modin-project/modin 0.24.0

Modin 0.24.0

on GitHub

This release upgrades the pandas version to 2.1, updates the minimum supported python version up to 3.9, introduces ModinDataLoader to improve interaction with PyTorch, fixes several issues with interchange protocol that solved known compatibility issues with Plotly, Seaborn and Altair, includes new version of HDK 0.8. It also includes some other new features, and many bug fixes.

Key Features and Updates Since 0.23.0

• Stability and Bugfixes
  • FIX-#0000: Don't test experimental xgboost with Ray nightly build (#6424)
  • FIX-#0000: Fix xgboost tests with ray>2.6.0 (#6425)
  • FIX-#1930: Fix one of the cases of heterogeneous data for read_csv (#5507)
  • FIX-#4347: read_excel: defaults to pandas for unsupported types of 'io' (#6462)
  • FIX-#4580: Fix access by row label in query and eval (#6488)
  • FIX-#4687: Change Column.null_count to return a built-in int instead of NumPy scalar (#6526)
  • FIX-#5164: Fix unwrap_partitions for virtual partitions when axis=None (#6560)
  • FIX-#5536: Remove branch disabling __getattribute__ for experimental mode (#6529)
  • FIX-#5627: Stop checking temp_df.dtype == 'category' (#6360)
  • FIX-#5972: Compute correct dtype for Series.str.find/index/rfind/rindex (#6426)
  • FIX-#6219: Don't default to pandas for copy on empty DataFrame/Series objects (#6371)
  • FIX-#6299: __array__ method always returns array of vanilla numpy (#6300)
  • FIX-#6334: Improve error message if HDK isn't installed in the environment (#6358)
  • FIX-#6347: Remove 'modin in the cloud' experimental feature (#6408)
  • FIX-#6364: Make reshuffling work with BenchmarkMode.put(True) (#6365)
  • FIX-#6367: Enable support for groupby.size() in reshuffling groupby (#6370)
  • FIX-#6368: Apply deferred indices before map-reduce groupby (#6369)
  • FIX-#6372: Precompute dtypes for sum operation (#6421)
  • FIX-#6375: Don't initialize engines at import time (#6374)
  • FIX-#6386: Don't make unnecessary astype calls for modin.array.sum op (#6395)
  • FIX-#6392: Compute dtypes for the DataFrame.mean() result (#6520)
  • FIX-#6394: Preserve dtypes for __setitem__ op when using not hashable key (#6547)
  • FIX-#6396: Set __factory to None in case of any problems during initialization (#6397)
  • FIX-#6402: Allow datetime and timedelta types in diff (#6403)
  • FIX-#6405: Apply disable_logging to __getattr__ (#6406)
  • FIX-#6410: Add a link to @modin_project twitter (#6411)
  • FIX-#6414: Fix read_feather with pyarrow<11.0 (#6415)
  • FIX-#6427: Make code compatible with flake8==6.1.0 (#6428)
  • FIX-#6429: Exclude pymssql==2.2.8 from environments (#6430)
  • FIX-#6436: Support ~ in paths in IO functions correctly (#6448)
  • FIX-#6443: Cast boolean columns before sum|mean|median groupby aggregations (#6444)
  • FIX-#6446: Stop requiring modin-xgboost approval (#6447)
  • FIX-#6456: Create fake xgboost module for building docs (#6457)
  • FIX-#6459: Support fastparquet>=2023.1.0 (#6458)
  • FIX-#6465: Fix groupby.apply() for UDFs that change the output's shape (#6506)
  • FIX-#6479: HDK CalciteBuilder: Do not call is_bool_dtype() for categorical (#6480)
  • FIX-#6483: Default to pandas for __array_ufunc__ (#6486)
  • FIX-#6509: Fix 'reshuffling' in case of a string key (#6510)
  • FIX-#6514: test_sort_cols_str from test_dataframe.py crashed on HDK 0.7.0 and python 3.9 (#6515)
  • FIX-#6516: HDK: test_dataframe.py is crashed if Calcite is disabled (#6517)
  • FIX-#6518: Fix interchange protocol for string columns (#6523)
  • FIX-#6519: Consider botocore as an optional dependency (#6521)
  • FIX-#6532: Fix read_excel so that it doesn't use rich_text param for old openpyxl (#6534)
  • FIX-#6535: Pin s3fs<2023.9.0 (#6536)
  • FIX-#6537: Unpin s3fs<2023.9.0 (#6544)
  • FIX-#6540: Correct handling of range indices and index names in read_parquet (#6545)
  • FIX-#6541: Fix ValueError: buffer source array is read-only for iloc (#6538)
  • FIX-#6549: Remove usage of dfsql module (#6550)
  • FIX-#6552: Avoid FutureWarnings in groupby unless necessary (#6595)
  • FIX-#6553: Fix read_csv with iterator=True (#6554)
  • FIX-#6558: Normalize the number of partitions after .read_parquet() (#6559)
  • FIX-#6561: Remove MODIN_OMNISCI_* env vars in favor of MODIN_HDK_* (#6562)
  • FIX-#6565: Don't implement map function via applymap (#6566)
  • FIX-#6572: Execute simple queries row-wise in pandas backend (#6575)
  • FIX-#6582: Avoid FutureWarnings in bfill/backfill/ffill/pad unless necessary (#6599)
  • FIX-#6587: Use different env files for unidist engine for windows and linux (#6588)
  • FIX-#6601: sort_values shouldn't affect source dataframe/series (#6603)
• Performance enhancements
  • PERF-#6332: Don't materialize axes in concat operation (#6381)
  • PERF-#6373: Preserve dtypes cache for _repartition (#6376)
  • PERF-#6378: Use numpy.array operations in internals of iloc/loc operation (#6393)
  • PERF-#6388: Avoid masking in __getitem__ when the number of rows to be taken > 90% (#6423)
  • PERF-#6398: Improved performance of list-like objects insertion into DataFrames (#6476)
  • PERF-#6433: Implement .dropna() using map-reduce pattern (#6472)
  • PERF-#6437: Preserve dtypes for reindex (#6438)
  • PERF-#6464: Improve reshuffling for multi-column groupby in low-cardinality cases (#6533)
  • PERF-#6466: Verify indices equality without triggering any computations (#6491)
  • PERF-#6478: Do not propagate new columns if they're identical to the previous ones (#6481)
  • PERF-#6524: Add a 'column' shape hint for the results of qc.to_datetime() (#6525)
  • PERF-#6583: Remove redundant index reassignment in query() (#6584)
  • PERF-#6590: Chunk axes independently in .from_pandas() (#6591)
• Refactor Codebase
  • REFACTOR-#4278: Remove unused arguments from BasePandasDataset.apply (#6451)
  • REFACTOR-#4902: Use isort (#6551)
  • REFACTOR-#6470: Remove Patcher internal class (#6471)
  • REFACTOR-#6489: Enforce API-layer bool/integer argument for __invert__ (#6490)
  • REFACTOR-#6569: Use contextlib.nullcontext instead of custom one (#6570)
  • REFACTOR-#6576: Don't use deprecated is_int64_dtype and is_period_dtype function (#6577)
• Update testing suite
  • TEST-#0000: Download ray wheel for python 3.9 (#6513)
  • TEST-#2008: Reduce runtime of CI checks a lot (#6356)
  • TEST-#4270: Revert disabling time_groupby_agg_nunique ASV bench (#6564)
  • TEST-#4348: Use psycopg2-binary for testing and developing purpose (#6573)
  • TEST-#4477: Add tests for df.eval with scalar and groupby.transofm call in the expr (#6546)
  • TEST-#4643: Add interchange test for empty dataframe (#6454)
  • TEST-#5008: Set benchmark mode within unit test instead of with environment variable (#6359)
  • TEST-#6349: Update minimum versions for test dependencies in general environments (#6350)
  • TEST-#6439: Create HDK environment manually for ASV (#6431)
  • TEST-#6449: Run tests in test_dmatrix.py only for Ray engine (#6450)
  • TEST-#6460: Don't use repr to force materialization (#6461)
  • TEST-#6469: Pin numexpr<2.8.5 (#6474)
  • TEST-#6477: Update ASV to 0.5.1 (#6432)
  • TEST-#6497: Remove boto3 from environments to speedup creation (#6496)
  • TEST-#6505: Update python version for ASV benchmarks on HDK (#6504)
  • TEST-#6593: Adapt tests for pandas 2.1.1 (#6592)
• Documentation improvements
  • DOCS-#0000: Update CI link in README to show only pushes (#6531)
  • DOCS-#6416: Fix import path for spreadsheet feature (#6581)
  • DOCS-#6419: Clarify read_parquet supported parameters (#6420)
  • DOCS-#6452: Update copyright year (#6453)
• New Features
  • FEAT-#1611: Add some datetime extraction functions for HDK (#6568)
  • FEAT-#5645: Add support for modin's numpy array in dataframe.insert function (#6400)
  • FEAT-#6139: DataLoader interplay. (#6140)
  • FEAT-#6377: HDK: Do not keep reference to arrow table imported to HDK (#6380)
  • FEAT-#6389: Make sure git ignores logs in .modin folder (#6390)
  • FEAT-#6401: Support compression param and more file extensions in to_parquet (#6404)
  • FEAT-#6407: Update minimum dependency versions (#6342)
  • FEAT-#6417: Add support for filters to read_parquet (#6442)
  • FEAT-#6434: HDK: Do not convert dictionary columns to string when importing arrow tables (#6435)
  • FEAT-#6440: Use different HDK parameters for different queries (#6441)
  • FEAT-#6484: HDK: Add support for nlargest/nsmallest groupby aggregation (#6485)
  • FEAT-#6500: HDK: Add support for datetime64 to int64 cast (#6501)
  • FEAT-#6502: HDK: Add enable_multifrag_execution_result=1 HDK launch parameter (#6503)
  • FEAT-#6511: Update the minimum supported python version up to 3.9 (#6508)
  • FEAT-#6522: Update to pandas 2.1.0 (#6512)
  • FEAT-#6527: HDK: Add support for the quantile group by aggregation. (#6528)
  • FEAT-#6597: Bump pyhdk version to 0.8 (#6598)

Contributors

@AndreyPavlenko
@RehanSD
@YarShev
@anmyachev
@dchigarev
@mvashishtha
@vnlitvinov
@abykovsk
@zmbc
@noloerino
@rentruewang