RUSLI: Real-time Updatable Spline Learned Index

Machine learning algorithms have accelerated data access through ‘learned index’, where a set of data items is indexed by a model learned on the pairs of data key and the corresponding record’s position in the memory. Most of the learned indexes require retraining of the model for new data insertions in the data set. The retraining is expensive and takes as much time as the model training. So, today, learned indexes are updated by retraining on batch inserts to amortize the cost. However, real-time applications, such as data-driven recommendation applications need to access users’ feature store in real-time both for reading data of existing users and adding new users as well. This motivates us to present a real-time updatable spline learned index, RUSLI, by learning the distribution of data keys with their positions in memory through splines. We have extended RadixSpline [8] to build the updatable learned index while supporting real-time inserts in a data set without affecting the lookup time on the updated data set. We have shown that RUSLI can update the index in constant time with an additional temporary memory of size proportional to the number of splines. We have discussed how to reduce the size of the presented index using the distribution of spline keys while building the radix table. RULSI is shown to incur 270ns for lookup and 50ns for insert operations. Further, we have shown that RUSLI supports concurrent lookup and insert operations with a throughput of 40 million ops/sec. We have presented and discussed performance numbers of RUSLI for single and concurrent inserts, lookup, and range queries on SOSD [9] benchmark.