Core Features¶

This section discusses the core features of the ASDF data format, and provides examples and use cases that are specific to the Python implementation.

Data Model¶

The fundamental data object in ASDF is the tree, which is a nested combination of basic data structures: dictionaries, lists, strings and numbers. In Python, these types correspond to dict, list, str, and int, float, and complex, respectively. The top-level tree object behaves like a Python dictionary and supports arbitrary nesting of data structures. For simple examples of creating and reading trees, see Overview.

Note

The ASDF Standard imposes a maximum size of 52 bits for integer literals in the tree (see the docs for details and justification). Attempting to store a larger value will result in a validation error.

Integers and floats of up to 64 bits can be stored inside of numpy arrays (see below).

For arbitrary precision integer support, see IntegerType.

One of the key features of ASDF is its ability to serialize numpy arrays. This is discussed in detail in Array Data.

While the core ASDF package supports serialization of basic data types and Numpy arrays, its true power comes from its ability to be extended to support serialization of a wide range of custom data types. Details on using ASDF extensions can be found in Using extensions. Details on creating custom ASDF extensions to support custom data types can be found in Writing ASDF extensions.

Array Data¶

Much of ASDF’s power and convenience comes from its ability to represent multidimensional array data. The asdf Python package provides native support for numpy arrays.

Using extensions¶

According to Wikipedia, serialization “is the process of translating data structures or object state into a format that can be stored…and reconstructed later” 1.

The power of ASDF is that it provides the ability to store, or serialize, the state of Python objects into a human-readable data format. The state of those objects can later be restored by another program in a process called deserialization.

While ASDF is capable of serializing basic Python types and Numpy arrays out of the box, it can also be extended to serialize arbitrary custom data types. This section discusses the extension mechanism from a user’s perspective. For documentation on creating extensions, see Writing ASDF extensions.

Even though this particular implementation of ASDF necessarily serializes Python data types, in theory an ASDF implementation in another language could read the resulting file and reconstruct an analogous type in that language. Conversely, this implementation can read ASDF files that were written by other implementations of ASDF as long as the proper extensions are available.

Schema validation¶

Schema validation is used to determine whether an ASDF file is well formed. All ASDF files must conform to the schemas defined by the ASDF Standard. Schema validation occurs when reading ASDF files (using asdf.open), and also when writing them out (using AsdfFile.write_to or AsdfFile.update).

Schema validation also plays a role when using custom extensions (see Using extensions and Writing ASDF extensions). Extensions must provide schemas for the types that they serialize. When writing a file with custom types, the output is validated against the schemas corresponding to those types. If the appropriate extension is installed when reading a file with custom types, then the types will be validated against the schemas provided by the corresponding extension.

Custom schemas¶

Every ASDF file is validated against the ASDF Standard, and also against any schemas provided by custom extensions. However, it is sometimes useful for particular applications to impose additional restrictions when deciding whether a given file is valid or not.

For example, consider an application that processes digital image data. The application expects the file to contain an image, and also some metadata about how the image was created. The following example schema reflects these expectations:

%YAML 1.1
---
$schema: "http://stsci.edu/schemas/yaml-schema/draft-01"
id: "http://stsci.edu/schemas/asdf/core/asdf-1.1.0"

tag: "tag:stsci.edu:asdf/core/asdf-1.1.0"
type: object
properties:
  image:
    description: An ndarray containing image data.
    $ref: "ndarray-1.0.0"

  metadata:
    type: object
    description: Metadata about the image
    properties:
      time:
        description: |
          A timestamp for when the image was created, in UTC.
        type: string
        format: date-time
      resolution:
        description: |
          A 2D array representing the resolution of the image (N x M).
        type: array
        items:
          type: integer
          number: 2

required: [image, metadata]
additionalProperties: true

This schema restricts the kinds of files that will be accepted as valid to those that contain a top-level image property that is an ndarray, and a top-level metadata property that contains information about the time the image was taken and the resolution of the image. Note that the schema uses the same id and tag as the top-level core schema from the ASDF Standard. This is because it is validating the file at the top level, but is imposing restrictions beyond what is normally required for an ASDF file.

In order to use this schema for a secondary validation pass, we pass the custom_schema argument to either asdf.open or the AsdfFile constructor. Assume that the schema file lives in image_schema.yaml, and we wish to open a file called image.asdf. We would open the file with the following code:

import asdf
af = asdf.open('image.asdf', custom_schema='image_schema.yaml')

Similarly, if we wished to use this schema when creating new files:

new_af = asdf.AsdfFile(custom_schema='image_schema.yaml')
...

Versioning and Compatibility¶

There are several different versions to keep in mind when discussing ASDF:

The software package version
The ASDF Standard version
The ASDF file format version
Individual tag and schema versions

Each ASDF file contains information about the various versions that were used to create the file. The most important of these are the ASDF Standard version and the ASDF file format version. A particular version of the ASDF software package will explicitly provide support for a specific combination of these versions.

Tag and schema versions are also important for serializing and deserializing data types that are stored in ASDF files. A detailed discussion of tag and schema versions from a user perspective can be found in Custom types, extensions, and versioning.

Since ASDF is designed to serve as an archival format, the software attempts to provide backwards compatibility when reading older versions of the ASDF Standard and ASDF file format. However, since deserializing ASDF types sometimes requires other software packages, backwards compatibility is often contingent on the available versions of such software packages.

In general, forward compatibility with newer versions of the ASDF Standard and ASDF file format is not supported by the software. However, if newer tag and schema versions are detected, the software will attempt to process them.

When creating new ASDF files, it is possible to control the version of the file format that is used. This can be specified by passing the version argument to either the AsdfFile constructor when the file object is created, or to the AsdfFile.write_to method when it is written. By default, the latest version of the file format will be used. Note that this option has no effect on the versions of tag types from custom extensions.

External References¶

Tree References¶

ASDF files may reference items in the tree in other ASDF files. The syntax used in the file for this is called “JSON Pointer”, but users of asdf can largely ignore that.

First, we’ll create a ASDF file with a couple of arrays in it:

import asdf
from asdf import AsdfFile
import numpy as np

tree = {
    'a': np.arange(0, 10),
    'b': np.arange(10, 20)
}

target = AsdfFile(tree)
target.write_to('target.asdf')

target.asdfb'#ASDF' 1.4.0
%YAML 1.1
%TAG ! tag:stsci.edu:asdf/
--- !core/asdf-1.1.0
asdf_library: !core/software-1.0.0 {author: Space Telescope Science Institute, homepage: 'http://github.com/spacetelescope/asdf',
  name: asdf, version: 2.5.2}
history:
  extensions:
  - !core/extension_metadata-1.0.0
    extension_class: asdf.extension.BuiltinExtension
    software: {name: asdf, version: 2.5.2}
a: !core/ndarray-1.0.0
  source: 0
  datatype: int64
  byteorder: little
  shape: [10]
b: !core/ndarray-1.0.0
  source: 1
  datatype: int64
  byteorder: little
  shape: [10]
...BLOCK 0:
    allocated_size: 80
    used_size: 80
    data_size: 80
    data: b'0000000000000000010000000000000002000000...'BLOCK 1:
    allocated_size: 80
    used_size: 80
    data_size: 80
    data: b'0a000000000000000b000000000000000c000000...'#ASDF BLOCK INDEX
%YAML 1.1
---
- 583
- 717
...

Then we will reference those arrays in a couple of different ways. First, we’ll load the source file in Python and use the make_reference method to generate a reference to array a. Second, we’ll work at the lower level by manually writing a JSON Pointer to array b, which doesn’t require loading or having access to the target file.

ff = AsdfFile()

with asdf.open('target.asdf') as target:
    ff.tree['my_ref_a'] = target.make_reference(['a'])

ff.tree['my_ref_b'] = {'$ref': 'target.asdf#b'}

ff.write_to('source.asdf')

source.asdfb'#ASDF' 1.4.0
%YAML 1.1
%TAG ! tag:stsci.edu:asdf/
--- !core/asdf-1.1.0
asdf_library: !core/software-1.0.0 {author: Space Telescope Science Institute, homepage: 'http://github.com/spacetelescope/asdf',
  name: asdf, version: 2.5.2}
history:
  extensions:
  - !core/extension_metadata-1.0.0
    extension_class: asdf.extension.BuiltinExtension
    software: {name: asdf, version: 2.5.2}
my_ref_a: {$ref: target.asdf#a}
my_ref_b: {$ref: target.asdf#b}
...

Calling find_references will look up all of the references so they can be used as if they were local to the tree. It doesn’t actually move any of the data, and keeps the references as references.

with asdf.open('source.asdf') as ff:
    ff.find_references()
    assert ff.tree['my_ref_b'].shape == (10,)

On the other hand, calling resolve_references places all of the referenced content directly in the tree, so when we write it out again, all of the external references are gone, with the literal content in its place.

with asdf.open('source.asdf') as ff:
    ff.resolve_references()
    ff.write_to('resolved.asdf')

resolved.asdfb'#ASDF' 1.4.0
%YAML 1.1
%TAG ! tag:stsci.edu:asdf/
--- !core/asdf-1.1.0
asdf_library: !core/software-1.0.0 {author: Space Telescope Science Institute, homepage: 'http://github.com/spacetelescope/asdf',
  name: asdf, version: 2.5.2}
history:
  extensions:
  - !core/extension_metadata-1.0.0
    extension_class: asdf.extension.BuiltinExtension
    software: {name: asdf, version: 2.5.2}
my_ref_a: !core/ndarray-1.0.0
  source: 0
  datatype: int64
  byteorder: little
  shape: [10]
my_ref_b: !core/ndarray-1.0.0
  source: 1
  datatype: int64
  byteorder: little
  shape: [10]
...BLOCK 0:
    allocated_size: 80
    used_size: 80
    data_size: 80
    data: b'0000000000000000010000000000000002000000...'BLOCK 1:
    allocated_size: 80
    used_size: 80
    data_size: 80
    data: b'0a000000000000000b000000000000000c000000...'#ASDF BLOCK INDEX
%YAML 1.1
---
- 597
- 731
...

A similar feature provided by YAML, anchors and aliases, also provides a way to support references within the same file. These are supported by asdf, however the JSON Pointer approach is generally favored because:

It is possible to reference elements in another file

Elements are referenced by location in the tree, not an identifier, therefore, everything can be referenced.

Anchors and aliases are handled automatically by asdf when the data structure is recursive. For example here is a dictionary that is included twice in the same tree:

d = {'foo': 'bar'}
d['baz'] = d
tree = {'d': d}

ff = AsdfFile(tree)
ff.write_to('anchors.asdf')

anchors.asdfb'#ASDF' 1.4.0
%YAML 1.1
%TAG ! tag:stsci.edu:asdf/
--- !core/asdf-1.1.0
asdf_library: !core/software-1.0.0 {author: Space Telescope Science Institute, homepage: 'http://github.com/spacetelescope/asdf',
  name: asdf, version: 2.5.2}
history:
  extensions:
  - !core/extension_metadata-1.0.0
    extension_class: asdf.extension.BuiltinExtension
    software: {name: asdf, version: 2.5.2}
d:
  baz:
    baz: &id001
      baz: *id001
      foo: bar
    foo: bar
  foo: bar
...

Array References¶

ASDF files can refer to array data that is stored in other files using the ExternalArrayReference type.

External files need not be ASDF files: ASDF is completely agnostic as to the format of the external file. The ASDF external array reference does not define how the external data file will be resolved; in fact it does not even check for the existence of the external file. It simply provides a way for ASDF files to refer to arrays that exist in external files.

Creating an external array reference is simple. Only four pieces of information are required:

The name of the external file. Since ASDF does not itself resolve the file or check for its existence, the format of the name is not important. In most cases the name will be a path relative to the ASDF file itself, or a URI for a network resource.
The data type of the array data. This is a string representing any valid numpy.dtype.
The shape of the data array. This is a tuple representing the dimensions of the array data.
The array data target. This is either an integer or a string that indicates to the user something about how the data array should be accessed in the external file. For example, if there are multiple data arrays in the external file, the target might be an integer index. Or if the external file is an ASDF file, the target might be a string indicating the key to use in the external file’s tree. The value and format of the target field is completely arbitrary since ASDF will not use it itself.

As an example, we will create a reference to an external CSV file. We will assume that one of the rows of the CSV file contains the array data we care about:

import asdf

csv_data_row = 10 # The row of the CSV file containing the data we want
csv_row_size = 100 # The size of the array
extref = asdf.ExternalArrayReference('data.csv', csv_data_row, "int64", (csv_row_size,))

tree = {'csv_data': extref}
af = asdf.AsdfFile(tree)
af.write_to('external_array.asdf')

external_array.asdfb'#ASDF' 1.4.0
%YAML 1.1
%TAG ! tag:stsci.edu:asdf/
--- !core/asdf-1.1.0
asdf_library: !core/software-1.0.0 {author: Space Telescope Science Institute, homepage: 'http://github.com/spacetelescope/asdf',
  name: asdf, version: 2.5.2}
history:
  extensions:
  - !core/extension_metadata-1.0.0
    extension_class: asdf.extension.BuiltinExtension
    software: {name: asdf, version: 2.5.2}
csv_data: !core/externalarray-1.0.0
  datatype: int64
  fileuri: data.csv
  shape: [100]
  target: 10
...

When reading a file containing external references, the user is responsible for using the information in the ExternalArrayReference type to open the external file and retrieve the associated array data.

Saving history entries¶

asdf has a convenience method for notating the history of transformations that have been performed on a file.

Given a AsdfFile object, call add_history_entry, given a description of the change and optionally a description of the software (i.e. your software, not asdf) that performed the operation.

from asdf import AsdfFile
import numpy as np

tree = {
    'a': np.random.rand(32, 32)
}

ff = AsdfFile(tree)
ff.add_history_entry(
    "Initial random numbers",
    {'name': 'asdf examples',
     'author': 'John Q. Public',
     'homepage': 'http://github.com/spacetelescope/asdf',
     'version': '0.1'})
ff.write_to('example.asdf')

example.asdfb'#ASDF' 1.4.0
%YAML 1.1
%TAG ! tag:stsci.edu:asdf/
--- !core/asdf-1.1.0
asdf_library: !core/software-1.0.0 {author: Space Telescope Science Institute, homepage: 'http://github.com/spacetelescope/asdf',
  name: asdf, version: 2.5.2}
history:
  entries:
  - !core/history_entry-1.0.0
    description: Initial random numbers
    software: !core/software-1.0.0 {author: John Q. Public, homepage: 'http://github.com/spacetelescope/asdf',
      name: asdf examples, version: '0.1'}
    time: 2020-02-28 20:46:32
  extensions:
  - !core/extension_metadata-1.0.0
    extension_class: asdf.extension.BuiltinExtension
    software: {name: asdf, version: 2.5.2}
a: !core/ndarray-1.0.0
  source: 0
  datatype: float64
  byteorder: little
  shape: [32, 32]
...BLOCK 0:
    allocated_size: 8192
    used_size: 8192
    data_size: 8192
    data: b'b907a1077d1ee63f567dc3253d32d23f005de14e...'#ASDF BLOCK INDEX
%YAML 1.1
---
- 767
...

ASDF automatically saves history metadata about the extensions that were used to create the file. This information is used when opening files to determine if the proper extensions are installed (see Extension checking for more details).

Saving ASDF in FITS¶

Note

This section is about packaging entire ASDF files inside of FITS data format files. This is probably only of interest to astronomers. Making use of this feature requires the astropy package to be installed.

Sometimes you may need to store the structured data supported by ASDF inside of a FITS file in order to be compatible with legacy tools that support only FITS.

First, create an HDUList object using astropy.io.fits. Here, we are building an HDUList from scratch, but it could also have been loaded from an existing file.

We will create a FITS file that has two image extensions, SCI and DQ respectively.

from astropy.io import fits

hdulist = fits.HDUList()
hdulist.append(fits.ImageHDU(np.arange(512, dtype=np.float), name='SCI'))
hdulist.append(fits.ImageHDU(np.arange(512, dtype=np.float), name='DQ'))

Next we make a tree structure out of the data in the FITS file. Importantly, we use the same array references in the FITS HDUList and store them in the tree. By doing this, ASDF will automatically refer to the data in the regular FITS extensions.

tree = {
    'model': {
        'sci': {
            'data': hdulist['SCI'].data,
        },
        'dq': {
            'data': hdulist['DQ'].data,
        }
    }
}

Now we take both the FITS HDUList and the ASDF tree and create an AsdfInFits object.

from asdf import fits_embed

ff = fits_embed.AsdfInFits(hdulist, tree)
ff.write_to('embedded_asdf.fits')

The special ASDF extension in the resulting FITS file contains the following data. Note that the data source of the arrays uses the fits: prefix to indicate that the data comes from a FITS extension:

content.asdfb'#ASDF' 1.4.0
%YAML 1.1
%TAG ! tag:stsci.edu:asdf/
--- !core/asdf-1.1.0
asdf_library: !core/software-1.0.0 {author: Space Telescope Science Institute, homepage: 'http://github.com/spacetelescope/asdf',
  name: asdf, version: 2.5.2}
history:
  extensions:
  - !core/extension_metadata-1.0.0
    extension_class: asdf.extension.BuiltinExtension
    software: {name: asdf, version: 2.5.2}
model:
  dq:
    data: !core/ndarray-1.0.0
      source: fits:DQ,1
      datatype: float64
      byteorder: big
      shape: [512]
  sci:
    data: !core/ndarray-1.0.0
      source: fits:SCI,1
      datatype: float64
      byteorder: big
      shape: [512]
...

To load an ASDF-in-FITS file, simply open it using asdf.open. The returned value will be an AsdfInFits object, which can be used in the same way as any other AsdfFile object.

with asdf.open('embedded_asdf.fits') as asdf_in_fits:
    science = asdf_in_fits.tree['model']['sci']

Footnotes

1: https://en.wikipedia.org/wiki/Serialization