This was a leftover from when we (or rather I) thought that the
old (now called "high-level") API would be removed. We deemed it
useful though, so let's remove that "legacy" language and make
the description more friendly.
The metric subsystem of Prosody has had some shortcomings from
the perspective of the current state-of-the-art in metric
observability.
The OpenMetrics standard [0] is a formalization of the data
model (and serialization format) of the well-known and
widely-used Prometheus [1] software stack.
The previous stats subsystem of Prosody did not map well to that
format (see e.g. [2] and [3]); the key reason is that it was
trying to do too much math on its own ([2]) while lacking
first-class support for "families" of metrics ([3]) and
structured metric metadata (despite the `extra` argument to
metrics, there was no standard way of representing common things
like "tags" or "labels").
Even though OpenMetrics has grown from the Prometheus world of
monitoring, it maps well to other popular monitoring stacks
such as:
- InfluxDB (labels can be mapped to tags and fields as necessary)
- Carbon/Graphite (labels can be attached to the metric name with
dot-separation)
- StatsD (see graphite when assuming that graphite is used as
backend, which is the default)
The util.statsd module has been ported to use the OpenMetrics
model as a proof of concept. An implementation which exposes
the util.statistics backend data as Prometheus metrics is
ready for publishing in prosody-modules (most likely as
mod_openmetrics_prometheus to avoid breaking existing 0.11
deployments).
At the same time, the previous measure()-based API had one major
advantage: It is really simple and easy to use without requiring
lots of knowledge about OpenMetrics or similar concepts. For that
reason as well as compatibility with existing code, it is preserved
and may even be extended in the future.
However, code relying on the `stats-updated` event as well as
`get_stats` from `statsmanager` will break because the data
model has changed completely; in case of `stats-updated`, the
code will simply not run (as the event was renamed in order
to avoid conflicts); the `get_stats` function has been removed
completely (so it will cause a traceback when it is attempted
to be used).
Note that the measure_*_event methods have been removed from
the module API. I was unable to find any uses or documentation
and thus deemed they should not be ported. Re-implementation is
possible when necessary.
[0]: https://openmetrics.io/
[1]: https://prometheus.io/
[2]: #959
[3]: #960
When set, no periodic statistics collection is done by
core.statsmanager, instead some module is expected to call collect()
when it suits. Obviously only one such module should be enabled.
Quoth jonas’
> correct way is to scrape the internal sources on each call to /metrics
> in the context of Prometheus
"manual" as opposed to "automatic", from the point of view of
statsmanager.
Zash> Btw, this conditional and loop, shouldn't it be covered by the timing measurement?
Zash> Isn't that where all the util.statistics work is done?
MattJ> Yeah, it should
Zash> ("the", but there's two ... which one‽)
MattJ> Yeah... not sure :)
MattJ> Processing I guess
