A Critical Analysis of Recursive Model Indexes

Learned indexes like the recursive model index (RMI) have recently been introduced as a machine-learned replacement for traditional indexes with possibly game-changing results for how database indexes are constructed and used. Has the time come to discard our good old hand-crafted index structures that have been invented over the past decades? We believe that such a bold claim -- with substantial impact on the database world -- is worth a deep examination that clarifies when RMIs have benefits and when not. We present the first inventor-independent study critically examining RMIs. To do so, we revisit the original paper and carefully reimplemented RMIs. We proceed by reproducing the most important experiments from the original paper and follow-up papers all involving the inventors. We extend the original experiments by adding more baselines and considering more configurations. Further, we give insight on why and when RMIs perform well. Our results show that while the general observation of the original work that "any index is a model of the underlying data" is truly inspiring, some conclusions drawn in the original work may mislead database architects to take unfortunate and too radical design decisions. In particular, we show that other types of indexes outperform RMIs in some situations. In addition, we will show that the performance of RMIs is surprisingly sensitive to different data distributions. We conclude by giving a clear guideline for database architects when to use RMIs, other learned indexes, or traditional indexes.