I think I have found a small typo-bug in tarfile.py, that seems to
be present in cpython upstream, which makes tarfile compression slower.
The issue can be seen here, in line 415 [1] of tarfile.py:

        self.cmp = self.zlib.compressobj(9, self.zlib.DEFLATED,
                                            -self.zlib.MAX_WBITS,
                                            self.zlib.DEF_MEM_LEVEL,
                                            0)

The minus sign doesn't make sense to me, because zlib.MAX_WBITS is 15,
and according to the documentation [2] wbits is by default 15 and
"This should be an integer from 8 to 15. Higher values give better
compression, but use more memory". -15 is not even in range. So I guess the expected value should be the zlib default, 15. Or maybe another value, but it should at least be in range. 

I marked it as "performance" type bug as this only really affects performance.

I might have gotten it all wrong and it's fine the way it is, but I thought it'd be best to report it, as it looked fishy to me.

--
[1] http://hg.python.org/cpython/file/94d0e842b9ea/Lib/tarfile.py#l415
[2] https://docs.python.org/3.4/library/zlib.html#zlib.compressobj

Is the thinking and hence the patch correct here?

I suspect the patch is wrong and the zlib documentation needs fixing instead. See <https://docs.python.org/dev/library/zlib.html#zlib.decompress> for a possible explanation of the negative sign, although it would make more sense for compress() which would actually create a “gzip” header.

Eduardo’s patch causes many tests to fail. Abbreviated test output:

$ ./python -bWall -m test -v test_tarfile
======================================================================
ERROR: test_compare_members (test.test_tarfile.GzipStreamReadTest)
ERROR: test_empty_tarfile (test.test_tarfile.GzipStreamReadTest)
ERROR: test_fileobj_regular_file (test.test_tarfile.GzipStreamReadTest)
ERROR: test_ignore_zeros (test.test_tarfile.GzipStreamReadTest)
ERROR: test_non_existent_tarfile (test.test_tarfile.GzipStreamReadTest)
ERROR: test_null_tarfile (test.test_tarfile.GzipStreamReadTest)
ERROR: test_provoke_stream_error (test.test_tarfile.GzipStreamReadTest)
ERROR: test_read_through (test.test_tarfile.GzipStreamReadTest)
----------------------------------------------------------------------
tarfile.ReadError: invalid compressed data

======================================================================
ERROR: test_stream_padding (test.test_tarfile.GzipStreamWriteTest)
----------------------------------------------------------------------
zlib.error: Error -3 while decompressing data: invalid stored block lengths

======================================================================
FAIL: test_detect_file (test.test_tarfile.GzipDetectReadTest)
FAIL: test_detect_fileobj (test.test_tarfile.GzipDetectReadTest)
AssertionError

----------------------------------------------------------------------
Ran 387 tests in 14.216s

FAILED (failures=2, errors=9, skipped=8)

There is already Issue 5784 open about clarifying the “wbits” parameter, so I think we can close this as a duplicate.