Deprecated extension API

This page documents the original asdf extension API, which has been deprecated in favor of Extensions. Since support for the deprecated API will be removed in asdf 3.0, we recommend that all new extensions be implemented with the new API.

Extensions provide a way for ASDF to represent complex types that are not defined by the ASDF standard. Examples of types that require custom extensions include types from third-party libraries, user-defined types, and complex types that are part of the Python standard library but are not handled in the ASDF standard. From ASDF’s perspective, these are all considered ‘custom’ types.

Supporting new types in ASDF is easy. Three components are required:

  1. A YAML Schema file for each new type.

  2. A tag class (inheriting from asdf.CustomType) corresponding to each new custom type. The class must override to_tree and from_tree from asdf.CustomType in order to define how ASDF serializes and deserializes the custom type.

  3. A Python class to define an “extension” to ASDF, which is a set of related types. This class must implement the asdf.extension.AsdfExtension abstract base class. In general, a third-party library that defines multiple custom types can group them all in the same extension.

Note

The mechanisms of tag classes and extension classes are specific to this particular implementation of ASDF. As of this writing, this is the only complete implementation of the ASDF Standard. However, other language implementations may use other mechanisms for processing custom types.

All implementations of ASDF, regardless of language, will make use of the same schemas for abstract data type definitions. This allows all ASDF files to be language-agnostic, and also enables interoperability.

An Example

As an example, we will write an extension for ASDF that allows us to represent Python’s standard fractions.Fraction class for representing rational numbers. We will call our new ASDF type fraction.

First, the YAML Schema, defining the type as a pair of integers:

%YAML 1.1
---
$schema: "http://stsci.edu/schemas/yaml-schema/draft-01"
id: "http://nowhere.org/schemas/custom/fraction-1.0.0"
title: An example custom type for handling fractions

tag: "tag:nowhere.org:custom/fraction-1.0.0"
type: array
items:
  type: integer
minItems: 2
maxItems: 2
...

Then, the Python implementation of the tag class and extension class. See the asdf.CustomType and asdf.extension.AsdfExtension documentation for more information:

Note that the method to_tree of the tag class FractionType defines how the library converts fractions.Fraction into a tree that can be stored by ASDF. Conversely, the method from_tree defines how the library reads a serialized representation of the object and converts it back into an instance of fractions.Fraction.

Note that the values of the name, organization, standard, and version fields are all reflected in the id and tag definitions in the schema.

Note also that the base of the tag value (up to the name and version components) is reflected in tag_mapping property of the FractionExtension type, which is used to map tags to URLs. The url_mapping is used to map URLs (of the same form as the id field in the schema) to the actual location of a schema file.

Once these classes and the schema have been defined, we can save an ASDF file using them:

Defining custom types

In the example above, we showed how to create an extension that is capable of serializing fractions.Fraction. The custom tag type that we created was defined as a subclass of asdf.CustomType.

Custom type attributes

We overrode the following attributes of CustomType in order to define FractionType (each bullet is also a link to the API documentation):

Each of these attributes is important, and each is described in more detail in the linked API documentation.

The choice of name should be descriptive of the custom type that is being serialized. The choice of organization, and standard is fairly arbitrary, but also important. Custom types that are provided by the same package should be grouped into the same standard and organization.

These three values, along with the version, are used to define the YAML tag that will mark the serialized type in ASDF files. In our example, the tag becomes tag:nowhere.org:custom/fraction-1.0.0. The tag is important when defining the asdf.extension.AsdfExtension subclass.

Critically, these values must all be reflected in the associated schema.

Custom type methods

In addition to the attributes mentioned above, we also overrode the following methods of CustomType (each bullet is also a link to the API documentation):

The to_tree method defines how an instance of a custom data type is converted into data structures that represent a YAML tree that can be serialized to a file.

The from_tree method defines how a YAML tree can be converted back into an instance of the original custom data type.

In the example above, we used a list to contain the important attributes of fractions.Fraction. However, this choice is fairly arbitrary, as long as it is consistent between the way that to_tree and from_tree are defined. For example, we could have also chosen to use a dict:

In this case, the associated schema would look like the following:

%YAML 1.1
---
$schema: "http://stsci.edu/schemas/yaml-schema/draft-01"
id: "http://nowhere.org/schemas/custom/fraction-1.0.0"
title: An example custom type for handling fractions

tag: "tag:nowhere.org:custom/fraction-1.0.0"
type: object
properties:
  numerator:
    type: integer
  denominator:
    type: integer
...

We can compare the output using this representation to the example above:

Serializing more complex types

Sometimes the custom types that we wish to represent in ASDF themselves have attributes which are also custom types. As a somewhat contrived example, consider a 2D cartesian coordinate that uses fraction.Fraction to represent each of the components. We will call this type Fractional2DCoordinate.

First we need to define a schema to represent this new type:

%YAML 1.1
---
$schema: "http://stsci.edu/schemas/yaml-schema/draft-01"
id: "http://nowhere.org/schemas/custom/fractional_2d_coord-1.0.0"
title: An example custom type for handling components

tag: "tag:nowhere.org:custom/fractional_2d_coord-1.0.0"
type: object
properties:
  x:
    $ref: fraction-1.0.0
  y:
    $ref: fraction-1.0.0
...

Note that in the schema, the x and y attributes are expressed as references to our fraction-1.0.0 schema. Since both of these schemas are defined under the same standard and organization, we can simply use the name and version of the fraction-1.0.0 schema to refer to it. However, if the reference type was defined in a different organization and standard, it would be necessary to use the entire YAML tag in the reference (e.g. tag:nowhere.org:custom/fraction-1.0.0). Relative tag references are also allowed where appropriate.

We also need to define the custom tag type that corresponds to our new type:

In previous versions of this library, it was necessary for our Fractional2DCoordinateType class to call yamlutil functions explicitly to convert the x and y components to and from their tree representations. Now, the library will automatically convert nested custom types before calling from_tree, and after receiving the result from to_tree.

Since Fractional2DCoordinateType shares the same organization and standard as FractionType, it can be added to the same extension class:

Now we can use this extension to create an ASDF file:

Note that in the resulting ASDF file, the x and y components of our new fraction_2d_coord type are tagged as fraction-1.0.0.

Serializing reference cycles

Special considerations must be made when deserializing a custom type that contains a reference to itself among its descendants. Consider a fractions.Fraction subclass that maintains a reference to its multiplicative inverse:

The inverse of the inverse of a fraction is the fraction itself, so you might wish to construct your objects in the following way:

Which creates an “infinite loop” between the two fractions. An ordinary CustomType wouldn’t be able to deserialize this, since each object requires that the other be deserialized first! Let’s see what happens when we define our from_tree method in a naive way:

After adding our type to the extension class, the tree will serialize correctly:

But upon deserialization, we notice a problem:

The presence of _PendingValue is asdf’s way of telling you that the value corresponding to the key inverse was not fully deserialized at the time that you retrieved it. We can handle this situation by making our from_tree a generator function:

The generator version of from_tree yields the partially constructed FractionWithInverse object before setting its inverse property. This allows asdf to proceed to constructing the inverse FractionWithInverse object, and resume the original from_tree execution only when the inverse is actually available.

With this new version of from_tree, we can successfully deserialize our ASDF file:

Assigning schema and tag versions

Authors of new tags and schemas should strive to use the conventions described by semantic versioning. Tags and schemas for types that have not been serialized before should begin at 1.0.0. Versions for a particular tag type need not move in lock-step with other tag types in the same extension.

The patch version should be bumped for bug fixes and other minor, backwards-compatible changes. New features can be indicated with increments to the minor version, as long as they remain backwards compatible with older versions of the schema. Any changes that break backwards compatibility must be indicated by a major version update.

Since ASDF is intended to be an archival file format, authors of tags and schemas should work to ensure that ASDF files created with older extensions can continue to be processed. This means that every time a schema version is bumped (with the possible exception of patch updates), a new schema file should be created.

For example, if we currently have a schema for xyz-1.0.0, and we wish to make changes and bump the version to xyz-1.1.0, we should leave the original schema intact. A new schema file should be created for xyz-1.1.0, which can exist in parallel with the old file. The version of the corresponding tag type should be bumped to 1.1.0.

For more details on the behavior of schema and tag versioning from a user perspective, see Versioning and Compatibility, and also Custom types, extensions, and versioning.

Explicit version support

To some extent schemas and tag classes will be closely tied to the custom data types that they represent. This means that in some cases API changes or other changes to the representation of the underlying types will force us to modify our schemas and tag classes. ASDF’s schema versioning allows us to handle changes in schemas over time.

Let’s consider an imaginary custom type called Person that we want to serialize in ASDF. The first version of Person was constructed using a first and last name:

person = Person("James", "Webb")
print(person.first, person.last)

Our version 1.0.0 YAML schema for Person might look like the following:

%YAML 1.1
---
$schema: "http://stsci.edu/schemas/yaml-schema/draft-01"
id: "http://nowhere.org/schemas/custom/person-1.0.0"
title: An example custom type for representing a Person

tag: "tag:nowhere.org:custom/person-1.0.0"
type: array
items:
  type: string
minItems: 2
maxItems: 2
...

And our tag implementation would look something like this:

import asdf
from people import Person


class PersonType(asdf.CustomType):
    name = "person"
    organization = "nowhere.org"
    version = (1, 0, 0)
    standard = "custom"
    types = [Person]

    @classmethod
    def to_tree(cls, node, ctx):
        return [node.first, node.last]

    @classmethod
    def from_tree(cls, tree, ctx):
        return Person(tree[0], tree[1])

However, a newer version of Person now requires a middle name in the constructor as well:

person = Person("James", "Edwin", "Webb")
print(person.first, person.middle, person.last)

So we update our YAML schema to version 1.1.0 in order to support newer versions of Person:

%YAML 1.1
---
$schema: "http://stsci.edu/schemas/yaml-schema/draft-01"
id: "http://nowhere.org/schemas/custom/person-1.1.0"
title: An example custom type for representing a Person

tag: "tag:nowhere.org:custom/person-1.1.0"
type: array
items:
  type: string
minItems: 3
maxItems: 3
...

We need to update our tag class implementation as well. However, we need to be careful. We still want to be able to read version 1.0.0 of our schema and be able to convert it to the newer version of Person objects. To accomplish this, we will make use of the supported_versions attribute for our tag class. This will allow us to declare explicit support for the schema versions our tag class implements.

Under the hood, asdf creates multiple copies of our PersonType tag class, each with a different version attribute corresponding to one of the supported versions. This means that in our new tag class implementation, we can condition our from_tree implementation on the value of version to determine which schema version should be used when reading:

import asdf
from people import Person


class PersonType(asdf.CustomType):
    name = "person"
    organization = "nowhere.org"
    version = (1, 1, 0)
    supported_versions = [(1, 0, 0), (1, 1, 0)]
    standard = "custom"
    types = [Person]

    @classmethod
    def to_tree(cls, node, ctx):
        return [node.first, node.middle, node.last]

    @classmethod
    def from_tree(cls, tree, ctx):
        # Handle the older version of the person schema
        if cls.version == (1, 0, 0):
            # Construct a Person object with an empty middle name field
            return Person(tree[0], "", tree[1])
        else:
            # The newer version of the schema stores the middle name too
            return person(tree[0], tree[1], tree[2])

Note that the implementation of to_tree is not conditioned on cls.version since we do not need to convert new Person objects back to the older version of the schema.

Handling subclasses

By default, if a custom type is serialized by an asdf tag class, then all subclasses of that type can also be serialized. However, no attributes that are specific to the subclass will be stored in the file. When reading the file, an instance of the base custom type will be returned instead of the subclass that was written.

To properly handle subclasses of custom types already recognized by asdf, it is necessary to implement a separate tag class that is specific to the subclass to be serialized.

Previous versions of this library implemented an experimental feature that allowed ADSF to serialize subclass attributes using the same tag class, but this feature was dropped as it produced files that were not portable.

Creating custom schemas

All custom types to be serialized by asdf require custom schemas. The best resource for creating ASDF schemas can be found in the ASDF Standard documentation.

In most cases, ASDF schemas will be included as part of a packaged software distribution. In these cases, it is important for the url_mapping of the corresponding AsdfExtension extension class to map the schema URL to an actual location on disk. However, it is possible for schemas to be hosted online as well, in which case the URL mapping can map (perhaps trivially) to an actual network location. See Defining custom extension classes for more information.

It is also important for packages that provide custom schemas to test them, both to make sure that they are valid, and to ensure that any examples they provide are also valid. See Testing custom schemas for more information.

Adding custom validators

A new type may also add new validation keywords to the schema language. This can be used to impose type-specific restrictions on the values in an ASDF file. This feature is used internally so a schema can specify the required datatype of an array.

To support custom validation keywords, set the validators member of a CustomType subclass to a dictionary where the keys are the validation keyword name and the values are validation functions. The validation functions are of the same form as the validation functions in the underlying jsonschema library, and are passed the following arguments:

  • validator: A jsonschema.Validator instance.

  • value: The value of the schema keyword.

  • instance: The instance to validate. This will be made up of basic datatypes as represented in the YAML file (list, dict, number, strings), and not include any object types.

  • schema: The entire schema that applies to instance. Useful to get other related schema keywords.

The validation function should either return None if the instance is valid or yield one or more jsonschema.ValidationError objects if the instance is invalid.

To continue the example from above, for the FractionType say we want to add a validation keyword “simplified” that, when true, asserts that the corresponding fraction is in simplified form:

from asdf import ValidationError


def validate_simplified(validator, simplified, instance, schema):
    if simplified:
        reduced = fraction.Fraction(instance[0], instance[1])
        if reduced.numerator != instance[0] or reduced.denominator != instance[1]:
            yield ValidationError("Fraction is not in simplified form.")


FractionType.validators = {"simplified": validate_simplified}

Defining custom extension classes

Extension classes are the mechanism that asdf uses to register custom tag types so that they can be used when processing ASDF files. Packages that define their own custom tag types must also define extensions in order for those types to be used.

All extension classes must implement the asdf.extension.AsdfExtension abstract base class. A custom extension will override each of the following properties of asdf.extension.AsdfExtension (the text in each bullet is also a link to the corresponding documentation):

Overriding built-in extensions

It is possible for externally defined extensions to override tag types that are provided by asdf’s built-in extension. For example, maybe an external package wants to provide a different implementation of NDArrayType. In this case, the external package does not need to provide custom schemas since the schema for the type to be overridden is already provided as part of the ASDF standard.

Instead, the extension class may inherit from asdf’s asdf.extension.BuiltinExtension and simply override the types property to indicate the type that is being overridden. Doing this preserves the tag_mapping and url_mapping that is used by the BuiltinExtension, which allows the schemas that are packaged by asdf to be located.

asdf will give precedence to the type that is provided by the external extension, effectively overriding the corresponding type in the built-in extension. Note that it is currently undefined if multiple external extensions are provided that override the same built-in type.

Packaging custom extensions

Packaging schemas

If a package provides custom schemas, the schema files must be installed as part of that package distribution. In general, schema files must be installed into a subdirectory of the package distribution. The asdf extension class must supply a url_mapping that maps to the installed location of the schemas. See Defining custom extension classes for more details.

Registering entry points

Packages that provide their own ASDF extensions can (and should!) install them so that they are automatically detectable by the asdf Python package. This is accomplished using Python’s setuptools entry points. Entry points are registered in a package’s setup.py file.

Consider a package that provides an extension class MyPackageExtension in the submodule mypackage.asdf.extensions. We need to register this class as an extension entry point that asdf will recognize. First, we create a dictionary:

entry_points = {}
entry_points["asdf_extensions"] = [
    "mypackage = mypackage.asdf.extensions:MyPackageExtension"
]

The key used in the entry_points dictionary must be 'asdf_extensions'. The value must be an array of one or more strings, each with the following format:

extension_name = fully.specified.submodule:ExtensionClass

The extension name can be any arbitrary string, but it should be descriptive of the package and the extension. In most cases the package itself name will suffice.

Note that depending on individual package requirements, there may be other entries in the entry_points dictionary.

The entry points must be passed to the call to setuptools.setup:

from setuptools import setup

entry_points = {}
entry_points["asdf_extensions"] = [
    "mypackage = mypackage.asdf.extensions:MyPackageExtension"
]

setup(
    # We omit other package-specific arguments that are not
    # relevant to this example
    entry_points=entry_points,
)

When running python setup.py install or python setup.py develop on this package, the entry points will be registered automatically. This allows the asdf package to recognize the extensions without any user intervention. Users of your package that wish to read ASDF files using types that you have registered will not need to use any extension explicitly. Instead, asdf will automatically recognize the types you have registered and will process them appropriately. See Extensions from other packages for more information on using extensions.

Testing custom schemas

Packages that provide their own schemas can test them using asdf’s pytest plugin for schema testing. Schemas are tested for overall validity, and any examples given within the schemas are also tested.

The schema tester plugin is automatically registered when the asdf package is installed. In order to enable testing, it is necessary to add the directory containing your schema files to the pytest section of your project’s build configuration (pyproject.toml or setup.cfg). If you do not already have such a file, creating one with the following should be sufficient:

The schema directory paths should be paths that are relative to the top of the package directory when it is installed. If this is different from the path in the source directory, then both paths can be used to facilitate in-place testing (see asdf’s own pyproject.toml for an example of this).

Note

Older versions of asdf (prior to 2.4.0) required the plugin to be registered in your project’s conftest.py file. As of 2.4.0, the plugin is now registered automatically and so this line should be removed from your conftest.py file, unless you need to retain compatibility with older versions of asdf.

The asdf_schema_skip_names configuration variable can be used to skip schema files that live within one of the asdf_schema_root directories but should not be tested. The names should be given as simple base file names (without directory paths or extensions). Again, see asdf’s own pyproject.toml file for an example.

The schema tests do not run by default. In order to enable the tests by default for your package, add asdf_schema_tests_enabled = 'true' to the [tool.pytest.ini_options] section of your pyproject.toml file (or [tool:pytest] in setup.cfg). If you do not wish to enable the schema tests by default, you can add the --asdf-tests option to the pytest command line to enable tests on a per-run basis.