While researching logs from a large set of audits, I noticed that nearly
all of them had streamIDs starting with 0 or 1. This seemed very odd,
because streamIDs are supposed to be pretty much entirely random, and
every hex digit from 0-f should have been represented with roughly equal
frequency.
It turned out that our A-Chao implementation of reservoir sampling is
flawed. As far as we can tell, so is the Wikipedia implementation. No
one has yet reviewed the original 1982 paper by Dr. Chao in enough
detail to know where the error originated, but we do know that we have
been auditing segments near the beginning of the segment loop (low
streamIDs) far more often than segments near the end of the segment loop
(high streamIDs).
This change uses an algorithm Wikipedia calls "A-Res" instead, and adds
a test to check for that sort of bias creeping back in somehow. A-Res
will be slightly slower than A-Chao, because of a few extra steps that
need to be done, but it does appear to be selecting items uniformly.
Change-Id: I45eba4c522bafc729cebe2aab6f3fe65cd6336be