pathlib and paths with arbitrary bytes

The pathlib module was added to the standard library in Python 3.4 and is one of the many nice improvements that Python 3 has gained over the past decade. In three weeks, Python 3.5 will be the oldest version of Python that still receives security patches. This means that the presence of pathlib can soon be taken for granted on all Python installations, and the quest towards replacing os.path can begin for real.

In this post, I’ll have a look at how pathlib can be used to handle file paths containing arbitrary bytes, as this is valid on most file systems.

Introduction to pathlib

pathlib provides an API for working with file system paths that works consistently across platforms and handles most corner cases well. If you replace all of your os.path usages with pathlib, you’ll probably have more readable code with fewer bugs.

Here are a few examples to get a feel for the pathlib API.

The entry point to the pathlib API is usually the Path class:

>>> from pathlib import Path
>>>

pathlib can replace the os.path module in most cases:

>>> p = Path("hello.txt")
>>> p
PosixPath('hello.txt')
>>> p.name
'hello.txt'
>>> p.parent
PosixPath('.')
>>> p.exists()
True
>>> p.is_file()
True
>>>

It has convenient shortcuts for getting the contents of a file, or writing to it:

>>> p.read_bytes()
b'Hello, world!\n'
>>> p.read_text()
'Hello, world!\n'
>>>

As well as the low level APIs:

>>> with p.open() as fh:
...     fh.seek(7)
...     print(fh.read(5))
...
7
world
>>>

It has replacements for os.rename() and friends:

>>> q = p.with_suffix(".md")
>>> q
PosixPath('hello.md')
>>> p.rename(q)
>>> p.exists()
False
>>> q.exists()
True
>>>

You can concatenate paths with a forward slash, which might be controversial, but reads a lot better than os.path.join(a, b) once you get used to it:

>>> d = p / ".."
>>> d
PosixPath('hello.txt/..')
>>> d = d.resolve()
>>> d
PosixPath('/home/jodal')
>>> d.is_dir()
True
>>>

When working with directories, you no longer need os.walk() or glob.glob():

>>> len(list(d.iterdir()))
112
>>> d.glob("*.md")
[PosixPath('/home/jodal/hello.md')]
>>>

Most APIs in the standard library, and many third party libraries, now accept os.PathLike in addition to bytes and strings as file paths, which means that you can use your Path instances where you used to use strings. If you’ve started to type annotate your Python programs, using Path will also be helpful because image: Path communicates much more about the indended usage than image: str would.

Path, text, and bytes

If you interface with an API that doesn’t accept Path instances, the convention is to convert the Path instance to a plain old string using str() or bytes():

>>> str(p)
'hello.txt'
>>> bytes(p)
b'hello.txt'
>>>

pathlib uses Unicode strings aka text instead of bytes wherever possible. In fact, pathlib requires you to instantiate Path using a Unicode string:

>>> Path(b'/tmp/foo')
Traceback (most recent call last):
  ...
TypeError: argument should be a str object or an os.PathLike object returning str, not <class 'bytes'>
>>> Path(b'/tmp/foo'.decode())
PosixPath('/tmp/foo')
>>>

If you provide pathlib with a non-ASCII text string and ask it to encode it as a URI, it will helpfully percent-encode it for you:

>>> p = Path('/tmp/æ')
>>> p
PosixPath('/tmp/æ')
>>> p.as_uri()
'file:///tmp/%C3%A6'
>>>

In the as_uri() output, we can see that æ became %C3%A6. It first encoded the character æ to bytes using UTF-8 encoding, resulting in the two bytes 0xC3 and 0xA6. It then encoded the bytes that were not safe for use in URIs with percent-encoding, resulting in %C3%A6.

Arbitrary bytes

Since pathlib implicitly uses UTF-8 for encoding, how well does it handle file systems that contain non-UTF-8 directory and file names?

The NTFS file system only allows file paths that can be represented as UTF-16, but most other file systems accept almost any sequence of bytes as a file path.

Let’s test with a couple of worst case bytes:

  • the null byte, 0x00 or \x00 when escaped in Python, and
  • the first half of a two-byte UTF-8 codepoint, 0xC3 or \xC3 in Python.
>>> name = b'/tmp/ab\x00cd\xC3'
>>> name
b'/tmp/ab\x00cd\xc3'
>>>

As this file path cannot be decoded as UTF-8, we can’t even construct a Path instance:

>>> p = Path(name)
Traceback (most recent call last):
  ...
TypeError: argument should be a str object or an os.PathLike object returning str, not <class 'bytes'>
>>> p = Path(name.decode())
Traceback (most recent call last):
  ...
UnicodeDecodeError: 'utf-8' codec can't decode byte 0xc3 in position 10: unexpected end of data
>>>

What if we handle the decoding errors by replacing the offending bytes?

>>> p = Path(name.decode(errors="replace"))
>>>

With the replace error handler, we manage to create a Path instance, but we’ve lost some information by replacing 0xC3 with the Unicode replacement character, U+FFFD, here encoded into UTF-8 as 0xEFBFBD:

>>> p
PosixPath('/tmp/ab\x00cd�')
>>> p.as_uri()
'file:///tmp/ab%00cd%EF%BF%BD'
>>> str(p)
'/tmp/ab\x00cd�'
>>> str(p).encode()
b'/tmp/ab\x00cd\xef\xbf\xbd'
>>>

A Path that cannot reconstruct the bytes it originated from is not useful, as we cannot use it to find and access the file represented by the Path instance.

>>> str(p).encode() == name
False
>>>

We need a way to convert arbitrary bytes to a Path instance without losing any information, so that given only a Path instance we can later reconstruct the bytes required to access that file.

Surrogate escape

Luckily, PEP 383 introduced just that in Python 3.1, released more than a decade ago: the surrogateescape error handler.

The surrogateescape error handler is not available in Python 2.7, so users of the pathlib2 backport are probably out of luck. With three weeks left until Python 2’s end-of-life, you probably have other things to worry about.

Using the surrogateescape error handler we get the following behavior from pathlib:

>>> name
b'/tmp/ab\x00cd\xc3'
>>> name.decode(errors="surrogateescape")
'/tmp/ab\x00cd\udcc3'
>>>

\udcc3 in the decoded string is a surrogate escape code point for the 0xC3 byte.

Passing the text string to pathlib, we get a Path like usual, and with str(p) we get the same text string back again, with the surrogate escape code point:

>>> p = Path(name.decode(errors="surrogateescape"))
>>> p
PosixPath('/tmp/ab\x00cd\udcc3')
>>> str(p)
'/tmp/ab\x00cd\udcc3'
>>>

Let’s try encoding this back to bytes and compare it to the bytes we started with:

>>> str(p).encode() == name
Traceback (most recent call last):
  ..
UnicodeEncodeError: 'utf-8' codec can't encode character '\udcc3' in position 10: surrogates not allowed
>>>

That failed because str.encode() does not allow encoding of surrogates by default. When we create a Unicode string by decoding with the surrogateescape error handler, we must also use the surrogateescape error handler to encode the text back to bytes:

>>> str(p).encode(errors="surrogateescape") == name
True
>>>

Interestingly, when rendering the path as a URI, pathlib correctly converts the surrogate to the original byte, 0xC3, percent-encoded as %C3.

>>> p.as_uri()
'file:///tmp/ab%00cd%C3'
>>>

In fact, pathlib is able to correctly encode paths with surrogates to bytes itself:

>>> bytes(p)
b'/tmp/ab\x00cd\xc3'
>>> bytes(p) == name
True
>>>

os.fsencode()

The reason pathlib correctly converts the path to bytes is because its __bytes__() implementation uses os.fseencode().

os.fsencode() explicitly encodes using the surrogateescape handler, except on Windows, where it uses the strict handler to not allow any file paths that cannot be used on NTFS. It lets bytes pass through unchanged.

>>> import os
>>> os.fsencode("/tmp/æ")
b'/tmp/\xc3\xa6'
>>> os.fsencode(b"/tmp/ab\x00cd\xc3")
b'/tmp/ab\x00cd\xc3'
>>>

os.fsencode() has a sibling in os.fsdecode(), which does the opposite. It decodes using the surrogateescape handler, except on Windows, where it uses the strict handler. It lets text, not bytes, pass through unchanged.

>>> os.fsdecode(b"/tmp/ab\x00cd\xc3")
'/tmp/ab\x00cd\udcc3'
>>> os.fsdecode("/tmp/æ")
'/tmp/æ'
>>>

With this, we can put together some recommendations.

To Path and back again

To create a Path instance that can hold a path with arbitrary byte encoding, use Path(os.fsdecode(...)):

>>> p1 = Path(os.fsdecode(b'/tmp/ab\x00cd\xc3'))
>>> p2 = Path(os.fsdecode('/tmp/æ'))
>>> p1
PosixPath('/tmp/ab\x00cd\udcc3')
>>> p2
PosixPath('/tmp/æ')
>>>

To get the exact bytes from a Path instance, use bytes(path):

>>> bytes(p1)
b'/tmp/ab\x00cd\xc3'
>>> bytes(p2)
b'/tmp/\xc3\xa6'

Presentation

When it comes to presenting the path to an end user, in a user interface or a log file, there is no perfect solution.

Using URIs yields a consistent representation of the file path, irrespective of its validity as UTF-8. The URIs are true to the actual bytes on the file system and do not contain any surrogates.

>>> p1.as_uri()
'file:///tmp/ab%00cd%C3'
>>> p2.as_uri()
'file:///tmp/%C3%A6'

Why you should not use str(path)

However, for anything that is valid UTF-8 but not valid ASCII, which includes most languages spoken by humans, the result of str(path) is a lot more readable than a URI.

For paths which are valid UTF-8 str(path) yields the cleanest result:

>>> str(p2)
'/tmp/æ'

But for names consisting of bytes that are invalid UTF-8, the use of surrogates leaks through. It’s undesirable to expose users to this implementation-specific detail:

>>> str(p1)
'/tmp/ab\x00cd\udcc3'

You might think that since invalid UTF-8 isn’t that common, maybe we can live with the surrogates leaking through in those rare cases?

Printing a text string involves encoding it to bytes before it is streamed to your output, e.g. a terminal or a file. And since encoding text containing surrogates isn’t allowed by default, print(str(path)) can crash your application:

>>> print(str(p2))
/tmp/æ
>>> print(str(p1))
Traceback (most recent call last):
  ...
UnicodeEncodeError: 'utf-8' codec can't encode character '\udcc3' in position 10: surrogates not allowed
>>>

It is possible to go down this path, but you’ll need to always use the Python repr() of the string, instead of simply printing the string:

>>> print(repr(str(p1)))
'/tmp/ab\x00de\udcc3'
>>>

Lesser of the evils

So, with my current knowledge of possible solutions, my recommendation for consistent representation of paths with arbitrary bytes in user interfaces, including logging and console output, is to use either repr(path) or path.as_uri(), with a preference for path.as_uri():

>>> print(repr(p1))
PosixPath('/tmp/ab\x00de\udcc3')
>>> print(p1.as_uri())
file:///tmp/ab%00de%C3
>>>

Summary

Let’s recap what we’ve learned here:

  • pathlib provides a nice, portable API to work with file system paths in Python.
  • With the help of the surrogateescape error handler, pathlib can handle paths with arbitrary bytes.
  • os.fsencode() and os.fsdecode() are convenient ways to use the surrogateescape error handler, with the bonus of correct behavior on Windows too.
  • Using str(path) to present Path objects in user interfaces might crash your application. The alternatives, repr(path) and path.as_uri(), are not perfect either, but at least they are safe.

Cheat sheet

>>> import os
>>> from pathlib import Path

>>> # To create paths:
>>> p = Path(os.fsdecode(b'/tmp/ab\x00cd\xc3'))
>>> p
PosixPath('/tmp/ab\x00cd\udcc3')

>>> # To use with APIs not supporting `os.PathLike`:
>>> bytes(p)
b'/tmp/ab\x00cd\xc3'

>>> # To show to users:
>>> print(p.as_uri())
file:///tmp/ab%00cd%C3

Thanks to Nick Steel for reviewing this blog post.