Feature request - public `registry._names_to_collectors` access

[Lancetnik]

Now we are working on Prometheus support in FastStream and faced with a problem. Shared registry between different middlewares leads to MetricsContainer duplication. So, we would like to have a public API to chech registry for already registered objects.

Can you please provide public access to this property some way?

https://github.com/prometheus/client_python/blob/master/prometheus_client/registry.py#L31

I think, the easiest way to save old API and provide us with a new one - add a public property like in the following snipper

class CollectorRegistry(Collector):
    def __init__(self):
        self._names_to_collectors: Dict[str, Collector] = {}

    @property
    def names_to_collectors(self):
        return self._names_to_collectors

[Lancetnik]

Related to - https://github.com/airtai/faststream/pull/1921

[csmarchbanks]

Is it not sufficient to call collect() and see what all is returned to get a list of all metrics?

My concern with adding that function is that not all registries are instances of CollectorRegistry, some applications may use their own implementation. Not to mention things like custom collectors.

[Lancetnik]

To clarify - we just want to avoid double metrics registration

counter_name = "..."

counter = (
    registry._names_to_collectors.get(counter_name)
) or Counter(
    name=counter_name,
    ...
    registry=registry,
)

https://github.com/airtai/faststream/blob/main/faststream/prometheus/container.py#L40

[Lancetnik]

Well, seems like we can't create any metric with custom CollectorRegistry implementation - all registry options are Optional[CollectorRegistry] = REGISTRY type

https://github.com/prometheus/client_python/blob/master/prometheus_client/metrics.py#L132

So, the initial point about CollectorRegistry property still actual

[draincoder]

Is it not sufficient to call collect() and see what all is returned to get a list of all metrics?

collect() returns only non-zero metrics and resets their state. In addition, we need to get an instance of the metric by the name of the metric.