Histogram issue with time unit duration_seconds

[prashanth057]

When I declare a histogram metric with a name that ends in duration_seconds, for example request_time_duration_seconds, and then use promethues_histogram:observe(request_time_duration_seconds, 30), I would expect this to be recorded as 30 seconds. But when I see the data recorded in the buckets, it is totally off. It seems to record this as 30 nano seconds. Is this a bug or am I missing something ?

[dadarakt]

I can confirm this behavior, had this problem when using a Histogram with a _duration_milliseconds suffix in an Elixir application. This also occurred when using _duration suffix name and setting the duration_unit manually when calling Histogram.new.

I think the intention of the authors is that clients use the macro Histogram.observe_duration/2, as documented in the example section for the histograms?

But I also ran into the same issue by accident, which was somewhat hard to debug. From my point of view this is relevant when developers would like to have a little more control over the computation of the duration and don't want to use the provided macro. Intuitively, when using Histogram.observe/1, I would expect the time unit to be respected as well, so bumping this issue.

If this is an issue with prometheus.ex I'm happy to move the discussion there.

[essen]

It expects native time and then converts to seconds automatically.

[dylan-chong]

Why was this implemented this way? This design decision to change behaviour based on a metric name cost me many hours.
PLEASE DOCUMENT THIS CLEARLY

[mtaylorsamsung]

This design decision and lack of clear documentation also cost me several hours. A more prominent call-out of this behavior or (better yet, at least in my opinion) a reconsideration of this design decision would maybe help usability a bit for new users.

[essen]

I don't know why it was made like this, but I'd be happy to merge a PR updating the documentation. Just ping me on the PR.