Summaries are similar to histograms in some ways, but present different trade-offs and are generally less useful. They are also used to track sizes and latencies, and also provide both a sum and a count of observed events. Additionally (and if the client library used supports it), summaries can also provide pre-calculated quantiles over a predetermined sliding time window. The main reason to use summary quantiles is when accurate quantile estimation is needed, irrespective of the distribution and range of the observed events.

Both quantiles and sliding window size are defined in the instrumentation code, so it's not possible to calculate other quantiles or window sizes on an ad hoc basis. Doing these calculations on the client side also means that the instrumentation and computational cost is a lot higher. The last downside to mention is that the resulting quantiles are not aggregable and thus of limited usefulness.

One benefit that summaries have is that, without quantiles, they are quite cheap to generate, collect, and store.

To help visualize this type of metric, here is an example of a summary and its graphical representation, based on the test environment we created in the previous chapter:

• The maximum duration of the Prometheus rule group in seconds by quantile: