Changelog#

This page contains the complete changelog for TanaT, documenting all notable changes, new features, bug fixes, and improvements across versions.

[v0.10.1] - June 2026#

Package structure cleanup

Subtype registration now relies on explicit __init__.py imports
type/__init__.py files own their subtype imports consistently across all modules

Polars Engine & Arrow Store - Full rewrite of the data layer for scalable, lazy sequence and trajectory manipulation

Repository moved to GitHub: TanaT is now hosted at Github. The Inria GitLab repository remains archived for historical versions up to v0.9.0.

Polars-based data engine: all internal data frames are now polars.LazyFrame, enabling lazy evaluation, predicate pushdown, and significant memory savings on large datasets
Arrow file store: sequences and trajectories are persisted as Apache Arrow files (SequenceStore, TrajectoryStore), replacing the previous in-memory/pandas approach
Virtual context layer (VirtualStore): lightweight copy-on-write mechanism for type conversions and derived views without materialising intermediate data
Ingestion sources: unified AbstractSource interface with DataFrameSource, CsvSource, ParquetSource, and SqlSource for SQL query ingestion
Cast recipe system (tanat.cast): structured type-cast overrides (SequenceCastRecipe, TrajectoryCastRecipe) applied lazily at view time, with validation (probe()) before execution
Frame assembler layer (SequenceFrameAssembler, TrajectoryFrameAssembler): centralises construction of view-schema LazyFrames, decoupling store layout from user-facing columns

Breaking: API redesign, v0.10.0 is not fully backward-compatible with v0.9.x

For historical versions below, please refer to the Inria GitLab archive.

Performance & Scalability - JIT parallel computing and memory-mapped storage

Numba-optimized metrics: JIT-compiled kernels
Parallel matrix computation: Pairwise distances computed in parallel with prange
Memory-mapped matrices: On-disk storage via MatrixStorageOptions for large dataset
Enhanced progress display: Structured, hierarchical progress tracking with timing for clustering and metric operations
Faceted plots: New facet() method creates grid visualizations from any static feature for multi-dimensional data exploration

Enhanced Validation & Analytics - Robust validation, metadata inference and statistical methods

Improve sequence/trajectory settings validation (Robust Pydantic-based validation, …)
Metadata inference/management system
Sequence type conversion
Clustering methods : Partitioning Around Medoids, Clustering Large Applications
Position/Rank-based methods: head(), tail(), slice() methods with negative indexing and step sampling
Statistical analysis methods: describe() method and statistics property for computing comprehensive sequence and trajectory statistics (entropy, vocabulary size, transition counts, etc.)

Foundation Release - Core architecture complete, ready for beta testing