latencyNative assumes metric is in seconds

[klubi]

Hi

Is there a way to say what units are used in metric that is being used with latencyNative?
I have a case where I would like to use metric that comes from AWS CloudWatch exporter (AWS ApiGateway latency).
Values for that metric are in milliseconds. But it looks like latencyNative assumes all metrics are in seconds.
Now, in order to work this around I can set latency to 200s instead of 200ms (what I actually want).
I can work this around, but for anyone looking at SLO definition this will be confusing.

Would it make more sense for latency to be unitless?

[metalmatze]

Would you mind sharing your config? The latencyNative indicator's latency field is indeed independent of the unit. 1ms, 2s, or 3m would work. However, for those you need a native histogram.

When it comes to the counter-based histograms (the ones we had for years in Prometheus) Prometheus works with seconds as base units. Thus Pyrra follows what Prometheus does.
OTel recently changed from milliseconds to seconds too.
I think in this case the cloudwatch exporter should be updated.

This is similar to the discussion in #667