(All credit for the idea in this comment is due to @aphyr)
Basically because the transactions modified keys selected from a uniform distribution, the probability of contention was extremely low. AKA this workload is basically a data-parallel problem, somewhat lessening the impressiveness of the high throughput. Would be interesting to see it with a Zipfian distribution (or even better, a Biebermark [0])
See this tweet by @aphyr: https://twitter.com/aphyr/status/542755074380791809
(All credit for the idea in this comment is due to @aphyr)
Basically because the transactions modified keys selected from a uniform distribution, the probability of contention was extremely low. AKA this workload is basically a data-parallel problem, somewhat lessening the impressiveness of the high throughput. Would be interesting to see it with a Zipfian distribution (or even better, a Biebermark [0])
[0] - http://smalldatum.blogspot.co.il/2014/04/biebermarks.html