Probability sampler is broken for low sample rates

[akoshochrein]

Hello everyone,

I've found an issue with the probability sampler not working for low sample rates. There are 2 piecens of code related to this:
https://github.com/census-instrumentation/opencensus-python/blob/master/opencensus/trace/samplers/probability.py#L56-L66
As the code says, it fetches the last 8 bytes of the trace id, that will later be used to compare to a constant value calculated using the sample rate. This algorithm is, in fact, flawless and propagates traces perfectly.
The issue comes from a bit deeper:
https://github.com/census-instrumentation/opencensus-python/blob/master/opencensus/trace/span_context.py#L165-L173
As a matter of fact, uuid4 is not an entirely random number according to the uuid specs.
Here's some empyrical data as proof:
https://gist.github.com/akoskaaa/a53c8b0201a8bffc1ae4df91e5965933

Source:
https://tools.ietf.org/html/rfc4122
https://en.wikipedia.org/wiki/Universally_unique_identifier

Solution could be to change the tracing id generation to a more random one. I've actually patched it in my systems with this function:

def patched_generate_trace_id():
    hexdigits = '0123456789abcdef'
    return ''.join([hexdigits[random.randint(0, 0xF)] for _ in range(32)])

Do you have any better ideas?

Also, I'd just like to say thanks to the owners and everyone who put time into making this happen. Super useful!

[reyang]

Should be fixed by #619, big thank you to @ewhauser!

[akoshochrein]

Huge thanks for the fix!