API of the
__dataframe__
protocol
¶
Specification for objects to be accessed, for the purpose of dataframe
interchange between libraries, via the
__dataframe__
method on a libraries’
data frame object.
For guiding requirements, see Protocol design requirements .
Concepts in this design ¶
-
A
Bufferclass. A buffer is a contiguous block of memory - this is the only thing that actually maps to a 1-D array in a sense that it could be converted to NumPy, CuPy, et al. -
A
Columnclass. A column has a single dtype. It can consist of multiple chunks . A single chunk of a column (which may be the whole column ifnum_chunks == 1) is modeled as again aColumninstance, and contains 1 data buffer and (optionally) one mask for missing data. -
A
DataFrameclass. A data frame is an ordered collection of columns , which are identified with names that are unique strings. All the data frame’s rows are the same length. It can consist of multiple chunks . A single chunk of a data frame is modeled as again aDataFrameinstance. -
A mask concept. A mask of a single-chunk column is a buffer .
-
A chunk concept. A chunk is a sub-dividing element that can be applied to a data frame or a column .
Note that the only way to access these objects is through a call to
__dataframe__
on a data frame object. This is NOT meant as public API;
only think of instances of the different classes here to describe the API of
what is returned by a call to
__dataframe__
. They are the concepts needed
to capture the memory layout and data access of a data frame.
Design decisions ¶
-
Use a separate column abstraction in addition to a dataframe interface.
Rationales:
-
This is how it works in R, Julia and Apache Arrow.
-
Semantically most existing applications and users treat a column similar to a 1-D array
-
We should be able to connect a column to the array data interchange mechanism(s)
Note that this does not imply a library must have such a public user-facing abstraction (ex.
pandas.Series) - it can only be accessed via__dataframe__. -
-
Use methods and properties on an opaque object rather than returning hierarchical dictionaries describing memory.
This is better for implementations that may rely on, for example, lazy computation.
-
No row names. If a library uses row names, use a regular column for them.
See discussion at wesm/dataframe-protocol/pull/1 Optional row names are not a good idea, because people will assume they’re present (see cuDF experience, forced to add because pandas has them). Requiring row names seems worse than leaving them out. Note that row labels could be added in the future - right now there’s no clear requirements for more complex row labels that cannot be represented by a single column. These do exist, for example Modin has has table and tree-based row labels.
Interface ¶
from abc import (
ABC,
abstractmethod,
)
import enum
from typing import (
Any,
Dict,
Iterable,
Optional,
Sequence,
Tuple,
TypedDict,
)
class DlpackDeviceType(enum.IntEnum):
"""Integer enum for device type codes matching DLPack."""
CPU = 1
CUDA = 2
CPU_PINNED = 3
OPENCL = 4
VULKAN = 7
METAL = 8
VPI = 9
ROCM = 10
class DtypeKind(enum.IntEnum):
"""
Integer enum for data types.
Attributes
----------
INT : int
Matches to signed integer data type.
UINT : int
Matches to unsigned integer data type.
FLOAT : int
Matches to floating point data type.
BOOL : int
Matches to boolean data type.
STRING : int
Matches to string data type (UTF-8 encoded).
DATETIME : int
Matches to datetime data type.
CATEGORICAL : int
Matches to categorical data type.
"""
INT = 0
UINT = 1
FLOAT = 2
BOOL = 20
STRING = 21 # UTF-8
DATETIME = 22
CATEGORICAL = 23
Dtype = Tuple[DtypeKind, int, str, str] # see Column.dtype
class ColumnNullType(enum.IntEnum):
"""
Integer enum for null type representation.
Attributes
----------
NON_NULLABLE : int
Non-nullable column.
USE_NAN : int
Use explicit float NaN value.
USE_SENTINEL : int
Sentinel value besides NaN.
USE_BITMASK : int
The bit is set/unset representing a null on a certain position.
USE_BYTEMASK : int
The byte is set/unset representing a null on a certain position.
"""
NON_NULLABLE = 0
USE_NAN = 1
USE_SENTINEL = 2
USE_BITMASK = 3
USE_BYTEMASK = 4
class ColumnBuffers(TypedDict):
# first element is a buffer containing the column data;
# second element is the data buffer's associated dtype
data: Tuple["Buffer", Dtype]
# first element is a buffer containing mask values indicating missing data;
# second element is the mask value buffer's associated dtype.
# None if the null representation is not a bit or byte mask
validity: Optional[Tuple["Buffer", Dtype]]
# first element is a buffer containing the offset values for
# variable-size binary data (e.g., variable-length strings);
# second element is the offsets buffer's associated dtype.
# None if the data buffer does not have an associated offsets buffer
offsets: Optional[Tuple["Buffer", Dtype]]
class CategoricalDescription(TypedDict):
# whether the ordering of dictionary indices is semantically meaningful
is_ordered: bool
# whether a dictionary-style mapping of categorical values to other objects exists
is_dictionary: bool
# Python-level only (e.g. ``{int: str}``).
# None if not a dictionary-style categorical.
categories: Optional[Column]
class Buffer(ABC):
"""
Data in the buffer is guaranteed to be contiguous in memory.
Note that there is no dtype attribute present, a buffer can be thought of
as simply a block of memory. However, if the column that the buffer is
attached to has a dtype that's supported by DLPack and ``__dlpack__`` is
implemented, then that dtype information will be contained in the return
value from ``__dlpack__``.
This distinction is useful to support both data exchange via DLPack on a
buffer and (b) dtypes like variable-length strings which do not have a
fixed number of bytes per element.
"""
@property
@abstractmethod
def bufsize(self) -> int:
"""
Buffer size in bytes.
"""
pass
@property
@abstractmethod
def ptr(self) -> int:
"""
Pointer to start of the buffer as an integer.
"""
pass
@abstractmethod
def __dlpack__(self):
"""
Produce DLPack capsule (see array API standard).
Raises:
- TypeError : if the buffer contains unsupported dtypes.
- NotImplementedError : if DLPack support is not implemented
Useful to have to connect to array libraries. Support optional because
it's not completely trivial to implement for a Python-only library.
"""
raise NotImplementedError("__dlpack__")
@abstractmethod
def __dlpack_device__(self) -> Tuple[DlpackDeviceType, Optional[int]]:
"""
Device type and device ID for where the data in the buffer resides.
Uses device type codes matching DLPack.
Note: must be implemented even if ``__dlpack__`` is not.
"""
pass
class Column(ABC):
"""
A column object, with only the methods and properties required by the
interchange protocol defined.
A column can contain one or more chunks. Each chunk can contain up to three
buffers - a data buffer, a mask buffer (depending on null representation),
and an offsets buffer (if variable-size binary; e.g., variable-length
strings).
TBD: there's also the "chunk" concept here, which is implicit in Arrow as
multiple buffers per array (= column here). Semantically it may make
sense to have both: chunks were meant for example for lazy evaluation
of data which doesn't fit in memory, while multiple buffers per column
could also come from doing a selection operation on a single
contiguous buffer.
Given these concepts, one would expect chunks to be all of the same
size (say a 10,000 row dataframe could have 10 chunks of 1,000 rows),
while multiple buffers could have data-dependent lengths. Not an issue
in pandas if one column is backed by a single NumPy array, but in
Arrow it seems possible.
Are multiple chunks *and* multiple buffers per column necessary for
the purposes of this interchange protocol, or must producers either
reuse the chunk concept for this or copy the data?
Note: this Column object can only be produced by ``__dataframe__``, so
doesn't need its own version or ``__column__`` protocol.
"""
@abstractmethod
def size(self) -> int:
"""
Size of the column, in elements.
Corresponds to DataFrame.num_rows() if column is a single chunk;
equal to size of this current chunk otherwise.
Is a method rather than a property because it may cause a (potentially
expensive) computation for some dataframe implementations.
"""
pass
@property
@abstractmethod
def offset(self) -> int:
"""
Offset of first element.
May be > 0 if using chunks; for example for a column with N chunks of
equal size M (only the last chunk may be shorter),
``offset = n * M``, ``n = 0 .. N-1``.
"""
pass
@property
@abstractmethod
def dtype(self) -> Dtype:
"""
Dtype description as a tuple ``(kind, bit-width, format string, endianness)``.
Bit-width : the number of bits as an integer
Format string : data type description format string in Apache Arrow C
Data Interface format.
Endianness : current only native endianness (``=``) is supported
Notes:
- Kind specifiers are aligned with DLPack where possible (hence the
jump to 20, leave enough room for future extension)
- Masks must be specified as boolean with either bit width 1 (for bit
masks) or 8 (for byte masks).
- Dtype width in bits was preferred over bytes
- Endianness isn't too useful, but included now in case in the future
we need to support non-native endianness
- Went with Apache Arrow format strings over NumPy format strings
because they're more complete from a dataframe perspective
- Format strings are mostly useful for datetime specification, and
for categoricals.
- For categoricals, the format string describes the type of the
categorical in the data buffer. In case of a separate encoding of
the categorical (e.g. an integer to string mapping), this can
be derived from ``self.describe_categorical``.
- Data types not included: complex, Arrow-style null, binary, decimal,
and nested (list, struct, map, union) dtypes.
"""
pass
@property
@abstractmethod
def describe_categorical(self) -> CategoricalDescription:
"""
If the dtype is categorical, there are two options:
- There are only values in the data buffer.
- There is a separate non-categorical Column encoding categorical values.
Raises TypeError if the dtype is not categorical
Returns the dictionary with description on how to interpret the data buffer:
- "is_ordered" : bool, whether the ordering of dictionary indices is
semantically meaningful.
- "is_dictionary" : bool, whether a mapping of
categorical values to other objects exists
- "categories" : Column representing the (implicit) mapping of indices to
category values (e.g. an array of cat1, cat2, ...).
None if not a dictionary-style categorical.
TBD: are there any other in-memory representations that are needed?
"""
pass
@property
@abstractmethod
def describe_null(self) -> Tuple[ColumnNullType, Any]:
"""
Return the missing value (or "null") representation the column dtype
uses, as a tuple ``(kind, value)``.
Value : if kind is "sentinel value", the actual value. If kind is a bit
mask or a byte mask, the value (0 or 1) indicating a missing value. None
otherwise.
"""
pass
@property
@abstractmethod
def null_count(self) -> Optional[int]:
"""
Number of null elements, if known.
Note: Arrow uses -1 to indicate "unknown", but None seems cleaner.
"""
pass
@property
@abstractmethod
def metadata(self) -> Dict[str, Any]:
"""
The metadata for the column. See `DataFrame.metadata` for more details.
"""
pass
@abstractmethod
def num_chunks(self) -> int:
"""
Return the number of chunks the column consists of.
"""
pass
@abstractmethod
def get_chunks(self, n_chunks: Optional[int] = None) -> Iterable["Column"]:
"""
Return an iterator yielding the chunks.
See `DataFrame.get_chunks` for details on ``n_chunks``.
"""
pass
@abstractmethod
def get_buffers(self) -> ColumnBuffers:
"""
Return a dictionary containing the underlying buffers.
The returned dictionary has the following contents:
- "data": a two-element tuple whose first element is a buffer
containing the data and whose second element is the data
buffer's associated dtype.
- "validity": a two-element tuple whose first element is a buffer
containing mask values indicating missing data and
whose second element is the mask value buffer's
associated dtype. None if the null representation is
not a bit or byte mask.
- "offsets": a two-element tuple whose first element is a buffer
containing the offset values for variable-size binary
data (e.g., variable-length strings) and whose second
element is the offsets buffer's associated dtype. None
if the data buffer does not have an associated offsets
buffer.
"""
pass
# def get_children(self) -> Iterable[Column]:
# """
# Children columns underneath the column, each object in this iterator
# must adhere to the column specification.
# """
# pass
class DataFrame(ABC):
"""
A data frame class, with only the methods required by the interchange
protocol defined.
A "data frame" represents an ordered collection of named columns.
A column's "name" must be a unique string.
Columns may be accessed by name or by position.
This could be a public data frame class, or an object with the methods and
attributes defined on this DataFrame class could be returned from the
``__dataframe__`` method of a public data frame class in a library adhering
to the dataframe interchange protocol specification.
"""
version = 0 # version of the protocol
@abstractmethod
def __dataframe__(
self, nan_as_null: bool = False, allow_copy: bool = True
) -> "DataFrame":
"""
Construct a new exchange object, potentially changing the parameters.
``nan_as_null`` is a DEPRECATED keyword that should not be used. See warning
below.
``allow_copy`` is a keyword that defines whether or not the library is
allowed to make a copy of the data. For example, copying data would be
necessary if a library supports strided buffers, given that this protocol
specifies contiguous buffers.
WARNING: the ``nan_as_null`` parameter will be removed from the API protocol.
Please avoid passing it as either a positional or keyword argument. Call this
method using ``.__dataframe__(allow_copy=...)``.
"""
pass
@property
@abstractmethod
def metadata(self) -> Dict[str, Any]:
"""
The metadata for the data frame, as a dictionary with string keys. The
contents of `metadata` may be anything, they are meant for a library
to store information that it needs to, e.g., roundtrip losslessly or
for two implementations to share data that is not (yet) part of the
interchange protocol specification. For avoiding collisions with other
entries, please add name the keys with the name of the library
followed by a period and the desired name, e.g, ``pandas.indexcol``.
"""
pass
@abstractmethod
def num_columns(self) -> int:
"""
Return the number of columns in the DataFrame.
"""
pass
@abstractmethod
def num_rows(self) -> Optional[int]:
# TODO: not happy with Optional, but need to flag it may be expensive
# why include it if it may be None - what do we expect consumers
# to do here?
"""
Return the number of rows in the DataFrame, if available.
"""
pass
@abstractmethod
def num_chunks(self) -> int:
"""
Return the number of chunks the DataFrame consists of.
"""
pass
@abstractmethod
def column_names(self) -> Iterable[str]:
"""
Return an iterator yielding the column names.
"""
pass
@abstractmethod
def get_column(self, i: int) -> Column:
"""
Return the column at the indicated position.
"""
pass
@abstractmethod
def get_column_by_name(self, name: str) -> Column:
"""
Return the column whose name is the indicated name.
"""
pass
@abstractmethod
def get_columns(self) -> Iterable[Column]:
"""
Return an iterator yielding the columns.
"""
pass
@abstractmethod
def select_columns(self, indices: Sequence[int]) -> "DataFrame":
"""
Create a new DataFrame by selecting a subset of columns by index.
"""
pass
@abstractmethod
def select_columns_by_name(self, names: Sequence[str]) -> "DataFrame":
"""
Create a new DataFrame by selecting a subset of columns by name.
"""
pass
@abstractmethod
def get_chunks(self, n_chunks: Optional[int] = None) -> Iterable["DataFrame"]:
"""
Return an iterator yielding the chunks.
By default (None), yields the chunks that the data is stored as by the
producer. If given, ``n_chunks`` must be a multiple of
``self.num_chunks()``, meaning the producer must subdivide each chunk
before yielding it.
Note that the producer must ensure that all columns are chunked the
same way.
"""
pass