# Copyright 2023 Iguazio
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import inspect
from abc import ABCMeta
from types import MethodType
from typing import Any, List, Tuple, Type, Union
import docstring_parser
from mlrun.artifacts import Artifact
from mlrun.datastore import DataItem
from mlrun.utils import logger
from ..errors import MLRunPackagePackingError, MLRunPackageUnpackingError
from ..packager import Packager
from ..utils import DEFAULT_PICKLE_MODULE, ArtifactType, Pickler, TypeHintUtils
class _DefaultPackagerMeta(ABCMeta):
"""
Metaclass for `DefaultPackager` to override `__doc__` attribute into a class property. This way sphinx will get a
dynamically generated docstring that will include a summary of the packager.
"""
def __new__(mcls, name: str, bases: tuple, namespace: dict, **kwargs):
"""
Create a new DefaultPackager metaclass that saves the original packager docstring to another attribute named
`_packager_doc`.
:param name: A string representing the name of the class being instantiated.
:param bases: A tuple of classes from which the class will inherit.
:param namespace: The namespace of the class holding its attributes (from here the docstring will be taken).
"""
# Save the original doc string to a separate class variable as it will be overriden later on by the metaclass
# property `__doc__`:
namespace["_packager_doc"] = namespace.get("__doc__", "")
# Continue creating the metaclass:
return super().__new__(mcls, name, bases, namespace, **kwargs)
@property
def __doc__(cls: Type["DefaultPackager"]) -> str:
"""
Override the `__doc__` attribute of a `DefaultPackager` to be a property in order to auto-summarize the
packager's class docstring. The summary is concatenated after the original class doc string.
The summary will be in the following structure:
<cls._packager_doc>
.. rubric:: Packager Summary
**Packing Type**: ``<cls.PACKABLE_OBJECT_TYPE>``
**Packing Sub-Classes**: True / False
* **Priority**: ...
**Default Artifact Types**:
* **Packing**: ...
* **Unpacking**: ...
**Artifact Types**:
* **type 1**: ...
* configuration 1 - ...
* configuration 2 - ...
* **type 2**: ...
* configuration 1: ...
* configuration 2: ...
:returns: The original docstring with the generated packager summary.
"""
# Create a packager instance:
packager = cls()
# Get the packager's name and module:
packager_name = packager.__class__.__name__
packager_module = packager.__module__
# Get the original packager class doc string:
packager_doc_string = cls._packager_doc.split("\n")
packager_doc_string = "\n".join(line[4:] for line in packager_doc_string)
# Parse the packable type section:
type_name = (
"Any type"
if packager.PACKABLE_OBJECT_TYPE is ...
else (
f"``{str(packager.PACKABLE_OBJECT_TYPE)}``"
if TypeHintUtils.is_typing_type(type_hint=packager.PACKABLE_OBJECT_TYPE)
else f"``{packager.PACKABLE_OBJECT_TYPE.__module__}.{packager.PACKABLE_OBJECT_TYPE.__name__}``"
)
)
packing_type = f"**Packing Type**: {type_name}"
# Subclasses support section:
packing_sub_classes = f"**Packing Sub-Classes**: {packager.PACK_SUBCLASSES}"
# Priority section:
priority_value = (
packager.priority
if packager.priority is not ...
else "Default priority (5)"
)
priority = f"**Priority**: {priority_value}"
# Default artifact types:
def get_default_artifact_type(pack_or_unpack: str) -> str:
pack_or_unpack = f"default_{pack_or_unpack}ing_artifact_type"
method_name = f"get_{pack_or_unpack}"
argument_name = pack_or_unpack.upper()
return (
getattr(packager, argument_name)
if packager_name == "DefaultPackager"
or method_name not in packager.__class__.__dict__
else (
f"Refer to the packager's "
f":py:meth:`~{packager_module}.{packager_name}.{method_name}` method."
)
)
default_artifact_types = (
f"**Default Artifact Types**:"
f"\n\n* **Packing**: {get_default_artifact_type('pack')}"
f"\n\n* **Unpacking**: {get_default_artifact_type('unpack')}"
)
# Artifact types section:
artifact_types = "**Artifact Types**:"
for artifact_type in packager.get_supported_artifact_types():
# Get the packing method docstring:
method_doc = docstring_parser.parse(
getattr(packager, f"pack_{artifact_type}").__doc__
)
# Add the artifact type bullet:
artifact_type_doc = f"{method_doc.short_description or ''}{method_doc.long_description or ''}".replace(
"\n", ""
)
artifact_types += (
f"\n\n* :py:meth:`{artifact_type}<{packager_module}.{packager_name}.pack_{artifact_type}>` - "
+ artifact_type_doc
)
# Add the artifact type configurations (ignoring the `obj` and `key` parameters):
configurations_doc = "\n\n * ".join(
"{} - {}".format(
parameter.arg_name, parameter.description.replace("\n", "")
)
for parameter in method_doc.params[2:]
)
if configurations_doc:
artifact_types += f"\n\n * {configurations_doc}"
# Construct the final doc string and return:
doc = (
f"{packager_doc_string}"
"\n\n.. rubric:: Packager Summary"
f"\n\n{packing_type}"
f"\n\n{packing_sub_classes}"
f"\n\n{priority}"
f"\n\n{default_artifact_types}"
f"\n\n{artifact_types}"
f"\n\n"
)
return doc
[docs]class DefaultPackager(Packager, metaclass=_DefaultPackagerMeta):
"""
A default packager that handles all types and packs them as pickle files.
The default packager implements all the required methods and has a default logic that should satisfy most
use cases. To work with this class, don't override the abstract class methods, but instead follow the
guidelines below:
* **The class variable** :py:meth:`PACKABLE_OBJECT_TYPE<PACKABLE_OBJECT_TYPE>`: The type of object this packager can
pack and unpack (used in the ``is_packable`` method).
* **The class variable** :py:meth:`PACK_SUBCLASSES<PACK_SUBCLASSES>`: A flag that indicates whether to pack all
subclasses of the ``PACKABLE_OBJECT_TYPE`` (used in the ``is_packable`` method). Default is False.
* **The class variable** :py:meth:`DEFAULT_PACKING_ARTIFACT_TYPE<DEFAULT_PACKING_ARTIFACT_TYPE>`: The default
artifact type to pack as. It is returned from the method ``get_default_packing_artifact_type``.
* **The class variable** :py:meth:`DEFAULT_UNPACKING_ARTIFACT_TYPE<DEFAULT_UNPACKING_ARTIFACT_TYPE>`: The default
artifact type to unpack from. It is returned from the method
``get_default_unpacking_artifact_type``.
* **The abstract class method** :py:meth:`pack`: This method is implemented to get the object and send it to the
relevant packing method by the given artifact type using the following naming: "pack_<artifact_type>". (If
the artifact type was not provided, it uses the default). For example: if the artifact type is `x` then
the class method ``pack_x`` must be implemented. The signature of each pack class method must be::
def pack_x(self, obj: Any, key: str, ...) -> Union[Tuple[Artifact, dict], dict]:
pass
Where 'x' is the artifact type, 'obj' is the object to pack, `key` is the key to name the artifact and `...` are
additional, custom, log hint configurations. The returned values are the packed artifact and the instructions
for unpacking it, or in the case of result, the dictionary of the result with its key and value. configurations
are sent by the user and shouldn't be mandatory, meaning they should have a default value (otherwise, the user
has to add them to every log hint).
* **The abstract class method** :py:meth:`unpack`: The method is implemented to get a
:py:meth:`DataItem<mlrun.datastore.base.DataItem>` and send it to the relevant unpacking method by the artifact
type using the following naming: `"unpack_<artifact_type>"`. (If the artifact type was not provided,
it uses the default).
For example: if the artifact type stored within the ``DataItem`` is `x` then the class method
``unpack_x`` must be implemented. The signature of each unpack class method must be::
def unpack_x(self, data_item: mlrun.DataItem, ...) -> Any:
pass
Where 'x' is the artifact type, 'data_item' is the artifact's data item to unpack, `...` are the instructions that
were originally returned from ``pack_x``. (Each instruction must be optional (have a default value) to support
objects from this type that were not packaged but custom-logged.) The returned value is the unpacked
object.
* **The abstract class method** :py:meth:`is_packable`: The method is implemented to automatically validate
the object type and artifact type by the following rules:
* **Object type validation**: Checks if the given object type matches the variable ``PACKABLE_OBJECT_TYPE``
with respect to the ``PACK_SUBCLASSES`` class variable.
* **Artifact type validation**: Checks if the given artifact type is in the list returned from
``get_supported_artifact_types``.
* **The abstract class method** :py:meth:`is_unpackable`: The method is left as implemented in ``Packager``.
* **The abstract class method** :py:meth:`get_supported_artifact_types`: The method is implemented to look for all
pack + unpack class methods implemented to collect the supported artifact types. If ``PackagerX`` has ``pack_y``,
``unpack_y`` and ``pack_z``, ``unpack_z`` that means the artifact types supported are `y` and `z`.
* **The abstract class method** :py:meth:`get_default_packing_artifact_type`: The method is implemented to return
the new class variable ``DEFAULT_PACKING_ARTIFACT_TYPE``. You can still override the method if the default
artifact type you need could change according to the object that's about to be packed.
* **The abstract class method** :py:meth:`get_default_unpacking_artifact_type`: The method is implemented to return
the new class variable ``DEFAULT_UNPACKING_ARTIFACT_TYPE``. You can still override the method if the default
artifact type you need could change according to the data item that's about to be unpacked.
.. rubric:: Important to remember
From the :py:meth:`Packager<mlrun.package.packager.Packager>` docstring:
* **Linking artifacts** ("extra data"): In order to link between packages (using the extra data or metrics spec
attributes of an artifact), use the key as if it exists and as value ellipses (...). The manager
links all packages once it is done packing.
For example, given extra data keys in the log hint as `extra_data`, set them to an artifact as follows::
artifact = Artifact(key="my_artifact")
artifact.spec.extra_data = {key: ... for key in extra_data}
* **Clearing outputs**: Some packagers may produce files and temporary directories that should be deleted after
the artifact is logged. The packager can mark paths of files and directories to delete after
logging using the class method ``add_future_clearing_path``.
For example, in the following packager's ``pack`` method, you can write a text file, create an artifact, and then
mark the text file to be deleted once the artifact is logged::
with open("./some_file.txt", "w") as file:
file.write("Pack me")
artifact = Artifact(key="my_artifact")
self.add_future_clearing_path(path="./some_file.txt")
return artifact, None
"""
#: The type of object this packager can pack and unpack.
PACKABLE_OBJECT_TYPE: Type = ...
#: A flag for indicating whether to also pack all subclasses of the `PACKABLE_OBJECT_TYPE`.
PACK_SUBCLASSES = False
#: The default artifact type to pack as.
DEFAULT_PACKING_ARTIFACT_TYPE = ArtifactType.OBJECT
#: The default artifact type to unpack from.
DEFAULT_UNPACKING_ARTIFACT_TYPE = ArtifactType.OBJECT
[docs] def get_default_packing_artifact_type(self, obj: Any) -> str:
"""
Get the default artifact type for packing an object of this packager.
:param obj: The about-to-be packed object.
:return: The default artifact type.
"""
return self.DEFAULT_PACKING_ARTIFACT_TYPE
[docs] def get_default_unpacking_artifact_type(self, data_item: DataItem) -> str:
"""
Get the default artifact type used for unpacking a data item holding an object of this packager. The method
is used when a data item is sent for unpacking without it being a package, but is a simple url or an old /
manually logged artifact.
:param data_item: The about-to-be unpacked data item.
:return: The default artifact type.
"""
return self.DEFAULT_UNPACKING_ARTIFACT_TYPE
[docs] def get_supported_artifact_types(self) -> List[str]:
"""
Get all the supported artifact types on this packager.
:return: A list of all the supported artifact types.
"""
# We look for pack + unpack method couples so there won't be a scenario where an object can be packed but not
# unpacked. Result has no unpacking so we add it separately.
return [
key[len("pack_") :]
for key in dir(self)
if key.startswith("pack_") and f"unpack_{key[len('pack_'):]}" in dir(self)
] + ["result"]
[docs] def pack(
self,
obj: Any,
key: str = None,
artifact_type: str = None,
configurations: dict = None,
) -> Union[Tuple[Artifact, dict], dict]:
"""
Pack an object as the given artifact type using the provided configurations.
:param obj: The object to pack.
:param key: The key of the artifact.
:param artifact_type: Artifact type to log to MLRun. If passing `None`, the default artifact type is used.
:param configurations: Log hints configurations to pass to the packing method.
:return: If the packed object is an artifact, a tuple of the packed artifact and unpacking instructions
dictionary. If the packed object is a result, a dictionary containing the result key and value.
"""
# Get default artifact type in case it was not provided:
if artifact_type is None:
artifact_type = self.get_default_packing_artifact_type(obj=obj)
# Set empty dictionary in case no configurations were given:
configurations = configurations or {}
# Get the packing method according to the artifact type:
pack_method = getattr(self, f"pack_{artifact_type}")
# Validate correct configurations were passed:
self._validate_method_arguments(
method=pack_method,
arguments=configurations,
is_packing=True,
)
# Call the packing method and return the package:
return pack_method(obj=obj, key=key, **configurations)
[docs] def unpack(
self,
data_item: DataItem,
artifact_type: str = None,
instructions: dict = None,
) -> Any:
"""
Unpack the data item's artifact by the provided type using the given instructions.
:param data_item: The data input to unpack.
:param artifact_type: The artifact type to unpack the data item as. If passing `None`, the default artifact type
is used.
:param instructions: Additional instructions noted in the package to pass to the unpacking method.
:return: The unpacked data item's object.
:raise MLRunPackageUnpackingError: In case the packager could not unpack the data item.
"""
# Get default artifact type in case it was not provided:
if artifact_type is None:
artifact_type = self.get_default_unpacking_artifact_type(
data_item=data_item
)
# Set empty dictionary in case no instructions were given:
instructions = instructions or {}
# Get the unpacking method according to the artifact type:
unpack_method = getattr(self, f"unpack_{artifact_type}")
# Validate correct instructions were passed:
self._validate_method_arguments(
method=unpack_method,
arguments=instructions,
is_packing=False,
)
# Call the unpacking method and return the object:
return unpack_method(data_item, **instructions)
[docs] def is_packable(
self, obj: Any, artifact_type: str = None, configurations: dict = None
) -> bool:
"""
Check if this packager can pack an object of the provided type as the provided artifact type.
The method is implemented to validate the object's type and artifact type by checking if the given object type
matches the variable ``PACKABLE_OBJECT_TYPE`` with respect to the ``PACK_SUBCLASSES`` class variable. If it
does, it checks if the given artifact type is in the list returned from ``get_supported_artifact_types``.
:param obj: The object to pack.
:param artifact_type: The artifact type to log the object as.
:param configurations: The log hint configurations passed by the user.
:return: True if packable and False otherwise.
"""
# Get the object's type:
object_type = type(obj)
# Check type (ellipses means any type):
if self.PACKABLE_OBJECT_TYPE is not ...:
if not TypeHintUtils.is_matching(
object_type=object_type,
type_hint=self.PACKABLE_OBJECT_TYPE,
include_subclasses=self.PACK_SUBCLASSES,
reduce_type_hint=False,
):
return False
# Check the artifact type:
if (
artifact_type is not None
and artifact_type not in self.get_supported_artifact_types()
):
return False
# Packable:
return True
[docs] def pack_object(
self,
obj: Any,
key: str,
pickle_module_name: str = DEFAULT_PICKLE_MODULE,
) -> Tuple[Artifact, dict]:
"""
Pack a python object, pickling it into a pkl file and store it in an artifact.
:param obj: The object to pack and log.
:param key: The artifact's key.
:param pickle_module_name: The pickle module name to use for serializing the object.
:return: The artifacts and its pickling instructions.
"""
# Pickle the object to file:
pickle_path, instructions = Pickler.pickle(
obj=obj, pickle_module_name=pickle_module_name
)
# Initialize an artifact to the pkl file:
artifact = Artifact(key=key, src_path=pickle_path)
# Add the pickle path to the clearing list:
self.add_future_clearing_path(path=pickle_path)
return artifact, instructions
[docs] def pack_result(self, obj: Any, key: str) -> dict:
"""
Pack an object as a result.
:param obj: The object to pack and log.
:param key: The result's key.
:return: The result dictionary.
"""
return {key: obj}
[docs] def unpack_object(
self,
data_item: DataItem,
pickle_module_name: str = DEFAULT_PICKLE_MODULE,
object_module_name: str = None,
python_version: str = None,
pickle_module_version: str = None,
object_module_version: str = None,
) -> Any:
"""
Unpack the data item's object, unpickle it using the instructions, and return.
Warnings of mismatching python and module versions between the original pickling interpreter and this one may be
raised.
:param data_item: The data item holding the pkl file.
:param pickle_module_name: Module to use for unpickling the object.
:param object_module_name: The original object's module. Used to verify that the current interpreter object
module version matches the pickled object version before unpickling the object.
:param python_version: The python version in which the original object was pickled. Used to verify that
the current interpreter python version matches the pickled object version before
unpickling the object.
:param pickle_module_version: The pickle module version. Used to verify that the current interpreter module
version matches the one that pickled the object before unpickling it.
:param object_module_version: The original object's module version to match to the interpreter's module version.
:return: The un-pickled python object.
"""
# Get the pkl file to local directory:
pickle_path = self.get_data_item_local_path(data_item=data_item)
# Unpickle and return:
return Pickler.unpickle(
pickle_path=pickle_path,
pickle_module_name=pickle_module_name,
object_module_name=object_module_name,
python_version=python_version,
pickle_module_version=pickle_module_version,
object_module_version=object_module_version,
)
def _validate_method_arguments(
self, method: MethodType, arguments: dict, is_packing: bool
):
"""
Validate keyword arguments to pass to a method. Used for validating log hint configurations for packing methods
and instructions for unpacking methods.
:param method: The method to validate the arguments for.
:param arguments: Keyword arguments to validate.
:param is_packing: Flag to know if the arguments came from packing or unpacking, to raise the correct exception
if validation failed.
:raise MLRunPackagePackingError: If there are missing configurations in the log hint.
:raise MLRunPackageUnpackingError: If there are missing instructions in the artifact's spec.
"""
# Get the possible and mandatory (arguments that has no default value) arguments from the functions:
possible_arguments = inspect.signature(method).parameters
mandatory_arguments = [
name
for name, parameter in possible_arguments.items()
# If default value is `empty` it is mandatory:
if parameter.default is inspect.Parameter.empty
# Ignore the *args and **kwargs parameters:
and parameter.kind
not in [inspect.Parameter.VAR_KEYWORD, inspect.Parameter.VAR_POSITIONAL]
]
# Remove non-configuration (or non-instructions in case of unpacking) arguments:
if is_packing:
mandatory_arguments.remove("obj")
if "key" in mandatory_arguments:
# Default keys may be an option in the future, so it is not necessary a mandatory argument:
mandatory_arguments.remove("key")
else:
mandatory_arguments.remove("data_item")
# Validate there are no missing arguments (only mandatory ones):
missing_arguments = [
mandatory_argument
for mandatory_argument in mandatory_arguments
if mandatory_argument not in arguments
]
if missing_arguments:
if is_packing:
raise MLRunPackagePackingError(
f"The packager '{self.__class__.__name__}' could not pack the package due to missing "
f"configurations: {', '.join(missing_arguments)}. Add the missing arguments to the log hint of "
f"this object in order to pack it. Make sure you pass a dictionary log hint and not a string in "
f"order to pass configurations in the log hint."
)
raise MLRunPackageUnpackingError(
f"The packager '{self.__class__.__name__}' could not unpack the package due to missing instructions: "
f"{', '.join(missing_arguments)}. Missing instructions are likely due to an update in the packager's "
f"code that not support the old implementation. This backward compatibility should not occur. To "
f"overcome it, try to edit the instructions in the artifact's spec to enable unpacking it again."
)
# Validate all given arguments are correct:
incorrect_arguments = [
argument for argument in arguments if argument not in possible_arguments
]
if incorrect_arguments:
arguments_type = "configurations" if is_packing else "instructions"
logger.warn(
f"Unexpected {arguments_type} given for {self.__class__.__name__}: {', '.join(incorrect_arguments)}. "
f"Possible {arguments_type} are: {', '.join(possible_arguments.keys())}. The packager tries to "
f"continue by ignoring the incorrect arguments."
)
