Original bug report: https://bugzilla.redhat.com/show_bug.cgi?id=1484497

It seems that on the development branch of Fedora, when we updated glibc from 2.26 to 2.26.90, test_float_with_comma started failing.

Details from the original bug report:

Under certain circumstances, when LC_NUMERIC is fr_FR.ISO8859-1 but LC_ALL is C.UTF-8, locale.localeconv() fails with
    UnicodeDecodeError: 'locale' codec can't decode byte 0xa0 in position 0: Invalid or incomplete multibyte or wide character

Apparently, the thousands separator (or something else) in the lconv is "\xa0" (unbreakable space in fr_FR.ISO8859-1), and it's being decoded with UTF-8.

This is tripped by Python's test suite, namely test_float.GeneralFloatCases.test_float_with_comma

I can reproduce the bug with Python 3.6 on Fedora 26 and these locales:

* LC_ALL = LC_CTYPE = fr_FR (encoding = ISO8859-1)
* LC_NUMERIC= es_MX.utf8 (encoding = UTF-8)


Good: LC_NUMERIC = LC_CTYPE = LC_ALL = "es_MX.utf8"

haypo@selma$ env -i python3 -c 'import locale; locale.setlocale(locale.LC_ALL, "es_MX.utf8"); print(ascii(locale.localeconv()["thousands_sep"]))'

=> '\u2009'


Bug: LC_NUMERIC = "es_MX.utf8" but LC_CTYPE = LC_ALL = "fr_FR"

haypo@selma$ env -i python3 -c 'import locale; locale.setlocale(locale.LC_ALL, "fr_FR"); locale.setlocale(locale.LC_NUMERIC, "es_MX.utf8"); print(ascii(locale.localeconv()["thousands_sep"]))'

=> '\xe2\x80\x89'

Tested the PR on a system with glibc 2.26.90 where the test was failing, and it successfully passed.

This is a duplicate of issue28604. See also issue25812.

Same as #7442, I think.

Oh wow, this bug is older than what I expected :-) I added support for non-ASCII thousands separator in 2012:

https://bugs.python.org/issue13706#msg151733

inconsistent_locale_encodings.py of closed issue #7442 is interesting: I copy it here.

Pinging here. Is there some way I can help to move the issue forward?

Oh. Another Python function is impacted by the bug, str.format:

$ env -i python3 -c 'import locale; locale.setlocale(locale.LC_ALL, "fr_FR"); locale.setlocale(locale.LC_NUMERIC, "es_MX.utf8"); print(ascii(f"{1000:n}"))'
'1\xe2\x80\x89000'

It should be '1\u2009000' ('1', '\u2009', '000').

I completed my change. It now fixes locale.localeconv(), str.format() for int, float, complex and decimal.Decimal:

vstinner@apu$ ./python lc_numeric.py 
LC_CTYPE: ('fr_FR', 'ISO8859-1')
LC_NUMERIC: ('es_MX', 'UTF-8')
decimal_point: '.'
thousands_sep: '\u2009'
grouping: [3, 3, 0]
int.__format__: '1\u2009234'
float.__format__: '1\u2009234'
complex.__format__: '1\u2009234+0j'
Decimal.__format__: '1\u2009234'

Update: I pushed a large change to fix locale encodings in bpo-29240: commit 7ed7aead9503102d2ed316175f198104e0cd674c.

Oops lc_numeric.py contains a typo:

d = decimal.Decimal(1234)
print("Decimal.__format__: %a" % f"{i:n}")

=> it should be f"{d:n}"

Just FYI: LC_ALL has precedence over all other more specific LC_* settings:

http://pubs.opengroup.org/onlinepubs/7908799/xbd/envvar.html
http://man7.org/linux/man-pages/man7/locale.7.html

Please confirm the bug without having LC_ALL or LANG set. Thanks.

> Please confirm the bug without having LC_ALL or LANG set.

lc_numeric.py uses:

  locale.setlocale(locale.LC_ALL, "fr_FR")

Are you talking about that? What is the problem with this configuration?

I'm sure that there is a bug :-) You aren't able to reproduce it? What is your operating system?

I just wanted to note that the description and title may cause a wrong interpretation of what should happen:

If you first set LC_ALL and then one of the other categories such as LC_NUMERIC, locale C functions will still use the LC_ALL setting for everything. LC_NUMERIC does not override the LC_ALL setting.

I tested this on OpenSUSE and get the same wrong results. Apparently, locale.localeconv() does not respect the above order. That's a bug.

I'm not sure whether the OP's quoted behavior is a bug, though, since if the locale encoding is not UTF-8, you cannot really expect using UTF-8 numeric separators to output correctly.

On Mon, Jan 15, 2018 at 12:37:28PM +0000, Marc-Andre Lemburg wrote:
> If you first set LC_ALL and then one of the other categories such as LC_NUMERIC, locale C functions will still use the LC_ALL setting for everything. LC_NUMERIC does not override the LC_ALL setting.

I have the exact same questions as Marc-Andre.  This is one of the reasons why I
blocked the _decimal change.  I don't fully understand the role of the new glibc,
since #7442 has existed for ages -- and it is a open question whether it is a bug
or not.

Both views are reasonable IMO.

Marc-Andre Lemburg: "If you first set LC_ALL and then one of the other categories such as LC_NUMERIC, locale C functions will still use the LC_ALL setting for everything. LC_NUMERIC does not override the LC_ALL setting."

The root of this issue is https://bugzilla.redhat.com/show_bug.cgi?id=1484497#c0:

Petr Viktorin reproducer scripts uses Python locale.setlocale(), not environment variables:
https://gist.github.com/encukou/70b3d3f9ef3e29ac1dbc23a5f7bd6431
---
locale.setlocale(locale.LC_ALL, 'C.UTF-8')
locale.setlocale(locale.LC_NUMERIC, 'fr_FR.ISO8859-1')
---

Example of Fedora 27 and Python 3.6:

vstinner@apu$ env -i LC_NUMERIC=uk_UA.koi8u python3 -c 'import locale; print(locale.setlocale(locale.LC_ALL, "")); print(locale.getpreferredencoding(), ascii(locale.localeconv()["thousands_sep"]))'
LC_CTYPE=C.UTF-8;LC_NUMERIC=uk_UA.koi8u;LC_TIME=C;LC_COLLATE=C;LC_MONETARY=C;LC_MESSAGES=C;LC_PAPER=C;LC_NAME=C;LC_ADDRESS=C;LC_TELEPHONE=C;LC_MEASUREMENT=C;LC_IDENTIFICATION=C
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/usr/lib64/python3.6/locale.py", line 110, in localeconv
    d = _localeconv()
UnicodeDecodeError: 'locale' codec can't decode byte 0x9a in position 0: Invalid or incomplete multibyte or wide character

"env -i" starts Python in an empty environment. It seems like LC_CTYPE defaults to C.UTF-8 in this case.

* LC_CTYPE = C.UTF-8, encoding = UTF-8
* LC_NUMERIC = uk_UA.koi8u, encoding = KOI8-U


With my PR, it works:

vstinner@apu$ env -i LC_NUMERIC=uk_UA.koi8u ./python -c 'import locale; print(locale.setlocale(locale.LC_ALL, "")); print(locale.getpreferredencoding(), ascii(locale.localeconv()["thousands_sep"]))'
LC_CTYPE=C.UTF-8;LC_NUMERIC=uk_UA.koi8u;LC_TIME=C;LC_COLLATE=C;LC_MONETARY=C;LC_MESSAGES=C;LC_PAPER=C;LC_NAME=C;LC_ADDRESS=C;LC_TELEPHONE=C;LC_MEASUREMENT=C;LC_IDENTIFICATION=C
UTF-8 '\xa0'

=> thousands_sep byte string b'\x9A' is decoded as the Uniode string '\xa0'.


vstinner@apu$ env -i LC_NUMERIC=uk_UA.koi8u ./python -c 'import locale; locale.setlocale(locale.LC_ALL, ""); print(ascii(f"{1234:n}"))'
'1\xa0234'

=> the number is properly formatted


vstinner@apu$ env -i LC_NUMERIC=uk_UA.koi8u ./python -c 'import locale; locale.setlocale(locale.LC_ALL, ""); print(f"{1234:n}")'
1 234

It's possible to display the result using print().

Ok, it seems that the C setlocale() itself does not follow the conventions set forth for environment variables:

http://pubs.opengroup.org/onlinepubs/7908799/xsh/setlocale.html

(see the example at the bottom)

So the behavior shown by Python's setlocale() is fine.

However, that still doesn't magically make this work:

locale.setlocale(locale.LC_ALL, 'C.UTF-8')
locale.setlocale(locale.LC_NUMERIC, 'fr_FR.ISO8859-1')

If LC_NUMERIC uses a different encoding than LC_ALL, there's really no surprise in having numeric formatting fail. localeconv() will output the set encoding for the numeric string conversion and Python will decode this using the locale encoding set by LC_ALL. If those two are different, you run into problems.

I would not consider this a bug in Python, but rather in the locale settings passed to setlocale().

The technical issue here is that the libc has no "stateless" function to process bytes and text with one specific locale. All functions rely on the *current* locales. To decode byte strings, we use mbstowcs(), and this function relies on the current LC_CTYPE locale, whereas decimal_point and thousands_sep should be decoded from the current LC_NUMERIC locale.

Indeed. The major problem with all libc locale functions is that they are not thread safe. The GIL does help a bit protecting against corrupted data, though.

> I would not consider this a bug in Python, but rather in the locale settings passed to setlocale().

Past 10 years, I repeated to every single user I met that "Python 3 is right, your system setup is wrong". But that's a waste of time. People continue to associate Python3 and Unicode to annoying bugs, because they don't understand how locales work.

Instead of having to repeat to each user that "hum, maybe your config is wrong", I prefer to support this non convential setup and work as expected ("it just works"). With my latest implementation, setlocale() is only done when LC_CTYPE and LC_NUMERIC are different, which is the corner case which "shouldn't occur in practice".

Sounds like a good compromise :-)

I tested localeconv() with PR 4174 on FreeBSD:
--
locale.setlocale(locale.LC_ALL, "C")
locale.setlocale(locale.LC_NUMERIC, "ar_SA.UTF-8")
--

It works as expected, result:
--
decimal_point: '\u066b'
thousands_sep: '\u066c'
--

Compare it to Python 3.6 which returns mojibake, it seems like bytes are decoded from Latin1:
--
decimal_point: '\xd9\xab'
thousands_sep: '\xd9\xac'
--

Raw byte strings, Python 2.7:

* decimal_point: b'\xd9\xab'
* thousands_sep: b'\xd9\xac'

Test on Linux (Fedora 27, glibc 2.26):

locale.setlocale(locale.LC_ALL, "fr_FR")
locale.setlocale(locale.LC_NUMERIC, "es_MX.utf8")

It works as expected, result:

decimal_point: '.'
thousands_sep: '\u2009'

Python 3.6 returns mojibake:

decimal_point: '.'
thousands_sep: '\xe2\x80\x89'

Python 2.7 raw strings, thousands_sep = b'\xE2\x80\x89'.

On macOS 10.13.2, I failed to find any non-ASCII decimal_point or thousands_sep in localeconv(). I wrote a script to find all non-ASCII data in all locales:
https://github.com/vstinner/misc/blob/master/python/all_locales.py

New changeset cb064fc2321ce8673fe365e9ef60445a27657f54 by Victor Stinner in branch 'master':
bpo-31900: Fix localeconv() encoding for LC_NUMERIC (#4174)
https://github.com/python/cpython/commit/cb064fc2321ce8673fe365e9ef60445a27657f54

lc_numeric.py contains a typo, used fixed lc_numeric2.py instead to test my PR 5191 which fixes decimal.Decimal.

New changeset 5f959c4f9eca404b8bc4bc6348fed27c4b907b89 by Victor Stinner in branch '3.6':
[3.6] bpo-31900: Fix localeconv() encoding for LC_NUMERIC (#4174) (#5192)
https://github.com/python/cpython/commit/5f959c4f9eca404b8bc4bc6348fed27c4b907b89

> The technical issue here is that the libc has no "stateless" function to process bytes and text with one specific locale.

That's not true.  There is a rich set of *_l functions that take a locale_t object and operate on that locale.

Victor:
> The technical issue here is that the libc has no "stateless" function to process bytes and text with one specific locale.

Andreas Schwab:
> That's not true.  There is a rich set of *_l functions that take a locale_t object and operate on that locale.

Oh. Do you want to work on a patch to use these functions? If yes, please open a new issue to enhance the code.

See also bpo-28604: localeconv() doesn't support LC_MONETARY encoding different than LC_CTYPE encoding.

The initial bug has been fixed, I close the issue.