2022-05-23 00:16:32 +04:00

390 lines
12 KiB
Python

# Copyright 2018-2022 Streamlit Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Hashing for st.memo and st.singleton."""
import collections
import functools
import hashlib
import inspect
import io
import os
import pickle
import sys
import tempfile
import threading
import unittest.mock
import weakref
from typing import Any, Pattern, Optional, Dict, List
from streamlit import type_util
from streamlit import util
from streamlit.logger import get_logger
from streamlit.uploaded_file_manager import UploadedFile
from .cache_errors import (
CacheType,
UnhashableTypeError,
)
_LOGGER = get_logger(__name__)
# If a dataframe has more than this many rows, we consider it large and hash a sample.
_PANDAS_ROWS_LARGE = 100000
_PANDAS_SAMPLE_SIZE = 10000
# Similar to dataframes, we also sample large numpy arrays.
_NP_SIZE_LARGE = 1000000
_NP_SAMPLE_SIZE = 100000
# Arbitrary item to denote where we found a cycle in a hashed object.
# This allows us to hash self-referencing lists, dictionaries, etc.
_CYCLE_PLACEHOLDER = b"streamlit-57R34ML17-hesamagicalponyflyingthroughthesky-CYCLE"
def update_hash(val: Any, hasher, cache_type: CacheType) -> None:
"""Updates a hashlib hasher with the hash of val.
This is the main entrypoint to hashing.py.
"""
ch = _CacheFuncHasher(cache_type)
ch.update(hasher, val)
class _HashStack:
"""Stack of what has been hashed, for debug and circular reference detection.
This internally keeps 1 stack per thread.
Internally, this stores the ID of pushed objects rather than the objects
themselves because otherwise the "in" operator inside __contains__ would
fail for objects that don't return a boolean for "==" operator. For
example, arr == 10 where arr is a NumPy array returns another NumPy array.
This causes the "in" to crash since it expects a boolean.
"""
def __init__(self):
self._stack: collections.OrderedDict[int, List[Any]] = collections.OrderedDict()
def __repr__(self) -> str:
return util.repr_(self)
def push(self, val: Any):
self._stack[id(val)] = val
def pop(self):
self._stack.popitem()
def __contains__(self, val: Any):
return id(val) in self._stack
class _HashStacks:
"""Stacks of what has been hashed, with at most 1 stack per thread."""
def __init__(self):
self._stacks: weakref.WeakKeyDictionary[
threading.Thread, _HashStack
] = weakref.WeakKeyDictionary()
def __repr__(self) -> str:
return util.repr_(self)
@property
def current(self) -> _HashStack:
current_thread = threading.current_thread()
stack = self._stacks.get(current_thread, None)
if stack is None:
stack = _HashStack()
self._stacks[current_thread] = stack
return stack
hash_stacks = _HashStacks()
def _int_to_bytes(i: int) -> bytes:
num_bytes = (i.bit_length() + 8) // 8
return i.to_bytes(num_bytes, "little", signed=True)
def _key(obj: Optional[Any]) -> Any:
"""Return key for memoization."""
if obj is None:
return None
def is_simple(obj):
return (
isinstance(obj, bytes)
or isinstance(obj, bytearray)
or isinstance(obj, str)
or isinstance(obj, float)
or isinstance(obj, int)
or isinstance(obj, bool)
or obj is None
)
if is_simple(obj):
return obj
if isinstance(obj, tuple):
if all(map(is_simple, obj)):
return obj
if isinstance(obj, list):
if all(map(is_simple, obj)):
return ("__l", tuple(obj))
if (
type_util.is_type(obj, "pandas.core.frame.DataFrame")
or type_util.is_type(obj, "numpy.ndarray")
or inspect.isbuiltin(obj)
or inspect.isroutine(obj)
or inspect.iscode(obj)
):
return id(obj)
return NoResult
class _CacheFuncHasher:
"""A hasher that can hash objects with cycles."""
def __init__(self, cache_type: CacheType):
self._hashes: Dict[Any, bytes] = {}
# The number of the bytes in the hash.
self.size = 0
self.cache_type = cache_type
def __repr__(self) -> str:
return util.repr_(self)
def to_bytes(self, obj: Any) -> bytes:
"""Add memoization to _to_bytes and protect against cycles in data structures."""
tname = type(obj).__qualname__.encode()
key = (tname, _key(obj))
# Memoize if possible.
if key[1] is not NoResult:
if key in self._hashes:
return self._hashes[key]
# Break recursive cycles.
if obj in hash_stacks.current:
return _CYCLE_PLACEHOLDER
hash_stacks.current.push(obj)
try:
# Hash the input
b = b"%s:%s" % (tname, self._to_bytes(obj))
# Hmmm... It's possible that the size calculation is wrong. When we
# call to_bytes inside _to_bytes things get double-counted.
self.size += sys.getsizeof(b)
if key[1] is not NoResult:
self._hashes[key] = b
finally:
# In case an UnhashableTypeError (or other) error is thrown, clean up the
# stack so we don't get false positives in future hashing calls
hash_stacks.current.pop()
return b
def update(self, hasher, obj: Any) -> None:
"""Update the provided hasher with the hash of an object."""
b = self.to_bytes(obj)
hasher.update(b)
def _to_bytes(self, obj: Any) -> bytes:
"""Hash objects to bytes, including code with dependencies.
Python's built in `hash` does not produce consistent results across
runs.
"""
if isinstance(obj, unittest.mock.Mock):
# Mock objects can appear to be infinitely
# deep, so we don't try to hash them at all.
return self.to_bytes(id(obj))
elif isinstance(obj, bytes) or isinstance(obj, bytearray):
return obj
elif isinstance(obj, str):
return obj.encode()
elif isinstance(obj, float):
return self.to_bytes(hash(obj))
elif isinstance(obj, int):
return _int_to_bytes(obj)
elif isinstance(obj, (list, tuple)):
h = hashlib.new("md5")
for item in obj:
self.update(h, item)
return h.digest()
elif isinstance(obj, dict):
h = hashlib.new("md5")
for item in obj.items():
self.update(h, item)
return h.digest()
elif obj is None:
return b"0"
elif obj is True:
return b"1"
elif obj is False:
return b"0"
elif type_util.is_type(obj, "pandas.core.frame.DataFrame") or type_util.is_type(
obj, "pandas.core.series.Series"
):
import pandas as pd
if len(obj) >= _PANDAS_ROWS_LARGE:
obj = obj.sample(n=_PANDAS_SAMPLE_SIZE, random_state=0)
try:
return b"%s" % pd.util.hash_pandas_object(obj).sum()
except TypeError:
# Use pickle if pandas cannot hash the object for example if
# it contains unhashable objects.
return b"%s" % pickle.dumps(obj, pickle.HIGHEST_PROTOCOL)
elif type_util.is_type(obj, "numpy.ndarray"):
h = hashlib.new("md5")
self.update(h, obj.shape)
if obj.size >= _NP_SIZE_LARGE:
import numpy as np
state = np.random.RandomState(0)
obj = state.choice(obj.flat, size=_NP_SAMPLE_SIZE)
self.update(h, obj.tobytes())
return h.digest()
elif inspect.isbuiltin(obj):
return bytes(obj.__name__.encode())
elif type_util.is_type(obj, "builtins.mappingproxy") or type_util.is_type(
obj, "builtins.dict_items"
):
return self.to_bytes(dict(obj))
elif type_util.is_type(obj, "builtins.getset_descriptor"):
return bytes(obj.__qualname__.encode())
elif isinstance(obj, UploadedFile):
# UploadedFile is a BytesIO (thus IOBase) but has a name.
# It does not have a timestamp so this must come before
# temproary files
h = hashlib.new("md5")
self.update(h, obj.name)
self.update(h, obj.tell())
self.update(h, obj.getvalue())
return h.digest()
elif hasattr(obj, "name") and (
isinstance(obj, io.IOBase)
# Handle temporary files used during testing
or isinstance(obj, tempfile._TemporaryFileWrapper)
):
# Hash files as name + last modification date + offset.
# NB: we're using hasattr("name") to differentiate between
# on-disk and in-memory StringIO/BytesIO file representations.
# That means that this condition must come *before* the next
# condition, which just checks for StringIO/BytesIO.
h = hashlib.new("md5")
obj_name = getattr(obj, "name", "wonthappen") # Just to appease MyPy.
self.update(h, obj_name)
self.update(h, os.path.getmtime(obj_name))
self.update(h, obj.tell())
return h.digest()
elif isinstance(obj, Pattern):
return self.to_bytes([obj.pattern, obj.flags])
elif isinstance(obj, io.StringIO) or isinstance(obj, io.BytesIO):
# Hash in-memory StringIO/BytesIO by their full contents
# and seek position.
h = hashlib.new("md5")
self.update(h, obj.tell())
self.update(h, obj.getvalue())
return h.digest()
elif type_util.is_type(obj, "numpy.ufunc"):
# For numpy.remainder, this returns remainder.
return bytes(obj.__name__.encode())
elif inspect.ismodule(obj):
# TODO: Figure out how to best show this kind of warning to the
# user. In the meantime, show nothing. This scenario is too common,
# so the current warning is quite annoying...
# st.warning(('Streamlit does not support hashing modules. '
# 'We did not hash `%s`.') % obj.__name__)
# TODO: Hash more than just the name for internal modules.
return self.to_bytes(obj.__name__)
elif inspect.isclass(obj):
# TODO: Figure out how to best show this kind of warning to the
# user. In the meantime, show nothing. This scenario is too common,
# (e.g. in every "except" statement) so the current warning is
# quite annoying...
# st.warning(('Streamlit does not support hashing classes. '
# 'We did not hash `%s`.') % obj.__name__)
# TODO: Hash more than just the name of classes.
return self.to_bytes(obj.__name__)
elif isinstance(obj, functools.partial):
# The return value of functools.partial is not a plain function:
# it's a callable object that remembers the original function plus
# the values you pickled into it. So here we need to special-case it.
h = hashlib.new("md5")
self.update(h, obj.args)
self.update(h, obj.func)
self.update(h, obj.keywords)
return h.digest()
else:
# As a last resort, hash the output of the object's __reduce__ method
h = hashlib.new("md5")
try:
reduce_data = obj.__reduce__()
except BaseException as e:
raise UnhashableTypeError() from e
for item in reduce_data:
self.update(h, item)
return h.digest()
class NoResult:
"""Placeholder class for return values when None is meaningful."""
pass