For environments where the response times are the least critical, the p50 latency SLI might be the most appropriate to use.

If half the requests that come into the application meet the set latency value, the SLO is not violated. That means the other half can have very slow response times and still be considered a healthy service.

Consider the sorted dataset used for Best Practice 1 and Best Practice 2. Now highlight 50 percent of the requests. Emphasize in bold the worst latency (slowest response time) in the highlighted data. The value in bold is the p50 latency for this dataset.



The above dataset has a p50 latency of 109 ms. Compare this to the p90 latency of 176 ms and p99 latency of 261 ms. It is much easier to meet quality standards if the standard is measured in p50 latencies because fewer requests must be under the threshold.

In other words, if a service can achieve a p99 latency value (for example, 261 ms), the service can pass a p90 latency standard of 261 ms and certainly can pass the p50 latency standard of 261 ms.