Track JVM gc time in histogram

[jack-berg]

Fixes #2902.

cc @open-telemetry/java-instrumentation-maintainers, @open-telemetry/java-maintainers

[bogdandrutu]

Why do you think that a default of histogram is better?

[jack-berg]

Why do you think that a default of histogram is better?

@bogdandrutu not sure I understand. Better than what?

In general, histogram is a good fit for this because it includes a lot of useful data for analysis:

• You can use the sum to analyze the total time spent in gc
• Can look at the count to analyze the number of gc events
• Can look at the max to analyze the maximum time spent in any single gc event (i.e. what was my longest "stop the world" gc event?)
• Can look at the distribution to analyze the distribution of time of gc events (i.e. how long do typical gc events take? how long do outlier gc events take?)

[trask]

@bogdandrutu just wanted to make sure you don't have any significant concerns before we request merging, thx!

[jack-berg]

FYI I've updated the metric name to process.runtime.jvm.gc.duration, prompted by this comment. I quite like @trask's suggestion to use the duration suffix instead of time. Makes it clear that this metric measures the distribution of individual gc event durations, rather than the total time spent in in garbage collection.

Please take another look approvers @breed-splk, @mateuszrzeszutek @jmacd @trask

[trask]

@bogdandrutu @jmacd @reyang @jsuereth - the Java folks would like to merge this if there's no further concerns

[reyang]

There is a wider discussion brought up by @jsuereth this morning regarding whether we should have different way of exposing metrics for the same concept or not. We should continue that discussion in the semantic convention workgroup, no need to block this PR since it has strong support from Java folks.