Data Types and In-Memory Data Model

Apache Arrow defines columnar array data structures by composing type metadata with memory buffers, like the ones explained in the documentation on Memory and IO. These data structures are exposed in Python through a series of interrelated classes:

  • Type Metadata: Instances of pyarrow.DataType, which describe a logical array type

  • Schemas: Instances of pyarrow.Schema, which describe a named collection of types. These can be thought of as the column types in a table-like object.

  • Arrays: Instances of pyarrow.Array, which are atomic, contiguous columnar data structures composed from Arrow Buffer objects

  • Record Batches: Instances of pyarrow.RecordBatch, which are a collection of Array objects with a particular Schema

  • Tables: Instances of pyarrow.Table, a logical table data structure in which each column consists of one or more pyarrow.Array objects of the same type.

We will examine these in the sections below in a series of examples.

Type Metadata

Apache Arrow defines language agnostic column-oriented data structures for array data. These include:

  • Fixed-length primitive types: numbers, booleans, date and times, fixed size binary, decimals, and other values that fit into a given number

  • Variable-length primitive types: binary, string

  • Nested types: list, map, struct, and union

  • Dictionary type: An encoded categorical type (more on this later)

Each logical data type in Arrow has a corresponding factory function for creating an instance of that type object in Python:

In [1]: import pyarrow as pa

In [2]: t1 = pa.int32()

In [3]: t2 = pa.string()

In [4]: t3 = pa.binary()

In [5]: t4 = pa.binary(10)

In [6]: t5 = pa.timestamp('ms')

In [7]: t1
Out[7]: DataType(int32)

In [8]: print(t1)
int32

In [9]: print(t4)
fixed_size_binary[10]

In [10]: print(t5)
timestamp[ms]

We use the name logical type because the physical storage may be the same for one or more types. For example, int64, float64, and timestamp[ms] all occupy 64 bits per value.

These objects are metadata; they are used for describing the data in arrays, schemas, and record batches. In Python, they can be used in functions where the input data (e.g. Python objects) may be coerced to more than one Arrow type.

The Field type is a type plus a name and optional user-defined metadata:

In [11]: f0 = pa.field('int32_field', t1)

In [12]: f0
Out[12]: pyarrow.Field<int32_field: int32>

In [13]: f0.name
Out[13]: 'int32_field'

In [14]: f0.type
Out[14]: DataType(int32)

Arrow supports nested value types like list, map, struct, and union. When creating these, you must pass types or fields to indicate the data types of the types’ children. For example, we can define a list of int32 values with:

In [15]: t6 = pa.list_(t1)

In [16]: t6
Out[16]: ListType(list<item: int32>)

A struct is a collection of named fields:

In [17]: fields = [
   ....:     pa.field('s0', t1),
   ....:     pa.field('s1', t2),
   ....:     pa.field('s2', t4),
   ....:     pa.field('s3', t6),
   ....: ]
   ....: 

In [18]: t7 = pa.struct(fields)

In [19]: print(t7)
struct<s0: int32, s1: string, s2: fixed_size_binary[10], s3: list<item: int32>>

For convenience, you can pass (name, type) tuples directly instead of Field instances:

In [20]: t8 = pa.struct([('s0', t1), ('s1', t2), ('s2', t4), ('s3', t6)])

In [21]: print(t8)
struct<s0: int32, s1: string, s2: fixed_size_binary[10], s3: list<item: int32>>

In [22]: t8 == t7
Out[22]: True

See Data Types API for a full listing of data type functions.

Schemas

The Schema type is similar to the struct array type; it defines the column names and types in a record batch or table data structure. The pyarrow.schema() factory function makes new Schema objects in Python:

In [23]: my_schema = pa.schema([('field0', t1),
   ....:                        ('field1', t2),
   ....:                        ('field2', t4),
   ....:                        ('field3', t6)])
   ....: 

In [24]: my_schema
Out[24]: 
field0: int32
field1: string
field2: fixed_size_binary[10]
field3: list<item: int32>
  child 0, item: int32

In some applications, you may not create schemas directly, only using the ones that are embedded in IPC messages.

Arrays

For each data type, there is an accompanying array data structure for holding memory buffers that define a single contiguous chunk of columnar array data. When you are using PyArrow, this data may come from IPC tools, though it can also be created from various types of Python sequences (lists, NumPy arrays, pandas data).

A simple way to create arrays is with pyarrow.array, which is similar to the numpy.array function. By default PyArrow will infer the data type for you:

In [25]: arr = pa.array([1, 2, None, 3])

In [26]: arr
Out[26]: 
<pyarrow.lib.Int64Array object at 0x7f4ebce06580>
[
  1,
  2,
  null,
  3
]

But you may also pass a specific data type to override type inference:

In [27]: pa.array([1, 2], type=pa.uint16())
Out[27]: 
<pyarrow.lib.UInt16Array object at 0x7f4ebcdb30a0>
[
  1,
  2
]

The array’s type attribute is the corresponding piece of type metadata:

In [28]: arr.type
Out[28]: DataType(int64)

Each in-memory array has a known length and null count (which will be 0 if there are no null values):

In [29]: len(arr)
Out[29]: 4

In [30]: arr.null_count
Out[30]: 1

Scalar values can be selected with normal indexing. pyarrow.array converts None values to Arrow nulls; we return the special pyarrow.NA value for nulls:

In [31]: arr[0]
Out[31]: <pyarrow.Int64Scalar: 1>

In [32]: arr[2]
Out[32]: <pyarrow.Int64Scalar: None>

Arrow data is immutable, so values can be selected but not assigned.

Arrays can be sliced without copying:

In [33]: arr[1:3]
Out[33]: 
<pyarrow.lib.Int64Array object at 0x7f4ebcdb3520>
[
  2,
  null
]

None values and NAN handling

As mentioned in the above section, the Python object None is always converted to an Arrow null element on the conversion to pyarrow.Array. For the float NaN value which is either represented by the Python object float('nan') or numpy.nan we normally convert it to a valid float value during the conversion. If an integer input is supplied to pyarrow.array that contains np.nan, ValueError is raised.

To handle better compatibility with Pandas, we support interpreting NaN values as null elements. This is enabled automatically on all from_pandas function and can be enable on the other conversion functions by passing from_pandas=True as a function parameter.

List arrays

pyarrow.array is able to infer the type of simple nested data structures like lists:

In [34]: nested_arr = pa.array([[], None, [1, 2], [None, 1]])

In [35]: print(nested_arr.type)
list<item: int64>

Struct arrays

pyarrow.array is able to infer the schema of a struct type from arrays of dictionaries:

In [36]: pa.array([{'x': 1, 'y': True}, {'z': 3.4, 'x': 4}])
Out[36]: 
<pyarrow.lib.StructArray object at 0x7f4ebcdb3700>
-- is_valid: all not null
-- child 0 type: int64
  [
    1,
    4
  ]
-- child 1 type: bool
  [
    true,
    null
  ]
-- child 2 type: double
  [
    null,
    3.4
  ]

Struct arrays can be initialized from a sequence of Python dicts or tuples. For tuples, you must explicitly pass the type:

In [37]: ty = pa.struct([('x', pa.int8()),
   ....:                 ('y', pa.bool_())])
   ....: 

In [38]: pa.array([{'x': 1, 'y': True}, {'x': 2, 'y': False}], type=ty)
Out[38]: 
<pyarrow.lib.StructArray object at 0x7f4ebcdb3be0>
-- is_valid: all not null
-- child 0 type: int8
  [
    1,
    2
  ]
-- child 1 type: bool
  [
    true,
    false
  ]

In [39]: pa.array([(3, True), (4, False)], type=ty)
Out[39]: 
<pyarrow.lib.StructArray object at 0x7f4ebcdb3ca0>
-- is_valid: all not null
-- child 0 type: int8
  [
    3,
    4
  ]
-- child 1 type: bool
  [
    true,
    false
  ]

When initializing a struct array, nulls are allowed both at the struct level and at the individual field level. If initializing from a sequence of Python dicts, a missing dict key is handled as a null value:

In [40]: pa.array([{'x': 1}, None, {'y': None}], type=ty)
Out[40]: 
<pyarrow.lib.StructArray object at 0x7f4ebcdb3a60>
-- is_valid:
  [
    true,
    false,
    true
  ]
-- child 0 type: int8
  [
    1,
    0,
    null
  ]
-- child 1 type: bool
  [
    null,
    false,
    null
  ]

You can also construct a struct array from existing arrays for each of the struct’s components. In this case, data storage will be shared with the individual arrays, and no copy is involved:

In [41]: xs = pa.array([5, 6, 7], type=pa.int16())

In [42]: ys = pa.array([False, True, True])

In [43]: arr = pa.StructArray.from_arrays((xs, ys), names=('x', 'y'))

In [44]: arr.type
Out[44]: StructType(struct<x: int16, y: bool>)

In [45]: arr
Out[45]: 
<pyarrow.lib.StructArray object at 0x7f4ebcdb3ee0>
-- is_valid: all not null
-- child 0 type: int16
  [
    5,
    6,
    7
  ]
-- child 1 type: bool
  [
    false,
    true,
    true
  ]

Map arrays

Map arrays can be constructed from lists of lists of tuples (key-item pairs), but only if the type is explicitly passed into array():

In [46]: data = [[('x', 1), ('y', 0)], [('a', 2), ('b', 45)]]

In [47]: ty = pa.map_(pa.string(), pa.int64())

In [48]: pa.array(data, type=ty)
Out[48]: 
<pyarrow.lib.MapArray object at 0x7f4ebcdd9400>
[
  keys:
  [
    "x",
    "y"
  ]
  values:
  [
    1,
    0
  ],
  keys:
  [
    "a",
    "b"
  ]
  values:
  [
    2,
    45
  ]
]

MapArrays can also be constructed from offset, key, and item arrays. Offsets represent the starting position of each map. Note that the MapArray.keys and MapArray.items properties give the flattened keys and items. To keep the keys and items associated to their row, use the ListArray.from_arrays() constructor with the MapArray.offsets property.

In [49]: arr = pa.MapArray.from_arrays([0, 2, 3], ['x', 'y', 'z'], [4, 5, 6])

In [50]: arr.keys
Out[50]: 
<pyarrow.lib.StringArray object at 0x7f4ebcdd93a0>
[
  "x",
  "y",
  "z"
]

In [51]: arr.items
Out[51]: 
<pyarrow.lib.Int64Array object at 0x7f4ebcdd94c0>
[
  4,
  5,
  6
]

In [52]: pa.ListArray.from_arrays(arr.offsets, arr.keys)
Out[52]: 
<pyarrow.lib.ListArray object at 0x7f4ebcdd9640>
[
  [
    "x",
    "y"
  ],
  [
    "z"
  ]
]

In [53]: pa.ListArray.from_arrays(arr.offsets, arr.items)
Out[53]: 
<pyarrow.lib.ListArray object at 0x7f4ebcdd9580>
[
  [
    4,
    5
  ],
  [
    6
  ]
]

Union arrays

The union type represents a nested array type where each value can be one (and only one) of a set of possible types. There are two possible storage types for union arrays: sparse and dense.

In a sparse union array, each of the child arrays has the same length as the resulting union array. They are adjuncted with a int8 “types” array that tells, for each value, from which child array it must be selected:

In [54]: xs = pa.array([5, 6, 7])

In [55]: ys = pa.array([False, False, True])

In [56]: types = pa.array([0, 1, 1], type=pa.int8())

In [57]: union_arr = pa.UnionArray.from_sparse(types, [xs, ys])

In [58]: union_arr.type
Out[58]: SparseUnionType(sparse_union<0: int64=0, 1: bool=1>)

In [59]: union_arr
Out[59]: 
<pyarrow.lib.UnionArray object at 0x7f4ebcdb3d60>
-- is_valid: all not null
-- type_ids:   [
    0,
    1,
    1
  ]
-- child 0 type: int64
  [
    5,
    6,
    7
  ]
-- child 1 type: bool
  [
    false,
    false,
    true
  ]

In a dense union array, you also pass, in addition to the int8 “types” array, a int32 “offsets” array that tells, for each value, at each offset in the selected child array it can be found:

In [60]: xs = pa.array([5, 6, 7])

In [61]: ys = pa.array([False, True])

In [62]: types = pa.array([0, 1, 1, 0, 0], type=pa.int8())

In [63]: offsets = pa.array([0, 0, 1, 1, 2], type=pa.int32())

In [64]: union_arr = pa.UnionArray.from_dense(types, offsets, [xs, ys])

In [65]: union_arr.type
Out[65]: DenseUnionType(dense_union<0: int64=0, 1: bool=1>)

In [66]: union_arr
Out[66]: 
<pyarrow.lib.UnionArray object at 0x7f4ebcdd9e20>
-- is_valid: all not null
-- type_ids:   [
    0,
    1,
    1,
    0,
    0
  ]
-- value_offsets:   [
    0,
    0,
    1,
    1,
    2
  ]
-- child 0 type: int64
  [
    5,
    6,
    7
  ]
-- child 1 type: bool
  [
    false,
    true
  ]

Dictionary Arrays

The Dictionary type in PyArrow is a special array type that is similar to a factor in R or a pandas.Categorical. It enables one or more record batches in a file or stream to transmit integer indices referencing a shared dictionary containing the distinct values in the logical array. This is particularly often used with strings to save memory and improve performance.

The way that dictionaries are handled in the Apache Arrow format and the way they appear in C++ and Python is slightly different. We define a special DictionaryArray type with a corresponding dictionary type. Let’s consider an example:

In [67]: indices = pa.array([0, 1, 0, 1, 2, 0, None, 2])

In [68]: dictionary = pa.array(['foo', 'bar', 'baz'])

In [69]: dict_array = pa.DictionaryArray.from_arrays(indices, dictionary)

In [70]: dict_array
Out[70]: 
<pyarrow.lib.DictionaryArray object at 0x7f4ebcd6e660>

-- dictionary:
  [
    "foo",
    "bar",
    "baz"
  ]
-- indices:
  [
    0,
    1,
    0,
    1,
    2,
    0,
    null,
    2
  ]

Here we have:

In [71]: print(dict_array.type)
dictionary<values=string, indices=int64, ordered=0>

In [72]: dict_array.indices
Out[72]: 
<pyarrow.lib.Int64Array object at 0x7f4ebcd683a0>
[
  0,
  1,
  0,
  1,
  2,
  0,
  null,
  2
]

In [73]: dict_array.dictionary
Out[73]: 
<pyarrow.lib.StringArray object at 0x7f4ebcdb36a0>
[
  "foo",
  "bar",
  "baz"
]

When using DictionaryArray with pandas, the analogue is pandas.Categorical (more on this later):

In [74]: dict_array.to_pandas()
Out[74]: 
0    foo
1    bar
2    foo
3    bar
4    baz
5    foo
6    NaN
7    baz
dtype: category
Categories (3, object): ['foo', 'bar', 'baz']

Record Batches

A Record Batch in Apache Arrow is a collection of equal-length array instances. Let’s consider a collection of arrays:

In [75]: data = [
   ....:     pa.array([1, 2, 3, 4]),
   ....:     pa.array(['foo', 'bar', 'baz', None]),
   ....:     pa.array([True, None, False, True])
   ....: ]
   ....: 

A record batch can be created from this list of arrays using RecordBatch.from_arrays:

In [76]: batch = pa.RecordBatch.from_arrays(data, ['f0', 'f1', 'f2'])

In [77]: batch.num_columns
Out[77]: 3

In [78]: batch.num_rows
Out[78]: 4

In [79]: batch.schema
Out[79]: 
f0: int64
f1: string
f2: bool

In [80]: batch[1]
Out[80]: 
<pyarrow.lib.StringArray object at 0x7f4ebcd68b20>
[
  "foo",
  "bar",
  "baz",
  null
]

A record batch can be sliced without copying memory like an array:

In [81]: batch2 = batch.slice(1, 3)

In [82]: batch2[1]
Out[82]: 
<pyarrow.lib.StringArray object at 0x7f4ebcd68820>
[
  "bar",
  "baz",
  null
]

Tables

The PyArrow Table type is not part of the Apache Arrow specification, but is rather a tool to help with wrangling multiple record batches and array pieces as a single logical dataset. As a relevant example, we may receive multiple small record batches in a socket stream, then need to concatenate them into contiguous memory for use in NumPy or pandas. The Table object makes this efficient without requiring additional memory copying.

Considering the record batch we created above, we can create a Table containing one or more copies of the batch using Table.from_batches:

In [83]: batches = [batch] * 5

In [84]: table = pa.Table.from_batches(batches)

In [85]: table
Out[85]: 
pyarrow.Table
f0: int64
f1: string
f2: bool
----
f0: [[1,2,3,4],[1,2,3,4],[1,2,3,4],[1,2,3,4],[1,2,3,4]]
f1: [["foo","bar","baz",null],["foo","bar","baz",null],["foo","bar","baz",null],["foo","bar","baz",null],["foo","bar","baz",null]]
f2: [[true,null,false,true],[true,null,false,true],[true,null,false,true],[true,null,false,true],[true,null,false,true]]

In [86]: table.num_rows
Out[86]: 20

The table’s columns are instances of ChunkedArray, which is a container for one or more arrays of the same type.

In [87]: c = table[0]

In [88]: c
Out[88]: 
<pyarrow.lib.ChunkedArray object at 0x7f4ebcd29d60>
[
  [
    1,
    2,
    3,
    4
  ],
  [
    1,
    2,
    3,
    4
  ],
  [
    1,
    2,
    3,
    4
  ],
  [
    1,
    2,
    3,
    4
  ],
  [
    1,
    2,
    3,
    4
  ]
]

In [89]: c.num_chunks
Out[89]: 5

In [90]: c.chunk(0)
Out[90]: 
<pyarrow.lib.Int64Array object at 0x7f4ebcd68e20>
[
  1,
  2,
  3,
  4
]

As you’ll see in the pandas section, we can convert these objects to contiguous NumPy arrays for use in pandas:

In [91]: c.to_pandas()
Out[91]: 
0     1
1     2
2     3
3     4
4     1
5     2
6     3
7     4
8     1
9     2
10    3
11    4
12    1
13    2
14    3
15    4
16    1
17    2
18    3
19    4
Name: f0, dtype: int64

Multiple tables can also be concatenated together to form a single table using pyarrow.concat_tables, if the schemas are equal:

In [92]: tables = [table] * 2

In [93]: table_all = pa.concat_tables(tables)

In [94]: table_all.num_rows
Out[94]: 40

In [95]: c = table_all[0]

In [96]: c.num_chunks
Out[96]: 10

This is similar to Table.from_batches, but uses tables as input instead of record batches. Record batches can be made into tables, but not the other way around, so if your data is already in table form, then use pyarrow.concat_tables.

Custom Schema and Field Metadata

Arrow supports both schema-level and field-level custom key-value metadata allowing for systems to insert their own application defined metadata to customize behavior.

Custom metadata can be accessed at Schema.metadata for the schema-level and Field.metadata for the field-level.

Note that this metadata is preserved in Streaming, Serialization, and IPC processes.

To customize the schema metadata of an existing table you can use Table.replace_schema_metadata():

In [97]: table.schema.metadata # empty

In [98]: table = table.replace_schema_metadata({"f0": "First dose"})

In [99]: table.schema.metadata
Out[99]: {b'f0': b'First dose'}

To customize the metadata of the field from the table schema you can use Field.with_metadata():

In [100]: field_f1 = table.schema.field("f1")

In [101]: field_f1.metadata # empty

In [102]: field_f1 = field_f1.with_metadata({"f1": "Second dose"})

In [103]: field_f1.metadata
Out[103]: {b'f1': b'Second dose'}

Both options create a shallow copy of the data and do not in fact change the Schema which is immutable. To change the metadata in the schema of the table we created a new object when calling Table.replace_schema_metadata().

To change the metadata of the field in the schema we would need to define a new schema and cast the data to this schema:

In [104]: my_schema2 = pa.schema([
   .....:    pa.field('f0', pa.int64(), metadata={"name": "First dose"}),
   .....:    pa.field('f1', pa.string(), metadata={"name": "Second dose"}),
   .....:    pa.field('f2', pa.bool_())],
   .....:    metadata={"f2": "booster"})
   .....: 

In [105]: t2 = table.cast(my_schema2)

In [106]: t2.schema.field("f0").metadata
Out[106]: {b'name': b'First dose'}

In [107]: t2.schema.field("f1").metadata
Out[107]: {b'name': b'Second dose'}

In [108]: t2.schema.metadata
Out[108]: {b'f2': b'booster'}

Metadata key and value pair are std::string objects in the C++ implementation and so they are bytes objects (b'...') in Python.