I tend to agree with you. For some large interconnected systems, long tail latencies have a high complexity cost because you need to code around the 1% cases when (for example) web service calls take much longer than normal to return results. Client code becomes more complex.