As discussed on python-ideas, swapping out the sys.std* streams can be fraught with peril. This can make writing code that wants to support an "--encoding" option for pipeline processing difficult.

The proposal [1] is that TextIOWrapper support a set_encoding() method that is only supported between creation of the stream and the first read or write operation. Once the stream is "dirty", you can no longer change the encoding - you must replace the stream (e.g. by passing stream.fileNo() to open())

[1] http://mail.python.org/pipermail/python-ideas/2012-June/015535.html

The proposal [1] is that TextIOWrapper support a set_encoding() method that is only supported between creation of the stream and the first read or write operation.

IMHO "encoding", "errors", "buffering", "line_buffering" and "newline" should be writable properties (unless first read/write).

The proposal [1] is that TextIOWrapper support a set_encoding() method
that is only supported between creation of the stream and the first
read or write operation.

That will be fragile. A bit of prematurate input or output (for whatever reason) and your program breaks.

Indeed. However, the current alternatives (based on detach() and fileNo()) are also problematic - using detach() breaks the corresponding sys.__std*__ entry, while using fileNo() means you now have two independent IO stacks using the same underlying descriptor.

My own preference is to have a set_encoding() method that *can* be used even if IO has already occurred. Yes, there's the chance of mixing encodings if data is processed before the encoding is changed, but that's also true for both of the alternatives.

Pipeline applications written in Python should have an easy way to implement "--encoding" parameters for the standard IO streams, and a method to update the settings is the obvious choice.

Would a "set_encoding" method be Pythonic? I would've preferred an "encoding" property which flushes the output when it's changed.

Would a "set_encoding" method be Pythonic? I would've preferred an
"encoding" property which flushes the output when it's changed.

I would prefer to have a method. The side-effect is too violent to be
hidden behind a property.
Besides, you want to set encoding and errors in a single operation.

I fail to see why this is a release blocker; no rationale is given in the original message, nor in the quoted message. So unblocking.

Here is a Python implementation of TextIOWrapper.set_encoding().

The main limitation is that it is not possible to set the encoding on a non-seekable stream after the first read (if the read buffer is not empty, ie. if there are pending decoded characters).

+ # flush read buffer, may require to seek backward in the underlying
+ # file object
+ if self._decoded_chars:
+ if not self.seekable():
+ raise UnsupportedOperation(
+ "It is not possible to set the encoding "
+ "of a non seekable file after the first read")
+ assert self._snapshot is not None
+ dec_flags, next_input = self._snapshot
+ offset = self._decoded_chars_used - len(next_input)
+ if offset:
+ self.buffer.seek(offset, SEEK_CUR)

--

I don't have an use case for setting the encoding of sys.stdout/stderr after startup, but I would like to be able to control the *error handler* after the startup! My implementation permits to change both (encoding, errors, encoding and errors).

For example, Lib/test/regrtest.py uses the following function to force the backslashreplace error handler on sys.stdout. It changes the error handler to avoid UnicodeEncodeError when displaying the result of tests.

def replace_stdout():
    """Set stdout encoder error handler to backslashreplace (as stderr error
    handler) to avoid UnicodeEncodeError when printing a traceback"""
    import atexit

    stdout = sys.stdout
    sys.stdout = open(stdout.fileno(), 'w',
        encoding=stdout.encoding,
        errors="backslashreplace",
        closefd=False,
        newline='\n')

    def restore_stdout():
        sys.stdout.close()
        sys.stdout = stdout
    atexit.register(restore_stdout)

The doctest module uses another trick to change the error handler:

        save_stdout = sys.stdout
        if out is None:
            encoding = save_stdout.encoding
            if encoding is None or encoding.lower() == 'utf-8':
                out = save_stdout.write
            else:
                # Use backslashreplace error handling on write
                def out(s):
                    s = str(s.encode(encoding, 'backslashreplace'), encoding)
                    save_stdout.write(s)
        sys.stdout = self._fakeout

That will be fragile. A bit of prematurate input or output
(for whatever reason) and your program breaks.

I agree that it is not the most pure solution, but sometimes practicality beats purity (rule #9) ;-) We can add an ugly big red warning in the doc ;-)

My implementation permits to change both (encoding, errors, encoding and errors).

We may also add a set_errors() method:

def set_errors(self, errors):
   self.set_encoding(self.encoding, errors)

The reason I marked this as a release blocker for 3.4 is because it's a key piece of functionality for writing command line apps which accept an encoding argument. I'll use "high" instead.

An interesting proposal was posted to the python-dev thread [1]: using self.detach() and self.__init__() to reinitialise the wrapper *in-place*.

With that approach, the pure Python version of set_encoding() would look something like this:

    _sentinel = object()
    def set_encoding(self, encoding=_sentinel, errors=_sentinel):
        if encoding is _sentinel:
            encoding = self.encoding
        if errors is _sentinel:
            errors = self.errors
        self.__init__(self.detach(),
                      encoding, errors,
                      self._line_buffering,
                      self._readnl,
                      self._write_through)

(The pure Python version currently has no self._write_through attribute - see bpo-15571)

Note that this approach addresses my main concern with the use of detach() for this: because the wrapper is reinitialised in place, old references (such as the sys.__std*__ attributes) will also see the change.

Yes, such a function would need a nice clear warning to say "Calling this may cause data loss or corruption if used without due care and attention", but it should *work*. (Without automatic garbage collection, the C implementation would need an explicit internal "reinitialise" function rather than being able to just use the existing init function directly, but that shouldn't be a major problem).

[1] http://mail.python.org/pipermail/python-ideas/2012-August/015898.html

To bring back Victor's comments from the list:

• stdout/stderr are fairly easy to handle, since the underlying buffers can be flushed before switching the encoding and error settings. Yes, there's a risk of creating mojibake, but that's unavoidable and, for this use case, trumped by the pragmatic need to support overriding the output encoding in a robust fashion (i.e. not breaking sys.__stdout__ or sys.__stderr__, and not crashing if something else displays output during startup, for example, when running under "python -v")

• stdin is more challenging, since it isn't entirely clear yet how to handle the case where data is already buffered internally. Victor proposes that it's acceptable to simply disallow changing the encoding of a stream that isn't seekable. My feeling is that such a restriction would largely miss the point, since the original use case that prompted the creation of this was shell pipeline processing, where stdin will often be a PIPE

I think the guiding use case here really needs to be this one: "How do I implement the equivalent of 'iconv' as a Python 3 script, without breaking internal interpreter state invariants?"

My current thought is that, instead of seeking, the input case can better be handled by manipulating the read ahead buffer directly. Something like (for the pure Python version):

   self._encoding = new_encoding
   if self._decoder is not None:
     old_data = self._get_decoded_chars().encode(old_encoding)
     old_data += self._decoder.getstate()[0]
     decoder = self._get_decoder()
     new_chars = ''
     if old_data:
         new_chars = decoder.decode(old_data)
     self._set_decoded_chars(new_chars)

(A similar mechanism could actually be used to support an "initial_data" parameter to TextIOWrapper, which would help in general encoding detection situations where changing encoding *in-place* isn't needed, but the application would like an easy way to "put back" the initial data for inclusion in the text stream without making assumptions about the underlying buffer implementation)

Also, StringIO should implement this new API as a no-op.

Victor proposes that it's acceptable to simply disallow changing the encoding of a stream that isn't seekable.

It is no what I said. My patch raises an exception if you already
started to read stdin. It should work fine if stdin is a pipe but the
read buffer is empty.

Ah, you're right - peeking into the underlying buffer would be enough to
handle encoding detection.

Oh, set_encoding.patch is wrong:

+ offset = self._decoded_chars_used - len(next_input)

self._decoded_chars_used is a number of Unicode characters, len(next_input) is a number of bytes. It only works with 7 and 8 bit encodings like ascii or latin1, but not with multibyte encodings like utf8 or ucs-4.

peeking into the underlying buffer would be enough to
handle encoding detection.

I wrote a new patch using this idea. It does not work (yet?) with non seekable streams. The raw read buffer (bytes string) is not stored in the _snapshot attribute if the stream is not seeakble. It may be changed to solve this issue.

set_encoding-2.patch is still a work-in-progress. It does not patch the _io module for example.

Note: it is not possible to reencode the buffer of decoded characters to compute the offset in bytes. Some codecs are not bijective.

Examples:

• b'\x00'.decode('utf7').encode('utf7') == b'+AAA-'
• b'\xff'.decode('ascii', 'replace').encode('ascii', 'replace') == b'?'
• b'\xff'.decode('ascii', 'ignore').encode('ascii', 'ignore') == b''

That's a fair point - I think it's acceptable to throw an error in the case of *already decoded* characters that haven't been read.

There's also a discussion on python-ideas about an explicit API for clearing the internal buffers, and pushing data back into a stream. If that is added, then set_encoding() would be free to error out if there was any already buffered data - it would be up to the application to call clear_buffer() before calling set_encoding(), and deal with an such data appropriately (such as calling push_data() with the results of the clear_buffer() call)

Oops, meant to link to my post in the thread about a buffer manipulation API: http://mail.python.org/pipermail/python-ideas/2013-March/019769.html

It would be good to have this addressed in Python 3.4 (I'm following up on a few other encoding related issues at the moment).

Is that a reasonable possibility before beta 1, or do we need to bump this one to 3.5?

Is that a reasonable possibility before beta 1, or do we need to bump this one to 3.5?

My patch was not reviewed yet and must be reimplemented in C. I will not have time before the beta1 to finish the work.

Just reviewed Victor's patch - aside from a trivial typo and not covering the C implementation or documentation yet, it looks good to me. Most importantly, the tests in that patch could be used to validate a C implementation as well.

I'll see if I can find anyone interested in creating a patch for the C io implementation, otherwise we can postpone the addition until 3.5.

I am interested in working on this, but I might need guidance at times. Is that acceptable? If yes, I'm willing to start as soon as possible.

Thanks, Andrei, that would be great.

The tests and the Python version in Victor's patch show the desired API and behaviour.

In theory, the C version should just be a matter of adding an equivalent textiowrapper_set_encoding method as a static function in
hg.python.org/cpython/file/default/Modules/_io/textio.c

I have to postpone this for a few weeks. Sorry if I stayed in anyone's way.

I am working on this now.

Question:

What is the point of the

        old_encoding = codecs.lookup(self._encoding).name
        encoding = codecs.lookup(encoding).name
        if encoding == old_encoding and errors == self._errors:
            # no change
            return

dance? Isn't this equivalent to

        if encoding == self.encoding and errors == self._errors:
            # no change
            return

except that it doesn't validate the given encoding? But if the encoding is invalid, isn't it enough that the exception is raised a few lines further down?

Second question:

The following looks as if a BOM might be written for writeable, non-seekable streams:

        # don't write a BOM in the middle of a file
        if self._seekable and self.writable():
            position = self.buffer.tell()
            if position != 0:
                try:
                    self._get_encoder().setstate(0)
                except LookupError:
                    # Sometimes the encoder doesn't exist
                    pass

Is that really desirable? It seems to me the safe choice is to *not* write the BOM, except when we know it's safe. Eg:

    # don't write a BOM unless we know we're at the beginning of a file
    if (self.writeable() and not
       (self._seekable and self.buffer.tell() == 0)):
        try:
            self._get_encoder().setstate(0)
        except LookupError:
            # Sometimes the encoder doesn't exist
            pass

It's a blunt instrument to be sure, but a necessary one, I think -
otherwise we end with a scattering of API specific solutions rather than a
single consistent approach.

Here's a thought, though: what if we went with Serhiy's "reconfigure"
idea, limited that to seekable streams (resetting then back to the start of
the file) and then also had a "force_reconfigure" that bypassed the safety
checks?

The main problem I see with that idea is that some changes (binary vs text,
universal newlines or not, buffered or not) would potentially require
adding or removing a class from the stream's IO stack, thus rendering it
difficult to handle as an in-place operation.

However the "seekable streams only by default, flag or separate method to
force an encoding change for a non-seekable stream" approach would be
applicable even for the basic "set_encoding" API.

I'm beginning to think that this is all complicated enough that it should
be written up in a PEP for 3.5 before we actually commit anything (even if
what we end up committing is just set_encoding with a "force=False" keyword
only parameter).

I think that any API that gives you a TextIOWrapper object with undesired properties is broken and needs to be fixed, rather than a workaround added to TextIOWrapper.

That said, I defer to the wisdow of the core developers, and I'll be happy to implement whatever is deemed to be the right thing(tm). However, my opinion on this feature probably makes me unsuitable to drive a PEP for it :-).

Newest version of the patch attached, the reference leak problem has been fixed.

If operating systems always exposed accurate metadata and configuration settings, I'd agree with you. They don't though, so sometimes developers need to be able to override whatever the interpreter figured out automatically.

In addition, needing to cope with badly designed APIs is an unfortunate fact of life - that's why monkeypatching is only *discouraged*, rather than disallowed :)

On 02/04/2014 03:28 AM, Nick Coghlan wrote:

Nick Coghlan added the comment:

If operating systems always exposed accurate metadata and configuration settings, I'd agree with you. They don't though, so sometimes developers need to be able to override whatever the interpreter figured out automatically.

Where does the operating system come into play? What I'm trying to say
is that any code of the form

fh = some_python_function()
fh.set_encoding(foo)

should really be written as

fh = some_python_function(encoding=foo)

or

fh = TextIOWrapper(some_python_function(raw=True), encoding=foo)

The only exception to this is sys.std{in,out,err}. But this special case
should, in my opinion, be solved with a corresponding special function
in sys.*, rather than a general purpose method in TextIOWrapper.

In addition, needing to cope with badly designed APIs is an unfortunate fact of life - that's why monkeypatching is only *discouraged*, rather than disallowed :)

Well, yes, but introducing a set_encoding() or reconfigure() function is
actively *encouraging* bad API design. "I'll just return my favorite
encoding from this function, after all, the caller can use set_encoding
afterwards".

Best,
Nikolaus

--
Encrypted emails preferred.
PGP fingerprint: 5B93 61F8 4EA2 E279 ABF6 02CF A9AD B7F8 AE4E 425C

         »Time flies like an arrow, fruit flies like a Banana.«

Metaclasses and monkeypatching encourage bad API design, too, but we still
provide them because they're essential tools for some use cases. In Python,
the system integrator is always right, and we provide them with power tools
to deal with the cases where reality doesn't want to cooperate with the
nice neat model of the world that we provide by default.

For the standard streams, the problem is that the stream is created
automatically by the interpreter, and sometimes we will get the encoding
choice wrong (because the info we retrieve from the OS doesn't match what
the user actually wants), but rebinding the names to a new IO stream causes
other problems. Now, we have two choices at this point:

1. Assume there is *no other API involving implicit text stream creation in
   any third party Python 3 library anywhere* that will ever have this problem
   and create a solution that only works with the standard streams

2. Acknowledge that while the implicit way we create the standard streams
   is unusual, it is unlikely to be *unique*, so it actually does make sense
   to provide the ability to change the encoding of an existing stream as a
   general purpose feature fenced about with as many "you probably don't want
   to use this unless you really know what you are doing, otherwise it may eat
   your data" warnings as seems appropriate.

Some more use cases for temporarily switching error handler in the middle of writing to a stream:

• Possibly simplify the implementation of sys.displayhook()
• I have done a similar hack at <https://bitbucket.org/Gfy/pyrescene/commits/42ad75e375d84e090a32d024acc865de341c22aa#Lrescene/srr.pyF132\>, to output a comment field with a permissive error handler, while other data is output with strict error handling.

A use case for changing a reader’s newline translation mode is to use standard input with the built-in “csv” module. My current idea is to do something like this:

encoding = sys.stdin.encoding
errors = sys.stdin.errors
line_buffering = sys.stdin.line_buffering
# No way to retain write_through mode, but shouldn’t matter for reading
sys.stdin = TextIOWrapper(sys.stdin.detach(), encoding, errors,
    newline="", line_buffering=line_buffering)

for row in csv.reader(sys.stdin):
    ...

On the other hand, I wonder about rewinding an input file after already having read and buffered text in the wrong encoding. From a distance, the Python native version of the code seems rather complex and full of dark magic. Is there a use case, or maybe it would be simpler to have it only work when nothing has been read or buffered?

Reviewing the items I had flagged as dependencies of bpo-22555 for personal tracking purposes, I suggest we defer further consideration of this idea to 3.6 after folks have had a chance to get some experience with the interpreter defaulting to using "surrogateescape" on the standard streams when an underlying *nix operating system (other than Mac OS X) claims that the system encoding is ASCII.

Revisiting the idea Nikolaus raised last year regarding whether or not this could be done using a dedicated API in the sys module, I realised today that even if we decided to use a separate public API, *that API* would still need a defined way to modify the stream encoding and error handling in place.

Since we're going to need the object level in-place modification API regardless, a separate sys module API for modifying the standard streams in particular would then introduce additional complexity without providing a proportional benefit.

quad, please don't change issue attributes to unrelated values.

Updated the patch for default branch.

Reviewing Inada-san's latest version of the patch, we seem to be in a somewhat hybrid state where:

1. The restriction to only being used with seekable() streams if there is currently unread data in the read buffer is in place

2. We don't actually call seek() anywhere to set the stream back to the beginning of the file. Instead, we try to shuffle data out of the old decoder and into the new one.

I'm starting to wonder if the best option here might be to attempt to make the API work for arbitrary codecs and non-seekable streams, and then simply accept that it may take a few maintenance releases before that's actually true. If we decide to go down that path, then I'd suggest the follow stress test:

• make a longish test string out of repeated copies of "ℙƴ☂ℌøἤ"
• pick a few pairs of multibyte non-universal/universal encodings for use with surrogateescape and strict as their respective error handlers (e.g. ascii/utf8, ascii/utf16le, ascii/utf32, ascii/shift_jis, ascii/iso2022_jp, ascii/gb18030, gbk/gb18030)
• for each pair, make the test data by encoding from str to bytes with the relevant universal encoding
• switch the encoding multiple times on the same stream at different points

Optionally:

• extract "codecs._switch_decoder" and "codecs._switch_encoder" helper functions to make this all a bit easier to test and debug (with a Python version in the codecs module and the C version accessible via the _codecs modules)

That way, confidence in the reliability of the feature (including across Python implementations) can be based on the strength of the test cases covering it.

I had not read the patch carefully when updating.
Now I'm reading pure Python part of the patch.

For writing side, this patch is very simple.
For reading side, this patch is very complex, because TextIOWrapper.tell()
and .seek() is complex already.

How about starting this API which works only when self._decoder is None?
Regardless acceptance of PEP-538 and PEP-540, this is important to change
stdio encoding.

We already failed to ship this important API in 3.6.

This patch doesn't change current behavior of "decode by chunk".
It's difficult to support changing encoding via header, like
Python's encoding cookie.

>>> import io
>>> buff = io.BytesIO("hello\nこんにちは".encode('utf-8'))
>>> f = io.TextIOWrapper(buff, encoding='ascii')
>>> f.readline()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/inada-n/work/python/cpython/Lib/encodings/ascii.py", line 26, in decode
    return codecs.ascii_decode(input, self.errors)[0]
UnicodeDecodeError: 'ascii' codec can't decode byte 0xe3 in position 6: ordinal not in range(128)
>>>
>>> buff = io.BytesIO("hello\nこんにちは".encode('utf-8'))
>>> f = io.TextIOWrapper(buff, encoding='ascii')
>>> f.read(6)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/inada-n/work/python/cpython/Lib/encodings/ascii.py", line 26, in decode
    return codecs.ascii_decode(input, self.errors)[0]
UnicodeDecodeError: 'ascii' codec can't decode byte 0xe3 in position 6: ordinal not in range(128)

Inada, I think you messed up the positioning of bits of the patch. E.g. there are now test methods declared inside a helper function (rather than a test class).

Since it seems other people are in favour of this API, I would like to expand it a bit to cover two uses cases (see set_encoding-newline.patch):

• change the error handler without affecting the main character encoding
• set the newline encoding (also suggested by Serhiy)

Regarding Serhiy’s other suggestion about buffering parameters, perhaps TextIOWrapper.line_buffering could become a writable attribute instead, and the class could grow a similar write_through attribute. I don’t think these affect encoding or decoding, so they could be treated independently.

The algorithm for rewinding unread data is complicated and can fail. What is the advantage of using it? What is the use case for reading from a stream and then changing the encoding, without a guarantee that it will work?

Even if it is enhanced to never “fail”, it will still have strange behaviour, such as data loss when a decoder is fed a single byte and produces multiple characters (e.g. CR newline, backslashreplace, UTF-7).

One step in the right direction IMO would be to only support calling set_encoding() when no extra read data has been buffered (or to explicitly say that any buffered data is silently dropped). So there is no support for changing the encoding halfway through a disk file, but it may be appropriate if you can regulate the bytes being read, e.g. from a terminal (user input), pipe, socket, etc.

But I would be happy enough without set_encoding(), and with something like my rewrap() function at the bottom of <https://github.com/vadmium/data/blob/master/data.py#L526\>. It returns a fresh TextIOWrapper, but when you exit the context manager you can continue to reuse the old stream with the old settings.

Inada, I think you messed up the positioning of bits of the patch. E.g. there are now test methods declared > inside a helper function (rather than a test class).

I'm sorry. patch -p1 merged previous patch into wrong place, and test passed accidently.

Since it seems other people are in favour of this API, I would like to expand it a bit to cover two uses cases (see set_encoding-newline.patch):

• change the error handler without affecting the main character encoding
• set the newline encoding (also suggested by Serhiy)

+1. Since stdio is configured before running Python program, TextIOWrapper should be configurable after creation, as possible.

Regarding Serhiy’s other suggestion about buffering parameters, perhaps TextIOWrapper.line_buffering could become a writable attribute instead, and the class could grow a similar write_through attribute. I don’t think these affect encoding or decoding, so they could be treated independently.

Could them go another new issue?
This issue is too long to read already.

The algorithm for rewinding unread data is complicated and can fail. What is the advantage of using it? What is the use case for reading from a stream and then changing the encoding, without a guarantee that it will work?

Even if it is enhanced to never “fail”, it will still have strange behaviour, such as data loss when a decoder is fed a single byte and produces multiple characters (e.g. CR newline, backslashreplace, UTF-7).

When I posted the set_encoding-7.patch, I hadn't read io module deeply. I just solved conflict and ran test.
After that, I read the code and I feel same thing (see msg285111 and msg285112).
Let's drop support changing encoding while reading.
It's significant step that allowing changing stdin encoding only before reading anything from it.

One step in the right direction IMO would be to only support calling set_encoding() when no extra read data has been buffered (or to explicitly say that any buffered data is silently dropped). So there is no support for changing the encoding halfway through a disk file, but it may be appropriate if you can regulate the bytes being read, e.g. from a terminal (user input), pipe, socket, etc.

Totally agree.

But I would be happy enough without set_encoding(), and with something like my rewrap() function at the bottom of <https://github.com/vadmium/data/blob/master/data.py#L526\>. It returns a fresh TextIOWrapper, but when you exit the context manager you can continue to reuse the old stream with the old settings.

I want one obvious way to control encoding and error handler from Python, (not from environment variable).
Rewrapping stream seems hacky way, rather than obvious way.

set_encoding-8.patch dropped support of changing encoding after read.
It is based on set_encoding-newline.patch

Antoine posted a simpler reconfiguration RFE over in https://bugs.python.org/issue30526 (for setting line buffering).

While technically orthogonal to this RFE, we're thinking it might be possible to expose them to users as a common "reconfigure()" API, rather than as independent methods.

TextIOWrapper.reconfigure() has been added for 3.7 as part of bpo-30526 (currently covering the line_buffering and write_through options), so I've updated the issue title here to reflect that that's now the relevant API to use to address this particular RFE.

New changeset 507434f by INADA Naoki in branch 'master':
bpo-15216: io: TextIOWrapper.reconfigure() accepts encoding, errors and newline (GH-2343)
507434f