"It's not a bug, you must fix your locale" is not an acceptable answer

Since Python 3.0 was released in 2008, the usual answer to users getting Unicode errors is to ask developers to fix their code to handle Unicode properly. Most applications and Python modules were fixed, but users keep reporting Unicode errors regulary: see the long list of issues in the Links section below.

In fact, a second class of bug comes from a locale which is not properly configured. The usual answer to such bug report is: "it is not a bug, you must fix your locale".

Technically, the answer is correct, but from a practical point of view, the answer is not acceptable. In many cases, "fixing the issue" is an hard task. Moreover, sometimes, the usage of the POSIX locale is deliberate.

A good example of a concrete issue are build systems which create a fresh environment for each build using a chroot, a container, a virtual machine or something else to get reproductible builds. Such setup usually uses the POSIX locale. To get 100% reproductible builds, the POSIX locale is a good choice: see the Locales section of reproducible-builds.org.

UNIX users don't expect Unicode errors, since the common command lines tools like cat, grep or sed never fail with Unicode errors. These users expect that Python 3 "just works" with any locale and don't bother them with encodings. From their point of the view, the bug is not their locale but is obviously Python 3.

Since Python 2 handles data as bytes, it's rarer in Python 2 compared to Python 3 to get Unicode errors. It also explains why users also perceive Python 3 as the root cause of their Unicode errors.

Some users expect that Python 3 just works with any locale and so don't bother with mojibake, whereas some developers are working hard to prevent mojibake and so expect that Python 3 fails early before creating mojibake.

Since different group of users have different expectations, there is no silver bullet which solves all issues at once. Last but not least, backward compatibility should be preserved whenever possible.

Locale and operating system data

Python uses an encoding called the "filesystem encoding" to decide how to encode and decode data from/to the operating system:

• file content
• command line arguments: sys.argv
• standard streams: sys.stdin, sys.stdout, sys.stderr
• environment variables: os.environ
• filenames: os.listdir(str) for example
• pipes: subprocess.Popen using subprocess.PIPE for example
• error messages: os.strerror(code) for example
• user and terminal names: os, grp and pwd modules
• host name, UNIX socket path: see the socket module
• etc.

At startup, Python calls setlocale(LC_CTYPE, "") to use the user LC_CTYPE locale and then store the locale encoding as the "filesystem error". It's possible to get this encoding using sys.getfilesystemencoding(). In the whole lifetime of a Python process, the same encoding and error handler are used to encode and decode data from/to the operating system.

The os.fsdecode() and os.fsencode() functions can be used to decode and encode operating system data. These functions use the filesystem error handler: sys.getfilesystemencodeerrors().

The POSIX locale and its encoding

The following environment variables are used to configure the locale, in this preference order:

• LC_ALL, most important variable
• LC_CTYPE
• LANG

The POSIX locale,also known as "the C locale", is used:

• if the first set variable is set to "C"
• if all these variables are unset, for example when a program is started in an empty environment.

The encoding of the POSIX locale must be ASCII or a superset of ASCII.

On Linux, the POSIX locale uses the ASCII encoding.

On FreeBSD and Solaris, nl_langinfo(CODESET) announces an alias of the ASCII encoding, whereas mbstowcs() and wcstombs() functions use the ISO 8859-1 encoding (Latin1) in practice. The problem is that os.fsencode() and os.fsdecode() use Python codec of the locale encoding. For example, if command line arguments are decoded by mbstowcs() and encoded back by os.fsencode(), an UnicodeEncodeError exception is raised instead of retrieving the original byte string.

To fix this issue, Python checks since Python 3.4 if mbstowcs() really uses the ASCII encoding if the the LC_CTYPE uses the the POSIX locale and nl_langinfo(CODESET) returns "ASCII" (or an alias to ASCII). If not (the effective encoding is not ASCII), Python uses its own ASCII codec instead of using mbstowcs() and wcstombs() functions for operating system data.

See the POSIX locale (2016 Edition).

POSIX locale used by mistake

In many cases, the POSIX locale is not really expected by users who get it by mistake. Examples:

• program started in an empty environment
• User forcing LANG=C to get messages in english
• LANG=C used for bad reasons, without being aware of the ASCII encoding
• SSH shell
• Linux installed with no configured locale
• chroot environment, Docker image, container, ... with no locale is configured
• User locale set to a non-existing locale, typo in the locale name for example

C.UTF-8 and C.utf8 locales

Some UNIX operating systems provide a variant of the POSIX locale using the UTF-8 encoding:

• Fedora 25: "C.utf8" or "C.UTF-8"
• Debian (eglibc 2.13-1, 2011), Ubuntu: "C.UTF-8"
• HP-UX: "C.utf8"

It was proposed to add a C.UTF-8 locale to the glibc: glibc C.UTF-8 proposal.

It is not planned to add such locale to BSD systems.

Popularity of the UTF-8 encoding

Python 3 uses UTF-8 by default for Python source files.

On Mac OS X, Windows and Android, Python always use UTF-8 for operating system data. For Windows, see the PEP 529 [2]: "Change Windows filesystem encoding to UTF-8".

On Linux, UTF-8 became the de facto standard encoding, replacing legacy encodings like ISO 8859-1 or ShiftJIS. For example, using different encodings for filenames and standard streams is likely to create mojibake, so UTF-8 is now used everywhere.

The UTF-8 encoding is the default encoding of XML and JSON file format. In January 2017, UTF-8 was used in more than 88% of web pages (HTML, Javascript, CSS, etc.).

See utf8everywhere.org for more general information on the UTF-8 codec.

Old data stored in different encodings and surrogateescape

Even if UTF-8 became the de facto standard, there are still systems in the wild which don't use UTF-8. And there are a lot of data stored in different encodings. For example, an old USB key using the ext3 filesystem with filenames encoded to ISO 8859-1.

The Linux kernel and the libc don't decode filenames: a filename is used as a raw array of bytes. The common solution to support any filename is to store filenames as bytes and don't try to decode them. When displayed to stdout, mojibake is displayed if the filename and the terminal don't use the same encoding.

Python 3 promotes Unicode everywhere including filenames. A solution to support filenames not decodable from the locale encoding was found: the surrogateescape error handler (PEP 383 [4]), store undecodable bytes as surrogate characters. This error handler is used by default for operating system data, by os.fsdecode() and os.fsencode() for example (except on Windows which uses the strict error handler).

Standard streams

Python uses the locale encoding for standard streams: stdin, stdout and stderr. The strict error handler is used by stdin and stdout to prevent mojibake.

The backslashreplace error handler is used by stderr to avoid Unicode encode error when displaying non-ASCII text. It is especially useful when the POSIX locale is used, because this locale usually uses the ASCII encoding.

The problem is that operating system data like filenames are decoded using the surrogateescape error handler (PEP 383 [4]). Displaying a filename to stdout raises a Unicode encode error if the filename contains an undecoded byte stored as a surrogate character.

Python 3.6 now uses surrogateescape for stdin and stdout if the POSIX locale is used: issue #19977. The idea is to passthrough operating system data even if it means mojibake, because most UNIX applications work like that. Most UNIX applications store filenames as bytes, usually simply because bytes are first-citizen class in the used programming language, whereas Unicode is badly supported.

Changes

Add a new UTF-8 mode, disabled by default, to ignore the locale and force the usage of the UTF-8 encoding with the surrogateescape error handler, instead using the locale encoding (with strict or surrogateescape error handler depending on the case).

Basically, the UTF-8 mode behaves as Python 2: it "just works" and don't bother users with encodings, but it can produce mojibake. It can be configured as strict to prevent mojibake: the UTF-8 encoding is used with the strict error handler for inputs and outputs, but the surrogateescape error handler is still used for operating system data.

New -X utf8 command line option and PYTHONUTF8 environment variable are added to control the UTF-8 mode. The UTF-8 mode is enabled by -X utf8 or PYTHONUTF8=1. The UTF-8 is configured as strict by -X utf8=strict or PYTHONUTF8=strict. Other option values fail with an error.

The POSIX locale enables the UTF-8 mode. In this case, the UTF-8 mode can be explicitly disabled by -X utf8=0 or PYTHONUTF8=0.

Options priority for the UTF-8 mode:

• PYTHONLEGACYWINDOWSFSENCODING
• -X utf8
• PYTHONUTF8
• POSIX locale

For example, PYTHONUTF8=0 python3 -X utf8 enables the UTF-8 mode, whereas LC_ALL=C python3.7 -X utf8=0 disables the UTF-8 mode and so use the encoding of the POSIX locale.

Encodings used by open(), highest priority first:

• encoding and errors parameters (if set)
• UTF-8 mode
• os.device_encoding(fd)
• os.getpreferredencoding(False)

Encoding and error handler

The UTF-8 mode changes the default encoding and error handler used by open(), os.fsdecode(), os.fsencode(), sys.stdin, sys.stdout and sys.stderr:

By comparison, Python 3.6 uses:

The UTF-8 mode uses the surrogateescape error handler instead of the strict mode for convenience: the idea is that data not encoded to UTF-8 are passed through "Python" without being modified, as raw bytes.

The PYTHONIOENCODING environment variable has the priority over the UTF-8 mode for standard streams. For example, PYTHONIOENCODING=latin1 python3 -X utf8 uses the Latin1 encoding for stdin, stdout and stderr.

Encoding and error handler on Windows

On Windows, the encodings and error handlers are different:

By comparison, Python 3.6 uses:

The "Legacy Windows FS encoding" is enabled by setting the PYTHONLEGACYWINDOWSFSENCODING environment variable to 1, see the PEP 529.

Enabling the legacy Windows filesystem encoding disables the UTF-8 mode (as -X utf8=0).

If stdin and/or stdout is redirected to a pipe, sys.stdin and/or sys.output uses mbcs encoding by default, rather than UTF-8. But with the UTF-8 mode, sys.stdin and sys.stdout always use the UTF-8 encoding.

There is no POSIX locale on Windows. The ANSI code page is used to the locale encoding, and this code page never uses the ASCII encoding.

Rationale

The UTF-8 mode is disabled by default to keep hard Unicode errors when encoding or decoding operating system data failed, and to keep the backward compatibility. The user is responsible to enable explicitly the UTF-8 mode, and so is better prepared for mojibake than if the UTF-8 mode would be enabled by default.

The UTF-8 mode should be used on systems known to be configured with UTF-8 where most applications speak UTF-8. It prevents Unicode errors if the user overrides a locale by mistake or if a Python program is started with no locale configured (and so with the POSIX locale).

Most UNIX applications handle operating system data as bytes, so LC_ALL, LC_CTYPE and LANG environment variables have a limited impact on how these data are handled by the application.

The Python UTF-8 mode should help to make Python more interoperable with the other UNIX applications in the system assuming that UTF-8 is used everywhere and that users expect UTF-8.

Ignoring LC_ALL, LC_CTYPE and LANG environment variables in Python is more convenient, since they are more commonly misconfigured by mistake (configured to use an encoding different than UTF-8, whereas the system uses UTF-8), rather than being misconfigured by intent.

Expected mojibake and surrogate character issues

The UTF-8 mode only affects code running directly in Python, especially code written in pure Python. The other code, called "external code" here, is not aware of this mode. Examples:

• C libraries called by Python modules like OpenSSL
• The application code when Python is embedded in an application

In the UTF-8 mode, Python uses the surrogateescape error handler which stores bytes not decodable from UTF-8 as surrogate characters.

If the external code uses the locale and the locale encoding is UTF-8, it should work fine.

List a directory into stdout

Script listing the content of the current directory into stdout:

import os
for name in os.listdir(os.curdir):
    print(name)

Result:

"No" means that the script can fail on decoding or encoding a filename depending on the locale or the filename.

To be able to always work, the program must be able to produce mojibake. Mojibake is more user friendly than an error with a truncated or empty output.

Example with a directory which contains the file called b'xxx\xff' (the byte 0xFF is invalid in UTF-8).

Default and UTF-8 Strict mode fail on print() with an encode error:

$ python3.7 ../ls.py
Traceback (most recent call last):
  File "../ls.py", line 5, in <module>
    print(name)
UnicodeEncodeError: 'utf-8' codec can't encode character '\udcff' ...

$ python3.7 -X utf8=strict ../ls.py
Traceback (most recent call last):
  File "../ls.py", line 5, in <module>
    print(name)
UnicodeEncodeError: 'utf-8' codec can't encode character '\udcff' ...

The UTF-8 mode, POSIX locale, Python 2 and the UNIX ls command work but display mojibake:

$ python3.7 -X utf8 ../ls.py
xxx�

$ LC_ALL=C /python3.6 ../ls.py
xxx�

$ python2 ../ls.py
xxx�

$ ls
'xxx'$'\377'

List a directory into a text file

Similar to the previous example, except that the listing is written into a text file:

import os
names = os.listdir(os.curdir)
with open("/tmp/content.txt", "w") as fp:
    for name in names:
        fp.write("%s\n" % name)

Result:

"Yes" involves that mojibake can be produced. "No" means that the script can fail on decoding or encoding a filename depending on the locale or the filename. Typical error:

$ LC_ALL=C python3 test.py
Traceback (most recent call last):
  File "test.py", line 5, in <module>
    fp.write("%s\n" % name)
UnicodeEncodeError: 'ascii' codec can't encode characters in position 0-1: ordinal not in range(128)

Display Unicode characters into stdout

Very basic example used to illustrate a common issue, display the euro sign (U+20AC: €):

print("euro: \u20ac")

Result:

The UTF-8 and UTF-8 Strict modes will always encode the euro sign as UTF-8. If the terminal uses a different encoding, we get mojibake.

Replace a word in a text

The following scripts replaces the word "apple" with "orange". It reads input from stdin and writes the output into stdout:

import sys
text = sys.stdin.read()
sys.stdout.write(text.replace("apple", "orange"))

Result:

Producer-consumer model using pipes

Let's say that we have a "producer" program which writes data into its stdout and a "consumer" program which reads data from its stdin.

On a shell, such programs are run with the command:

producer | consumer

The question if these programs will work with any data and any locale. UNIX users don't expect Unicode errors, and so expect that such programs "just works".

If the producer only produces ASCII output, no error should occur. Let's say the that producer writes at least one non-ASCII character (at least one byte in the range 0x80..0xff).

To simplify the problem, let's say that the consumer has no output (don't write result into a file or stdout).

A "Bytes producer" is an application which cannot fail with a Unicode error and produces bytes into stdout.

Let's say that a "Bytes consumer" does not decode stdin but stores data as bytes: such consumer always work. Common UNIX command line tools like cat, grep or sed are in this category. Many Python 2 applications are also in this category.

"Python producer" and "Python consumer" are producer and consumer implemented in Python.

print("euro: \u20ac")

import sys
text = sys.stdin.read()
result = text.replace("apple", "orange")
# ignore the result

print("euro: \u20ac")

import sys
text = sys.stdin.read()
result = text.replace("apple", "orange")
# ignore the result

Don't modify the encoding of the POSIX locale

A first version of the PEP did not change the encoding and error handler used of the POSIX locale.

The problem is that adding the -X utf8 command line option or setting the PYTHONUTF8 environment variable is not possible in some cases, or at least not convenient.

Moreover, many users simply expect that Python 3 behaves as Python 2: don't bother them with encodings and "just works" in all cases. These users don't worry about mojibake, or even expect mojibake because of complex documents using multiple incompatibles encodings.

Always use UTF-8

Python already always use the UTF-8 encoding on Mac OS X, Android and Windows. Since UTF-8 became the de facto encoding, it makes sense to always use it on all platforms with any locale.

The risk is to introduce mojibake if the locale uses a different encoding, especially for locales other than the POSIX locale.

Force UTF-8 for the POSIX locale

An alternative to always using UTF-8 in any case is to only use UTF-8 when the LC_CTYPE locale is the POSIX locale.

The PEP 538 [1] "Coercing the legacy C locale to C.UTF-8" of Nick Coghlan proposes to implement that using the C.UTF-8 locale.

Use the strict error handler for operating system data

Using the surrogateescape error handler for operating system data creates surprising surrogate characters. No Python codec (except of utf-7) accept surrogates, and so encoding text coming from the operating system is likely to raise an error error. The problem is that the error comes late, very far from where the data was read.

The strict error handler can be used instead to decode (os.fsdecode()) and encode (os.fsencode()) operating system data, to raise encoding errors as soon as possible. It helps to find bugs more quickly.

The main drawback of this strategy is that it doesn't work in practice. Python 3 is designed on top on Unicode strings. Most functions expect Unicode and produce Unicode. Even if many operating system functions have two flavors, bytes and Unicode, the Unicode flavar is used is most cases. There are good reasons for that: Unicode is more convenient in Python 3 and using Unicode helps to support the full Unicode Character Set (UCS) on Windows (even if Python now uses UTF-8 since Python 3.6, see the PEP 528 [3] and the PEP 529 [2]).

For example, if os.fsdecode() uses utf8/strict, os.listdir(str) fails to list filenames of a directory if a single filename is not decodable from UTF-8. As a consequence, shutil.rmtree(str) fails to remove a directory. Undecodable filenames, environment variables, etc. are simply too common to make this alternative viable.