Adaptive histogram based on something like streaming k-means crossed with Q-digest.
The special characteristics of this algorithm are:
a) smaller summaries than Q-digest
b) works on doubles as well as integers.
c) provides part per million accuracy for extreme quantiles and typically <1000 ppm accuracy for middle quantiles
d) fast
e) simple
f) test coverage > 90%
g) easy to adapt for use with map-reduce