The Arrow C data interface

Rationale

Apache Arrow is designed to be a universal in-memory format for the representation of tabular (“columnar”) data. However, some projects may face a difficult choice between either depending on a fast-evolving project such as the Arrow C++ library, or having to reimplement adapters for data interchange, which may require significant, redundant development effort.

The Arrow C data interface defines a very small, stable set of C definitions that can be easily copied in any project’s source code and used for columnar data interchange in the Arrow format. For non-C/C++ languages and runtimes, it should be almost as easy to translate the C definitions into the corresponding C FFI declarations.

Applications and libraries can therefore work with Arrow memory without necessarily using Arrow libraries or reinventing the wheel. Developers can choose between tight integration with the Arrow software project (benefitting from the growing array of facilities exposed by e.g. the C++ or Java implementations of Apache Arrow, but with the cost of a dependency) or minimal integration with the Arrow format only.

Goals

  • Expose an ABI-stable interface.

  • Make it easy for third-party projects to implement support for (including partial support where sufficient), with little initial investment.

  • Allow zero-copy sharing of Arrow data between independent runtimes and components running in the same process.

  • Match the Arrow array concepts closely to avoid the development of yet another marshalling layer.

  • Avoid the need for one-to-one adaptation layers such as the limited JPype-based bridge between Java and Python.

  • Enable integration without an explicit dependency (either at compile-time or runtime) on the Arrow software project.

Ideally, the Arrow C data interface can become a low-level lingua franca for sharing columnar data at runtime and establish Arrow as the universal building block in the columnar processing ecosystem.

Non-goals

  • Expose a C API mimicking operations available in higher-level runtimes (such as C++, Java…).

  • Data sharing between distinct processes or storage persistence.

Comparison with the Arrow IPC format

Pros of the C data interface vs. the IPC format:

  • No dependency on Flatbuffers.

  • No buffer reassembly (data is already exposed in logical Arrow format).

  • Zero-copy by design.

  • Easy to reimplement from scratch.

  • Minimal C definition that can be easily copied into other codebases.

  • Resource lifetime management through a custom release callback.

Pros of the IPC format vs. the data interface:

  • Works across processes and machines.

  • Allows data storage and persistence.

  • Being a streamable format, the IPC format has room for composing more features (such as integrity checks, compression…).

  • Does not require explicit C data access.

Data type description – format strings

A data type is described using a format string. The format string only encodes information about the top-level type; for nested type, child types are described separately. Also, metadata is encoded in a separate string.

The format strings are designed to be easily parsable, even from a language such as C. The most common primitive formats have one-character format strings:

Format string

Arrow data type

Notes

n

null

b

boolean

c

int8

C

uint8

s

int16

S

uint16

i

int32

I

uint32

l

int64

L

uint64

e

float16

f

float32

g

float64

Format string

Arrow data type

Notes

z

binary

Z

large binary

u

utf-8 string

U

large utf-8 string

d:19,10

decimal128 [precision 19, scale 10]

d:19,10,NNN

decimal bitwidth = NNN [precision 19, scale 10]

w:42

fixed-width binary [42 bytes]

Temporal types have multi-character format strings starting with t:

Format string

Arrow data type

Notes

tdD

date32 [days]

tdm

date64 [milliseconds]

tts

time32 [seconds]

ttm

time32 [milliseconds]

ttu

time64 [microseconds]

ttn

time64 [nanoseconds]

tss:...

timestamp [seconds] with timezone “…”

(1)

tsm:...

timestamp [milliseconds] with timezone “…”

(1)

tsu:...

timestamp [microseconds] with timezone “…”

(1)

tsn:...

timestamp [nanoseconds] with timezone “…”

(1)

tDs

duration [seconds]

tDm

duration [milliseconds]

tDu

duration [microseconds]

tDn

duration [nanoseconds]

tiM

interval [months]

tiD

interval [days, time]

tin

interval [month, day, nanoseconds]

Dictionary-encoded types do not have a specific format string. Instead, the format string of the base array represents the dictionary index type, and the value type can be read from the dependent dictionary array (see below “Dictionary-encoded arrays”).

Nested types have multiple-character format strings starting with +. The names and types of child fields are read from the child arrays.

Format string

Arrow data type

Notes

+l

list

+L

large list

+w:123

fixed-sized list [123 items]

+s

struct

+m

map

(2)

+ud:I,J,...

dense union with type ids I,J…

+us:I,J,...

sparse union with type ids I,J…

Notes:

  1. The timezone string is appended as-is after the colon character :, without any quotes. If the timezone is empty, the colon : must still be included.

  2. As specified in the Arrow columnar format, the map type has a single child type named entries, itself a 2-child struct type of (key, value).

Examples

  • A dictionary-encoded decimal128(precision = 12, scale = 5) array with int16 indices has format string s, and its dependent dictionary array has format string d:12,5.

  • A list<uint64> array has format string +l, and its single child has format string L.

  • A struct<ints: int32, floats: float32> has format string +s; its two children have names ints and floats, and format strings i and f respectively.

  • A map<string, float64> array has format string +m; its single child has name entries and format string +s; its two grandchildren have names key and value, and format strings u and g respectively.

  • A sparse_union<ints: int32, floats: float32> with type ids 4, 5 has format string +us:4,5; its two children have names ints and floats, and format strings i and f respectively.

Structure definitions

The following free-standing definitions are enough to support the Arrow C data interface in your project. Like the rest of the Arrow project, they are available under the Apache License 2.0.

#define ARROW_FLAG_DICTIONARY_ORDERED 1
#define ARROW_FLAG_NULLABLE 2
#define ARROW_FLAG_MAP_KEYS_SORTED 4

struct ArrowSchema {
  // Array type description
  const char* format;
  const char* name;
  const char* metadata;
  int64_t flags;
  int64_t n_children;
  struct ArrowSchema** children;
  struct ArrowSchema* dictionary;

  // Release callback
  void (*release)(struct ArrowSchema*);
  // Opaque producer-specific data
  void* private_data;
};

struct ArrowArray {
  // Array data description
  int64_t length;
  int64_t null_count;
  int64_t offset;
  int64_t n_buffers;
  int64_t n_children;
  const void** buffers;
  struct ArrowArray** children;
  struct ArrowArray* dictionary;

  // Release callback
  void (*release)(struct ArrowArray*);
  // Opaque producer-specific data
  void* private_data;
};

The ArrowSchema structure

The ArrowSchema structure describes the type and metadata of an exported array or record batch. It has the following fields:

const char *ArrowSchema.format

Mandatory. A null-terminated, UTF8-encoded string describing the data type. If the data type is nested, child types are not encoded here but in the ArrowSchema.children structures.

Consumers MAY decide not to support all data types, but they should document this limitation.

const char *ArrowSchema.name

Optional. A null-terminated, UTF8-encoded string of the field or array name. This is mainly used to reconstruct child fields of nested types.

Producers MAY decide not to provide this information, and consumers MAY decide to ignore it. If omitted, MAY be NULL or an empty string.

const char *ArrowSchema.metadata

Optional. A binary string describing the type’s metadata. If the data type is nested, child types are not encoded here but in the ArrowSchema.children structures.

This string is not null-terminated but follows a specific format:

int32: number of key/value pairs (noted N below)
int32: byte length of key 0
key 0 (not null-terminated)
int32: byte length of value 0
value 0 (not null-terminated)
...
int32: byte length of key N - 1
key N - 1 (not null-terminated)
int32: byte length of value N - 1
value N - 1 (not null-terminated)

Integers are stored in native endianness. For example, the metadata [('key1', 'value1')] is encoded on a little-endian machine as:

\x01\x00\x00\x00\x04\x00\x00\x00key1\x06\x00\x00\x00value1

On a big-endian machine, the same example would be encoded as:

\x00\x00\x00\x01\x00\x00\x00\x04key1\x00\x00\x00\x06value1

If omitted, this field MUST be NULL (not an empty string).

Consumers MAY choose to ignore this information.

int64_t ArrowSchema.flags

Optional. A bitfield of flags enriching the type description. Its value is computed by OR’ing together the flag values. The following flags are available:

  • ARROW_FLAG_NULLABLE: whether this field is semantically nullable (regardless of whether it actually has null values).

  • ARROW_FLAG_DICTIONARY_ORDERED: for dictionary-encoded types, whether the ordering of dictionary indices is semantically meaningful.

  • ARROW_FLAG_MAP_KEYS_SORTED: for map types, whether the keys within each map value are sorted.

If omitted, MUST be 0.

Consumers MAY choose to ignore some or all of the flags. Even then, they SHOULD keep this value around so as to propagate its information to their own consumers.

int64_t ArrowSchema.n_children

Mandatory. The number of children this type has.

ArrowSchema **ArrowSchema.children

Optional. A C array of pointers to each child type of this type. There must be ArrowSchema.n_children pointers.

MAY be NULL only if ArrowSchema.n_children is 0.

ArrowSchema *ArrowSchema.dictionary

Optional. A pointer to the type of dictionary values.

MUST be present if the ArrowSchema represents a dictionary-encoded type. MUST be NULL otherwise.

void (*ArrowSchema.release)(struct ArrowSchema*)

Mandatory. A pointer to a producer-provided release callback.

See below for memory management and release callback semantics.

void *ArrowSchema.private_data

Optional. An opaque pointer to producer-provided private data.

Consumers MUST not process this member. Lifetime of this member is handled by the producer, and especially by the release callback.

The ArrowArray structure

The ArrowArray describes the data of an exported array or record batch. For the ArrowArray structure to be interpreted type, the array type or record batch schema must already be known. This is either done by convention – for example a producer API that always produces the same data type – or by passing a ArrowSchema on the side.

It has the following fields:

int64_t ArrowArray.length

Mandatory. The logical length of the array (i.e. its number of items).

int64_t ArrowArray.null_count

Mandatory. The number of null items in the array. MAY be -1 if not yet computed.

int64_t ArrowArray.offset

Mandatory. The logical offset inside the array (i.e. the number of items from the physical start of the buffers). MUST be 0 or positive.

Producers MAY specify that they will only produce 0-offset arrays to ease implementation of consumer code. Consumers MAY decide not to support non-0-offset arrays, but they should document this limitation.

int64_t ArrowArray.n_buffers

Mandatory. The number of physical buffers backing this array. The number of buffers is a function of the data type, as described in the Columnar format specification.

Buffers of children arrays are not included.

const void **ArrowArray.buffers

Mandatory. A C array of pointers to the start of each physical buffer backing this array. Each void* pointer is the physical start of a contiguous buffer. There must be ArrowArray.n_buffers pointers.

The producer MUST ensure that each contiguous buffer is large enough to represent length + offset values encoded according to the Columnar format specification.

It is recommended, but not required, that the memory addresses of the buffers be aligned at least according to the type of primitive data that they contain. Consumers MAY decide not to support unaligned memory.

The pointer to the null bitmap buffer, if the data type specifies one, MAY be NULL only if ArrowArray.null_count is 0.

Buffers of children arrays are not included.

int64_t ArrowArray.n_children

Mandatory. The number of children this array has. The number of children is a function of the data type, as described in the Columnar format specification.

ArrowArray **ArrowArray.children

Optional. A C array of pointers to each child array of this array. There must be ArrowArray.n_children pointers.

MAY be NULL only if ArrowArray.n_children is 0.

ArrowArray *ArrowArray.dictionary

Optional. A pointer to the underlying array of dictionary values.

MUST be present if the ArrowArray represents a dictionary-encoded array. MUST be NULL otherwise.

void (*ArrowArray.release)(struct ArrowArray*)

Mandatory. A pointer to a producer-provided release callback.

See below for memory management and release callback semantics.

void *ArrowArray.private_data

Optional. An opaque pointer to producer-provided private data.

Consumers MUST not process this member. Lifetime of this member is handled by the producer, and especially by the release callback.

Dictionary-encoded arrays

For dictionary-encoded arrays, the ArrowSchema.format string encodes the index type. The dictionary value type can be read from the ArrowSchema.dictionary structure.

The same holds for ArrowArray structure: while the parent structure points to the index data, the ArrowArray.dictionary points to the dictionary values array.

Extension arrays

For extension arrays, the ArrowSchema.format string encodes the storage type. Information about the extension type is encoded in the ArrowSchema.metadata string, similarly to the IPC format. Specifically, the metadata key ARROW:extension:name encodes the extension type name, and the metadata key ARROW:extension:metadata encodes the implementation-specific serialization of the extension type (for parameterized extension types). The base64 encoding of metadata values ensures that any possible serialization is representable.

The ArrowArray structure exported from an extension array simply points to the storage data of the extension array.

Memory management

The ArrowSchema and ArrowArray structures follow the same conventions for memory management. The term “base structure” below refers to the ArrowSchema or ArrowArray that is passed between producer and consumer – not any child structure thereof.

Member allocation

It is intended for the base structure to be stack- or heap-allocated by the consumer. In this case, the producer API should take a pointer to the consumer-allocated structure.

However, any data pointed to by the struct MUST be allocated and maintained by the producer. This includes the format and metadata strings, the arrays of buffer and children pointers, etc.

Therefore, the consumer MUST not try to interfere with the producer’s handling of these members’ lifetime. The only way the consumer influences data lifetime is by calling the base structure’s release callback.

Released structure

A released structure is indicated by setting its release callback to NULL. Before reading and interpreting a structure’s data, consumers SHOULD check for a NULL release callback and treat it accordingly (probably by erroring out).

Release callback semantics – for consumers

Consumers MUST call a base structure’s release callback when they won’t be using it anymore, but they MUST not call any of its children’s release callbacks (including the optional dictionary). The producer is responsible for releasing the children.

In any case, a consumer MUST not try to access the base structure anymore after calling its release callback – including any associated data such as its children.

Release callback semantics – for producers

If producers need additional information for lifetime handling (for example, a C++ producer may want to use shared_ptr for array and buffer lifetime), they MUST use the private_data member to locate the required bookkeeping information.

The release callback MUST not assume that the structure will be located at the same memory location as when it was originally produced. The consumer is free to move the structure around (see “Moving an array”).

The release callback MUST walk all children structures (including the optional dictionary) and call their own release callbacks.

The release callback MUST free any data area directly owned by the structure (such as the buffers and children members).

The release callback MUST mark the structure as released, by setting its release member to NULL.

Below is a good starting point for implementing a release callback, where the TODO area must be filled with producer-specific deallocation code:

static void ReleaseExportedArray(struct ArrowArray* array) {
  // This should not be called on already released array
  assert(array->format != NULL);

  // Release children
  for (int64_t i = 0; i < array->n_children; ++i) {
    struct ArrowArray* child = array->children[i];
    if (child->release != NULL) {
      child->release(child);
      assert(child->release == NULL);
    }
  }

  // Release dictionary
  struct ArrowArray* dict = array->dictionary;
  if (dict != NULL && dict->release != NULL) {
    dict->release(dict);
    assert(dict->release == NULL);
  }

  // TODO here: release and/or deallocate all data directly owned by
  // the ArrowArray struct, such as the private_data.

  // Mark array released
  array->release = NULL;
}

Moving an array

The consumer can move the ArrowArray structure by bitwise copying or shallow member-wise copying. Then it MUST mark the source structure released (see “released structure” above for how to do it) but without calling the release callback. This ensures that only one live copy of the struct is active at any given time and that lifetime is correctly communicated to the producer.

As usual, the release callback will be called on the destination structure when it is not needed anymore.

Moving child arrays

It is also possible to move one or several child arrays, but the parent ArrowArray structure MUST be released immediately afterwards, as it won’t point to valid child arrays anymore.

The main use case for this is to keep alive only a subset of child arrays (for example if you are only interested in certain columns of the data), while releasing the others.

Note

For moving to work correctly, the ArrowArray structure has to be trivially relocatable. Therefore, pointer members inside the ArrowArray structure (including private_data) MUST not point inside the structure itself. Also, external pointers to the structure MUST not be separately stored by the producer. Instead, the producer MUST use the private_data member so as to remember any necessary bookkeeping information.

Record batches

A record batch can be trivially considered as an equivalent struct array with additional top-level metadata.

Example use case

A C++ database engine wants to provide the option to deliver results in Arrow format, but without imposing themselves a dependency on the Arrow software libraries. With the Arrow C data interface, the engine can let the caller pass a pointer to a ArrowArray structure, and fill it with the next chunk of results.

It can do so without including the Arrow C++ headers or linking with the Arrow DLLs. Furthermore, the database engine’s C API can benefit other runtimes and libraries that know about the Arrow C data interface, through e.g. a C FFI layer.

C producer examples

Exporting a simple int32 array

Export a non-nullable int32 type with empty metadata. In this case, all ArrowSchema members point to statically-allocated data, so the release callback is trivial.

static void release_int32_type(struct ArrowSchema* schema) {
   // Mark released
   schema->release = NULL;
}

void export_int32_type(struct ArrowSchema* schema) {
   *schema = (struct ArrowSchema) {
      // Type description
      .format = "i",
      .name = "",
      .metadata = NULL,
      .flags = 0,
      .n_children = 0,
      .children = NULL,
      .dictionary = NULL,
      // Bookkeeping
      .release = &release_int32_type
   };
}

Export a C-malloc()ed array of the same type as a Arrow array, transferring ownership to the consumer through the release callback:

static void release_int32_array(struct ArrowArray* array) {
   assert(array->n_buffers == 2);
   // Free the buffers and the buffers array
   free((void *) array->buffers[1]);
   free(array->buffers);
   // Mark released
   array->release = NULL;
}

void export_int32_array(const int32_t* data, int64_t nitems,
                        struct ArrowArray* array) {
   // Initialize primitive fields
   *array = (struct ArrowArray) {
      // Data description
      .length = nitems,
      .offset = 0,
      .null_count = 0,
      .n_buffers = 2,
      .n_children = 0,
      .children = NULL,
      .dictionary = NULL,
      // Bookkeeping
      .release = &release_int32_array
   };
   // Allocate list of buffers
   array->buffers = (const void**) malloc(sizeof(void*) * array->n_buffers);
   assert(array->buffers != NULL);
   array->buffers[0] = NULL;  // no nulls, null bitmap can be omitted
   array->buffers[1] = data;
}

Exporting a struct<float32, utf8> array

Export the array type as a ArrowSchema with C-malloc()ed children:

static void release_malloced_type(struct ArrowSchema* schema) {
   int i;
   for (i = 0; i < schema->n_children; ++i) {
      struct ArrowSchema* child = schema->children[i];
      if (child->release != NULL) {
         child->release(child);
      }
   }
   free(schema->children);
   // Mark released
   schema->release = NULL;
}

void export_float32_utf8_type(struct ArrowSchema* schema) {
   struct ArrowSchema* child;

   //
   // Initialize parent type
   //
   *schema = (struct ArrowSchema) {
      // Type description
      .format = "+s",
      .name = "",
      .metadata = NULL,
      .flags = 0,
      .n_children = 2,
      .dictionary = NULL,
      // Bookkeeping
      .release = &release_malloced_type
   };
   // Allocate list of children types
   schema->children = malloc(sizeof(struct ArrowSchema*) * schema->n_children);

   //
   // Initialize child type #0
   //
   child = schema->children[0] = malloc(sizeof(struct ArrowSchema));
   *child = (struct ArrowSchema) {
      // Type description
      .format = "f",
      .name = "floats",
      .metadata = NULL,
      .flags = ARROW_FLAG_NULLABLE,
      .n_children = 0,
      .dictionary = NULL,
      .children = NULL,
      // Bookkeeping
      .release = &release_malloced_type
   };

   //
   // Initialize child type #1
   //
   child = schema->children[1] = malloc(sizeof(struct ArrowSchema));
   *child = (struct ArrowSchema) {
      // Type description
      .format = "u",
      .name = "strings",
      .metadata = NULL,
      .flags = ARROW_FLAG_NULLABLE,
      .n_children = 0,
      .dictionary = NULL,
      .children = NULL,
      // Bookkeeping
      .release = &release_malloced_type
   };
}

Export C-malloc()ed arrays in Arrow-compatible layout as an Arrow struct array, transferring ownership to the consumer:

static void release_malloced_array(struct ArrowArray* array) {
   int i;
   // Free children
   for (i = 0; i < array->n_children; ++i) {
      struct ArrowArray* child = array->children[i];
      if (child->release != NULL) {
         child->release(child);
      }
   }
   free(array->children);
   // Free buffers
   for (i = 0; i < array->n_buffers; ++i) {
      free((void *) array->buffers[i]);
   }
   free(array->buffers);
   // Mark released
   array->release = NULL;
}

void export_float32_utf8_array(
      int64_t nitems,
      const uint8_t* float32_nulls, const float* float32_data,
      const uint8_t* utf8_nulls, const int32_t* utf8_offsets, const uint8_t* utf8_data,
      struct ArrowArray* array) {
   struct ArrowArray* child;

   //
   // Initialize parent array
   //
   *array = (struct ArrowArray) {
      // Data description
      .length = nitems,
      .offset = 0,
      .null_count = 0,
      .n_buffers = 1,
      .n_children = 2,
      .dictionary = NULL,
      // Bookkeeping
      .release = &release_malloced_array
   };
   // Allocate list of parent buffers
   array->buffers = malloc(sizeof(void*) * array->n_buffers);
   array->buffers[0] = NULL;  // no nulls, null bitmap can be omitted
   // Allocate list of children arrays
   array->children = malloc(sizeof(struct ArrowArray*) * array->n_children);

   //
   // Initialize child array #0
   //
   child = array->children[0] = malloc(sizeof(struct ArrowArray));
   *child = (struct ArrowArray) {
      // Data description
      .length = nitems,
      .offset = 0,
      .null_count = -1,
      .n_buffers = 2,
      .n_children = 0,
      .dictionary = NULL,
      .children = NULL,
      // Bookkeeping
      .release = &release_malloced_array
   };
   child->buffers = malloc(sizeof(void*) * array->n_buffers);
   child->buffers[0] = float32_nulls;
   child->buffers[1] = float32_data;

   //
   // Initialize child array #1
   //
   child = array->children[1] = malloc(sizeof(struct ArrowArray));
   *child = (struct ArrowArray) {
      // Data description
      .length = nitems,
      .offset = 0,
      .null_count = -1,
      .n_buffers = 3,
      .n_children = 0,
      .dictionary = NULL,
      .children = NULL,
      // Bookkeeping
      .release = &release_malloced_array
   };
   child->buffers = malloc(sizeof(void*) * array->n_buffers);
   child->buffers[0] = utf8_nulls;
   child->buffers[1] = utf8_offsets;
   child->buffers[2] = utf8_data;
}

Why two distinct structures?

In many cases, the same type or schema description applies to multiple, possibly short, batches of data. To avoid paying the cost of exporting and importing the type description for each batch, the ArrowSchema can be passed once, separately, at the beginning of the conversation between producer and consumer.

In other cases yet, the data type is fixed by the producer API, and may not need to be communicated at all.

However, if a producer is focused on one-shot exchange of data, it can communicate the ArrowSchema and ArrowArray structures in the same API call.

Updating this specification

Once this specification is supported in an official Arrow release, the C ABI is frozen. This means the ArrowSchema and ArrowArray structure definitions should not change in any way – including adding new members.

Backwards-compatible changes are allowed, for example new ArrowSchema.flags values or expanded possibilities for the ArrowSchema.format string.

Any incompatible changes should be part of a new specification, for example “Arrow C data interface v2”.

Inspiration

The Arrow C data interface is inspired by the Python buffer protocol, which has proven immensely successful in allowing various Python libraries exchange numerical data with no knowledge of each other and near-zero adaptation cost.