DataFrame Serialization

Serialise Pandas DataFrames to/from bytes

Serialisation to bytes

For the serialsation, we need to pick a format serializer, you either use default_serializer() or explicitly select a serializer, e.g. ParquetSerializer.

from kartothek.api.serialization import ParquetSerializer

serializer = ParquetSerializer()
df = ..., "storage_key", df)


For deserialisation, you don’t have to instantiate any serializer as the correct one is determined from the filename.

from kartothek.api.serialization import DataFrameSerializer

df = DataFrameSerializer.restore_dataframe(store, "file.parquet")

Supported data types

Kartothek generally does not impose any restrictions on the data types to be used as long as they are compatible and in alignment with the pyarrow pandas integration.

For a detailed explanation about how types are handled, please consult Table type system.

Filtering / Predicate pushdown

You can provide a filter expression in a DNF in a format of a nested list where every inner list is interpreted as a logical conjunction (AND) whereas the entire expression is interpreted as one disjunction (OR)

prediactes = [
    [("ColumnA", "==", 5),],
    [("ColumnA", ">", 5), ("ColumnB", "<=",, 1, 1)),],

The above list of predicates can be interpreted as the following whereclause:

ColumnA = 5 OR (ColumnA > 5 AND ColumnB < '2021-01-01')

The predicate expression can be provided to the predicates keyword argument of the serializer and/or full dataset read interfaces.


All kartothek reading pipelines are exposing this predicates argument as well where it is not only used for predicate pushdown but also for partition pruning. See Efficient Querying for details.

The kartothek cube interface, see Kartothek Cubes, exposes a similar mechanism via the Condition and Conjunction classes.

Literals, operators and typing

The literals used for building the predicates are tuples with three elements.

(<FieldName: str>, <Operator: str>, <Value: Any>)
  • FieldName is a str identifying the column this literal describes.

  • Operator is a string for the logical operation applied to the field. Available operators are ==, !=, <=, >=, <, >, in

  • Value is the actual value for the query. The type of this value is always required to be identical to the fields data type. We apply the same type normalization for the predicates as described in Table type system.

Filtering for missing values / nulls is supported with operators ==, != and in and values np.nan and None for float and string columns respectively.