I have released PyAthena 3.39.0, a minor release in the 3.x maintenance line. It backports bug fixes from master for Arrow credentials, result conversion, and SQLAlchemy DDL. The release notes contain the complete list of changes.

Arrow credentials

ArrowCursor, AsyncArrowCursor, and AioArrowCursor now read result files with the connection session's credentials. This fixes profiles backed by credential_process, IAM Identity Center, or an assumed role, which could previously fall back to PyArrow's default credential chain. An explicitly supplied session= also takes precedence over separate credential arguments.

With the role_arn connection argument, PyArrow reuses the role credentials already obtained by the connection, including MFA, instead of assuming the role again. Those credentials expire after duration_seconds, as they do for the connection's boto3 clients. Credential-resolution errors now propagate instead of silently switching to PyArrow's default credential chain.

Result conversion fixes

Native-format arrays now preserve empty strings and whitespace-only items, and stop splitting a value such as a,b into separate elements. For example, ARRAY['x', '', 'y'] returns ['x', '', 'y'], and ARRAY['a,b', 'c'] returns ['a,b', 'c']. Leading and trailing spaces within items are preserved too.

With array(json), map(..., json), or row(... json) type hints, JSON string elements such as "123" stay strings. The fixes also prevent JSONDecodeError for JSON fields in rows and JSON arrays whose first element is not a string.

For PandasCursor with chunksize, PandasDataFrameIterator.get_chunk() now applies the same TIME conversion as iteration. NULL TIME values read from result files become None instead of NaT, including in as_pandas(). With managed query result storage, pandas and Polars results containing TIME columns no longer fail during conversion. Closing a Polars result set also stops an existing iter_chunks() iterator.

SQLAlchemy table detection

DDL now treats a table as Iceberg only when its table_type property is ICEBERG, compared case-insensitively. Unrelated mentions of table_type and iceberg in table properties no longer cause a Hive table to compile with the Iceberg layout. On an S3 Tables catalog, such a table now raises CompileError instead of compiling as Iceberg.

Upgrading

Install this release with:

python -m pip install --upgrade 'PyAthena==3.39.0'

Keep any extras used by the application, for example 'PyAthena[SQLAlchemy]==3.39.0'. The minimum tenacity version is now 8.0.1, up from 4.1.0. In retry_api_call(), versions before 6.1.0 fail on Python 3.11+, and 8.0.0 raises TypeError on the first retry. The 3.x line continues to support Python 3.10 and retains its sqlalchemy>=1.0.0 dependency requirement.

The full changelog lists the backported pull requests.