Does this crate implement Prometheus or OpenMetrics? about client_rust

[mxinden]

and are definitely not fully compatible with each other.

Can you expand in which way recent Prometheus versions are not compatible with the OpenMetrics specification?

Note that this has also been discussed in #66 (comment) and in #51. I don't understand why this requires a new GitHub issue.

[08d2]

Can you expand in which way recent Prometheus versions are not compatible with the OpenMetrics specification?

Consider a counter metric exported via curl <hostport>/metrics, and specifically maybe e.g. the mxinden/kademlia-exporter metric kademlia_exporter_meta_random_node_lookup_triggered. Running that program and then fetching the /metrics endpoint returns

# HELP kademlia_exporter_meta_random_node_lookup_triggered Number of times a random Kademlia node lookup was triggered.
# TYPE kademlia_exporter_meta_random_node_lookup_triggered counter
kademlia_exporter_meta_random_node_lookup_triggered_total 1

Note that the HELP and TYPE lines have metric name "kademlia_exporter_xxx" and that the actual metric data line has a metric name "kadmelia_exporter_xxx_total". This may be valid for OpenMetrics, but it is not valid for Prometheus. Prometheus requires consistent metric names, e.g.

# HELP kademlia_exporter_meta_random_node_lookup_triggered_total Number of times a random Kademlia node lookup was triggered.
# TYPE kademlia_exporter_meta_random_node_lookup_triggered_total counter
kademlia_exporter_meta_random_node_lookup_triggered_total 1

[08d2]

I don't understand why this requires a new GitHub issue.

Consider this issue a superset of the other issues. OpenMetrics and Prometheus define specs which are not equivalent. If this repo lives in the Prometheus org, it should implement the Prometheus spec. If it doesn't implement the Prometheus spec, then it shouldn't live in the Prometheus org. That's my only point.

[kalabukdima]

I was also trying to understand why the metrics exported by this crate can't be scraped by Prometheus.

In particular, the issue is with the Info metric that is not part of the default set and leads to the error invalid metric type "info" preventing all other metrics to be scraped too (tested with prometheus 2.51.2 and some other verions — none of them could deal with this format).

So I think it is confusing indeed that the official Prometheus client doesn't work with Prometheus. I can see the following possible solutions:

• rename this crate to something like open-metrics-client or
• add a possibility to export metrics in Prometheus format as done in Python client

Upd: after setting the correct header with the response, Prometheus could parse the output, but it's still confusing that those formats are not compatible and the naming doesn't suggest which one is used.

[mxinden]

@08d2 and @kalabukdima I understand your concerns.

@RichiH do you have additional input here as one of the authors of the open metrics specification? How should the Rust client proceed to be consistent within the Prometheus ecosystem.

In the meantime, @08d2 and @kalabukdima, how about very prominently calling out on the main docs.rs page (i.e. in src/lib.rs) that application/openmetrics-text needs to be set as a content type? Does either of you want to create a pull request?

[RichiH]

FYI, OpenMetrics is now part of Prometheus officially

I am on the road right now, but this feels like a bug in Prometheus. Potentially dev summit material to discuss?

[mxinden]

FYI, OpenMetrics is now part of Prometheus officially

Can you point us to some docs @RichiH? Is OpenMetrics the default? Is setting the HTTP content type still needed?

Related: #194 and prometheus/prometheus#13944