The algorithm's accuracy is known. From the wiki[1]:

    The HyperLogLog algorithm is able to estimate 
    cardinalities of > 10^9 with a typical error rate of 2%

[1] https://en.wikipedia.org/wiki/HyperLogLog

But what about the implementation accuracy? :)

Tests against both historical and synthetic datasets.

Reddit probably has enough analytics to be able to show mathematically that it will be accurate without simulating any requests.

Can't you just use Apache Benchmark and some proxies?