Histograms

[talex5]

This is the core part of #10 (which was a bit big to review), with a few changes:

• When a metric has an invalid label, give the reason rather than saying the metric name is wrong.
• Change the validation to reject only the exact strings, not any string starting with the reserved label.
• Replace the tuple with a sample record type (added in Add Sample_set module #13).
• Separate the histogram spec from the runtime values and put those functions in their own Histogram_spec module.
• Fixed of_exponential so the first bucket is at start, not start * factor.
• Add doc-string for DefaultHistogram.

@stijn-devriendt : does this look OK to you?

[stijn-devriendt]

Looks OK to me. Thanks for taking the time to refine this.

I saw that you stripped out the getters. Is this something you'd prefer leaving out forever? The Java client for example does have these methods, so it's not like adding it would make it a one-off.
So does it make sense for me to send a pull request later, re-adding them (perhaps when we've had more internal exposure over the next few weeks about which bits we're actually finding useful, leaving out others) or do you prefer not having in-process retrieval?

[talex5]

Splitting it is to make it simpler to review. I need to think a bit more about the getters.

Ideally, I'd prefer that code doesn't have access to any mutable global state, so that the behaviour of a component can't depend on other components to which it wasn't explicitly given access. However, OCaml does not make it possible to enforce this, and we have to expose CollectorRegistry anyway. So it's really just to discourage people from trying to read the values, since that's probably the wrong thing to do. e.g. if a web-server library reads the counters to make decisions, then you can't run two web-servers in the same process. But on the other hand, there may be valid uses for it and it's really up to library authors to write sensible code.