Python md5 hashes of same gzipped file are inconsistent

Question

I am trying to zip a file using the python module gzip, and then hash the gzipped filed using hashlib. I have the following code:

import hashlib
import gzip

f_name = 'read_x.fastq'

for x in range(0,3):

    file = open(f_name, 'rb')

    myzip = gzip.open('test.gz', 'wb', compresslevel=1)

    n = 100000000
    try:
        print 'zipping ' + str(x)
        for chunk in iter(lambda: file.read(n), ''):
            myzip.write(chunk)
    finally:
        file.close()
        myzip.close()

    md5 = hashlib.md5()
    print 'hashing ' + str(x)
    with open('test.gz', 'r') as f:
        for chunk in iter(lambda: f.read(n), ''):
            md5.update(chunk)

    print md5.hexdigest()
    print '\n'

which I thought should simply zip the file, hash it and display the same output hash three times in a row. However, the output I get is:

zipping 0
hashing 0
7bd80798bce074c65928e0cf9d66cae4


zipping 1
hashing 1
a3bd4e126e0a156c5d86df75baffc294


zipping 2
hashing 2
85812a39f388c388cb25a35c4fac87bf

If I leave out the gzip step, and just hash the same gzipped file three times in a row, I do indeed get the same output three times:

hashing 0
ccfddd10c8fd1140db0b218124e7e9d3


hashing 1
ccfddd10c8fd1140db0b218124e7e9d3


hashing 2
ccfddd10c8fd1140db0b218124e7e9d3

Can anyone explain what is going on here? The issue must be that the gzip process is different each time. But as far as I knew, the DEFLATE algorithm is Huffman coding followed by LZ77 (a form of run-length-encoding) or LZ77 followed by Huffman, and therefore given identical input should produce identical output.

Answer 1

There are several reasons why compressing the exact same content will produce different gzip outputs:

• compression level. This you can control via the compress level parameter.
• The name of the original file which is in the header. This you can control if you use the gzip.GzipFile api rather than the gzip.open api.
• The modification time which is also in the header and can also be controlled with the gzip.GzipFile api.

So here is a piece of code that demonstrated the wrong and the right way to get reproducible output from python gzip:

import hashlib
import gzip

f_name = '/etc/passwd'
output_template = '/tmp/test{}.gz'

def digest(filename: str) -> str:
    md5 = hashlib.md5()
    with open(output_filename, 'rb') as f:
        for chunk in iter(lambda: f.read(block_size), b''):
            md5.update(chunk)
    return md5.hexdigest()

print("The default way - non identical outputs")
for x in range(0,3):
    input_handle = open(f_name, 'rb')
    output_filename = output_template.format(x)
    myzip = gzip.open(output_filename, 'wb')
    block_size = 4096
    try:
        for chunk in iter(lambda: input_handle.read(block_size), b''):
            myzip.write(chunk)
    finally:
        input_handle.close()
        myzip.close()
    print(digest(output_filename))

print("The right way to get identical outputs")
for x in range(3,6):
    input_handle = open(f_name, 'rb')
    output_filename = output_template.format(x)
    myzip = gzip.GzipFile(
        filename='',  # do not emit filename into the output gzip file
        mode='wb',
        fileobj=open(output_filename, 'wb'),
        mtime=0,
    )
    block_size = 4096
    try:
        for chunk in iter(lambda: input_handle.read(block_size), b''):
            myzip.write(chunk)
    finally:
        input_handle.close()
        myzip.close()
    print(digest(output_filename))

Answer 2

Oh but wait....apparently gzip adds timestamp info to the header of the gzip file, so the hash would be different.