homepage

The attached ZIP file contains "test.bat" which runs "test.py" with
Python 2.6 and Python 3.0.
Python 2.6 behaves as expected (see "py26.out"), since it returns
strings from both "mike.txt" and "mike.txt.gz". However, the same test
with Python 3.0 returns bytes from "mike.txt.gz", as shown in "py30.out":
Output: Hello from Mike.
Output: This is the second line.
Output: Why did the robot cross the road?
Output: b'Hello from Mike.'
Output: b'This is the second line.'
Output: b'Why did the robot cross the road?'
For reference, I tested this on Python versions:
Python 2.6.1 (r261:67517, Dec 4 2008, 16:51:00) [MSC v.1500 32 bit
(Intel)] on win32
Python 3.0.1 (r301:69561, Feb 13 2009, 20:04:18) [MSC v.1500 32 bit
(Intel)] on win32

gzip.open() only implements the "rb" mode, and returns bytes.
fileinput could certainly wrap it with a io.TextIOWrapper.
Then the encoding issue arises.
fileinput.FileInput should grow an "encoding" parameter instead of always relying on the default encoding.

Working on this

I had a lot of help of R. David Murray, could someone review our patch, please?
We are specially concerned about backward compatibility.

Thanks for the patch. I commented on not-so-important things on the code review site (which should have sent an email to Tatiana) and wanted to raise an issue here.
Is it best that the encoding used defaults to Python’s default encoding (UTF-8) or the I/O default encoding (depends on the OS and locale), like what open does?

I think it should use the same default as open, but frankly I couldn't remember the right way to achieve that. Suggestions welcome :)

We can use locale.getpreferredencoding, or be smarter and just pass down encoding=encoding to TextIOWrapper and let *it* call getpreferredencoding if it’s given a None :)

Excellent! Thanks for all remarks, Éric! I'll follow your suggestions and will submit a new patch.

Duh. I should have remembered that the io module used None to represent the default encoding. Thanks, Éric.

Nobody can remember everything: I tried in a shell and looked at the docs to find it :)

Improved patch, according to eric.araujo's suggestions and mnewman's guidance.

Thanks. Unless another core dev wants to do a complementary review I will slightly tweak the patch and commit it. I need to finish waking up and eat some food before I do that :)
Technically adding a new argument means that this is a new feature and cannot be applied to the stable 3.2 version, but something needs to be done for this bug in 3.2 too, like a recipe in the docs for a hook_compressed that returns strings (i.e. a function that calls fileinput.hook_compressed and wraps it in a TextIOWrapper), or at least a note to warn about this bug.

For 3.2 could we use the same fix, but without exposing the ability to *change* the encoding? That is, we use TextIOWrapper but always with the default None for encoding.
It also occurs to me that this really exposes a weakness in the design. What if the user wants to specify other open parameters? I wonder if we should say that for better future-proofing openhooks should always take **kw. You could even envision fileinput accepting **kw and passing them along to the openhook. I think charset is the most important open paramenter in this context, though, so I don't think we have to solve the general problem in this fix.

It also occurs to me that this fix makes the charset hook look rather odd. We could render it redundant by passing charset to open in the non-openhook case, and mark it deprecated.
There is also a bug in the hook_encoding docs. It says the file is opened with codecs.open, but that is not the case, regular open is used.

> For 3.2 could we use the same fix, but without exposing the ability to *change* the encoding?
> That is, we use TextIOWrapper but always with the default None for encoding.
Yes!
> It also occurs to me that this really exposes a weakness in the design. What if the user wants to
> specify other open parameters? I wonder if we should say that for better future-proofing openhooks
> should always take **kw. You could even envision fileinput accepting **kw and passing them along
> to the openhook. I think charset is the most important open paramenter in this context, though, so
> I don't think we have to solve the general problem in this fix.
I concur. I’ve never had to care about buffering for example, but mode is another parameter of open that people may want to give. I’ll commit the minimal fix to 3.2 and merge in 3.3, and then we can discuss on a new RFE bug about adding encoding vs. **kwargs for 3.3.
Agreed on deprecating the charset hook when it becomes redundant.
Will fix the doc bug about codecs.open too.

This issue is not newcomer friendly, I remove the easy keyword.

The patch for this change has broken code that relied on the old behavior. For example:
```
import fileinput
import bz2
import pathlib
target = pathlib.Path('data.bz2')
target.write_bytes(bz2.compress(b'Foo\nBar\nBiz'))
inp = fileinput.FileInput([str(target)], openhook=fileinput.hook_compressed)
for line in inp:
 print(line.decode().strip())
```
That code passes on Python3.10a6 but on 3.10b1 fails with:
```
Traceback (most recent call last):
 File "/Users/jaraco/code/main/cmdix/text.py", line 10, in <module>
 print(line.decode().strip())
AttributeError: 'str' object has no attribute 'decode'. Did you mean: 'encode'?
```
I encountered this issue in [this function and its test](https://github.com/jaraco/cmdix/blob/dc5fac3817ff9815b2f8d9a1dfad4258c14b1693/cmdix/lib.py#L165-L193).
I'm struggling with how to adapt the code to provide a uniform interface on Python 3.6+. Perhaps a backport of hook_compressed is in order.

A backport now exists (https://pypi.org/project/backports.hook_compressed) and addresses the issue (https://github.com/jaraco/cmdix/actions/runs/873404846).

> I'm struggling with how to adapt the code to provide a uniform interface on Python 3.6+.
It's easy. Just passing `mode="rb"`.
```
FileInput(filelist, mode="rb", openhook=fileinput.hook_compressed)
```
This is better than `backports.hook_compressed`, because it returns bytes always regardless file is compressed or uncompressed.

I did consider and confirm that mode="rb" does also provide a uniform solution, but it also regresses the behavior of uncompressed inputs, making them bytes where they were text. This approach feels like the "Python 1" compatibility approach.
At least in my case, the preference was to migrate to the future behavior, where all content is decoded to Unicode text, without having to wait for Python 3.9 to sunset.

> I did consider and confirm that mode="rb" does also provide a uniform solution, but it also regresses the behavior of uncompressed inputs, making them bytes where they were text.
Of course, I suggested to use "rb" when you want to read bytes.
* When reading binary from uncompressed file, mode="rb" just works for 3.6+.
* When reading text from uncompressed file, mode="r" just works for 3.6+.
* When reading binary from compressed file, mode="rb" just works for 3.6+.
* Reading text from compressed file is not supported until Python 3.10.