Plateau

===========================

This will be the final release in the 3.X series. Please ensure your existing
codebase does not raise any DeprecationWarning from kartothek and migrate your
import paths ahead of time to the new ``kartothek.api`` modules to ensure a
smooth migration to 4.X.

* Introduce ``kartothek.api`` as the public definition of the API. See also :doc:`versioning`.
* Introduce `DatasetMetadataBase.schema` to prepare deprecation of `table_meta`
* ``kartothek.io.eager.read_dataset_as_dataframes`` and
``kartothek.io.iter.read_dataset_as_dataframes__iterator`` now correctly return
categoricals as requested for misaligned categories.

===========================

* Allow ``pyarrow==3`` as a dependency.
* Fix a bug in ``kartothek.io_components.utils.align_categories`` for dataframes
with missings and of non-categorical dtype.
* Fix an issue with the cube index validation introduced in v3.19.0 (413).

===========================

* Fix an issue where updates on cubes or updates on datatsets using
dask.dataframe might not update all secondary indices, resulting in a corrupt
state after the update
* Expose compression type and row group chunk size in Cube interface via optional
parameter of type ``kartothek.serialization.ParquetSerializer``.
* Add retries to ``kartothek.serialization._parquet.ParquetSerializer.restore_dataframe``
IOErrors on long running ktk + dask tasks have been observed. Until the root cause is fixed,
the serialization is retried to gain more stability.

===========================

* Add ``cube.suppress_index_on`` to switch off the default index creation for dimension columns
* Fixed the import issue of zstd module for ``kartothek.core _zmsgpack``.
* Fix a bug in ``kartothek.io_components.read.dispatch_metapartitions_from_factory`` where
`dispatch_by=[]` would be treated like `dispatch_by=None`, not merging all dataset partitions into
a single partitions.

===========================

* Allow ``pyarrow==2`` as a dependency.

===========================

* 378 Improve logging information for potential buffer serialization errors