This is a concept I've had to explain to entirely too many teams over the years, that 0.001% of requests failing as a (mostly) random distribution of all requests is very different than a 0.001% subset of requests that will fail (nearly) every time until the underlying issue is mitigated. They look the same on a high level dashboard but they are completely different conditions in terms of how the customer will feel it, and understanding which kind of problem you have also guides the investigation and troubleshooting process.