The case for a learned sorting algorithm

In recent years, machine learning (ML) has become an integral part of application development, permeating web giants, large consumer companies, and even traditional enterprises. While ML has made significant strides in these areas, its impact on systems software has been more limited. However, recent advancements like "The case for learned index structures" and SageDB are beginning to change this dynamic. Today's focus is on a classic computer science problem: sorting.

The paper "The case for a learned sorting algorithm" by Kristo, Vaidya, et al., presented at SIGMOD’20, explores the potential of using machine learning to enhance sorting performance. The authors propose a novel approach called Learned Sort, which leverages a model to predict the position of each data item in a sorted list. This model is trained on a subset of the data, learning an approximation of the cumulative distribution function (CDF) of the dataset.

The core idea behind Learned Sort is that if a model can predict the position of an item in a sorted list with reasonable accuracy, it can significantly speed up the sorting process. For instance, if an item is predicted to be at position 287 in the sorted list, it can be placed there directly. While a model with 100% accuracy would essentially need to memorize the positions of all items, making it slower than traditional sorting algorithms, the authors demonstrate that a sufficiently accurate model trained on a subset of the data can be both fast to train and effective in sorting.

To evaluate the performance of Learned Sort, the authors tested it on a 1 billion-item dataset. The results were impressive: Learned Sort outperformed the next best competitor, RadixSort, by a factor of 1.49x. Notably, this performance improvement includes the time taken to train the model. This means that the overall process of training the model and using it for sorting is still faster than traditional methods.

The authors achieved this breakthrough by combining the model's predictions with a sorting algorithm that works well with nearly-sorted arrays, such as Insertion Sort. The process involves two steps: first, the dataset is scanned, and each item is placed in its approximate position based on the model's predictions. Then, Insertion Sort is applied to the nearly-sorted list to produce a fully sorted result.

The key to the success of Learned Sort lies in the efficiency of training the model. The authors trained their model on a small subset of the data, which allowed them to learn an approximation of the CDF without requiring excessive computational resources. This approach ensures that the model can be trained quickly, making the overall sorting process faster than traditional methods.

In conclusion, "The case for a learned sorting algorithm" demonstrates that machine learning can be effectively applied to traditional systems software problems like sorting. By leveraging a model to predict item positions and combining it with a suitable sorting algorithm, Learned Sort achieves significant performance improvements over established methods. This innovation not only highlights the potential of ML in systems software but also opens the door to further exploration of learned algorithms for other classic problems. As ML continues to evolve, it will be interesting to see how these ideas are extended and applied to various domains, further enhancing the efficiency and performance of software systems.