I present the cache-oblivious streaming B-tree, a high performance alternative to the traditional B-tree. Modern databases and file systems are based on B-trees. As a result, they can exhibit performance cliffs and unpredictable run times. Replacing B-trees with cache-oblivious streaming B-trees can remove some of these performance deficiencies.
I explain some of the technical issues that we needed to address to turn the streaming B-tree prototype into an industrial-strength product. Then I explain how legacy and established practice influenced our engineering decisions.