Python 3 - Can pickle handle byte objects larger than 4GB?

Question

Based on this comment and the referenced documentation, Pickle 4.0+ from Python 3.4+ should be able to pickle byte objects larger than 4 GB.

However, using python 3.4.3 or python 3.5.0b2 on Mac OS X 10.10.4, I get an error when I try to pickle a large byte array:

>>> import pickle
>>> x = bytearray(8 * 1000 * 1000 * 1000)
>>> fp = open("x.dat", "wb")
>>> pickle.dump(x, fp, protocol = 4)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
OSError: [Errno 22] Invalid argument

Is there a bug in my code or am I misunderstanding the documentation?

Answer 1

Here is a simple workaround for issue 24658. Use pickle.loads or pickle.dumps and break the bytes object into chunks of size 2**31 - 1 to get it in or out of the file.

import pickle
import os.path

file_path = "pkl.pkl"
n_bytes = 2**31
max_bytes = 2**31 - 1
data = bytearray(n_bytes)

## write
bytes_out = pickle.dumps(data)
with open(file_path, 'wb') as f_out:
    for idx in range(0, len(bytes_out), max_bytes):
        f_out.write(bytes_out[idx:idx+max_bytes])

## read
bytes_in = bytearray(0)
input_size = os.path.getsize(file_path)
with open(file_path, 'rb') as f_in:
    for _ in range(0, input_size, max_bytes):
        bytes_in += f_in.read(max_bytes)
data2 = pickle.loads(bytes_in)

assert(data == data2)

Answer 2

To sum up what was answered in the comments:

Yes, Python can pickle byte objects bigger than 4GB. The observed error is caused by a bug in the implementation (see Issue24658).

Answer 3

Here is the full workaround, though it seems pickle.load no longer tries to dump a huge file anymore (I am on Python 3.5.2) so strictly speaking only the pickle.dumps needs this to work properly.

import pickle

class MacOSFile(object):

    def __init__(self, f):
        self.f = f

    def __getattr__(self, item):
        return getattr(self.f, item)

    def read(self, n):
        # print("reading total_bytes=%s" % n, flush=True)
        if n >= (1 << 31):
            buffer = bytearray(n)
            idx = 0
            while idx < n:
                batch_size = min(n - idx, 1 << 31 - 1)
                # print("reading bytes [%s,%s)..." % (idx, idx + batch_size), end="", flush=True)
                buffer[idx:idx + batch_size] = self.f.read(batch_size)
                # print("done.", flush=True)
                idx += batch_size
            return buffer
        return self.f.read(n)

    def write(self, buffer):
        n = len(buffer)
        print("writing total_bytes=%s..." % n, flush=True)
        idx = 0
        while idx < n:
            batch_size = min(n - idx, 1 << 31 - 1)
            print("writing bytes [%s, %s)... " % (idx, idx + batch_size), end="", flush=True)
            self.f.write(buffer[idx:idx + batch_size])
            print("done.", flush=True)
            idx += batch_size


def pickle_dump(obj, file_path):
    with open(file_path, "wb") as f:
        return pickle.dump(obj, MacOSFile(f), protocol=pickle.HIGHEST_PROTOCOL)


def pickle_load(file_path):
    with open(file_path, "rb") as f:
        return pickle.load(MacOSFile(f))

Answer 4

You can specify the protocol for the dump. If you do pickle.dump(obj,file,protocol=4) it should work.

Answer 5

Reading a file by 2GB chunks takes twice as much memory as needed if bytes concatenation is performed, my approach to loading pickles is based on bytearray:

class MacOSFile(object):
    def __init__(self, f):
        self.f = f

    def __getattr__(self, item):
        return getattr(self.f, item)

    def read(self, n):
        if n >= (1 << 31):
            buffer = bytearray(n)
            pos = 0
            while pos < n:
                size = min(n - pos, 1 << 31 - 1)
                chunk = self.f.read(size)
                buffer[pos:pos + size] = chunk
                pos += size
            return buffer
        return self.f.read(n)

Usage:

with open("/path", "rb") as fin:
    obj = pickle.load(MacOSFile(fin))

Answer 6

Had the same issue and fixed it by upgrading to Python 3.6.8.

This seems to be the PR that did it: https://github.com/python/cpython/pull/9937

Answer 7

I also found this issue, to solve this problem i chunk the code into several iteration. Let say in this case i have 50.000 data which i have to calc tf-idf and do knn classfication. When i run and directly iterate 50.000 it give me "that error". So, to solve this problem i chunk it.

tokenized_documents = self.load_tokenized_preprocessing_documents()
    idf = self.load_idf_41227()
    doc_length = len(documents)
    for iteration in range(0, 9):
        tfidf_documents = []
        for index in range(iteration, 4000):
            doc_tfidf = []
            for term in idf.keys():
                tf = self.term_frequency(term, tokenized_documents[index])
                doc_tfidf.append(tf * idf[term])
            doc = documents[index]
            tfidf = [doc_tfidf, doc[0], doc[1]]
            tfidf_documents.append(tfidf)
            print("{} from {} document {}".format(index, doc_length, doc[0]))

        self.save_tfidf_41227(tfidf_documents, iteration)