bcaine / learned_indices

A C++11 implementation of the B-Tree part of "The Case for Learned Index Structures"

A research proof of concept that implements the B-Tree section of The Case for Learned Index Structures paper in C++.

The general design is to have a single lookup structure that you can parameterize with a KeyType and a ValueType, and an overflow list that keeps new inserts until you retrain. There is a value in the constructor of the RMI that triggers a retrain when the overflow array reaches a certain size.

The basic API:

// [first/second]StageParams are network parameters
int maxAllowedError = 256;
int maxBufferBeforeRetrain = 10001;
auto modelIndex = RecursiveModelIndex<int, int, 128> recursiveModelIndex(firstStageParams, 
                                                                         secondStageParams, 
                                                                         maxAllowedError, 
                                                                         maxBufferBeforeRetrain);

for (int ii = 0; ii < 10000; ++ii) {
    modelIndex.insert(ii, ii * 2);
}

// Since we still have one more insert before retraining, retrain before searching...
modelIndex.train();
 
auto result = modelIndex.find(5);
 
if (result) {
    std::cout << "Yay! We got: " << result.get().first << ", " << result.get().second << std::endl;
} else {
    std::cout << "Value not found." << std::endl; // This shouldn't happen in the above usage...
}

See src/main.cpp for a usage example where it stores scaled log normal data.

Dependencies

• nn_cpp - Eigen based minimalistic C++ Neural Network library
• cpp-btree - A fast C++ implementation of a B+ Tree

TODO:

• Lots of code cleanup
• Profiling of where the slowdowns are. On small tests, the cpp_btree lib beats it by 10-100x
  • Eigen::TensorFixed in nn_cpp would definitely help
  • Increasing dataset size may lead to more of an advantage to the RMI
  • Being much, much more efficient with memory and conversions (lots of casting)
• Very non-linear data the second stage tends to break down or stop performing on.
• A non-trivial amount of our second stage "dies" in the sense that we don't use it for predictions.
  • The larger the dataset, or the more second stage nodes, the more likely this is. Bug somewhere?
• Experimenting/tuning of training parameters
  • Still more learning rate sensitive than I'd like
• Checking, and failing if there are non-integer keys
• Tests on the actual RMI code (instead of using tests for experiments)
• Move retrain to non-blocking thread
• Logging