1. 1 exemplar per bucket is fine; "le +Inf" is also a bucket. (OpenMetrics insists "A bucket MUST NOT have more than one exemplar.", but I don't think the Prometheus code enforces this rule) 2. Prometheus will remember as many exemplars as it can, in a ring buffer. 2b You can't always predict when Prometheus will next scrape. There might be multiple scrapers, e.g. an HA pair.
I'm not sure about 3; maybe some other client authors will comment. As far as I know the Go client retains up to 1 exemplar per series forever. On Tuesday, 27 May 2025 at 22:10:14 UTC+1 [email protected] wrote: > Hi, I'm working on adding exemplar support for histograms to a Haskell > prometheus client. I wanted to run my implementation by Prometheus folks to > see if this all sounds correct: > > 1. For a given histogram with 5 buckets, store up to 1 exemplar per > bucket. Don't store exemplars for +Inf samples. > 2. There should be a minimum retention period, such that if you store two > exemplars for the same bucket back-to-back, only the first exemplar is > stored. My understanding is there are two main purposes of this: > > a. Store fewer exemplars in prometheus > b. Letting callers know up-front if an exemplar is likely to make it to > prometheus, so that they can sample the trace. > > 3. There should be a maximum retention period, such that an exemplar is no > longer sent to prometheus if it's e.g. older than a minute. I'm less clear > on why this is desired, but want to give a good explanation in docs. > > WIP PR here: https://github.com/fimad/prometheus-haskell/pull/74 > -- 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 visit https://groups.google.com/d/msgid/prometheus-developers/3e293706-f3ed-413b-8eb0-68f0a4bfa4aan%40googlegroups.com.
