load. He set each threshold from traffic he had already watched.
The rest is drills. Back at the payments platform, the disaster-recovery runbooks ran to 30+ documented failure scenarios, exercised every month with automated cross-region failover tests and held to a sub-minute recovery time objective. Nobody should be finding out mid-incident whether the runbook works. You give us a month of traffic data and access to the failover path; you get back an SLO set against what your users actually did, a runbook per failure mode, and a failover you have watched run.
If nothing is in production yet, there is nothing to measure, and an SLO guessed in advance is a number neither of us would believe.