counters and logarithmically bucketed histograms for distributed systems

 A metric system for high performance counters and histograms. Unlike
 popular metric systems today, this does not destroy the accuracy of
 histograms by sampling. Instead, a logarithmic bucketing function
 compresses values, generally within 1% of their true value (although
 between 0 and 1 the precision loss may not be within this boundary). This
 allows for extreme compression, which allows us to calculate arbitrarily
 high percentiles with no loss of accuracy - just a small amount of
 precision. This is particularly useful for highly-clustered events that
 are tolerant of a small precision loss, but for which you REALLY care
 about what the tail looks like, such as measuring latency across a
 distributed system.