ducktools-lazyimporter 0.7.1

LazyImporter class to manage imports on attribute access.

ducktools: lazyimporter

Create an object to handle lazily importing from other modules.

Nearly every form of "lazyimporter" module name is taken on PyPI so this is namespaced.

Intended to help save on start time where some modules are only needed for specific functions while allowing information showing the import information to appear at the top of a module where expected.

This form of import works by creating a specific LazyImporter object that lazily imports modules or module attributes when the module or attribute is accessed on the object.

How to download

Download from PyPI: python -m pip install ducktools-lazyimporter

Example

Example using the packaging module.

__version__ = "v0.1.5"

from ducktools.lazyimporter import LazyImporter, FromImport

laz = LazyImporter([
    FromImport("packaging.version", "Version")
])

def is_newer_version(version_no: str) -> bool:
    """Check if a version number given indicates 
    a newer version than this package."""
    this_ver = laz.Version(__version__) 
    new_ver = laz.Version(version_no)
    return new_ver > this_ver

# Import will only occur when the function is called and 
# laz.Version is accessed
print(is_newer_version("v0.2.0"))

Why use a lazy importer?

One obvious use case is if you are creating a simple CLI application that you wish to feel fast. If the application has multiple pathways a lazy importer can improve performance by avoiding loading the modules that are only needed for heavier pathways. (It may also be worth looking at what library you are using for CLI argument parsing.)

I created this so I could use it on my own projects so here's an example of the performance of getting the help menu for ducktools-env with and without lazy imports.

With lazy imports:

hyperfine -w3 -r20 "python -m ducktools.env --help"

Benchmark 1: python -m ducktools.env --help
  Time (mean ± σ):      41.4 ms ±   1.0 ms    [User: 21.1 ms, System: 15.9 ms]
  Range (min … max):    40.0 ms …  44.1 ms    20 runs

Without lazy imports (by setting DUCKTOOLS_EAGER_IMPORT=true):

hyperfine -w3 -r20 "python -m ducktools.env --help"

Benchmark 1: python -m ducktools.env --help
  Time (mean ± σ):     112.8 ms ±   2.6 ms    [User: 78.1 ms, System: 35.9 ms]
  Range (min … max):   109.2 ms … 117.8 ms    20 runs

Hasn't this already been done

Yes.

But...

Most implementations rely on stdlib modules that are themselves slow to import (for example: typing, importlib.util, logging, inspect, ast). By contrast ducktools-lazyimporter only uses modules that python imports on launch.

ducktools-lazyimporter does not attempt to propagate laziness, only the modules provided to ducktools-lazyimporter directly will be imported lazily. Any subdependencies of those modules will be imported eagerly as if the import statement is placed where the importer attribute is first accessed.

Use Case

There are two main use cases this is designed for.

Replacing in-line imports used in a module

Sometimes it is useful to use tools from a module that has a significant import time. If this is part of a function/method that won't necessarily always be used it is common to delay the import and place it inside the function/method.

Regular import within function:

def get_copy(obj):
    from copy import deepcopy
    return deepcopy(obj)

With a LazyImporter:

from ducktools.lazyimporter import LazyImporter, FromImport

laz = LazyImporter([FromImport("copy", "deepcopy")])

def get_copy(obj):
    return laz.deepcopy(obj)

While the LazyImporter is more verbose, it only invokes the import mechanism once when first accessed, while placing the import within the function invokes it every time the function is called. This can be a significant overhead if the function ends up used in a loop.

This also means that if the attribute is accessed anywhere it will be imported and in place wherever it is used.

Delaying the import of parts of a module's public API

Eager import:

from .submodule import useful_tool

__all__ = [..., "useful_tool"]

Lazy import:

from ducktools.lazyimporter import LazyImporter, FromImport, get_module_funcs

__all__ = [..., "useful_tool"]

laz = LazyImporter(
    [FromImport(".submodule", "useful_tool")],
    globs=globals(),  # globals() is used for relative imports, LazyImporter will attempt to infer it if not provided
)
__getattr__, __dir__ = get_module_funcs(laz, __name__)  # __name__ will also be inferred if not given

The import classes

In all of these instances modules is intended as the first argument to LazyImporter and all attributes would be accessed from the LazyImporter instance and not in the global namespace.

eg:

from ducktools.lazyimporter import LazyImporter, ModuleImport

modules = [ModuleImport("functools")]
laz = LazyImporter(modules)
laz.functools  # provides access to the module "functools"

ModuleImport

ModuleImport is used for your basic module style imports.

from ducktools.lazyimporter import ModuleImport

modules = [
    ModuleImport("module"),
    ModuleImport("other_module", "other_name"),
    ModuleImport("base_module.submodule", asname="short_name"),
]

is equivalent to

import module
import other_module as other_name
import base_module.submodule as short_name

when provided to a LazyImporter.

FromImport and MultiFromImport

FromImport is used for standard 'from' imports, MultiFromImport for importing multiple items from the same module. By using a MultiFromImport, when the first attribute is accessed, all will be assigned on the LazyImporter.

from ducktools.lazyimporter import FromImport, MultiFromImport

modules = [
    FromImport("dataclasses", "dataclass"),
    FromImport("functools", "partial", "partfunc"),
    MultiFromImport("collections", ["namedtuple", ("defaultdict", "dd")]),
]

is equivalent to

from dataclasses import dataclass
from functools import partial as partfunc
from collections import namedtuple, defaultdict as dd

when provided to a LazyImporter.

TryExceptImport, TryExceptFromImport and TryFallbackImport

TryExceptImport is used for compatibility where a module may not be available and so a fallback module providing the same functionality should be used. For example when a newer version of python has a stdlib module that has replaced a third party module that was used previously.

from ducktools.lazyimporter import TryExceptImport, TryExceptFromImport, TryFallbackImport

modules = [
    TryExceptImport("tomllib", "tomli", "tomllib"),
    TryExceptFromImport("tomllib", "loads", "tomli", "loads", "loads"),
    TryFallbackImport("tomli", None),
]

is roughly equivalent to

try:
    import tomllib as tomllib
except ImportError:
    import tomli as tomllib

try:
    from tomllib import loads as loads
except ImportError:
    from tomli import loads as loads

try:
    import tomli
except ImportError:
    tomli = None

when provided to a LazyImporter.

Experimental import statement capture

There is an experimental mode that can capture import statements within a context block.

This is currently in a separate 'capture' submodule but may be merged (or lazily imported itself) in the future.

from ducktools.lazyimporter import LazyImporter, get_importer_state
from ducktools.lazyimporter.capture import capture_imports

laz = LazyImporter()

with capture_imports(laz, auto_export=True):
    # Inside this block, imports are captured and converted to lazy imports on laz
    import functools
    from collections import namedtuple as nt

print(get_importer_state(laz))

# Note that the captured imports are *not* available in the module namespace
try:
    functools
except NameError:
    print("functools is not here")

Imports are placed on the lazy importer object as with the explicit syntax. Unlike the regular syntax, these imports are exported by default.

This works by replacing and restoring the builtin __import__ function that is called by the import statement while in the block.

Context Manager Caveats

• This only supports Module imports and From imports
  • The actual statement executes immediately and returns a placeholder, so a try/except can't work.
• Imports triggered inside functions or classes while within the block will still occur eagerly
• Imports triggered in other modules while within the block will still occur eagerly
• The context manager must be used at the module level
  • It will error if you use it inside a class or function scope
• As with the ModuleImport class, submodule imports without an assigned name are not supported.
• If other modules are also replacing __import__ simultaneously this will probably fail.
  • In a library you may not be able to guarantee this.
  • Hopefully this will be resolvable.

Environment Variables

There are two environment variables that can be used to modify the behaviour for debugging purposes.

If DUCKTOOLS_EAGER_PROCESS is set to any value other than 'False' (case insensitive) the initial processing of imports will be done on instance creation.

Similarly if DUCKTOOLS_EAGER_IMPORT is set to any value other than 'False' all imports will be performed eagerly on instance creation (this will also force processing on import).

If they are unset this is equivalent to being set to False.

If there is a lazy importer where it is known this will not work (for instance if it is managing a circular dependency issue) these can be overridden for an importer by passing values to eager_process and/or eager_import arguments to the LazyImporter constructer as keyword arguments.

How does it work

The following lazy importer:

from ducktools.lazyimporter import LazyImporter, FromImport

laz = LazyImporter([FromImport("functools", "partial")])

Generates an object that's roughly equivalent to this:

class SpecificLazyImporter:
    def __getattr__(self, name):
        if name == "partial":
            from functools import partial
            setattr(self, name, partial)
            return partial
        
        raise AttributeError(...)

laz = SpecificLazyImporter()

The first time the attribute is accessed the import is done and the output is stored on the instance, so repeated access immediately gets the desired object and the import mechanism is only invoked once.

(The actual __getattr__ function uses a dictionary lookup and delegates importing to the FromImport class. Names are all dynamic and imports are done through the __import__ function.)