The Unzip module is always unbuffered (tested v.3.1.2 Windows XP, 32-bit). This means that if one has to do many small reads it is a lot slower than reading a chunk of data to a buffer and then reading from that buffer. It seems logical that the unzip module should default to buffered reading and/or have a buffered argument. Likewise, the documentation should clarify that there is no buffering involved when doing a read, which runs contrary to the default behavior of a normal read.

start Zipfile read
done
27432 reads done
took 0.859 seconds
start buffered Zipfile read
done
27432 reads done
took 0.072 seconds
start normal read (default buffer)
done
27432 reads done
took 0.139 seconds
start buffered normal read
done
27432
took 0.137 seconds

I should clarify that this is the zipfile constructor I am using:

zipfile.ZipFile(filename, mode='r', allowZip64=True);

Actually reading from the zip file is buffered (at least 4 KiB of uncompressed data at a time).

Can you give tests, scripts and data, which show the problem?

See attached, which will open a zipfile that contains one file and reads it a bunch of times using unbuffered and buffered idioms. This was tested on windows using python 3.2

You're in charge of coming up with a file to test it on. Sorry.

Example output:

Enter filename: test.zip
Timing unbuffered read, 5 bytes at a time. 10 loops
took 6.671999931335449
Timing buffered read, 5 bytes at a time (4000 byte buffer). 10 loops
took 0.7350001335144043

This is not because zipfile module is unbuffered. This is the difference between expensive function call and cheap bytes slicing. Replace zf.open(namelist [0]) to io.BufferedReader(zf.open(namelist [0])) to see the effect of a good buffering. In 3.2 zipfile read() implemented not optimal, so it slower (twice), but in 3.3 it will be almost as fast as using io.BufferedReader. It is still several times more slowly than bytes slicing, but there's nothing you can do with it.

Here is a patch, which is speeds up (+20%) the reading from a zip file by small chunks. Microbenchmark:

./python -m zipfile -c test.zip python
./python -m timeit -n 1 -s "import zipfile;zf=zipfile.ZipFile('test.zip')" "with zf.open('python') as f:" " while f.read(1):pass"

Python 3.3 (vanilla): 1 loops, best of 3: 36.4 sec per loop
Python 3.3 (patched): 1 loops, best of 3: 30.1 sec per loop
Python 3.3 (with io.BufferedReader): 1 loops, best of 3: 30.2 sec per loop
And, for comparison, Python 3.2: 1 loops, best of 3: 74.5 sec per loop

Thank you, Martin, now I understood why not work Rietveld review.

The patch updated to reflect Martin's stylistic comments.

Sorry for the delay, Martin. I have not received an email with your review from 2012-05-13, and only today accidentally discovered your comments in Rietveld. It seems to have been some bug in Rietveld.

Martin, now the patch is good?

Any chance to commit the patch before final feature freeze?

Patch looks fine to me.

Antoine, can you commit this? I'm currently away from the computer that
has my SSH key on it.

New changeset 0e8285321659 by Antoine Pitrou in branch 'default':
On behalf of Nadeem Vawda: issue bpo-10376: micro-optimize reading from a Zipfile.
http://hg.python.org/cpython/rev/0e8285321659

Antoine, can you commit this?

Ok, done.