On 06.09.22 02:50, 'Fabian Stäber' via Prometheus Developers wrote: > > Looking at client_golang, it seems you can observe math.Inf(), and bucket > index math.MaxInt32 is used to represent the Inf bucket. > > https://github.com/prometheus/client_golang/blob/95cf173f1965388665dcb2a28971f35af280e3a5/prometheus/histogram.go#L589-L590 > > I'm wondering how to represent the Inf bucket as a BucketSpan in protobuf. > Initially I set the offset to current index minus previous index, but > obviously that doesn't work if the current index is MaxInt32. > > Any ideas?
Yeah, very good question. And definitely something that needs to get ironed out before coming up with a final spec for Native Histograms. In practice, I think, observations of ±Inf will be irrelevant. The set the sum of observations to ±Inf, too (or even to NaN if it was +Inf before and then -Inf is observed or vice versa), thereby rendering the sum useless. My idea so far was to put observations of ±Inf and even NaN in no bucket at all, let them "ruin" the sum of observations (setting it to ±Inf or NaN as appropriate), and increment the count of observations as usual. In that way, the difference between observations in buckets and observations in the count would account for all those observations. The downside is that you cannot distinguish between the three types of "weird" observations (+Inf, -Inf, NaN). On the other hand, I don't think we should add a whole lot of costly plumbing throughout the stack to store them separately. >From a completionist's perspective, observations of very large positive or negative numbers should be treated similarly as very small observations, i.e. adding an "overflow bucket" (or even two, for negative and positive observations separately) similarly to the zero bucket we already have. The reason for not doing it so far is mainly pragmatic: While it is easy to accidentally create values close to zero (may it come from some calculation or from actual physical measurements), it is far less likely (but not impossible, of course) to accidentally create numbers with a very large absolute value of up to ±Inf. This assumption might not hold, and that's exactly why the Native Histograms are marked as experimental. We can still correct those things if needed. > Not sure if this is covered in client_golang either > https://github.com/prometheus/client_golang/blob/95cf173f1965388665dcb2a28971f35af280e3a5/prometheus/histogram.go#L1272-L1280 Yeah, that's weird. I filed https://github.com/prometheus/client_golang/issues/1131 to investigate more closely. -- Björn Rabenstein [PGP-ID] 0x851C3DA17D748D03 [email] [email protected] -- You received this message because you are subscribed to the Google Groups "Prometheus Developers" 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-developers/Yxjfqp8TaxOuxtjx%40mail.rabenste.in.
