For details of the versioning schema, see Versioning.


Version 5.3.0 (2021-12-10)

  • Add Deprecation warnings and migration helpers in order to facilitate the Kartothek version 6.0.0 migration.

  • Removed warning for distinct categoricals (#501)

Version 5.2.0 (2021-11-22)

  • Remove support for Python 3.6

  • Allow pyarrow<7 as a dependency.

Version 5.1.0 (2021-07-05)

Version 5.0.0 (2021-06-23)

This release rolls all the changes introduced with 4.x back to 3.20.0.

As the incompatibility between 4.0 and 5.0 will be an issue for some customers, we encourage you to use the very stable kartothek 3.20.0 and not version 4.x.

Please refer the Issue #471 for further information.

Kartothek 4.0.3 (2021-06-10)

  • Pin dask to not use 2021.5.1 and 2020.6.0 (#475)

Kartothek 4.0.2 (2021-06-07)

  • Fix a bug in MetaPartition._reconstruct_index_columns that would raise an IndexError when loading few columns of a dataset with many primary indices.

Kartothek 4.0.1 (2021-04-13)

  • Fixed dataset corruption after updates when table names other than “table” are used (#445).

Kartothek 4.0.0 (2021-03-17)

This is a major release of kartothek with breaking API changes.

  • Removal of complex user input (see gh427)

  • Removal of multi table feature

  • Removal of module

  • class DatasetMetadata now has an attribute called schema which replaces the previous attribute table_meta and returns only a single schema

  • All outputs which previously returned a sequence of dictionaries where each key-value pair would correspond to a table-data pair now returns only one pandas.DataFrame

  • All read pipelines will now automatically infer the table to read such that it is no longer necessary to provide table or table_name as an input argument

  • All writing pipelines which previously supported a complex user input type now expose an argument table_name which can be used to continue usage of legacy datasets (i.e. datasets with an intrinsic, non-trivial table name). This usage is discouraged and we recommend users to migrate to a default table name (i.e. leave it None / table)

  • All pipelines which previously accepted an argument tables to select the subset of tables to load no longer accept this keyword. Instead the to-be-loaded table will be inferred

  • Trying to read a multi-tabled dataset will now cause an exception telling users that this is no longer supported with kartothek 4.0

  • The dict schema for to_dict() and from_dict() changed replacing a dictionary in table_meta with the simple schema

  • All pipeline arguments which previously accepted a dictionary of sequences to describe a table specific subset of columns now accept plain sequences (e.g. columns, categoricals)

  • Remove the following list of deprecated arguments for io pipelines * label_filter * central_partition_metadata * load_dynamic_metadata * load_dataset_metadata * concat_partitions_on_primary_index

  • Remove output_dataset_uuid and df_serializer from since these arguments didn’t have any effect

  • Remove metadata, df_serializer, overwrite, metadata_merger from

  • store_dataframes_as_dataset() now requires a list as an input

  • Default value for argument date_as_object is now universally set to True. The behaviour for False will be deprecated and removed in the next major release

  • No longer allow to pass delete_scope as a delayed object to update_dataset_from_ddf()

  • update_dataset_from_ddf() and store_dataset_from_ddf() now return a dd.core.Scalar object. This enables all dask.DataFrame graph optimizations by default.

  • Remove argument table_name from collect_dataset_metadata()

Version 3.20.0 (2021-03-15)

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.

Version 3.19.1 (2021-02-24)

  • Allow pyarrow==3 as a dependency.

  • Fix a bug in 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).

Version 3.19.0 (2021-02-12)

  • 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 ParquetSerializer.

  • Add retries to 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.

Version 3.18.0 (2021-01-25)

  • 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 where dispatch_by=[] would be treated like dispatch_by=None, not merging all dataset partitions into a single partitions.

Version 3.17.3 (2020-12-04)

  • Allow pyarrow==2 as a dependency.

Version 3.17.2 (2020-12-01)

  • #378 Improve logging information for potential buffer serialization errors

Version 3.17.1 (2020-11-24)


  • Fix GitHub #375 by loosening checks of the supplied store argument

Version 3.17.0 (2020-11-23)


  • Improve performance for “in” predicate literals using long object lists as values

  • commit_dataset() now allows to modify the user metadata without adding new data.


Version 3.16.0 (2020-09-29)

New functionality

  • Allow filtering of nans using “==”, “!=” and “in” operators


  • Fix a regression which would not allow the usage of non serializable stores even when using factories

Version 3.15.1 (2020-09-28)

  • Fix a packaging issue where typing_extensions was not properly specified as a requirement for python versions below 3.8

Version 3.15.0 (2020-09-28)

New functionality

  • Add store_dataset_from_ddf() to offer write support of a dask dataframe without update support. This forbids or explicitly allows overwrites and does not update existing datasets.

  • The sort_partitions_by feature now supports multiple columns. While this has only marginal effect for predicate pushdown, it may be used to improve the parquet compression.

  • build_cube_from_dataframe now supports the shuffle methods offered by store_dataset_from_ddf() and update_dataset_from_ddf() but writes the output in the cube format


  • Reduce memory consumption during index write.

  • Allow simplekv stores and storefact URLs to be passed explicitly as input for the store arguments

Version 3.14.0 (2020-08-27)

New functionality

  • Add hash_dataset functionality


  • Expand pandas version pin to include 1.1.X

  • Expand pyarrow version pin to include 1.x

  • Large addition to documentation for multi dataset handling (Kartothek Cubes)

Version 3.13.1 (2020-08-04)

  • Fix evaluation of “OR”-connected predicates (#295)

Version 3.13.0 (2020-07-30)


  • Update timestamp related code into Ktk Discover Cube functionality.

  • Support backward compatibility to old cubes and fix for cli entry point.

Version 3.12.0 (2020-07-23)

New functionality

  • Introduction of cube Functionality which is made with multiple Kartothek datasets.

  • Basic Features - Extend, Query, Remove(Partitions), Delete (can delete entire datasets/cube), API, CLI, Core and IO features.

  • Advanced Features - Multi-Dataset with Single Table, Explicit physical Partitions, Seed based join system.

Version 3.11.0 (2020-07-15)

New functionality

Bug fixes

  • Performance of dataset update with delete_scope significantly improved for datasets with many partitions (#308)

Version 3.10.0 (2020-07-02)


  • Dispatch performance improved for large datasets including metadata

  • Introduction of dispatch_metadata kwarg to metapartitions read pipelines to allow for transition for future breaking release.

Bug fixes

  • Ensure that the empty (sentinel) DataFrame used in read_table() also has the correct behaviour when using the categoricals argument.

Breaking changes in

  • The dispatch_metapartitions and dispatch_metapartitions_from_factory will no longer attach index and metadata information to the created MP instances, unless explicitly requested.

Version 3.9.0 (2020-06-03)


  • Arrow 0.17.X support

  • Significant performance improvements for shuffle operations in update_dataset_from_ddf() for large dask.DataFrames with many payload columns by using in-memory compression during the shuffle operation.

  • Allow calling update_dataset_from_ddf() without partition_on when shuffle=True.

  • read_dataset_as_ddf() supports kwarg dispatch_by to control the internal partitioning structure when creating a dataframe.

  • read_dataset_as_ddf() and update_dataset_from_ddf() now allow the keyword table to be optional, using the default SINGLE_TABLE identifier. (recommended since the multi table dataset support is in sunset).

Version 3.8.2 (2020-04-09)


  • Read performance improved for, especially for partitioned datasets and queries with empty payload columns.

Bug fixes

  • GH262: Raise an exception when trying to partition on a column with null values to prevent silent data loss

  • Fix multiple index creation issues (cutting data, crashing) for uint data

  • Fix index update issues for some types resulting in TypeError: Trying to update an index with different types... messages.

  • Fix issues where index creation with empty partitions can lead to ValueError: Trying to create non-typesafe index

Version 3.8.1 (2020-03-20)


  • Only fix column odering when restoring DataFrame if the ordering is incorrect.

Bug fixes

  • GH248 Fix an issue causing a ValueError to be raised when using dask_index_on on non-integer columns

  • GH255 Fix an issue causing the python interpreter to shut down when reading an empty file (see also

Version 3.8.0 (2020-03-12)


  • Add keyword argument dask_index_on which reconstructs a dask index from an kartothek index when loading the dataset

  • Add method observed_values() which returns an array of all observed values of the index column

  • Updated and improved documentation w.r.t. guides and API documentation

Bug fixes

  • GH227 Fix a Type error when loading categorical data in dask without specifying it explicitly

  • No longer trigger the SettingWithCopyWarning when using bucketing

  • GH228 Fix an issue where empty header creation from a pyarrow schema would not normalize the schema which causes schema violations during update.

  • Fix an issue where create_empty_dataset_header() would not accept a store factory.

Version 3.7.0 (2020-02-12)


Version 3.6.2 (2019-12-17)


Bug fixes

Version 3.6.1 (2019-12-11)

Bug fixes

  • Fix a regression introduced in 3.5.0 where predicates which allow multiple values for a field would generate duplicates

Version 3.6.0 (2019-12-03)

New functionality

  • The partition on shuffle algorithm in update_dataset_from_ddf() now supports producing deterministic buckets based on hashed input data.

Bug fixes

  • Fix addition of bogus index columns to Parquet files when using sort_partitions_by.

  • Fix bug where partition_on in write path drops empty DataFrames and can lead to datasets without tables.

Version 3.5.1 (2019-10-25)

Version 3.5.0 (2019-10-21)

New functionality

  • Add support for pyarrow 0.15.0

  • Additional functions in kartothek.serialization module for dealing with predicates * check_predicates() * filter_predicates_by_column() * columns_in_predicates()

  • Added available types for type annotation when dealing with predicates * ~kartothek.serialization.PredicatesType * ~kartothek.serialization.ConjunctionType * ~kartothek.serialization.LiteralType

  • Make*read_table* methods use default table name if unspecified

  • MetaPartition.parse_input_to_metapartition accepts dicts and list of tuples equivalents as obj input

  • Added secondary_indices as a default argument to the write pipelines

Bug fixes

  • Input to normalize_args is properly normalized to list

  • MetaPartition.load_dataframes now raises if table in columns argument doesn’t exist

  • require urlquote>=1.1.0 (where urlquote.quoting was introduced)

  • Improve performance for some cases where predicates are used with the in operator.

  • Correctly preserve ExplicitSecondaryIndex dtype when index is empty

  • Fixed DeprecationWarning in pandas CategoricalDtype

  • Fixed broken docstring for store_dataframes_as_dataset

  • Internal operations no longer perform schema validations. This will improve performance for batched partition operations (e.g. partition_on) but will defer the validation in case of inconsistencies to the final commit. Exception messages will be less verbose in these cases as before.

  • Fix an issue where an empty dataframe of a partition in a multi-table dataset would raise a schema validation exception

  • Fix an issue where the dispatch_by keyword would disable partition pruning

  • Creating dataset with non existing columns as explicit index to raise a ValueError

Breaking changes

  • Remove support for pyarrow < 0.13.0

  • Move the docs module from io_components to core

Version 3.4.0 (2019-09-17)

  • Add support for pyarrow 0.14.1

  • Use urlquote for faster quoting/unquoting

Version 3.3.0 (2019-08-15)

  • Fix rejection of bool predicates in filter_array_like() when bool columns contains None

  • Streamline behavior of store_dataset_from_ddf when passing empty ddf.

  • Fix an issue where a segmentation fault may be raised when comparing MetaPartition instances

  • Expose a date_as_object flag in kartothek.core.index.as_flat_series

Version 3.2.0 (2019-07-25)

  • Fix gh:66 where predicate pushdown may evalute false results if evaluated using improper types. The behavior now is to raise in these situations.

  • Predicate pushdown and filter_array_like() will now properly handle pandas Categoricals.

  • Add read_dataset_as_dataframe_bag()

  • Add

Version 3.1.1 (2019-07-12)

Version 3.1.0 (2019-07-10)


Version 3.0.0 (2019-05-02)

  • Initial public release