Apache SystemDS 3.3.0 Release Notes

Release notes for SystemDS 3.3.0

SystemDS 3.3 is a minor release. Release 3.3 contains new features and major improvements to existing features.

New Features

Built-in Functions

  • Added sliceLineExtract for efficient slicing operations.
  • Introduced adasyn for synthetic data generation in imbalanced datasets.
  • Added GELU Activation Function (approximation) for deep learning tasks.
  • New built-in for vectorized simple exponential smoothing.
  • Python API Enhancements:
    • Added built-ins for solve, median, quantile, lu, qr, cov, and more.
    • Introduced cumulative aggregate functions like cumsum, cumprod, etc.
    • Added support for matrix operations like diag, trace, sqrt, exp, floor, and ceil.

Performance Enhancements

  • Optimized matrix multiplication, including multi-threaded dot-product operations.
  • Improved performance for CSV and MatrixMarket readers with parallelism.
  • Enhanced incremental slice-line pruning for better efficiency.
  • Optimized federated operations with improved threading and cost estimation.

Federated Learning

  • Fixed threading bugs in federated operations.
  • Improved federated planner memoization and added support for weighted covariance.

Compression

  • Introduced workload-aware compression for intermediates.
  • Enhanced compressed matrix operations with new tests and optimizations.

Scuro Framework

  • Added support for multi-head attention layers and LSTM representations.
  • Introduced new modalities and data representations for Scuro.

Bug Fixes

  • Fixed robustness issues in built-in functions:
    • Corrected trace handling for non-square matrices.
    • Fixed seq() length computation for edge cases.
    • Resolved issues in transformencode for federated backends.
  • Addressed edge cases in matrix operations:
    • Sparse-dense matrix multiplication.
    • Incremental slice-line pruning.
  • Fixed federated operation bugs:
    • Improved robustness for federated weighted covariance.
    • Resolved flaky tests in federated primitives.

Code Quality and Testing

  • Improved test coverage:
    • Simplification rewrites, federated operations, and Python APIs.
    • Added new tests for matrix operations and compression.
  • Refactored and cleaned up code for better maintainability and performance.

Documentation

  • Updated Python API documentation with examples for new built-ins.
  • Improved overall documentation for better usability.

Dependency Updates

  • Updated library versions and dependencies:
    • Bumped codecov-action to the latest version.
    • Updated other dependencies for compatibility and security.

Miscellaneous

  • Added new rewrite rules for simplifications:
    • pushdown-sum and rm-reorg for better performance.
  • Introduced new I/O support for cloud-optimized GeoTIFF files.
  • Added new determinant function and kernels for matrix operations.