feat: implement gauge and counter support for OpenMetrics 2.0

[dashpole]

Part 1 of #893

Histograms and summaries are left for follow-ups.

Open Questions:

• How do we indicate that OM 2.0 support is experimental?

@ywwg @bwplotka @krajorama

[bwplotka]

For common, given v0 I would just comment on the method. If we want to be fancy we can generate comment in the exposition format too (:

Looks good so far, good start (see comments) - only question is do we iterate on this on main or some feature branch - no strong opinion, assuming commentary is clear.

Thanks

[ywwg]

looks good to me so far. My general opinion is that if some scraper asks for openmetrics 2.0, they should get it... Hypothetically I could see requiring the version number "2.0.0rc" or "1.9.9" just in case?

[dashpole]

Lots of discussion at the OM wg. We will add a NegotiateIncluding() function that takes a list of supported protocols so we don't end up with NegotiateIncludingOpenMetricsAndOpenMetrics2 ... 😆 . Then the client can decide when it wants to add support, and how it wants to gate it (behind HandlerOpts.EnableOpenMetrics?).

[dashpole]

I added a NegotiateIncluding function. I didn't deprecate the old functions (e.g. NegotiateIncludingOpenMetrics), but we can do that in a follow-up if we want to.

[bwplotka]

I think that's great, the experimental factor at this level should be a commentary.

As per discussion on Slack I think we decided to follow:

1. Exposer should eventually switch to OM2 default, but only once ready (which means we recommend opt-in "exposed protocol" option)
2. Scrapers should be opt-in for now, while OM2 is experimental

[dashpole]

In the spirit of @krajorama's feedback, i've updated the implementation to perform some additional validation. It checks for newline characters, NaN and negative counter values, valid timestamps, empty exemplar label keys. It does not validate UTF-8 for strings, or duplicate label keys or duplicate label sets because those are expensive to validate. This does more validation than OM 1.0, but still leaves some things not validated. I also ended up writing benchmarks as part of that work, which i've split into #943.

The benchmark results from #943 run against this PR are:

goos: linux
goarch: amd64
pkg: github.com/prometheus/common/expfmt
cpu: Intel(R) Xeon(R) CPU @ 2.20GHz
                                             │ /tmp/bench.txt │
                                             │     sec/op     │
ConvertMetricFamily/TEXT/COUNTER-24               775.5n ± 1%
ConvertMetricFamily/OM1.0/COUNTER-24              1.664µ ± 1%
ConvertMetricFamily/OM2.0/COUNTER-24              1.617µ ± 1%
ConvertMetricFamily/TEXT/GAUGE-24                 761.5n ± 1%
ConvertMetricFamily/OM1.0/GAUGE-24                1.091µ ± 6%
ConvertMetricFamily/OM2.0/GAUGE-24                1.003µ ± 3%
ConvertMetricFamily/TEXT/UNTYPED-24               680.6n ± 1%
ConvertMetricFamily/OM1.0/UNTYPED-24              923.8n ± 3%
ConvertMetricFamily/OM2.0/UNTYPED-24              883.1n ± 1%
ConvertMetricFamily/TEXT/SUMMARY-24               3.207µ ± 1%
ConvertMetricFamily/OM1.0/SUMMARY-24              3.965µ ± 1%
ConvertMetricFamily/TEXT/HISTOGRAM-24             4.532µ ± 2%
ConvertMetricFamily/OM1.0/HISTOGRAM-24            5.582µ ± 2%
ConvertMetricFamily/TEXT/GAUGE_HISTOGRAM-24       3.414µ ± 1%
ConvertMetricFamily/OM1.0/GAUGE_HISTOGRAM-24      3.852µ ± 1%
geomean                                           1.741µ

                                             │ /tmp/bench.txt │
                                             │      B/op      │
ConvertMetricFamily/TEXT/COUNTER-24              0.000 ± 0%
ConvertMetricFamily/OM1.0/COUNTER-24             33.00 ± 0%
ConvertMetricFamily/OM2.0/COUNTER-24             0.000 ± 0%
ConvertMetricFamily/TEXT/GAUGE-24                0.000 ± 0%
ConvertMetricFamily/OM1.0/GAUGE-24               1.000 ± 0%
ConvertMetricFamily/OM2.0/GAUGE-24               0.000 ± 0%
ConvertMetricFamily/TEXT/UNTYPED-24              0.000 ± 0%
ConvertMetricFamily/OM1.0/UNTYPED-24             1.000 ± 0%
ConvertMetricFamily/OM2.0/UNTYPED-24             0.000 ± 0%
ConvertMetricFamily/TEXT/SUMMARY-24              56.00 ± 0%
ConvertMetricFamily/OM1.0/SUMMARY-24             57.00 ± 0%
ConvertMetricFamily/TEXT/HISTOGRAM-24            288.0 ± 0%
ConvertMetricFamily/OM1.0/HISTOGRAM-24           289.0 ± 0%
ConvertMetricFamily/TEXT/GAUGE_HISTOGRAM-24      160.0 ± 0%
ConvertMetricFamily/OM1.0/GAUGE_HISTOGRAM-24     161.0 ± 0%
geomean                                                     ¹
¹ summaries must be >0 to compute geomean

                                             │ /tmp/bench.txt │
                                             │   allocs/op    │
ConvertMetricFamily/TEXT/COUNTER-24              0.000 ± 0%
ConvertMetricFamily/OM1.0/COUNTER-24             2.000 ± 0%
ConvertMetricFamily/OM2.0/COUNTER-24             0.000 ± 0%
ConvertMetricFamily/TEXT/GAUGE-24                0.000 ± 0%
ConvertMetricFamily/OM1.0/GAUGE-24               1.000 ± 0%
ConvertMetricFamily/OM2.0/GAUGE-24               0.000 ± 0%
ConvertMetricFamily/TEXT/UNTYPED-24              0.000 ± 0%
ConvertMetricFamily/OM1.0/UNTYPED-24             1.000 ± 0%
ConvertMetricFamily/OM2.0/UNTYPED-24             0.000 ± 0%
ConvertMetricFamily/TEXT/SUMMARY-24              2.000 ± 0%
ConvertMetricFamily/OM1.0/SUMMARY-24             3.000 ± 0%
ConvertMetricFamily/TEXT/HISTOGRAM-24            6.000 ± 0%
ConvertMetricFamily/OM1.0/HISTOGRAM-24           7.000 ± 0%
ConvertMetricFamily/TEXT/GAUGE_HISTOGRAM-24      5.000 ± 0%
ConvertMetricFamily/OM1.0/GAUGE_HISTOGRAM-24     6.000 ± 0%
geomean                                                     ¹
¹ summaries must be >0 to compute geomean