Probability Sampler

[nholbrook]

Implements ProbabilitySampler based on the sampler interface in #118.

Relevant snippet from specification:

Probability
The default behavior should be to trust the parent SampledFlag. However there should be configuration to change this.
The default behavior is to apply the sampling probability only for Spans that are root spans (no parent) and Spans with remote parent. However there should be configuration to change this to "root spans only", or "all spans".
Description MUST be ProbabilitySampler{0.000100}.
TODO: Add details about how the probability sampler is implemented as a function of the TraceID. TODO: Split out the parent handling.

[codecov]

Codecov Report

Merging #136 into master will increase coverage by 0.40%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #136      +/-   ##
==========================================
+ Coverage   94.18%   94.59%   +0.40%     
==========================================
  Files          80       83       +3     
  Lines        1893     2034     +141     
==========================================
+ Hits         1783     1924     +141     
  Misses        110      110              

Impacted Files	Coverage Δ
...ude/opentelemetry/sdk/trace/samplers/probability.h	100.00% <100.00%> (ø)
sdk/src/trace/samplers/probability.cc	100.00% <100.00%> (ø)
sdk/test/trace/probability_sampler_test.cc	100.00% <100.00%> (ø)

[pyohannes]

I think it would make sense to simplify this for the first go, and then discuss what of the SHOULD parts of the spec we want to add:

• We can get rid of SamplingBehavior and just implement the default DETACHED_SPANS_ONLY logic.
• We can get rid of defer_parent and just implement the default defer_parent = true.

If deemed necessary, we can add those parts in separate PRs (no of the other SDKs I looked at have this implemented now). This way we can reduce complexity.

You can refer to some other implementations:
Go
Java
Python

[nholbrook]

The current implementation of this Sampler handles invalid parameter values (i.e. any probability out of the bounds [0, 1]) silently. The intended functionality is to raise an exception to alert the user, however, this is currently not allowed by CI tests (see #156). This PR can be merged now and the error handing can be added in a later PR, or this will have to wait for #156 to be resolved.

[pyohannes]

This looks good, I added some thoughts.

[reyang]

LGTM.

[reyang]

@nholbrook would you please rebase? Thanks.

[nholbrook]

@reyang Rebased

[neopaf]

Friends, please shed some light on necessity of CalculateThreshold

It moves us from double to int64 space, which is sometimes good.

But is it really necessary in this case?

What is wrong with converting trace_id to ratio and using it directly, without double-to-int magic?

//  if (CalculateThresholdFromBuffer(trace_id) <= threshold_)
  if (CalculateRatioFromBuffer(trace_id) <= ratio_)