pypi duckdb 2.0.1
v2.0.1 Bugfix Release

one hour ago

DuckDB Python 2.0.1 is a bugfix release for the 2.0 line. It fixes the hangs and crashes at interpreter exit that 2.0.0 introduced, a number of wrong results when scanning pandas, numpy, Arrow and polars data, and several long-standing conversion bugs. It ships DuckDB v2.0.1.

Exit and shutdown

  • A script that ends with an open result over a pandas object column no longer hangs at exit. Every live result is now ended by the module's exit hook, before the interpreter finalizes, including Arrow readers and capsules exported from a result (#709).
  • A process whose daemon threads still run queries over Python data no longer crashes or hangs at exit on Python 3.14. A query that starts after the exit handlers ran fails with InvalidInputException: the Python interpreter is shutting down instead (#719, fixes #714).
  • The nanobind: leaked N instances! report at exit is gone from release builds. Debug and assertion builds keep it (#784, fixes #732 and #736). We're following up with better object tracking to prevent the leaks from happening.
  • A map() function that calls sys.exit() no longer ends or hangs the process; it raises an engine error like every other exception from the function (#746).

Results and conversion

  • A fetch whose conversion fails, such as invalid UTF-8 in a VARCHAR or a Python error from a datetime conversion, now ends its result. The next statement runs instead of failing with ResourceInUseException (#712).
  • TIMESTAMPTZ values are converted to Python through zoneinfo instead of pytz. A clean install no longer fails with "Required module 'pytz' failed to import", and instants from 2038 on get the right DST offset (#710, #406 and #579).
  • Comparing a DuckDB type with a Python class no longer constructs the class. A class is read as a numpy scalar type only when it is a concrete subclass of numpy.generic; a KeyboardInterrupt raised while an operand converts reaches the caller instead of being swallowed. Lone surrogates in identifiers and SQL text are refused with a clear error (#747).

Scanning pandas, numpy and Arrow data

  • A pandas 2 categorical read through a strided view (df[::2], iloc[1::2]) returned the rows of the full column. Strided and reversed views, tuples of arrays and pandas 2 Series now read correctly or are refused with a clear error instead of std::bad_cast (#711).
  • A sliced numeric numpy column (iloc[1::3]) could read wrong values (#724).
  • con.append and Spark createDataFrame with a schema now convert an object column into the declared column type, cell by cell, instead of sampling the column and silently truncating values the sample missed (#739).
  • con.appender accepts numpy scalars (#748).
  • pandas 2 dictionary-encoded string columns are read (#745).
  • The Arrow importer reads three layouts it refused before: a struct or union below an offset, empty lists over a dictionary child, and a run-end encoded child inside a fixed size list. Sliced pyarrow Tables with nested struct columns are the common case (#730). Arrow refusals now name the column instead of its position (#745).

map()

  • map() keeps its input's column types: an ENUM no longer comes back as TINYINT with NULL as -1, and STRUCT, MAP, ARRAY, UNION, TIME_NS and nested columns no longer fail with a type mismatch. A whole chunk of NULLs in an object column no longer trips the type check (#744, #746).

Views, polars and Parquet pruning

  • A view created with to_view or create_view over a relation that reads a file-like object (read_csv(StringIO(...))) keeps working after the relation is collected (#741, fixes #477).
  • A polars lazy frame over a DuckDB relation pushes a filter to DuckDB only where it reads the same as polars. A filter that casts the column before comparing, such as a timestamp cast to milliseconds, no longer returns fewer rows than the eager frame (#740).
  • Parquet row group pruning works for TIMESTAMPTZ and TIME constants that lie between two points of a coarser column, and for points before 1970 (#745).

Expression API

  • Aggregate calls take an ORDER BY inside the call: col("x").first(order_by=col("t")) and fn("first", col("x")).order_by(col("t")) render "first"("x" ORDER BY "t"). Works with FILTER, IGNORE NULLS and over() (#713).

Engine

  • Ships DuckDB v2.0.1 (8d78114e9). Behavior that follows the engine: a parse error in SQL that runs aborts the active transaction; EXECUTE of a prepared statement over a register()ed source fails after unregister(); every finite timestamp renders as text; read_json reads DECIMAL from the number's text (#721, #729, #734, #770, #772, #778 and the submodule bumps).

Build and CI

  • The macOS wheels no longer carry a 10.7 deployment minimum on the module's objects (#749).
  • CI test jobs install the test group with uv sync (#728); the release workflow's S3 staging upload pins the classic transfer client and retries (#750).

Full Changelog: v2.0.0...v2.0.1

Don't miss a new duckdb release

NewReleases is sending notifications on new releases.