High level overview of the basic ASDF library¶
This document is an attempt to make it easier to understand the design and workings of the python asdf library for those unfamiliar with it. This is expected to grow organically so at the moment it should not be considered complete or comprehensive.
Understanding the design is complicated by the fact that the library effectively inserts custom methods or classes into the objects that the pyyaml and jsonschema libraries use. Understanding what is going on thus means having some understanding of the relevant parts of the internals of both of those libraries. This overview will try to provide a small amount of context for these packages to illuminate how the code in asdf interacts with them.
There are at least two ways of outlining the design. One is to give high level overviews of the various modules and how they interact with other modules. The other is to illustrate how code is actually invoked in common operations, this often being much more informative on a practical level (at least some find that to be the case). This document will attempt to do both.
We will start with a high-level review of concepts and terms and point to where these are handled in the asdf modules.
Because of the complexity, this initial design overview will focus on issues of validation and tree construction when reading.
Some terminology and definitions¶
URI vs URL (Universal Resource Identifier). This is distinguished from URL (Universal Resource Locator) primarily in that URI is a mechanism for a unique name that follows a particular syntax, but itself may not indicate where the resource is. Generally URLs are expected to be used on the web for the HTTP protocol, though for asdf, this isn’t necessarily the case as mentioned next.
Resolver: Tools to map URIs and tags into actual locations of schema files, which may be local directories (the usual approach) or an actual URL for retrieval over the network. This is more complicated that it may seem for reasons explained later.
Validator: Tool to confirm that the YAML conforms to the schemas that apply. A lot goes on in this area and it is pretty complex in the implementation.
Tree building: The YAML content is built into a tree in two stages. The YAML parser converts the raw YAML into a custom Python structure. It is that structure that is validated. Then if no errors are found, the tree is converted into a tree where tagged nodes get converted into corresponding Python objects (usually, an option exists to prevent this from happening, which is useful for some applications), e.g., WCS object or numpy arrays (well, not quite that simply for numpy arrays).
The above is a simplified view of what happens when an ASDF file is read.
Most of resolver tools and code is in
resolver.py (but not all).
Most of the validation code is in
The code that builds the trees is spread in many places:
types.py as well as all the extension code that supplies
code to handle the tags within (and often the the associated schemas).
A note on the location of schemas and tag code; there is a bit of schizophrenic aspect to this since schema should be language agnostic and in that view, not bundled with specific language library code. But currently nearly all of the implementation is in Python so while the long-term goal is to keep them separate, it is more convenient to keep them together for now. You will see cases where they are separate and some where they are bundled].
Actions that happen when ASDF is imported¶
The entry points for all asdf extensions are obtained in
_DefaultExtensions) which is instantiated at the end of the module
default_extensions, but the entry points are only found when
default_extensions.extensions is accessed (it’s a property)
The effect of this is to load all the specified entry point classes for all the
extensions that have registered through the entry point mechanism. (see
[https://packaging.python.org/specifications/entry-points/]) The list of classes
so loaded is what
default_extensions.extensions returns along with all the
built-in extensions part of ASDF.
AsdfFile class is instantiated, one thing that happens on the
__init__ is that
self._process_extensions() is called with an empty
list. That results in
default_extensions.extension_list being accessed,
which then results in
extension.AsdfExtensionList being instantiated with
the created extensions property.
This class populates the
url_mapping lists and the
validators dictionary, as well as populating the
_type_index attribute with
AsdfTypes subclasses defined in the extensions.
As a last step, the
url_mapping methods are generated
resolver.Resolver with the initial lists. These lists consist of
2-tuples. In the first case it is a mechanism to map the tag string to a url
string, typically with an expected prefix or suffix to the tag (suffix is
typical) so that given a full tag, it generates a url that includes the suffix
This permits one mapping to cover many tag variants. (The details of mapping
machinery with examples are given in a later section since understanding this is
essential to defining new tags and corresponding schemas.)
The URL mapping works in a similar way, except that it consists of 2-tuples where the first element is the common elements of the url, and the second part maps it to an actual location (url or file path). Again the second part may include a place holder for the suffix or prefix, and code to generate the path to the schema file.
The use of the resolver object turns these lists into functions so that supplied the appropriate input that matches something in the list, it gives the corresponding output.
Outline of how an ASDF file is opened and read into the corresponding Python object. ————————————————————————————
The starting point can be found in
asdf.py essentially through the following
chain (many calls and steps left out to keep it simpler to follow)
asdf.open("myasdffile.asdf") is called, it is aliased to
asdf.open_asdf which first creates an instance of
call the instance
af), then calls
af._open_impl() and then
af._open_asdf. That invokes a call to
generic.py basically contains code to handle all the variants of I/O
possible (files, streaming, http access, etc). In this case it returns a
RealFile instance that wraps a local file system file.
Next the file is examined to see if it is an ASDF file (first by examining the first few lines in the header). If it passes those checks, the header (yaml) section of the file is extracted through a proxy mechanism that signals an end of file when the end of the yaml is reached, but otherwise looks like a file object.
The yaml parsing phase described below normally returns a “tagged_tree”. That is (somewhat simplified), it returns the data structure that yaml would normally return without any object conversion (i.e., all nodes are either dicts, lists, or scalar values), except that they are objects that now support a tag attribute that indicates if a tag was associated with that node and what the tag was.
This reader object is passed to the yaml parser by calling
yamlutil.load_tree. A simple explanation for what goes on here is necessary
to understand how this all works. Yaml supports various kinds of loaders. For
security reasons, the “safe” loader is used (note that both C and python
versions are supported through an indirection of the
defined at the beginning of that module that determines whether the C version is
available). The loaders are recursive mechanisms that build the tree structure.
yamlutil.load_tree creates a temporary subclass of
and attaches a reference to the AsdfFile instance as the
.ctx attribute of
that temporary subclass.
One of the hooks that pyyaml supplies is the ability to overload the method
construct_object. That’s what the class
yamlutil.AsdfLoader does. pyyaml
calls this method at each node in the tree to see if anything special should be
done. One could perform conversion to predefined objects here, but instead it
does the following: it sees if the node.tag attribute is handled by yaml itself
(examples?) it calls that constructor which returns the type yaml converts it
it converts the node to the type indicated (dict, list, or scalar type) by yaml for that node.
it obtains the appropriate tag class (an AsdfType subclass) from the AsdfFile instance (using
ctx.type_index.fix_yaml_tagto deal with version issues to match the most appropriate tag class).
it wraps all the node alternatives in a special asdf
Taggedclass instance variant where that object contains a ._tag attribute that is a reference to the corresponding Tag class.
The loading process returns a tree of these Tagged object instances. This
tagged_tree is then returned to the
af instance (still running the
_open_asdf() method) this tree is passed to to the
(This is the major reason that the tree isn’t directly converted to an object
tree since jsonschema would not be able to use the final object tree for
validation, besides issues relate to the fact that things that don’t validate
may not be convertable to the designated object.)
The validate machinery is a bit confusing since there are essentially two basic approaches to how validation is done. One type of validation is for validation of schema files themselves, and the other for schemas for tags.
The schema.py file is fairly involved and the details are covered elsewhere. When the validator machinery is constructed, it uses the fundamental validation files (schemas). But this doesn’t handle the fact that the file being validated is yaml, not json and that there are items in yaml not part of json so special handling is needed. And the way it is handled is through a internal mechanism of the jsonschema library. There is a method that jsonschema calls recursively for a validator and it is called iter_errors. The subclass of the jsonschema validator class is defined as schema.ASDFValidator and this method is overloaded in this class. Despite its name, it’s primary purpose is to validate the special features that yaml has, namely applying schemas associated with tags (this is not part of the normal jsonschema scheme [ahem]). It is in this method that it looks for a tag for a node and if it exists and in the tag_index, loads the appropriate schema and applies it to the node. (jsonschemas are normally only associated with a whole json entity rather than specific nodes). While the purpose of this method is to iteratively handle errors that jsonschema detects, it has essentially been repurposed as the means of interjecting handling tag schemas.
In order to prevent repeated loading of the same schema, the lru caching scheme is used (from functools in the standard library) where the last n cached schemas are saved (details of how this works were recently changed to prevent a serious memory leak)
In any event, a lot is going on behind the scenes in validation and it deserves its own description elsewhere.
After validation, the tagged tree is then passed to yamlutil.tagged_tree_to_custom_tree() where the nodes in the tree that have special tag code convert the nodes into the appropriate Python objects that the base asdf and extensions are aware of. This is accomplished by that function defining a walker “callback” function (defined within that function as to pick up the af object intrinsically). The function then passes the callback walker to treeutil.walk_and_modify() where the tree will be traversed recursively applying the tag code associated with the tag to the more primitive tree representation replacing such nodes with Python objects. The tree traversal starts from the top, but the objects are created from the bottom up due to recursion (well, not quite that simple).
Understanding how this works is described more fully later on.
The result is what af.tree is set to, after doing another tree traversal looking for special type hooks for each node. It isn’t clear if there is yet any use of that feature.
Not quite that simple¶
Outline of schema.py¶
This module is somewhat confusing due to the many functions and methods with some variant of validate in their name. This will try to make clear what they do (a renaming of these may be in order).
Here is a list of the functions/classes in
schema.py and their purpose and
where they sit in the order of things
_type_to_tag: Handles mapping python types to yaml_tags, with the addition of support for OrderedDicts.
The next 5 functions are put in the
YAML_VALIDATORS dictionary to ultimately
be used by
_create_validator to create the json validator object
validate_tag: Obtain the relevant tag for the supplied instance (either built ins or custom objects) and check that it matches the tag supplied to the function.
validate_propertyOrder: Not really a validator but rather as a trick to indicate that properties should retain their order.
validate_flowStyle: Not really a validator but rather as a trick to store what style to use to write the elements (for yaml objects and arrays)
validate_style: Not really a validator but rather as a trick to store info on what style to use to write the string.
validate_type: Used to deal with date strings
(It may make sense to rename the above to be more descriptive of the action than where
they are stuck in the validation machinery; e.g.,
validate_fill_default: Set the default values for all properties that have a
subschema that defines a default. Called indirectly in
validate_remove_default: does the opposite; remove all properties where
value equals subschema default. Called indirectly in
this and the above, validate in the name mostly confuses although it is used by
the json validator.)
[these could be renamed as well since they do more than validate]
_create_validator: Creates an
ASDFValidator class on the fly that uses
jsonchema.validators class created. This
iter_errors method that is used to handle yaml tag cases
._tag attribute of the node to obtain the corresponding schema
for that tag; e.g., it calls
load_schema to obtain the right schema when
called for each node in the jsonschema machinery). What isn’t clear to me is why
this is done on the fly and at least cached since it really only handles two
variants of calls (basically which JSONSCHEMA version is to be used). Otherwise
it doesn’t appear to vary except for that. Admittedly, this is only created at
the top level. This is called by
class OrderedLoader: Inherits from the
_yaml_base_loader, but otherwise
does nothing new in the definition. But the following code defines
construct_mapping, and then adds it as a method.
construct_mapping: Defined outside the
OrderedLoader class but to be
added to the
OrderedLoader class by use of the base class add_constructor
method. This function flattens the mapping and returns an
OrderedDict of the
property attributes (This needs some deep understanding of how the yaml parser
actually works, which is not covered here. Apparently mappings can be
represented as nested trees as the yaml is originally parsed. Or something like
_load_schema: Loads json or yaml schemas (using the
_make_schema_loader: Defines the function load_schema using the provided resolver and _load_schema.
_make_resolver: Sets the schema loader for http, https, file, tag using a dictionary where these access methods are the keys and the schema loader returning only the schema (and not the uri). These all appear to use the same schema loader.
load_custom_schema: Deals with custom schemas.
load_schema: Loads a schema from the specified location (this is cached).
Called for every tag encountered (uses resolver machinery). Most of the
complexity is in resolving json references. Calls
resolver, reference.resolve_fragment, load_schema
_create_validator. Is called by validate to return
the created validator.
validate_large_literals: Ensures tree has no large literals (raises error if it does)
get_validator to get a validator object and then calls
its validate method, and validates any large literals using
fill_defaults: Inserts attributes missing with the default value
remove_defaults: Where the tree has attributes with value equal to the default, strip the attribute.
check_schema: Checks schema against the metaschema.
Illustration of the where these are called:
af.validate which calls
af._validate which then
schema.validate with the tagged tree as the first argument (it can be
called again if there is a custom schema).
validate -> get_validator -> _create_validator (returns
There are two levels of validation, those passed to the json_validation
machinery for the schemas themselves, and those that the tag machinery triggers
when the jsonschema validator calls through
iter_errors. The first level
handles all the tricks at the top. the
which in turn calls
_load_schema uses the
OrderedLoader to load the schemas.
How the ASDF library works with pyyaml¶
A Tree Identifier¶
There are three flavors of trees in the process of reading ASDF files, one will see many references to each in the code and description below.
pyyaml native tree. This consists of standard Python containers like dict and list, and primitive values like string, integer, float, etc.
Tagged tree. These are similar to pyyaml native trees, but with the basic types wrapped in a class that has has an attribute that identifies the tag associated with that node so that later processing can apply the appropriate conversion code to convert to the final Python object.
Custom tree. This is a tree where all nodes are converted to the destination Python objects. For example, a numpy array or GWCS object.
Brief overview of how pyyaml constructs a Python tree¶
Understanding the process of creating Python objects from yaml requires some understanding of how pyyaml works. We will not go into all the details of pyyaml, but instead concentrate on one phase of its loading process. First an outline of the phases of processing that pyyaml goes through in loading a yaml file:
scanning: Converting the text into lexical tokens. Done in scanner.py
parsing: Converting the lexical tokens into parsing events. Done in parser.py.
composing: Converting the parsing events into a tree structure of pyyaml objects. Done in composer.py
loading: Converting the pyyaml tree into a Python object tree. Done in constructor.py
We will focus on the last step since that is where asdf integrates with how pyyaml works.
The key object in that module is
BaseConstructor and its subclasses (asdf
SafeConstructor for security purposes). Note that the pyyaml code is
severely deficient in docstrings and comments. The key method that kicks
off the conversion is
construct_document(). Its responsibilities are to call
construct_object() method on the top node, “drain” any generators
produced by construction (more on this later), and finally reset internal
data structures once construction is complete.
The actual process seems somewhat mysterious because what is going on is
that it is using generators in place of vanilla code to construct the
children for mutable items. The general scheme is that each constructor
for mutable elements (see as an example the
SafeConstructor.construct_yaml_seq() method) is written
as a generator that is expected to be asked a value twice. The first value
returned is an empty object of the expected type (e.g., empty dict or
list) and when asked a second time, it populates the previous object
returned (and returns None, which is not used). (In rare exceptions,
when called with
deep=True, it does immediately populate the child nodes.)
Normally the generator is appended to the loader’s state_generators
attribute (a list) for later use. Any generators not handled in the
recursive chain are handled when contruct_object returns to
construct_document, where it iteratively asks each generator to complete
populating its referenced object. Since that step of populating the object
may in turn create new generators on the
state_generator list, it only
stops when no more generators appear on the list.
Why is this done? One reason is to handle references (anchors and aliases) that may be circular.
Suppose one had the following yaml source:
A: &a x: 1 B: item1: 42 item2: life, the universe, and everything circular: *a
Without generators, it would not be possible to handle this case since the node
identified by anchor
a has not been fully constructed when pyyaml encounters
a reference to that anchor among the same node’s descendants. The use
of the generator allows creation of the container object to reference
to before it is populated so that the above construction will work when
constructing the tree. To follow the above example in more detail, the
construction creates a dictionary for
a and then returns to the
construct_document() method, which then starts handling the generators put on
the list (there is only one in this case). The generator then populates
the contents of
a. For the attribute
B it encounters a new
mutable container, and puts its generator on the list to handle, and then
makes a reference to
a which now is defined. One last time it
handles the generator for
B and since each item in that is not
a container, the construction completes.
Pyyaml tracks pending objects in a recursive objects dict and throws an exception if generators fail to handle reference cycles. (The conversion of the tagged tree to the custom tree, performed later does not use the same technique; explained later)
How ASDF hooks into pyyaml construction¶
ASDF makes use of this by adding generators to this process by defining
a new construct method
construct_undefined() that handles all ASDF tag
cases. This is added to the pyyaml dict of construct methods under the
None. When pyyaml doesn’t find a tag, that is what it uses as
a key to handle unknown tags. Thus the construction is redirected to
ASDF code. That code returns a generator in the case of mutable ASDF
objects in line with how yaml works with mutable objects.
Historical note: Versions older than 2.6.0 did not work this way. Instead,
those versions completely replaced the pyyaml method
their own version that did not use generators as pyyaml did.
How conversion to ASDF objects is done¶
The current means of conversion is simpler to use by tag code, but also more subtle to understand how it actually works (for many, that means harder ;-)
The YAML loading process produces a tagged tree of basic Python types.
The conversion of these into ASDF types is kicked off when the
This function defines a walker function that is to be used with
treeutil.walk_and_modify(). Most of what the walker function does is
handle tag issues (e.g., can the tag be appropriately mapped to the
tag creation code) and then returns the appropriate ASDF type by calling
A note on tree traversal. One can traverse a tree in three ways:
inorder, preorder, and postorder (
asdf.info() uses a breadth-first
traversal, yet another exciting option, which we won’t describe here).
These respectively mean whether
nodes are visited in the horizontal ordering of the nodes displayed on
a graphs (inorder), descending the tree from the root, doing the left
node first, before the right node (preorder), or from the bottom up, doing
both leaf nodes before the parent node (postorder). In generating the
pyyaml tree, preorder works since it builds the tree from the root
as one would expect in constructing the tree. But in converting the
tagged tree into the custom tree, postorder is the natural course, where
the children are generated first so that the parent node can refer to
the final objects.
An important part of this conversion process is handled by an instance
of the class
treeutil._TreeModificationContext. This class does much the
same trick that pyyaml does with generators. Although pyyaml creates
references between basic python objects, these references must be
converted to references between ASDF objects, and doing so requires
a similar mechanism for building the ASDF objects. The
_TreeModificationContext object (hereafter context object)
holds the incomplete generators in a way similar to the pyyaml
There are differences though. The class
methods to indicate if nodes are pending (i.e., incomplete), and there
is a special value
PendingValue that is a signal that the node hasn’t
been handled yet (e.g., it may be referencing something yet to be done).
PendingValue persists to the end, it indicates a failure to handle
circular references in the tag code. This approach was taken because
one of the earlier prototype implementations did something like this,
passing dict and list subclasses that would throw an exception if a
PendingValue element was accessed. That would have been more friendly
to extension developers, but it was discarded because it wasn’t thought
it was worth turning all those high performance containers into slower
asdf subclasses. We may want to revisit this if we decide to implement
a tree that tracks “dirty” nodes and only writes to disk those that
have changed, since in that case we’ll need custom container subclasses
anyway. We could also consider writing our own dict/list subclass in C
so we could have our cake and eat it too.
walk_and_modify code handles the case where the tag code returns
a generator instead of a value. This generator is expected to be a
similar kind of generator to what pyyaml uses, but differing in that instead
of returning an empty container object it will populate whatever elements
it can complete (e.g, all non-mutable ones), and complete the
population of all the mutable members on the second iteration
(which may, in turn, generate new generators for mutable elements
contained within). When it detects a generator, the
code retrieves the first yielded value, then saves the generator in the
context. When the
top level of the context is reached (it handles nesting by indicating
how many times it has been entered as a context), it starts “draining”
the saved generators by doing the second iteration on them. Like
pyyaml, this second iteration may produce yet more generators that
get saved, and thus keeps iterating on the saved generators until none
It is not possible to construct reference cycles in immutable objects within pure Python code, and thus the generators are only needed for mutable constructs (e.g., dicts and lists).
Historical note: versions of the ASDF library prior to 2.6.0 required
tag code when converting from a tagged object to a custom object to
tagged_tree_to_custom_tree on any values of attributes that may be
arbitrarily nested objects. That no longer is needed with the latest code
since any attribute that contains a mapping or sequence object automatically
uses a generator, so population of that attribute is automatically
deferred until the context is exited. Thus there is no need to explicitly
call a function to populate it.
More explicitly, the
_recurse function defined within
(in this postorder case) calls
_handle_children() on the node
in question first. If the node contains children, they are each fed back into
_recurse and transformed into their final objects. A new node is populated
with these transformed children, and that is the node that gets handed to
tag.from_tree_tagged(). The effect is that the tag class receives
a structure containing only transformed children, so it has no need to
tagged_tree_to_custom_tree on its own.
Thus reader, your mind shall now be drained.