Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

ENH some steps to make cloudpickle dynamic function/classes more deterministic #524

Merged
merged 20 commits into from
Nov 23, 2023
Merged
Show file tree
Hide file tree
Changes from 17 commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions ci/install_coverage_subprocess_pth.py
Original file line number Diff line number Diff line change
Expand Up @@ -9,8 +9,8 @@
import coverage; coverage.process_startup()
"""

filename = op.join(get_path('purelib'), 'coverage_subprocess.pth')
with open(filename, 'wb') as f:
f.write(FILE_CONTENT.encode('ascii'))
filename = op.join(get_path("purelib"), "coverage_subprocess.pth")
with open(filename, "wb") as f:
f.write(FILE_CONTENT.encode("ascii"))

print('Installed subprocess coverage support: %s' % filename)
print("Installed subprocess coverage support: %s" % filename)
96 changes: 72 additions & 24 deletions cloudpickle/cloudpickle.py
Original file line number Diff line number Diff line change
Expand Up @@ -126,7 +126,7 @@ def _lookup_class_or_track(class_tracker_id, class_def):


def register_pickle_by_value(module):
"""Register a module to make it functions and classes picklable by value.
"""Register a module to make its functions and classes picklable by value.

By default, functions and classes that are attributes of an importable
module are to be pickled by reference, that is relying on re-importing
Expand Down Expand Up @@ -369,7 +369,7 @@ def func():
# sys.modules.
if name is not None and name.startswith(prefix):
# check whether the function can address the sub-module
tokens = set(name[len(prefix) :].split("."))
tokens = set(name[len(prefix):].split("."))
if not tokens - set(code.co_names):
subimports.append(sys.modules[name])
return subimports
Expand Down Expand Up @@ -409,7 +409,12 @@ def _walk_global_ops(code):

def _extract_class_dict(cls):
"""Retrieve a copy of the dict of a class without the inherited method."""
clsdict = dict(cls.__dict__) # copy dict proxy to a dict
# copy dict proxy to a dict. Sort the keys to make the pickle deterministic
# Also create a copy of the dict's keys, to avoid its memoization.
# This is necessary as memoization happens only if all string are interned,
# which is not the case in reconstructed dynamic classes.
tomMoral marked this conversation as resolved.
Show resolved Hide resolved
clsdict = {"".join(k): cls.__dict__[k] for k in sorted(cls.__dict__)}

if len(cls.__bases__) == 1:
inherited_dict = cls.__bases__[0].__dict__
else:
Expand Down Expand Up @@ -533,9 +538,15 @@ class id will also reuse this class definition.
The "extra" variable is meant to be a dict (or None) that can be used for
forward compatibility shall the need arise.
"""
# We need to intern the keys of the type_kwargs dict to avoid having
# different pickles for the same dynamic class depending on whether it was
# dynamically created or reconstructed from a pickled stream.
type_kwargs = {sys.intern(k): v for k, v in type_kwargs.items()}
ogrisel marked this conversation as resolved.
Show resolved Hide resolved

skeleton_class = types.new_class(
name, bases, {"metaclass": type_constructor}, lambda ns: ns.update(type_kwargs)
)

return _lookup_class_or_track(class_tracker_id, skeleton_class)


Expand Down Expand Up @@ -694,7 +705,13 @@ def _function_getstate(func):
# unpickling time by iterating over slotstate and calling setattr(func,
# slotname, slotvalue)
slotstate = {
"__name__": func.__name__,
# Create a copy of the function name, to avoid memoization. This is necessary
# to ensure deterministic pickles (when doing rountrips with a remote Python
# process): the behavior of the pickler's memoizer depends on the string
# physical identity and therefore on whether the name is interned or not
# The name of reconstructed dynamic function is typically not interned, so
# we make sure it is not interned prior to pickling as well.
tomMoral marked this conversation as resolved.
Show resolved Hide resolved
"__name__": "".join(func.__name__),
"__qualname__": func.__qualname__,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Any idea why we don't need this treatment for the __qualname__ value nor for the keys of func.__dict__?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actually, at least for the latter, there are still things to solver. I will push a slight change to the test to show what I mean.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The issue with string interning is mostly due to collision between the string defined in two different places: typically method names that are interned as class attributes, vs the dynamic func name.
So I would say this won't happen for __qualname__ as this is only present here but maybe I am wrong.

"__annotations__": func.__annotations__,
"__kwdefaults__": func.__kwdefaults__,
Expand All @@ -721,7 +738,10 @@ def _function_getstate(func):
)
slotstate["__globals__"] = f_globals

state = func.__dict__
# copy dict proxy to a dict. Create a copy of the dict's keys, to avoid their
# memoization. This is necessary as memoization happens only if all string
# are interned, which is not the case in reconstructed dynamic functions.
tomMoral marked this conversation as resolved.
Show resolved Hide resolved
state = {"".join(k): v for k, v in func.__dict__.items()}
return state, slotstate


Expand Down Expand Up @@ -802,6 +822,22 @@ def _code_reduce(obj):
# of the specific type from types, for example:
# >>> from types import CodeType
# >>> help(CodeType)

# Create a copy of the object name, to avoid memoization. This is necessary
# to ensure deterministic pickles, that depends wheter the name is interned
# or not. The name of code objects of reconstructed dynamic functions or
# methods is typically not interned, so we make sure it is not interned
# either prior to pickling.
tomMoral marked this conversation as resolved.
Show resolved Hide resolved
co_name = "".join(obj.co_name)

# Create shallow copies of these tuple to make cloudpickle payload deterministic.
# When creating a code object during load, copies of these four tuples are
# created, while in the main process, these tuples can be shared.
# By always creating copies, we make sure the resulting payload is deterministic.
co_names = tuple(name for name in obj.co_names)
co_varnames = tuple(name for name in obj.co_varnames)
co_freevars = tuple(name for name in obj.co_freevars)
co_cellvars = tuple(name for name in obj.co_cellvars)
if hasattr(obj, "co_exceptiontable"):
# Python 3.11 and later: there are some new attributes
# related to the enhanced exceptions.
Expand All @@ -814,16 +850,16 @@ def _code_reduce(obj):
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_names,
co_varnames,
obj.co_filename,
obj.co_name,
co_name,
obj.co_qualname,
obj.co_firstlineno,
obj.co_linetable,
obj.co_exceptiontable,
obj.co_freevars,
obj.co_cellvars,
co_freevars,
co_cellvars,
)
elif hasattr(obj, "co_linetable"):
# Python 3.10 and later: obj.co_lnotab is deprecated and constructor
Expand All @@ -837,14 +873,14 @@ def _code_reduce(obj):
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_names,
co_varnames,
obj.co_filename,
obj.co_name,
co_name,
obj.co_firstlineno,
obj.co_linetable,
obj.co_freevars,
obj.co_cellvars,
co_freevars,
co_cellvars,
)
elif hasattr(obj, "co_nmeta"): # pragma: no cover
# "nogil" Python: modified attributes from 3.9
Expand All @@ -859,15 +895,15 @@ def _code_reduce(obj):
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_varnames,
co_varnames,
obj.co_filename,
obj.co_name,
co_name,
obj.co_firstlineno,
obj.co_lnotab,
obj.co_exc_handlers,
obj.co_jump_table,
obj.co_freevars,
obj.co_cellvars,
co_freevars,
co_cellvars,
obj.co_free2reg,
obj.co_cell2reg,
)
Expand All @@ -882,14 +918,14 @@ def _code_reduce(obj):
obj.co_flags,
obj.co_code,
obj.co_consts,
obj.co_names,
obj.co_varnames,
co_names,
co_varnames,
obj.co_filename,
obj.co_name,
co_name,
obj.co_firstlineno,
obj.co_lnotab,
obj.co_freevars,
obj.co_cellvars,
co_freevars,
co_cellvars,
)
return types.CodeType, args

Expand Down Expand Up @@ -1127,6 +1163,18 @@ def _class_setstate(obj, state):
if attrname == "_abc_impl":
registry = attr
else:
# Note: setting attribute names on a class automatically triggers their
# interning in CPython:
# https://github.com/python/cpython/blob/v3.12.0/Objects/object.c#L957
#
# This means that to get deterministic pickling for a dynamic class that
# was initially defined in a different Python process, the pickler
# needs to ensure that dynamic class and function attribute names are
# systematically copied into a non-interned version to avoid
# unpredictable pickle payloads.
#
# Indeed the Pickler's memoizer relies on physical object identity to break
# cycles in the reference graph of the object being serialized.
setattr(obj, attrname, attr)
if registry is not None:
for subclass in registry:
Expand Down
1 change: 1 addition & 0 deletions cloudpickle/cloudpickle_fast.py
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@

See: tests/test_backward_compat.py
"""

from . import cloudpickle


Expand Down
3 changes: 3 additions & 0 deletions tests/__init__.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,3 @@
import pytest

pytest.register_assert_rewrite("tests.testutils")
Loading