A lot of the metrics for outage impact are available, and if you really want to Do Science you can use whatever A/B Canary/experiment framework you have to simu

A lot of the metrics for outage impact are available, and if you really want to Do Science you can use whatever A/B Canary/experiment framework you have to simulate outages (which, arguably, are not perfect because of down-detectors) on users to understand their behavior. I tend to agree that most people don't care about transient downtime beyond 2 or 3 nines for reputation. Covid has proven how bad people are at reasoning about 0.5% probabilities so there's essentially no chance that they are noticing (en mass) the difference between 2.5 or 3 nines of reliability. Know your customer, as the article points out; more complex customers will know and measure actual availability because they need it for their own reliability calculations. This brings me to my final guess, that SLOs are primarily useful because they provide a risk model that upstream and downstream consumers can reason about. SLOs are a way of building even larger distributed and interdependent systems that can still be somewhat reliable. Precisely what SLIs are chosen and the limits are a little less important than the fact that they are published publicly. There's an economy of reliability in large systems and like any market information about the market is key to its efficiency. If everyone just trusted that everyone was doing their best at maintaining dependencies most of those beliefs would be very unanchored. A side benefit of SLOs is postmortems and communication; people working on codependent products and systems need to talk to each other from time to time and understand the true nature of the interfaces between their systems. Postmortems and SLO negotiations give a concrete framework to do that within, metrics drive quantitative understanding of the systems instead of handwaving.