http://hg.python.org/cpython/file/8d5df9602a72/Lib/ConfigParser.py#l529

RawConfigParser has a special case of replacing '""' with '' in reader.
However the writer does not do the same.

This may cause problems, for example if you merge multiple ini files into one and the 3rd party program that should use the output segfaults on an ini that has `key=` instead of `key=value`, as it is in my case. :)

I think that the reasoning with the reader maybe was that an empty string must be an empty python string. 
However there is no option to control how it gets written back to the output. One can of course walkaround it by replacing the '' values with '""', but I think the standard library is ought to be consistent, walkarounds are unpythonic, a raw parser should be transparent either way, and there should preferrably be one obvious way to do it.

I think this is not just unintuitive (especially in a parser called "raw"), but goes against the philosophy of "refusing the temptation to guess". 

As I know, the syntax of .ini files is not exactly well-defined, so the best option in my opinion would be allowing to configure the parser how to treat these values.

Whether or not adding this option is viable, I think the problem still remains that the writer now works inconsistently of the reader.

typo: 
    `key=` instead of `key=value` 
was meant as
    `key=` instead of `key=""`

A configurable option is probably the only way forward, due to backward compatibility reasons, but Łukasz will know for sure.  If so, it can only go in the *next* version of Python, and while we haven't hit beta yet on 3.4 it may be too late for anyone to have time to get it in before then.

Also, I was under the impression that 'key=' was the canonical way to indicate an empty value in an ini file.  As you say, the format is not well defined (no surprise).

I guess it's logical to have the value "" turned to '' in the reader, but I would expect the same transformation in the writer.

If the write function would write "" for an empty string, will it solve your issue?

This bug has been fixed in Python 3.2 (see 892236137db9). Please use the configparser backport for Python 2.7.