The main problem the article is talking about is tracking 99th percentile issues (e.g. tail latency) which, given the nature of sampling is difficult to capture

The main problem the article is talking about is tracking 99th percentile issues (e.g. tail latency) which, given the nature of sampling is difficult to capture otherwise. Overhead is an issue with instrumentation, but as the article mentions, there are a number of clever ways to minimize it. Some simpler examples are physical probing and dynamic instrumentation. FTrace might be a nice example of instrumentation on Linux worth looking at.