Author:

Pablo Galindo Salgado , Germán Méndez Bravo <german.mb at gmail.com>, Thomas Wouters , Dino Viehland , Brittany Reynoso , Noah Kim , Tim Stumbaugh

Discussions-To:

Discourse thread

Status:

Accepted

Type:

Standards Track

Created:

02-Oct-2025

Python-Version:

3.15

Post-History:

03-Oct-2025

Resolution:

03-Nov-2025

Table of Contents

• Abstract
• Motivation
• Rationale
  • Other design decisions
• Specification
  • Grammar
    • Syntax restrictions
  • Semantics
    • Note on .pth files
  • Lazy objects
  • Lazy import mechanism
  • Reification
• Reference Implementation
  • Bytecode and adaptive specialization
  • Lazy imports filter
  • Global lazy imports control
• Backwards Compatibility
  • Unchanged semantics
  • Observable behavioral shifts (opt-in only)
  • Thread-safety and reification
  • Performance
    • Runtime performance
    • Filter function performance
    • Startup time improvements
  • Typing and tools
• Security Implications
• How to Teach This
• FAQ
  • How does this differ from the rejected PEP 690?
  • What changes at reification time? What stays the same?
  • What happens when lazy imports encounter errors?
  • How do lazy imports affect modules with import-time side effects?
  • Can I use lazy imports with from ... import ... statements?
  • Does lazy from module import Class load the entire module or just the class?
  • What about type annotations and TYPE_CHECKING imports?
  • What’s the performance overhead of lazy imports?
  • Can I mix lazy and eager imports of the same module?
  • How do I migrate existing code to use lazy imports?
  • What about star imports (from module import *)?
  • How do lazy imports interact with import hooks and custom loaders?
  • What happens in multi-threaded environments?
  • Can I force reification of a lazy import without using it?
  • Why not use importlib.util.LazyLoader instead?
  • Will this break tools like isort or black?
  • How do I know if a library is compatible with lazy imports?
  • What happens if I globally enable lazy imports mode and a library doesn’t work correctly?
  • Can I use lazy imports inside functions?
  • What about forwards compatibility with older Python versions?
  • How do explicit lazy imports interact with PEP 649 and PEP 749?
  • How do lazy imports interact with dir(), getattr(), and module introspection?
  • Do lazy imports work with circular imports?
  • Will lazy imports affect the performance of my hot paths?
  • What about sys.modules? When does a lazy import appear there?
  • Does lazy from __future__ import feature work?
  • Why did you choose lazy as the keyword name?
• Deferred Ideas
  • Alternative syntax and ergonomic improvements
  • Automatic lazy imports for if TYPE_CHECKING blocks
  • Module-level lazy import mode
  • Package metadata for lazy-safe declarations
  • C API for lazy imports
• Alternate Implementation Ideas
  • Leveraging a subclass of dict
  • Alternate keyword names
• Rejected Ideas
  • Making the new behavior the default
  • Supporting __lazy_modules__ = ["*"] as built-in syntax
  • Disallowing lazy imports inside with blocks
  • Forcing eager imports in with blocks under the global flag
  • Modification of the dict object
  • Transforming lazy objects via __class__ mutation
  • Making lazy imports find the module without loading it
  • Placing the lazy keyword in the middle of from imports
  • Placing the lazy keyword at the end of import statements
  • Adding an explicit eager keyword
  • Allowing the filter to force lazy imports even when globally disabled
  • Using underscore-prefixed names for advanced features
  • Using a decorator syntax for lazy imports
  • Using a context manager instead of a new soft keyword
  • Returning a proxy dict from globals()
  • Automatically reifying on __dict__ or globals() access
• Acknowledgements
• Footnotes
• Copyright

Abstract

This PEP introduces syntax for lazy imports as an explicit language feature:

lazy import json lazy from json import dumps

Lazy imports defer the loading and execution of a module until the first time the imported name is used, in contrast to ‘normal’ imports, which eagerly load and execute a module at the point of the import statement.

By allowing developers to mark individual imports as lazy with explicit syntax, Python programs can reduce startup time, memory usage, and unnecessary work. This is particularly beneficial for command-line tools, test suites, and applications with large dependency graphs.

This proposal preserves full backwards compatibility: normal import statements remain unchanged, and lazy imports are enabled only where explicitly requested.

Motivation

The dominant convention in Python code is to place all imports at the module level, typically at the beginning of the file. This avoids repetition, makes import dependencies clear and minimizes runtime overhead by only evaluating an import statement once per module.

A major drawback with this approach is that importing the first module for an execution of Python (the “main” module) often triggers an immediate cascade of imports, and optimistically loads many dependencies that may never be used. The effect is especially costly for command-line tools with multiple subcommands, where even running the command with --help can load dozens of unnecessary modules and take several seconds. This basic example demonstrates what must be loaded just to get helpful feedback to the user on how to run the program at all. Inefficiently, the user incurs this overhead again when they figure out the command they want and invoke the program “for real.”

A somewhat common way to delay imports is to move the imports into functions (inline imports), but this practice requires more work to implement and maintain, and can be subverted by a single inadvertent top-level import. Additionally, it obfuscates the full set of dependencies for a module. Analysis of the Python standard library shows that approximately 17% of all imports outside tests (nearly 3500 total imports across 730 files) are already placed inside functions or methods specifically to defer their execution. This demonstrates that developers are already manually implementing lazy imports in performance-sensitive code, but doing so requires scattering imports throughout the codebase and makes the full dependency graph harder to understand at a glance.

The standard library provides the LazyLoader class to solve some of these inefficiency problems. It permits imports at the module level to work mostly like inline imports do. Many scientific Python libraries have adopted a similar pattern, formalized in SPEC 1. There’s also the third-party lazy_loader package, yet another implementation of lazy imports. Imports used solely for static type checking are another source of potentially unneeded imports, and there are similarly disparate approaches to minimizing the overhead. The various approaches used here to defer or remove eager imports do not cover all potential use-cases for a general lazy import mechanism. There is no clear standard, and there are several drawbacks including runtime overhead in unexpected places, or worse runtime introspection.

This proposal introduces syntax for lazy imports with a design that is local, explicit, controlled, and granular. Each of these qualities is essential to making the feature predictable and safe to use in practice.

The behavior is local: laziness applies only to the specific import marked with the lazy keyword, and it does not cascade recursively into other imports. This ensures that developers can reason about the effect of laziness by looking only at the line of code in front of them, without worrying about whether imported modules will themselves behave differently. A lazy import is an isolated decision each time it is used, not a global shift in semantics.

The semantics are explicit. When a name is imported lazily, the binding is created in the importing module immediately, but the target module is not loaded until the first time the name is accessed. After this point, the binding is indistinguishable from one created by a normal import. This clarity reduces surprises and makes the feature accessible to developers who may not be deeply familiar with Python’s import machinery.

Lazy imports are controlled, in the sense that lazy loading is only triggered by the importing code itself. In the general case, a library will only experience lazy imports if its own authors choose to mark them as such. This avoids shifting responsibility onto downstream users and prevents accidental surprises in library behavior. Since library authors typically manage their own import subgraphs, they retain predictable control over when and how laziness is applied.

The mechanism is also granular. It is introduced through explicit syntax on individual imports, rather than a global flag or implicit setting. This allows developers to adopt it incrementally, starting with the most performance-sensitive areas of a codebase. As this feature is introduced to the community, we want to make the experience of onboarding optional, progressive, and adaptable to the needs of each project.

Lazy imports provide several concrete advantages:

• Command-line tools are often invoked directly by a user, so latency – in particular startup latency – is quite noticeable. These programs are also typically short-lived processes (contrasted with, e.g., a web server). With lazy imports, only the code paths actually reached will import a module. This can reduce startup time by 50-70% in practice, providing a significant improvement to a common user experience and improving Python’s competitiveness in domains where fast startup matters most.
• Type annotations frequently require imports that are never used at runtime. The common workaround is to wrap them in if TYPE_CHECKING: blocks [1]. With lazy imports, annotation-only imports impose no runtime penalty, eliminating the need for such guards and making annotated codebases cleaner.
• Large applications often import thousands of modules, and each module creates function and type objects, incurring memory costs. In long-lived processes, this noticeably raises baseline memory usage. Lazy imports defer these costs until a module is needed, keeping unused subsystems unloaded. Memory savings of 30-40% have been observed in real workloads.

Rationale

The design of this proposal is centered on clarity, predictability, and ease of adoption. Each decision was made to ensure that lazy imports provide tangible benefits without introducing unnecessary complexity into the language or its runtime.

It is also worth noting that while this PEP outlines one specific approach, we list alternate implementation strategies for some of the core aspects and semantics of the proposal. If the community expresses a strong preference for a different technical path that still preserves the same core semantics or there is fundamental disagreement over the specific option, we have included the brainstorming we have already completed in preparation for this proposal as reference.

The choice to introduce a new lazy keyword reflects the need for explicit syntax. Lazy imports have different semantics from normal imports: errors and side effects occur at first use rather than at the import statement. This semantic difference makes it critical that laziness is visible at the import site itself, not hidden in global configuration or distant module-level declarations. The lazy keyword provides local reasoning about import behavior, avoiding the need to search elsewhere in the code to understand whether an import is deferred. The rest of the import semantics remain unchanged: the same import machinery, module finding, and loading mechanisms are used.

Another important decision is to represent lazy imports with proxy objects in the module’s namespace, rather than by modifying dictionary lookup. Earlier approaches experimented with embedding laziness into dictionaries, but this blurred abstractions and risked affecting unrelated parts of the runtime. The dictionary is a fundamental data structure in Python – literally every object is built on top of dicts – and adding hooks to dictionaries would prevent critical optimizations and complicate the entire runtime. The proxy approach is simpler: it behaves like a placeholder until first use, at which point it resolves the import and rebinds the name. From then on, the binding is indistinguishable from a normal import. This makes the mechanism easy to explain and keeps the rest of the interpreter unchanged.

Compatibility for library authors was also a key concern. Many maintainers need a migration path that allows them to support both new and old versions of Python at once. For this reason, the proposal includes the __lazy_modules__ global as a transitional mechanism. A module can declare which imports should be treated as lazy (by listing the module names as strings), and on Python 3.15 or later those imports will become lazy automatically, as if they were imported with the lazy keyword. On earlier versions the declaration is ignored, leaving imports eager. This gives authors a practical bridge until they can rely on the keyword as the canonical syntax.

Finally, the feature is designed to be adopted incrementally. Nothing changes unless a developer explicitly opts in, and adoption can begin with just a few imports in performance-sensitive areas. This mirrors the experience of gradual typing in Python: a mechanism that can be introduced progressively, without forcing projects to commit globally from day one. Notably, the adoption can also be done from the “outside in”, permitting CLI authors to introduce lazy imports and speed up user-facing tools, without requiring changes to every library the tool might use.

Other design decisions

• The scope of laziness is deliberately local and non-recursive. A lazy import only affects the specific statement where it appears; it does not cascade into other modules or submodules. This choice is crucial for predictability. When developers read code, they can reason about import behavior line by line, without worrying about hidden laziness deeper in the dependency graph. The result is a feature that is powerful but still easy to understand in context.
• In addition, it is useful to provide a mechanism to activate or deactivate lazy imports for all code running in the interpreter (referred to in this PEP as the ‘global lazy imports flag’). While the primary design centers the explicit lazy import syntax, there are scenarios – such as large applications, testing environments, or frameworks – where enabling laziness consistently across many modules provides the most benefit. A global switch makes it easy to experiment with or enforce consistent behavior, while still working in combination with the filtering API to respect exclusions or tool-specific configuration. This ensures that global adoption can be practical without reducing flexibility or control.

Specification

Grammar

A new soft keyword lazy is added. A soft keyword is a context-sensitive keyword that only has special meaning in specific grammatical contexts; elsewhere it can be used as a regular identifier (e.g., as a variable name). The lazy keyword only has special meaning when it appears before import statements:

import_name: | 'lazy'? 'import' dotted_as_names

import_from: | 'lazy'? 'from' ('.' | '...')* dotted_name 'import' import_from_targets | 'lazy'? 'from' ('.' | '...')+ 'import' import_from_targets

Syntax restrictions

The soft keyword is only allowed at the global (module) level, not inside functions, class bodies, try blocks, or import *. Import statements that use the soft keyword are potentially lazy. Imports that can’t be lazy are unaffected by the global lazy imports flag, and instead are always eager. Additionally, from __future__ import statements cannot be lazy.

Examples of syntax errors:

# SyntaxError: lazy import not allowed inside functions def foo(): lazy import json

# SyntaxError: lazy import not allowed inside classes class Bar: lazy import json

# SyntaxError: lazy import not allowed inside try/except blocks try: lazy import json except ImportError: pass

# SyntaxError: lazy from ... import * is not allowed lazy from json import *

# SyntaxError: lazy from __future__ import is not allowed lazy from __future__ import annotations

Semantics

When the lazy keyword is used, the import becomes potentially lazy (see Lazy imports filter for advanced override mechanisms). The module is not loaded immediately at the import statement; instead, a lazy proxy object is created and bound to the name. The actual module is loaded on first use of that name.

When using lazy from ... import, each imported name is bound to a lazy proxy object. The first access to any of these names triggers loading of the entire module and reifies only that specific name to its actual value. Other names remain as lazy proxies until they are accessed. The interpreter’s adaptive specialization will optimize away the lazy checks after a few accesses.

Example with lazy import:

import sys

lazy import json

print('json' in sys.modules) # False - module not loaded yet

# First use triggers loading result = json.dumps({"hello": "world"})

print('json' in sys.modules) # True - now loaded

Example with lazy from ... import:

import sys

lazy from json import dumps, loads

print('json' in sys.modules) # False - module not loaded yet

# First use of 'dumps' triggers loading json and reifies ONLY 'dumps' result = dumps({"hello": "world"})

print('json' in sys.modules) # True - module now loaded

# Accessing 'loads' now reifies it (json already loaded, no re-import) data = loads(result)

A module may define a __lazy_modules__ variable in its global scope, which specifies which module names should be made potentially lazy (as if the lazy keyword was used). This variable is checked on each import statement to determine whether the import should be made potentially lazy. The check is performed by calling __contains__ on the __lazy_modules__ object with a string containing the fully qualified module name being imported. Typically, __lazy_modules__ is a set of fully qualified module name strings. When a module is made lazy this way, from-imports using that module are also lazy, but not necessarily imports of sub-modules.

The normal (non-lazy) import statement will check the global lazy imports flag. If it is “all”, all imports are potentially lazy (except for imports that can’t be lazy, as mentioned above.)

Example:

__lazy_modules__ = ["json"] import json print('json' in sys.modules) # False result = json.dumps({"hello": "world"}) print('json' in sys.modules) # True

If the global lazy imports flag is set to “none”, no potentially lazy import is ever imported lazily, and the behavior is equivalent to a regular import statement: the import is eager (as if the lazy keyword was not used).

Finally, the application may use a custom filter function on all potentially lazy imports to determine if they should be lazy or not (this is an advanced feature, see Lazy imports filter). If a filter function is set, it will be called with the name of the module doing the import, the name of the module being imported, and (if applicable) the fromlist. An import remains lazy only if the filter function returns True. If no lazy import filter is set, all potentially lazy imports are lazy.

Note on .pth files

The lazy import mechanism does not apply to .pth files processed by the site module. While .pth files have special handling for lines that begin with import followed by a space or tab, this special handling will not be adapted to support lazy imports. Imports specified in .pth files remain eager as they always have been.

Lazy objects

Lazy modules, as well as names lazy imported from modules, are represented by types.LazyImportType instances, which are resolved to the real object (reified) before they can be used. This reification is usually done automatically (see below), but can also be done by calling the lazy object’s resolve method.

Lazy import mechanism

When an import is lazy, __lazy_import__ is called instead of __import__. __lazy_import__ has the same function signature as __import__. It adds the module name to sys.lazy_modules, a set of fully-qualified module names which have been lazily imported at some point (primarily for diagnostics and introspection), and returns a types.LazyImportType object for the module.

The implementation of from ... import (the IMPORT_FROM bytecode implementation) checks if the module it’s fetching from is a lazy module object, and if so, returns a types.LazyImportType for each name instead.

The end result of this process is that lazy imports (regardless of how they are enabled) result in lazy objects being assigned to global variables.

Lazy module objects do not appear in sys.modules, they’re just listed in the sys.lazy_modules set. Under normal operation lazy objects should only end up stored in global variables, and the common ways to access those variables (regular variable access, module attributes) will resolve lazy imports (reify) and replace them when they’re accessed.

It is still possible to expose lazy objects through other means, like debuggers. This is not considered a problem.

Reification

When a lazy object is used, it needs to be reified. This means resolving the import at that point in the program and replacing the lazy object with the concrete one. Reification imports the module at that point in the program. Notably, reification still calls __import__ to resolve the import, which uses the state of the import system (e.g. sys.path, sys.meta_path, sys.path_hooks and __import__) at reification time, not the state when the lazy import statement was evaluated.

When the module is reified, it’s removed from sys.lazy_modules (even if there are still other unreified lazy references to it). When a package is reified and submodules in the package were also previously lazily imported, those submodules are not automatically reified but they are added to the reified package’s globals (unless the package already assigned something else to the name of the submodule).

If reification fails (e.g., due to an ImportError), the lazy object is not reified or replaced. Subsequent uses of the lazy object will re-try the reification. Exceptions that happen during reification are raised as normal, but the exception is enhanced with chaining to show both where the lazy import was defined and where it was accessed (even though it propagates from the code that triggered reification). This provides clear debugging information:

# app.py - has a typo in the import lazy from json import dumsp # Typo: should be 'dumps'

print("App started successfully") print("Processing data...")

# Error occurs here on first use result = dumsp({"key": "value"})

The traceback shows both locations:

App started successfully Processing data... Traceback (most recent call last): File "app.py", line 2, in lazy from json import dumsp ImportError: lazy import of 'json.dumsp' raised an exception during resolution

The above exception was the direct cause of the following exception:

Traceback (most recent call last): File "app.py", line 8, in result = dumsp({"key": "value"}) ^^^^^ ImportError: cannot import name 'dumsp' from 'json'. Did you mean: 'dump'?

This exception chaining clearly shows:

1. where the lazy import was defined,
2. that the module was not eagerly imported, and
3. where the actual access happened that triggered the error.

Reification does not automatically occur when a module that was previously lazily imported is subsequently eagerly imported. Reification does not immediately resolve all lazy objects (e.g. lazy from statements) that referenced the module. It only resolves the lazy object being accessed.

Accessing a lazy object (from a global variable or a module attribute) reifies the object.

However, calling globals() or accessing a module’s __dict__ does not trigger reification – they return the module’s dictionary, and accessing lazy objects through that dictionary still returns lazy proxy objects that need to be manually reified upon use. A lazy object can be resolved explicitly by calling the resolve method. Calling dir() at the global scope will not reify the globals, nor will calling dir(mod) (through special-casing in mod.__dir__.) Other, more indirect ways of accessing arbitrary globals (e.g. inspecting frame.f_globals) also do not reify all the objects.

Example using globals() and __dict__:

# my_module.py import sys lazy import json

# Calling globals() does NOT trigger reification g = globals() print('json' in sys.modules) # False - still lazy print(type(g['json'])) # <class 'LazyImport'>

# Accessing __dict__ also does NOT trigger reification d = __dict__ print(type(d['json'])) # <class 'LazyImport'>

# Explicitly reify using the resolve() method resolved = g['json'].resolve() print(type(resolved)) # <class 'module'> print('json' in sys.modules) # True - now loaded

Reference Implementation

A reference implementation is available at: https://github.com/LazyImportsCabal/cpython/tree/lazy

A demo is available (not necessarily synced with the latest PEP) for evaluation purposes at: https://lazy-import-demo.pages.dev/

Bytecode and adaptive specialization

Lazy imports are implemented through modifications to four bytecode instructions: IMPORT_NAME, IMPORT_FROM, LOAD_GLOBAL, and LOAD_NAME.

The lazy syntax sets a flag in the IMPORT_NAME instruction’s oparg (oparg & 0x01). The interpreter checks this flag and calls _PyEval_LazyImportName() instead of _PyEval_ImportName(), creating a lazy import object rather than executing the import immediately. The IMPORT_FROM instruction checks whether its source is a lazy import (PyLazyImport_CheckExact()) and creates a lazy object for the attribute rather than accessing it immediately.

When a lazy object is accessed, it must be reified. The LOAD_GLOBAL instruction (used in function scopes) and LOAD_NAME instruction (used at module and class level) both check whether the object being loaded is a lazy import. If so, they call _PyImport_LoadLazyImportTstate() to perform the actual import and store the module in sys.modules.

This check incurs a very small cost on each access. However, Python’s adaptive interpreter can specialize LOAD_GLOBAL after observing that a lazy import has been reified. After several executions, LOAD_GLOBAL becomes LOAD_GLOBAL_MODULE, which accesses the module dictionary directly without checking for lazy imports.

Examples of the bytecode generated:

lazy import json # IMPORT_NAME with flag set

Generates:

IMPORT_NAME 1 (json + lazy)

lazy from json import dumps # IMPORT_NAME + IMPORT_FROM

Generates:

IMPORT_NAME 1 (json + lazy) IMPORT_FROM 1 (dumps)

lazy import json x = json # Module-level access

Generates:

LOAD_NAME 0 (json)

lazy import json

def use_json(): return json.dumps({}) # Function scope

Before any calls:

LOAD_GLOBAL 0 (json) LOAD_ATTR 2 (dumps)

After several calls, LOAD_GLOBAL specializes to LOAD_GLOBAL_MODULE:

LOAD_GLOBAL_MODULE 0 (json) LOAD_ATTR_MODULE 2 (dumps)

Lazy imports filter

Note: This is an advanced feature. These are intended for specialized/advanced users who need fine-grained control over lazy import behavior when using the global flags. Library developers are discouraged from using these functions as they can affect the runtime execution of applications (similar to ``sys.setrecursionlimit()``, ``sys.setswitchinterval()``, or ``gc.set_threshold()``).

This PEP adds the following new functions to the sys module to manage the lazy imports filter:

• sys.set_lazy_imports_filter(func) - Sets the filter function. If func=None then the import filter is removed. The func parameter must have the signature: func(importer: str, name: str, fromlist: tuple[str, ...] | None) -> bool
• sys.get_lazy_imports_filter() - Returns the currently installed filter function, or None if no filter is set.
• sys.set_lazy_imports(mode, /) - Programmatic API for controlling lazy imports at runtime. The mode parameter can be "normal" (respect lazy keyword only), "all" (force all imports to be potentially lazy), or "none" (force all imports to be eager).
• sys.get_lazy_imports() - Returns the current lazy imports mode as a string: "normal", "all", or "none".

The filter function is called for every potentially lazy import, and must return True if the import should be lazy. This allows for fine-grained control over which imports should be lazy, useful for excluding modules with known side-effect dependencies or registration patterns. The filter function is called at the point of execution of the lazy import or lazy from import statement, not at the point of reification. The filter function may be called concurrently.

The filter mechanism serves as a foundation that tools, debuggers, linters, and other ecosystem utilities can leverage to provide better lazy import experiences. For example, static analysis tools could detect modules with side effects and automatically configure appropriate filters. In the future (out of scope for this PEP), this foundation may enable better ways to declaratively specify which modules are safe for lazy importing, such as package metadata, type stubs with lazy-safety annotations, or configuration files. The current filter API is designed to be flexible enough to accommodate such future enhancements without requiring changes to the core language specification.

Example:

import sys

def exclude_side_effect_modules(importer, name, fromlist): """ Filter function to exclude modules with import-time side effects.

Args:
    importer: Name of the module doing the import
    name: Name of the module being imported
    fromlist: Tuple of names being imported (for 'from' imports), or None

Returns:
    True to allow lazy import, False to force eager import
"""
\# Modules known to have important import-time side effects
side\_effect\_modules \= {'legacy\_plugin\_system', 'metrics\_collector'}

if name in side\_effect\_modules:
    return False  \# Force eager import

return True  \# Allow lazy import

# Install the filter sys.set_lazy_imports_filter(exclude_side_effect_modules)

# These imports are checked by the filter lazy import data_processor # Filter returns True -> stays lazy lazy import legacy_plugin_system # Filter returns False -> imported eagerly

print('data_processor' in sys.modules) # False - still lazy print('legacy_plugin_system' in sys.modules) # True - loaded eagerly

# First use of data_processor triggers loading result = data_processor.transform(data) print('data_processor' in sys.modules) # True - now loaded

Global lazy imports control

Note: This is an advanced feature. This is intended for application developers and framework authors who need to control lazy imports across their entire application. Library developers are discouraged from using the global activation mechanism as it can affect the runtime execution of applications (similar to ``sys.setrecursionlimit()``, ``sys.setswitchinterval()``, or ``gc.set_threshold()``).

The global lazy imports flag can be controlled through:

• The -X lazy_imports=<mode> command-line option
• The PYTHON_LAZY_IMPORTS=<mode> environment variable
• The sys.set_lazy_imports(mode) function (primarily for testing)

The precedence order for setting the lazy imports mode follows the standard Python pattern: sys.set_lazy_imports() takes highest precedence, followed by -X lazy_imports=<mode>, then PYTHON_LAZY_IMPORTS=<mode>. If none are specified, the mode defaults to "normal".

Where <mode> can be:

• "normal" (or unset): Only explicitly marked lazy imports are lazy
• "all": All module-level imports (except in try blocks and import *) become potentially lazy
• "none": No imports are lazy, even those explicitly marked with lazy keyword

When the global flag is set to "all", all imports at the global level of all modules are potentially lazy except for those inside a try block or any wild card (from ... import *) import.

If the global lazy imports flag is set to "none", no potentially lazy import is ever imported lazily, the import filter is never called, and the behavior is equivalent to a regular import statement: the import is eager (as if the lazy keyword was not used).

Python code can run the sys.set_lazy_imports() function to override the state of the global lazy imports flag inherited from the environment or CLI. This is especially useful if an application needs to ensure that all imports are evaluated eagerly, via sys.set_lazy_imports("none").

Backwards Compatibility

Lazy imports are opt-in. Existing programs continue to run unchanged unless a project explicitly enables laziness (via lazy syntax, __lazy_modules__, or an interpreter-wide switch).

Unchanged semantics

• Regular import and from ... import ... statements remain eager unless explicitly made potentially lazy by the local or global mechanisms provided.
• Dynamic import APIs remain eager and unchanged: __import__() and importlib.import_module().
• Import hooks and loaders continue to run under the standard import protocol when a lazy object is reified.

Observable behavioral shifts (opt-in only)

These changes are limited to bindings explicitly made lazy:

• Error timing. Exceptions that would have occurred during an eager import (for example ImportError or AttributeError for a missing member) now occur at the use of the lazy name.

  # With eager import - error at import statement import broken_module # ImportError raised here

  # With lazy import - error deferred lazy import broken_module print("Import succeeded") broken_module.foo() # ImportError raised here on use

• Side-effect timing. Import-time side effects in lazily imported modules occur at first use of the binding, not at module import time.

• Import order. Because modules are imported on first use, the order in which modules are imported may differ from how they appear in code.

• Presence in ``sys.modules``. A lazily imported module does not appear in sys.modules until first use. After reification, it must appear in sys.modules. If some other code eagerly imports the same module before first use, the lazy binding resolves to that existing (lazy) module object when it is first used.

• Proxy visibility. Before first use, the bound name refers to a lazy proxy. Indirect introspection that touches the value may observe a proxy lazy object representation. After first use (provided the module was imported successfully), the name is rebound to the real object and becomes indistinguishable from an eager import.

Thread-safety and reification

Reification follows the existing import-lock discipline. Exactly one thread performs the import and atomically rebinds the importing module’s global to the resolved object. Concurrent readers thereafter observe the real object.

Lazy imports are thread-safe and have no special considerations for free-threading. A module that would normally be imported in the main thread may be imported in a different thread if that thread triggers the first access to the lazy import. This is not a problem: the import lock ensures thread safety regardless of which thread performs the import.

Subinterpreters are supported. Each subinterpreter maintains its own sys.lazy_modules and import state, so lazy imports in one subinterpreter do not affect others.

Performance

Lazy imports have no measurable performance overhead. The implementation is designed to be performance-neutral for both code that uses lazy imports and code that doesn’t.

Runtime performance

After reification (provided the import was successful), lazy imports have zero overhead. The adaptive interpreter specializes the bytecode (typically after 2-3 accesses), eliminating any checks. For example, LOAD_GLOBAL becomes LOAD_GLOBAL_MODULE, which directly accesses the module identically to normal imports.

The pyperformance suite confirms the implementation is performance-neutral.

Filter function performance

The filter function (set via sys.set_lazy_imports_filter()) is called for every potentially lazy import to determine whether it should actually be lazy. When no filter is set, this is simply a NULL check (testing whether a filter function has been registered), which is a highly predictable branch that adds essentially no overhead. When a filter is installed, it is called for each potentially lazy import, but this still has almost no measurable performance cost. To measure this, we benchmarked importing all 278 top-level importable modules from the Python standard library (which transitively loads 392 total modules including all submodules and dependencies), then forced reification of every loaded module to ensure everything was fully materialized.

Note that these measurements establish the baseline overhead of the filter mechanism itself. Of course, any user-defined filter function that performs additional work beyond a trivial check will add overhead proportional to the complexity of that work. However, we expect that in practice this overhead will be dwarfed by the performance benefits gained from avoiding unnecessary imports. The benchmarks below measure the minimal cost of the filter dispatch mechanism when the filter function does essentially nothing.

We compared four different configurations:

The four configurations:

1. Eager imports (baseline): Normal Python imports with no lazy machinery. Standard Python behavior.
2. Lazy + filter forcing eager: Filter function returns False for all imports, forcing eager execution, then all imports are reified at script end. Measures pure filter calling overhead since every import goes through the filter but executes eagerly.
3. Lazy + filter allowing lazy + reification: Filter function returns True for all imports, allowing lazy execution. All imports are reified at script end. Measures filter overhead when imports are actually lazy.
4. Lazy + no filter + reification: No filter installed, imports are lazy and reified at script end. Baseline for lazy behavior without filter.

The benchmarks used hyperfine, testing 278 standard library modules. Each ran in a fresh Python process. All configurations force the import of exactly the same set of modules (all modules loaded by the eager baseline) to ensure a fair comparison.

The benchmark environment used CPU isolation with 32 logical CPUs (0-15 at 3200 MHz, 16-31 at 2400 MHz), the performance scaling governor, Turbo Boost disabled, and full ASLR randomization. The overhead error bars are computed using standard error propagation for the formula (value - baseline) / baseline, accounting for uncertainties in both the measured value and the baseline.

Startup time improvements

The primary performance benefit of lazy imports is reduced startup time by loading only the modules actually used at runtime, rather than optimistically loading entire dependency trees at startup.

Real-world deployments at scale have demonstrated that the benefits can be massive, though of course this depends on the specific codebase and usage patterns. Organizations with large, interconnected codebases have reported substantial reductions in server reload times, ML training initialization, command-line tool startup, and Jupyter notebook loading. Memory usage improvements have also been observed as unused modules remain unloaded.

For detailed case studies and performance data from production deployments, see:

• Python Lazy Imports With Cinder (Meta Instagram Server)
• Lazy is the new fast: How Lazy Imports and Cinder accelerate machine learning at Meta (Meta ML Workloads)
• Inside HRT’s Python Fork (Hudson River Trading)
• Create an On-Demand Initializer for PySide (Qt for Python/PySide) - Christian Tismer’s implementation of lazy initialization for PySide6 based on ideas from PEP 690, showing 10-20% startup time improvement for PySide applications. This demonstrates the particular value of lazy imports for frameworks with extensive initialization at import time.

The benefits scale with codebase complexity: the larger and more interconnected the codebase, the more dramatic the improvements. The PySide implementation particularly highlights how frameworks with heavy initialization overhead can benefit significantly from opt-in lazy loading.

Typing and tools

Type checkers and static analyzers may treat lazy imports as ordinary imports for name resolution. At runtime, annotation-only imports can be marked lazy to avoid startup overhead. IDEs and debuggers should be prepared to display lazy proxies before first use and the real objects thereafter.

Security Implications

Tools that install packages while performing imports from that the same environment should ensure all modules are imported eagerly, or reified, before the installation step, to avoid newly installed distributions from shadowing them.

Such tools can use sys.set_lazy_imports() with "none" to force eager evaluation, or provide a sys.set_lazy_imports_filter() function for fine-grained control.

How to Teach This

The new lazy keyword will be documented as part of the language standard.

As this feature is opt-in, new Python users should be able to continue using the language as they are used to. For experienced developers, we expect them to leverage lazy imports for the variety of benefits listed above (decreased latency, decreased memory usage, etc) on a case-by-case basis. Developers interested in the performance of their Python binary will likely leverage profiling to understand the import time overhead in their codebase and mark the necessary imports as lazy. In addition, developers can mark imports that will only be used for type annotations as lazy.

Additional documentation will be added to the Python documentation, including guidance, a dedicated how-to guide, and updates to the import system documentation covering: identifying slow-loading modules with profiling tools (such as -X importtime), migration strategies for existing codebases, best practices for avoiding common pitfalls with import-time side effects, and patterns for using lazy imports effectively with type annotations and circular imports.

Below is guidance on how to best take advantage of lazy imports and how to avoid incompatibilities:

• When adopting lazy imports, users should be aware that eliding an import until it is used will result in side effects not being executed. In turn, users should be wary of modules that rely on import time side effects. Perhaps the most common reliance on import side effects is the registry pattern, where population of some external registry happens implicitly during the importing of modules, often via decorators but sometimes implemented via metaclasses or __init_subclass__. Instead, registries of objects should be constructed via explicit discovery processes (e.g. a well-known function to call).

  # Problematic: Plugin registers itself on import # my_plugin.py from plugin_registry import register_plugin

  @register_plugin("MyPlugin") class MyPlugin: pass # In main code: lazy import my_plugin # Plugin NOT registered yet - module not loaded!

  # Better: Explicit discovery # plugin_registry.py def discover_plugins(): from my_plugin import MyPlugin register_plugin(MyPlugin) # In main code: plugin_registry.discover_plugins() # Explicit loading

• Always import needed submodules explicitly. It is not enough to rely on a different import to ensure a module has its submodules as attributes. Plainly, unless there is an explicit from . import bar in foo/__init__.py, always use import foo.bar; foo.bar.Baz, not import foo; foo.bar.Baz. The latter only works (unreliably) because the attribute foo.bar is added as a side effect of foo.bar being imported somewhere else.

• Users who are moving imports into functions to improve startup time, should instead consider keeping them where they are but adding the lazy keyword. This allows them to keep dependencies clear and avoid the overhead of repeatedly re-resolving the import but will still speed up the program.

  # Before: Inline import (repeated overhead) def process_data(data): import json # Re-resolved on every call return json.dumps(data) # After: Lazy import at module level lazy import json

  def process_data(data): return json.dumps(data) # Loaded once on first call

• Avoid using wild card (star) imports, as those are always eager.

FAQ

How does this differ from the rejected PEP 690?

PEP 810 takes an explicit, opt-in approach instead of PEP 690’s implicit global approach. The key differences are:

• Explicit syntax: lazy import foo clearly marks which imports are lazy.
• Local scope: Laziness only affects the specific import statement, not cascading to dependencies.
• Simpler implementation: Uses proxy objects instead of modifying core dictionary behavior.

What changes at reification time? What stays the same?

What changes (the timing):

• When the module is imported - deferred to first use instead of at the import statement
• When import errors occur - at first use rather than at import time
• When module-level side effects execute - at first use rather than at import time

What stays the same (everything else):

• The import machinery used - same __import__, same hooks, same loaders
• The module object created - identical to an eagerly imported module
• Import state consulted - sys.path, sys.meta_path, etc. at reification time (not at import statement time)
• Module attributes and behavior - completely indistinguishable after reification
• Thread safety - same import lock discipline as normal imports

In other words: lazy imports only change when something happens, not what happens. After reification, a lazy-imported module is indistinguishable from an eagerly imported one.

What happens when lazy imports encounter errors?

Import errors (ImportError, ModuleNotFoundError, syntax errors) are deferred until first use of the lazy name. This is similar to moving an import into a function. The error will occur with a clear traceback pointing to the first access of the lazy object.

The implementation provides enhanced error reporting through exception chaining. When a lazy import fails during reification, the original exception is preserved and chained, showing both where the import was defined and where it was first used:

Traceback (most recent call last): File "test.py", line 1, in lazy import broken_module ImportError: lazy import of 'broken_module' raised an exception during resolution

The above exception was the direct cause of the following exception:

Traceback (most recent call last): File "test.py", line 3, in broken_module.foo() ^^^^^^^^^^^^^ File "broken_module.py", line 2, in 1/0 ZeroDivisionError: division by zero

Exceptions during reification prevent the replacement of the lazy object, and subsequent uses of the lazy object will retry the whole reification.

How do lazy imports affect modules with import-time side effects?

Side effects are deferred until first use. This is generally desirable for performance, but may require code changes for modules that rely on import-time registration patterns. We recommend:

• Use explicit initialization functions instead of import-time side effects
• Call initialization functions explicitly when needed
• Avoid relying on import order for side effects

Can I use lazy imports with from ... import ... statements?

Yes, as long as you don’t use from ... import *. Both lazy import foo and lazy from foo import bar are supported. The bar name will be bound to a lazy object that resolves to foo.bar on first use.

Does lazy from module import Class load the entire module or just the class?

It loads the entire module, not just the class. This is because Python’s import system always executes the complete module file – there’s no mechanism to execute only part of a .py file. When you first access Class, Python:

1. Loads and executes the entire module.py file
2. Extracts the Class attribute from the resulting module object
3. Binds Class to the name in your namespace

This is identical to eager from module import Class behavior. The only difference with lazy imports is that steps 1-3 happen on first use instead of at the import statement.

# heavy_module.py print("Loading heavy_module") # This ALWAYS runs when module loads

class MyClass: pass

class UnusedClass: pass # Also gets defined, even though we don't import it

# app.py lazy from heavy_module import MyClass

print("Import statement done") # heavy_module not loaded yet obj = MyClass() # NOW "Loading heavy_module" prints # (and UnusedClass gets defined too)

Key point: Lazy imports defer when a module loads, not what gets loaded. You cannot selectively load only parts of a module – Python’s import system doesn’t support partial module execution.

What about type annotations and TYPE_CHECKING imports?

Lazy imports eliminate the common need for TYPE_CHECKING guards. You can write:

lazy from collections.abc import Sequence, Mapping # No runtime cost

def process(items: Sequence[str]) -> Mapping[str, int]: ...

Instead of:

from typing import TYPE_CHECKING if TYPE_CHECKING: from collections.abc import Sequence, Mapping

def process(items: Sequence[str]) -> Mapping[str, int]: ...

What’s the performance overhead of lazy imports?

The overhead is minimal:

• Zero overhead after first use (provided the import doesn’t fail) thanks to the adaptive interpreter optimizing the slow path away.
• Small one-time cost to create the proxy object.
• Reification (first use) has the same cost as a regular import.
• No ongoing performance penalty.

Benchmarking with the pyperformance suite shows the implementation is performance neutral when lazy imports are not used.

Can I mix lazy and eager imports of the same module?

Yes. If module foo is imported both lazily and eagerly in the same program, the eager import takes precedence and both bindings resolve to the same module object.

How do I migrate existing code to use lazy imports?

Migration is incremental:

1. Identify slow-loading modules using profiling tools.
2. Add lazy keyword to imports that aren’t needed immediately.
3. Test that side-effect timing changes don’t break functionality.
4. Use __lazy_modules__ for compatibility with older Python versions.

What about star imports (from module import *)?

Wild card (star) imports cannot be lazy - they remain eager. This is because the set of names being imported cannot be determined without loading the module. Using the lazy keyword with star imports will be a syntax error. If lazy imports are globally enabled, star imports will still be eager.

How do lazy imports interact with import hooks and custom loaders?

Import hooks and loaders work normally. When a lazy object is used, the standard import protocol runs, including any custom hooks or loaders that were in place at reification time.

What happens in multi-threaded environments?

Lazy import reification is thread-safe. Only one thread will perform the actual import, and the binding is atomically updated. Other threads will see either the lazy proxy or the final resolved object.

Can I force reification of a lazy import without using it?

Yes, individual lazy objects can be resolved by calling their resolve() method.

Why not use importlib.util.LazyLoader instead?

The standard library’s LazyLoader was designed for specific use cases but has fundamental limitations as a general-purpose lazy import mechanism.

Most critically, LazyLoader does not support from ... import statements. There is no straightforward mechanism to lazily import specific attributes from a module - users would need to manually wrap and proxy individual attributes, which is both error-prone and defeats the performance benefits.

Additionally, LazyLoader must resolve the module spec before creating the lazy loader, which introduces overhead that reduces the performance benefits of lazy loading. The spec resolution involves filesystem operations and path searching that this PEP’s approach defers until actual module use.

LazyLoader also operates at the import machinery level rather than providing language-level syntax, which means there’s no canonical way for tools like linters and type checkers to recognize lazy imports. A dedicated syntax enables ecosystem-wide standardization and allows compiler and runtime optimizations that would be impossible with a purely library-based approach.

Finally, LazyLoader requires significant boilerplate, involving manual manipulation of module specs, loaders, and sys.modules, making it impractical for common use cases where multiple modules need to be lazily imported.

Will this break tools like isort or black?

Linters, formatters, and other tools will need updates to recognize the lazy keyword, but the changes should be minimal since the import structure remains the same. The keyword appears at the beginning, making it easy to parse.

How do I know if a library is compatible with lazy imports?

Most libraries should work fine with lazy imports. Libraries that might have issues:

• Those with essential import-time side effects (registration, monkey-patching).
• Those that expect specific import ordering.
• Those that modify global state during import.

When in doubt, test lazy imports with your specific use cases.

What happens if I globally enable lazy imports mode and a library doesn’t work correctly?

Note: This is an advanced feature. You can use the lazy imports filter to exclude specific modules that are known to have problematic side effects:

import sys

def my_filter(importer, name, fromlist): # Don't lazily import modules known to have side effects if name in {'problematic_module', 'another_module'}: return False # Import eagerly return True # Allow lazy import

sys.set_lazy_imports_filter(my_filter)

The filter function receives the importer module name, the module being imported, and the fromlist (if using from ... import). Returning False forces an eager import.

Alternatively, set the global mode to "none" via -X lazy_imports=none to turn off all lazy imports for debugging.

Can I use lazy imports inside functions?

No, the lazy keyword is only allowed at module level. For function-level lazy loading, use traditional inline imports or move the import to module level with lazy.

What about forwards compatibility with older Python versions?

Use the __lazy_modules__ global for compatibility:

# Works on Python 3.15+ as lazy, eager on older versions __lazy_modules__ = ['expensive_module', 'expensive_module_2'] import expensive_module from expensive_module_2 import MyClass

The __lazy_modules__ attribute is a list of module name strings. When an import statement is executed, Python checks if the module name being imported appears in __lazy_modules__. If it does, the import is treated as if it had the lazy keyword (becoming potentially lazy). On Python versions before 3.15 that don’t support lazy imports, the __lazy_modules__ attribute is simply ignored and imports proceed eagerly as normal.

This provides a migration path until you can rely on the lazy keyword. For maximum predictability, it’s recommended to define __lazy_modules__ once, before any imports. But as it is checked on each import, it can be modified between import statements.

How do explicit lazy imports interact with PEP 649 and PEP 749?

Python 3.14 implemented deferred evaluation of annotations, as specified by PEP 649 and PEP 749. If an annotation is not stringified, it is an expression that is evaluated at a later time. It will only be resolved if the annotation is accessed. In the example below, the fake_typing module is only loaded when the user inspects the __annotations__ dictionary. The fake_typing module would also be loaded if the user uses annotationlib.get_annotations() or getattr to access the annotations.

lazy from fake_typing import MyFakeType def foo(x: MyFakeType): pass print(foo.__annotations__) # Triggers loading the fake_typing module

How do lazy imports interact with dir(), getattr(), and module introspection?

Accessing lazy imports through normal attribute access or getattr() will trigger reification of the accessed attribute. Calling dir() on a module will be special cased in mod.__dir__ to avoid reification.

lazy import json

# Before any access # json not in sys.modules

# Any of these trigger reification: dumps_func = json.dumps dumps_func = getattr(json, 'dumps') # Now json is in sys.modules

Do lazy imports work with circular imports?

Lazy imports don’t automatically solve circular import problems. If two modules have a circular dependency, making the imports lazy might help only if the circular reference isn’t accessed during module initialization. However, if either module accesses the other during import time, you’ll still get an error.

Example that works (deferred access in functions):

# user_model.py lazy import post_model

class User: def get_posts(self): # OK - post_model accessed inside function, not during import return post_model.Post.get_by_user(self.name)

# post_model.py lazy import user_model

class Post: @staticmethod def get_by_user(username): return f"Posts by {username}"

This works because neither module accesses the other at module level – the access happens later when get_posts() is called.

Example that fails (access during import):

# module_a.py lazy import module_b

result = module_b.get_value() # Error! Accessing during import

def func(): return "A"

# module_b.py lazy import module_a

result = module_a.func() # Circular dependency error here

def get_value(): return "B"

This fails because module_a tries to access module_b at import time, which then tries to access module_a before it’s fully initialized.

The best practice is still to avoid circular imports in your code design.

Will lazy imports affect the performance of my hot paths?

After first use (provided the import succeed), lazy imports have zero overhead thanks to the adaptive interpreter. The interpreter specializes the bytecode (e.g., LOAD_GLOBAL becomes LOAD_GLOBAL_MODULE) which eliminates the lazy check on subsequent accesses. This means once a lazy import is reified, accessing it is just as fast as a normal import.

lazy import json

def use_json(): return json.dumps({"test": 1})

# First call triggers reification use_json()

# After 2-3 calls, bytecode is specialized use_json() use_json()

You can observe the specialization using dis.dis(use_json, adaptive=True):

=== Before specialization === LOAD_GLOBAL 0 (json) LOAD_ATTR 2 (dumps)

=== After 3 calls (specialized) === LOAD_GLOBAL_MODULE 0 (json) LOAD_ATTR_MODULE 2 (dumps)

The specialized LOAD_GLOBAL_MODULE and LOAD_ATTR_MODULE instructions are optimized fast paths with no overhead for checking lazy imports.

What about sys.modules? When does a lazy import appear there?

A lazily imported module does not appear in sys.modules until it’s reified (first used). Once reified, it appears in sys.modules just like any eager import.

import sys lazy import json

print('json' in sys.modules) # False

result = json.dumps({"key": "value"}) # First use

print('json' in sys.modules) # True

Does lazy from __future__ import feature work?

No, future imports can’t be lazy because they’re parser/compiler directives. It’s technically possible for the runtime behavior to be lazy but there’s no real value in it.

Why did you choose lazy as the keyword name?

Not “why”… memorize! :)

Deferred Ideas

The following ideas have been considered but are deliberately deferred to focus on delivering a stable, usable core feature first. These may be considered for future enhancements once we have real-world experience with lazy imports.

Alternative syntax and ergonomic improvements

Several alternative syntax forms have been suggested to improve ergonomics:

• Type-only imports: A specialized syntax for imports used exclusively in type annotations (similar to the type keyword in other contexts) could be added, such as type from collections.abc import Sequence. This would make the intent clearer than using lazy for type-only imports and would signal to readers that the import is never used at runtime. However, since lazy imports already solve the runtime cost problem for type annotations, we prefer to start with the simpler, more general mechanism and evaluate whether specialized syntax adds sufficient value after gathering usage data.

• Block-based syntax: Grouping multiple lazy imports in a block, such as:

  as lazy: import foo from bar import baz This could reduce repetition when marking many imports as lazy. However, it would require introducing an entirely new statement form (as lazy: blocks) that doesn’t fit into Python’s existing grammar patterns. It’s unclear how this would interact with other language features or what the precedent would be for similar block-level modifiers. This approach also makes it less clear when scanning code whether a particular import is lazy, since you must look at the surrounding context rather than the import line itself.

While these alternatives could provide different ergonomics in certain contexts, they share similar drawbacks: they would require introducing new statement forms or overloading existing syntax in non-obvious ways, and they open the door to many other potential uses of similar syntax patterns that would significantly expand the language. We prefer to start with the explicit lazy import syntax and gather real-world feedback before considering additional syntax variations. Any future ergonomic improvements should be evaluated based on actual usage patterns rather than speculative benefits.

Automatic lazy imports for if TYPE_CHECKING blocks

A future enhancement could automatically treat all imports inside if TYPE_CHECKING: blocks as lazy:

from typing import TYPE_CHECKING

if TYPE_CHECKING: from foo import Bar # Could be automatically lazy

However, this would require significant changes to make this work at compile time, since TYPE_CHECKING is currently just a runtime variable. The compiler would need special knowledge of this pattern, similar to how from __future__ import statements are handled. Additionally, making TYPE_CHECKING a built-in would be required for this to work reliably. Since lazy imports already solve the runtime cost problem for type-only imports, we prefer to start with the explicit syntax and evaluate whether this optimization adds sufficient value.

Module-level lazy import mode

A module-level declaration to make all imports in that module lazy by default:

from __future__ import lazy_imports import foo # Automatically lazy

This was discussed but deferred because it raises several questions. Using from __future__ import implies this would become the default behavior in a future Python version, which is unclear and not currently planned. It also raises questions about how such a mode would interact with the global flag and what the transition path would look like. The current explicit syntax and __lazy_modules__ provide sufficient control for initial adoption.

Package metadata for lazy-safe declarations

Future enhancements could allow packages to declare in their metadata whether they are safe for lazy importing (e.g., no import-time side effects). This could be used by the filter mechanism or by static analysis tools. The current filter API is designed to accommodate such future additions without requiring changes to the core language specification.

C API for lazy imports

No dedicated C API is planned for creating or resolving lazy imports. This feature is designed as a purely Python-facing mechanism, as C extensions typically need immediate access to modules and cannot benefit from deferred loading. Existing C API functions like PyImport_ImportModule() remain unchanged and continue to perform eager imports. If compelling use cases emerge, this could be revisited in future versions.

Alternate Implementation Ideas

Here are some alternative design decisions that were considered during the development of this PEP. While the current proposal represents what we believe to be the best balance of simplicity, performance, and maintainability, these alternatives offer different trade-offs that may be valuable for implementers to consider or for future refinements.

Leveraging a subclass of dict

Instead of updating the internal dict object to directly add the fields needed to support lazy imports, we could create a subclass of the dict object to be used specifically for Lazy Import enablement. This would still be a leaky abstraction though - methods can be called directly such as dict.__getitem__ and it would impact the performance of globals lookup in the interpreter.

Alternate keyword names

For this PEP, we decided to propose lazy for the explicit keyword as it felt the most familiar to those already focused on optimizing import overhead. We also considered a variety of other options to support explicit lazy imports. The most compelling alternates were defer and delay.

Rejected Ideas

Making the new behavior the default

Changing import to be lazy by default is outside of the scope of this PEP. From the discussion on PEP 690 it is clear that this is a fairly contentious idea, although perhaps once we have wide-spread use of lazy imports this can be reconsidered.

Supporting __lazy_modules__ = ["*"] as built-in syntax

The suggestion to support __lazy_modules__ = ["*"] as a convenient way to make all imports in a module lazy without explicit enumeration has been considered. This approach was rejected because __lazy_modules__ already represents implicit action-at-a-distance behavior that is tolerated solely as a backwards compatibility mechanism. Extending support to wildcard patterns would significantly increase implementation complexity and invite scope creep into pattern matching and globbing functionality. As __lazy_modules__ is a permanent language feature that cannot be removed in future versions, the design prioritizes minimalism and restricts its scope to serving as a transitional tool for backwards compatibility.

It is worth noting that the implementation performs membership checks by calling __contains__ on the __lazy_modules__ object. Consequently, users requiring wildcard behavior may provide a custom object implementing __contains__ to return True for all queries or other desired patterns. This design provides the necessary flexibility for advanced use cases while maintaining a simple, focused specification for the primary mechanism. If this PEP is accepted, adding such helper objects to the standard library can be discussed in a future issue. Presently, it is out of scope for this PEP.

Disallowing lazy imports inside with blocks

An earlier version of this PEP proposed disallowing lazy import statements inside with blocks, similar to the restriction on try blocks. The concern was that certain context managers (like contextlib.suppress(ImportError)) could suppress import errors in confusing ways when combined with lazy imports.

However, this restriction was rejected because with statements have much broader semantics than try/except blocks. While try/except is explicitly about catching exceptions, with blocks are commonly used for resource management, temporary state changes, or scoping – contexts where lazy imports work perfectly fine. The lazy import syntax is explicit enough that developers who write it inside a with block are making an intentional choice, aligning with Python’s “consenting adults” philosophy. For genuinely problematic cases like with suppress(ImportError): lazy import foo, static analysis tools and linters are better suited to catch these patterns than hard language restrictions.

Forcing eager imports in with blocks under the global flag

Another rejected idea was to make imports inside with blocks remain eager even when the global lazy imports flag is set to "all". The rationale was to be conservative: since with statements can affect how imports behave (e.g., by modifying sys.path or suppressing exceptions), forcing imports to remain eager could prevent subtle bugs. However, this would create inconsistent behavior where lazy import is allowed explicitly in with blocks, but normal imports remain eager when the global flag is enabled. This inconsistency between explicit and implicit laziness is confusing and hard to explain.

The simpler, more consistent rule is that the global flag affects imports everywhere that explicit lazy import syntax is allowed. This avoids having three different sets of rules (explicit syntax, global flag behavior, and filter mechanism) and instead provides two: explicit syntax rules match what the global flag affects, and the filter mechanism provides escape hatches for edge cases. For users who need fine-grained control, the filter mechanism (sys.set_lazy_imports_filter()) already provides a way to exclude specific imports or patterns. Additionally, there’s no inverse operation: if the global flag forces imports eager in with blocks but a user wants them lazy, there’s no way to override it, creating an asymmetry.

In summary: imports in with blocks behave consistently whether marked explicitly with lazy import or implicitly via the global flag, creating a simple rule that’s easy to explain and reason about.

Modification of the dict object

The initial PEP for lazy imports (PEP 690) relied heavily on the modification of the internal dict object to support lazy imports. We recognize that this data structure is highly tuned, heavily used across the codebase, and very performance sensitive. Because of the importance of this data structure and the desire to keep the implementation of lazy imports encapsulated from users who may have no interest in the feature, we’ve decided to invest in an alternate approach.

The dictionary is the foundational data structure in Python. Every object’s attributes are stored in a dict, and dicts are used throughout the runtime for namespaces, keyword arguments, and more. Adding any kind of hook or special behavior to dicts to support lazy imports would:

1. Prevent critical interpreter optimizations including future JIT compilation.
2. Add complexity to a data structure that must remain simple and fast.
3. Affect every part of Python, not just import behavior.
4. Violate separation of concerns – the hash table shouldn’t know about the import system.

Past decisions that violated this principle of keeping core abstractions clean have caused significant pain in the CPython ecosystem, making optimization difficult and introducing subtle bugs.

Transforming lazy objects via __class__ mutation

An alternative implementation approach was proposed where lazy import objects would be transformed into their final form by mutating their internal state, rather than replacing the object entirely. Under this approach, a lazy object would be transformed in-place after the actual import completes.

This approach was rejected for several reasons:

1. This technique could potentially work for module objects, but breaks down completely for arbitrary objects imported via from statements. When a user writes lazy from foo import bar, the object bar could be any Python object (a function, class, constant, etc.), not just a module. Any transformation approach would require that the lazy proxy object have compatible memory layout and other considerations with the target object, which is impossible to know before loading the module. This creates a fundamental asymmetry where lazy import x and lazy from x import y would require completely different implementation strategies, with the latter still needing the proxy replacement mechanism.
2. Even for module objects, the approach has fundamental limitations. Some implementations substitute custom classes in sys.modules that inherit from or replace the standard module type. These custom module classes can have different memory layouts and sizes than PyModuleObject. The transformation approach cannot work with such generic custom module implementations, creating fragility and maintenance burden across the ecosystem.
3. While the transformation approach might appear simpler in some respects, this is somewhat subjective. It introduces a bifurcated implementation: one path for modules and a completely different path for non-module objects. Whether this is simpler than a unified proxy mechanism depends on perspective and implementation details.
4. Any code holding a reference to the object before transformation will see a different type after transformation. This can break code that checks object types or relies on type stability, particularly C extensions that cache type pointers or use PyObject_TypeCheck. The transformation also requires careful coordination between the lazy import machinery and the type system to ensure that the object remains valid throughout the transformation process. The current proxy-based design avoids these issues by maintaining clear boundaries between the lazy proxy and the actual imported object.

The current design, which uses object replacement through the LazyImportType proxy pattern, provides a consistent mechanism that works uniformly for both import and from ... import statements while maintaining cleaner separation between the lazy import machinery and Python’s core object model.

Making lazy imports find the module without loading it

The Python import machinery separates out finding a module and loading it, and the lazy import implementation could technically defer only the loading part. However, this approach was rejected for several critical reasons.

A significant part of the performance win comes from skipping the finding phase. The issue is particularly acute on NFS-backed filesystems and distributed storage, where each stat() call incurs network latency. In these kinds of environments, stat() calls can take tens to hundreds of milliseconds depending on network conditions. With dozens of imports each doing multiple filesystem checks traversing sys.path, the time spent finding modules before executing any Python code can become substantial. In some measurements, spec finding accounts for the majority of total import time. Skipping only the loading phase would leave most of the performance problem unsolved.

More critically, separating finding from loading creates the worst of both worlds for error handling. Some exceptions from the import machinery (e.g., ImportError from a missing module, path resolution failures, ModuleNotFoundError) would be raised at the lazy import statement, while others (e.g., SyntaxError, ImportError from circular imports, attribute errors from from module import name) would be raised later at first use. This split is both confusing and unpredictable: developers would need to understand the internal import machinery to know which errors happen when. The current design is simpler: with full lazy imports, all import-related errors occur at first use, making the behavior consistent and predictable.

Additionally, there are technical limitations: finding the module does not guarantee the import will succeed, nor even that it will not raise ImportError. Finding modules in packages requires that those packages are loaded, so it would only help with lazy loading one level of a package hierarchy. Since “finding” attributes in modules requires loading them, this would create a hard to explain difference between from package import module and from module import function.

Placing the lazy keyword in the middle of from imports

While we found from foo lazy import bar to be a really intuitive placement for the new explicit syntax, we quickly learned that placing the lazy keyword here is already syntactically allowed in Python. This is because from . lazy import bar is legal syntax (because whitespace does not matter.)

Placing the lazy keyword at the end of import statements

We discussed appending lazy to the end of import statements like such import foo lazy or from foo import bar, baz lazy but ultimately decided that this approach provided less clarity. For example, if multiple modules are imported in a single statement, it is unclear if the lazy binding applies to all of the imported objects or just a subset of the items.

Adding an explicit eager keyword

Since we’re not changing the default behavior, and we don’t want to encourage use of the global flags, it’s too early to consider adding superfluous syntax for the common, default case. It would create too much confusion about what the default is, or when the eager keyword would be necessary, or whether it affects lazy imports in the explicitly eagerly imported module.

Allowing the filter to force lazy imports even when globally disabled

As lazy imports allow some forms of circular imports that would otherwise fail, as an intentional and desirable thing (especially for typing-related imports), there was a suggestion to add a way to override the global disable and force particular imports to be lazy, for instance by calling the lazy imports filter even if lazy imports are globally disabled.

This approach could introduce a complex hierarchy of the different “override” systems, making it much harder to analyze and reason about the code. Additionally, this may require additional complexity to introduce finer-grained systems to enable or disable particular imports as the use of lazy imports evolves. The global disable is not expected to see commonplace use, but be more of a debugging and selective testing tool for those who want to tightly control their dependency on lazy imports. We think it’s reasonable for package maintainers, as they update packages to adopt lazy imports, to decide to not support running with lazy imports globally disabled.

It may be that this means that in time, as more and more packages embrace both typing and lazy imports, the global disable becomes mostly unused and unusable. Similar things have happened in the past with other global flags, and given the low cost of the flag this seems acceptable. It’s also easier to add more specific re-enabling mechanisms later, when we have a clearer picture of real-world use and patterns, than it is to remove a hastily added mechanism that isn’t quite right.

Using underscore-prefixed names for advanced features

The global activation and filter functions (sys.set_lazy_imports, sys.set_lazy_imports_filter, sys.get_lazy_imports_filter) could be marked as “private” or “advanced” by using underscore prefixes (e.g., sys._set_lazy_imports_filter). This was rejected because branding as advanced features through documentation is sufficient. These functions have legitimate use cases for advanced users, particularly operators of large deployments. Providing an official mechanism prevents divergence from upstream CPython. The global mode is intentionally documented as an advanced feature for operators running huge fleets, not for day-to-day users or libraries. Python has precedent for advanced features that remain public APIs without underscore prefixes - for example, gc.disable(), gc.get_objects(), and gc.set_threshold() are advanced features that can cause issues if misused, yet they are not underscore-prefixed.

Using a decorator syntax for lazy imports

A decorator-based syntax could mark imports as lazy:

@lazy import json

@lazy from foo import bar

This approach was rejected because it introduces too many open questions and complications. Decorators in Python are designed to wrap and transform callable objects (functions, classes, methods), not statements. Allowing decorators on import statements would open the door to many other potential statement decorators (@cached, @traced, @deprecated, etc.), significantly expanding the language’s syntax in ways we don’t want to explore. Furthermore, this raises the question of where such decorators would come from: they would need to be either imported or built-in, creating a bootstrapping problem for import-related decorators. This is far more speculative and generic than the focused lazy import syntax.

Using a context manager instead of a new soft keyword

A backward compatible syntax, for example in the form of a context manager, has been proposed:

with lazy_imports(...): import json

This would replace the need for __lazy_modules__, and allow libraries to use one of the existing lazy imports implementations in older Python versions. However, adding magic with statements with that kind of effect would be a significant change to Python and with statements in general, and it would not be easy to combine with the implementation for lazy imports in this proposal. Adding standard library support for existing lazy importers without changes to the implementation amounts to the status quo, and does not solve the performance and usability issues with those existing solutions.

Returning a proxy dict from globals()

An alternative to reifying on globals() or exposing lazy objects would be to return a proxy dictionary that automatically reifies lazy objects when they’re accessed through the proxy. This would seemingly give the best of both worlds: globals() returns immediately without reification cost, but accessing items through the result would automatically resolve lazy imports.

However, this approach is fundamentally incompatible with how globals() is used in practice. Many standard library functions and built-ins expect globals() to return a real dict object, not a proxy:

• exec(code, globals()) requires a real dict.
• eval(expr, globals()) requires a real dict.
• Functions that check type(globals()) is dict would break.
• Dictionary methods like .update() would need special handling.
• Performance would suffer from the indirection on every access.

The proxy would need to be so transparent that it would be indistinguishable from a real dict in almost all cases, which is extremely difficult to achieve correctly. Any deviation from true dict behavior would be a source of subtle bugs.

Automatically reifying on __dict__ or globals() access

Three options were considered for how globals() and mod.__dict__ should behave with lazy imports:

1. Calling globals() or mod.__dict__ traverses and resolves all lazy objects before returning.
2. Calling globals() or mod.__dict__ returns the dictionary with lazy objects present (chosen).
3. Calling globals() returns the dictionary with lazy objects, but mod.__dict__ reifies everything.

We chose option 2: both globals() and __dict__ return the raw namespace dictionary without triggering reification. This provides a clean, predictable model where low-level introspection APIs don’t trigger side effects.

Having globals() and __dict__ behave identically creates symmetry and a simple mental model: both expose the raw namespace view. Low-level introspection APIs should not automatically trigger imports, which would be surprising and potentially expensive. Real-world experience implementing lazy imports in the standard library (such as the traceback module) showed that automatic reification on __dict__ access was cumbersome and forced introspection code to load modules it was only examining.

Option 1 (always reifying) was rejected because it would make globals() and __dict__ access surprisingly expensive and prevent introspecting the lazy state of a module. Option 3 was initially considered to “protect” external code from seeing lazy objects, but real-world usage showed this created more problems than it solved, particularly for stdlib code that needs to introspect modules without triggering side effects.

Acknowledgements

We would like to thank Paul Ganssle, Yury Selivanov, Łukasz Langa, Lysandros Nikolaou, Pradyun Gedam, Mark Shannon, Hana Joo and the Python Google team, the Python team(s) @ Meta, the Python @ HRT team, the Bloomberg Python team, the Scientific Python community, everyone who participated in the initial discussion of PEP 690, and many others who provided valuable feedback and insights that helped shape this PEP.

Footnotes

Copyright

This document is placed in the public domain or under the CC0-1.0-Universal license, whichever is more permissive.