Monitoring Architecture and Practice - The Practice of Cloud System Administration - page 358

utilization, nearly 2.5 hours each day. More importantly for the colocation facility, it best
represents the cost of the bandwidth used.
Percentiles are also commonly used when discussing latency or delay per unit. For ex-
ample,asitethatistryingtoimproveitspage-loadtimemightsetagoalofhavingthe80th
percentile page load time be 300 ms or less. A particular API call might have a lot of vari-
ation in how long it takes to execute due to caching, type of request, and so on. One might
set a goal of having its 99th percentile be less than some amount, such as 100 ms.
17.5.2 Stack Ranking
Stack ranking makes it easy to compare data by sorting the elements by value. Figure 17.2
isanexampleof stack ranking .Ifthecitieshadbeensortedalphabetically,itwouldnotbe
clear how, for example, Atlanta compares to New York. With the stack rank we can clearly
see the difference.
Figure 17.2: Cities stack ranked by number of tacos consumed
Stack ranking is best used when data is apples-to-apples comparable. One way to
achievethisistonormalizethedatatobeperunit(e.g.,percapita,permachine,perservice,