我有一个长期运行的Python服务器,希望能够在不重新启动服务器的情况下升级服务。最好的方法是什么?

if foo.py has changed:
    unimport foo  <-- How do I do this?
    import foo
    myfoo = foo.Foo()

当前回答

重新加载(模块),但前提是它是完全独立的。如果其他任何东西引用了该模块(或属于该模块的任何对象),那么您将会得到一些微妙而奇怪的错误,这些错误是由旧代码存在的时间比您预期的要长,并且像isinstance这样的错误在相同代码的不同版本之间无法工作。

如果有单向依赖关系,还必须重新加载依赖于重新加载的模块的所有模块,以消除对旧代码的所有引用。然后递归地重新加载依赖于重新加载模块的模块。

如果你有循环依赖关系,这是非常常见的,例如当你处理重新加载一个包时,你必须一次性卸载组中的所有模块。使用reload()无法做到这一点,因为它会在刷新依赖项之前重新导入每个模块,从而允许旧的引用渗透到新模块中。

在这种情况下,唯一的方法是入侵sys。模块,这是不受支持的。你必须仔细检查并删除每个系统。您希望在下次导入时重新加载的模块项,还可以删除值为None的项,以处理缓存失败的相对导入的实现问题。这并不是很好,但只要你有一个完全自包含的依赖集,并且不会在代码库之外留下引用,它就是可行的。

最好是重新启动服务器。:-)

其他回答

对于像我这样想卸载所有模块的人(当在Emacs下的Python解释器中运行时):

   for mod in sys.modules.values():
      reload(mod)

更多信息请参见重新加载Python模块。

其他选项。请参阅Python默认importlib。Reload将只是重新导入作为参数传递的库。它不会重新加载库导入的库。如果您更改了很多文件,并且有一个有点复杂的包要导入,那么您必须进行深度重载。

如果你安装了IPython或Jupyter,你可以使用一个函数来深度重载所有库:

from IPython.lib.deepreload import reload as dreload
dreload(foo)

如果你没有Jupyter,在你的shell中使用以下命令安装它:

pip3 install jupyter

在Python 3.0-3.3中,您将使用:

BDFL已经回答了这个问题。

然而,imp在3.4中被弃用了,取而代之的是importlib(谢谢@Stefan!)

因此,我认为您现在应该使用importlib.reload(module),尽管我不确定。

编辑(答案V2)

前面的解决方案只适用于获取重置信息,但它不会更改所有引用(超过重新加载,但小于所需)。为了设置所有的引用,我必须进入垃圾收集器,并在那里重写引用。现在它就像一个咒语!

请注意,如果GC被关闭,或者如果重新加载不受GC监视的数据,这将不起作用。如果您不想打乱GC,原始答案可能就足够了。

新代码:

import importlib
import inspect
import gc
from enum import EnumMeta
from weakref import ref


_readonly_attrs = {'__annotations__', '__call__', '__class__', '__closure__', '__code__', '__defaults__', '__delattr__',
               '__dict__', '__dir__', '__doc__', '__eq__', '__format__', '__func__', '__ge__', '__get__',
               '__getattribute__', '__globals__', '__gt__', '__hash__', '__init__', '__init_subclass__',
               '__kwdefaults__', '__le__', '__lt__', '__module__', '__name__', '__ne__', '__new__', '__qualname__',
               '__reduce__', '__reduce_ex__', '__repr__', '__self__', '__setattr__', '__sizeof__', '__str__',
               '__subclasshook__', '__weakref__', '__members__', '__mro__', '__itemsize__', '__isabstractmethod__',
               '__basicsize__', '__base__'}


def reset_module(module, inner_modules_also=True):
    """
    This function is a stronger form of importlib's `reload` function. What it does, is that aside from reloading a
    module, it goes to the old instance of the module, and sets all the (not read-only) attributes, functions and classes
    to be the reloaded-module's
    :param module: The module to reload (module reference, not the name)
    :param inner_modules_also: Whether to treat ths module as a package as well, and reload all the modules within it.
    """

    # For the case when the module is actually a package
    if inner_modules_also:
        submods = {submod for _, submod in inspect.getmembers(module)
                   if (type(submod).__name__ == 'module') and (submod.__package__.startswith(module.__name__))}
        for submod in submods:
            reset_module(submod, True)

    # First, log all the references before reloading (because some references may be changed by the reload operation).
    module_tree = _get_tree_references_to_reset_recursively(module, module.__name__)

    new_module = importlib.reload(module)
    _reset_item_recursively(module, module_tree, new_module)


def _update_referrers(item, new_item):
    refs = gc.get_referrers(item)

    weak_ref_item = ref(item)
    for coll in refs:
        if type(coll) == dict:
            enumerator = coll.keys()
        elif type(coll) == list:
            enumerator = range(len(coll))
        else:
            continue

        for key in enumerator:

            if weak_ref_item() is None:
                # No refs are left in the GC
                return

            if coll[key] is weak_ref_item():
                coll[key] = new_item

def _get_tree_references_to_reset_recursively(item, module_name, grayed_out_item_ids = None):
    if grayed_out_item_ids is None:
        grayed_out_item_ids = set()

    item_tree = dict()
    attr_names = set(dir(item)) - _readonly_attrs
    for sub_item_name in attr_names:

        sub_item = getattr(item, sub_item_name)
        item_tree[sub_item_name] = [sub_item, None]

        try:
            # Will work for classes and functions defined in that module.
            mod_name = sub_item.__module__
        except AttributeError:
            mod_name = None

        # If this item was defined within this module, deep-reset
        if (mod_name is None) or (mod_name != module_name) or (id(sub_item) in grayed_out_item_ids) \
                or isinstance(sub_item, EnumMeta):
            continue

        grayed_out_item_ids.add(id(sub_item))
        item_tree[sub_item_name][1] = \
            _get_tree_references_to_reset_recursively(sub_item, module_name, grayed_out_item_ids)

    return item_tree


def _reset_item_recursively(item, item_subtree, new_item):

    # Set children first so we don't lose the current references.
    if item_subtree is not None:
        for sub_item_name, (sub_item, sub_item_tree) in item_subtree.items():

            try:
                new_sub_item = getattr(new_item, sub_item_name)
            except AttributeError:
                # The item doesn't exist in the reloaded module. Ignore.
                continue

            try:
                # Set the item
                _reset_item_recursively(sub_item, sub_item_tree, new_sub_item)
            except Exception as ex:
                pass

    _update_referrers(item, new_item)

原来的答案

正如在@bobince的回答中所写的,如果在另一个模块中已经有对该模块的引用(特别是如果它是用As关键字导入的,如import numpy As np),该实例将不会被覆盖。

当应用需要配置模块的“全新”状态的测试时,这对我来说是相当有问题的,所以我写了一个名为reset_module的函数,它使用importlib的reload函数并递归地覆盖所有声明的模块属性。它已经用Python版本3.6进行了测试。

import importlib
import inspect
from enum import EnumMeta

_readonly_attrs = {'__annotations__', '__call__', '__class__', '__closure__', '__code__', '__defaults__', '__delattr__',
               '__dict__', '__dir__', '__doc__', '__eq__', '__format__', '__func__', '__ge__', '__get__',
               '__getattribute__', '__globals__', '__gt__', '__hash__', '__init__', '__init_subclass__',
               '__kwdefaults__', '__le__', '__lt__', '__module__', '__name__', '__ne__', '__new__', '__qualname__',
               '__reduce__', '__reduce_ex__', '__repr__', '__self__', '__setattr__', '__sizeof__', '__str__',
               '__subclasshook__', '__weakref__', '__members__', '__mro__', '__itemsize__', '__isabstractmethod__',
               '__basicsize__', '__base__'}


def reset_module(module, inner_modules_also=True):
    """
    This function is a stronger form of importlib's `reload` function. What it does, is that aside from reloading a
    module, it goes to the old instance of the module, and sets all the (not read-only) attributes, functions and classes
    to be the reloaded-module's
    :param module: The module to reload (module reference, not the name)
    :param inner_modules_also: Whether to treat ths module as a package as well, and reload all the modules within it.
    """

    new_module = importlib.reload(module)

    reset_items = set()

    # For the case when the module is actually a package
    if inner_modules_also:
        submods = {submod for _, submod in inspect.getmembers(module)
                   if (type(submod).__name__ == 'module') and (submod.__package__.startswith(module.__name__))}
        for submod in submods:
            reset_module(submod, True)

    _reset_item_recursively(module, new_module, module.__name__, reset_items)


def _reset_item_recursively(item, new_item, module_name, reset_items=None):
    if reset_items is None:
        reset_items = set()

    attr_names = set(dir(item)) - _readonly_attrs

    for sitem_name in attr_names:

        sitem = getattr(item, sitem_name)
        new_sitem = getattr(new_item, sitem_name)

        try:
            # Set the item
            setattr(item, sitem_name, new_sitem)

            try:
                # Will work for classes and functions defined in that module.
                mod_name = sitem.__module__
            except AttributeError:
                mod_name = None

            # If this item was defined within this module, deep-reset
            if (mod_name is None) or (mod_name != module_name) or (id(sitem) in reset_items) \
                    or isinstance(sitem, EnumMeta):  # Deal with enums
                continue

            reset_items.add(id(sitem))
            _reset_item_recursively(sitem, new_sitem, module_name, reset_items)
        except Exception as ex:
            raise Exception(sitem_name) from ex

注意事项:请谨慎使用!在非外围模块(例如,定义外部使用的类的模块)上使用这些可能会导致Python中的内部问题(例如pickle / unpickling问题)。

2018-02-01

模块foo必须提前导入成功。 从importlib导入重载,重载(foo)

31.5. import - Python 3.6.4文档的实现