HI to all,
Brian pointed to me to that link [1] which basically claims that in
Prometheus - compared to other TimeSeries databases - there is no necessity
of prefixing the names of the metrics. And TBH I like the idea of not
having to prefix the metric name with for example the service name, I can
see some potential use cases on that, for example being able to gather all
metrics seen from upstream servcies when they are calling your service, by
prefixing the metrics it would become a painful process.
But having the feeling that my original question is still applicable, but
for labels. So let me rephrase the question that I've originally opened
here [2].
Seems that Prometheus Python client prescribes a pattern for declaring the
metrics at the module level, which has the good benefit of making it
available quite straightforward by just importing that specific symbol,
also allows the system to keep everything in memory - not deallocated by
the GC - for the whole process time. And its there when the labels epxected
for a specific metric are enumerated, for example:
REQUESTS = Counter(
"starlette_requests_total", "Total count of requests by method and
path.", ["method", "path_template"]
)
Later on in your code, you do something like this:
REQUESTS_IN_PROGRESS.labels(method=method,
path_template=path_template).inc()
My questions is, how could we manage the necessity of having to add the
same label value for all of the metrics that are generated within an
application? By following the prescribed pattern we would be forced to
pollute all of the metrics with something like this:
REQUESTS = Counter(
"starlette_requests_total", "Total count of requests by method and
path.", ["method", "path_template", "service_name"]
)
And later on in our code do something like this:
REQUESTS_IN_PROGRESS.labels(method=method, path_template=path_template,
service_name=get_service_name()).inc()
I guess that if you have just a bunch of metrics and a few services might
be manageable, but it quickly becomes a lot of repeated code that not even
the developer should take care. Is there any alternative to do so in a
generic and none friction way?
I'm wondering if we could create our own Registry - I do not know if
`Resgistry` is a first citizen class that can be used by the developers -
and use the same instance in all of our metrics and when the data is
gathered by the `generate_latest` do something like this:
from myregistry import RegistryWithServiceName
if RegistryWithServiceName.get_service_name() is None:
RegistryWithServiceName.set_service_name(get_service_name())
generate_latest(REGISTRY)
Behind the scenes, RegistryWithServiceName would need to update the
collected metrics by adding a new label to all of them, so having as a
result a list of metrics with their original labels plus the one that we
inserted during the `generate_latest` execution
WDYT?
[1] https://www.robustperception.io/target-labels-not-metric-name-prefixes
[2] https://github.com/prometheus/client_python/issues/571
--
You received this message because you are subscribed to the Google Groups
"Prometheus Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To view this discussion on the web visit
https://groups.google.com/d/msgid/prometheus-users/6f760079-348b-44d6-a98e-1f4329d52bf7o%40googlegroups.com.
