I'm not sure that "making the graph smooth" is a good goal in itself; the variability you're seeing could well be a better representation of your input data.
Imagine that in one scrape your request durations samples were: - 60ms - 62ms - 65ms - 63ms - 61ms and in a later scrape they were: - 93ms - 85ms - 87ms - 89ms - 92ms With a histogram, all these points fall into the "50-100ms" bucket. Both cases will be treated as identical (5 entries in the 50-100ms bucket), and hence all the variation has been lost. Any function which derives data from the histogram. including quantile estimation, will give the same results for both scrapes. With a summary, a streaming calculation takes place on the client side to return the 50th, 90th and 99th percentile values, based on the current value and recent history. These percentiles adapt dynamically based on the input values received. It can therefore capture those percentiles more accurately, tracking variations in the input data. If you have a look at the raw logs of the put request durations which are being fed in, it may become clearer what's going on. -- You received this message because you are subscribed to the Google Groups "Prometheus Users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/prometheus-users/0824ccfc-955e-4483-abc0-1a33ff5af2d2o%40googlegroups.com.
