BentoML

Do you have a change in mind ?

I faced the same the issue when I was trying out bentoml with fastapi where I had a middleware which was using `prometheus_client` .

A work around or probably a solution for problem was to use `bentoml.metrics`  instead of `prometheus_client`  in the code, only drawback is not all the methods are implemented as part of bentoml metrics.

I think the docs were clear to use `bentoml.metrics` but it took a bit of time to make my services apply that

Maybe a fallback approach with a warning instead of stopping `bentoml serve` completely?

<@U04M1MH05DL> great suggestion! I vaguely remember this had to do with the default prometheus client having issue in a multi-process environment like BentoML. Note that in traditional FastAPI/Flask app, all processes are equivalent fork of each other, but BentoML differentiates each Runner process group separately than the serving logic, which makes the metrics/tracing implementation more challenging. I will look into it a bit more and see if it’s possible to support using `prometheus_client` directly tho

what methods that you want to use? `bentoml.metrics` is mostly a passthrough to the prometheus_client. Note that some of the functions won’t be supported in multiprocess mode, per prometheus_client documentation

Thanks for the reply guys, my service is very simple, I was not really using any `bentoml.metrics` all what was happening is that I had two other services on my machine using prometheus, I understand in production this most likely wont be the case. But during active dev it is.

Can you send your service definition here if possible?