>>> import locale
>>> locale.setlocale(locale.LC_NUMERIC, 'uk_UA')
'uk_UA'
>>> locale.getlocale(locale.LC_NUMERIC)
('uk_UA', 'KOI8-U')
>>> locale.nl_langinfo(locale.THOUSEP)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
UnicodeDecodeError: 'locale' codec can't decode byte 0x9a in position 0: Invalid or incomplete multibyte or wide character

Looks as locale.nl_langinfo() always uses the UTF-8 encoding (or may be locale.getpreferredencoding()).

Adding the test below to test__locale.py will reproduce the issue under the following conditions:

• The locale uk_UA is installed on your system.

• 'uk_UA': (',', '\xa0') is added to the known_numerics dictionary in this test file

  @unittest.skipUnless(nl_langinfo, "nl_langinfo is not available")
  def test_lc_numeric_not_char_nl_langinfo(self):
  # Test nl_langinfo against known values/
  # It should still work if there's a mismatch between
  # String & Numeric Locales
  tested = False
  for loc in candidate_locales:
  try:
  setlocale(LC_NUMERIC, loc)
  except Error:
  continue
  for li, lc in ((RADIXCHAR, "decimal_point"),
  (THOUSEP, "thousands_sep")):
  if self.numeric_tester('nl_langinfo', nl_langinfo(li), lc, loc):
  tested = True
  if not tested:
  self.skipTest('no suitable locales')

Thanks, Nina. We do have support in test.support for running a test with a specific locale (run_with_locale), so this could be turned into a unit test patch if you or someone else is willing to do that.

Since this bug has been reported, locale.localeconv() has been fixed in bpo-31900 to temporarily set LC_CTYPE to LC_NUMERIC to decode numeric fields of localeconv() from the proper encoding. I guess that a similar fix can be applied to locale.nl_langinfo(): set LC_CTYPE to LC_NUMERIC if the parameter is a numeric field.

I only knew locale.nl_langinfo(locale.CODESET). I didn't know that this function accepted other arguments :-)

I even wrote an article about these locale bugs :-)
#5191

See also bpo-35697: "decimal: formatter error if LC_NUMERIC uses a different encoding than LC_CTYPE".