The sys.argv docs [1] currently contain no mention of the fact that they are Unicode strings decoded from bytes provided by the OS. They also don't explain how to correct a decoding error by reversing Python's implicit conversion and redoing it based on the application's knowledge of the correct encoding, as described at [2]

[1] http://docs.python.org/3/library/sys#sys.argv
[2] http://stackoverflow.com/questions/6981594/sys-argv-as-bytes-in-python-3k/

I tried running with Python 3.4 the following code

import sys

print(sys.argv[1])
print(b'bytes')

And I ran as follows trying to run with a different encoding. 
$ python ~/a.py `echo priya|iconv -t latin1`
priya
bytes

There was no unicode encode error generated! Is it because the problem is fixed?

> There was no unicode encode error generated! Is it because the problem 
> is fixed?

No, it's not fixed.
First, it seems you are testing with Python 2 (otherwise you would get "b'bytes'", not "bytes"). Python 2 won't have a problem here, since it treats everything as bytestrings.
Second, to evidence the issue you must pass a non-ASCII string. For example:

$ ./python a.py `echo éléphant|iconv -t latin1`
Traceback (most recent call last):
  File "a.py", line 4, in <module>
    print(sys.argv[1])
UnicodeEncodeError: 'utf-8' codec can't encode character '\udce9' in position 0: surrogates not allowed

You are right. Instead of running ./python inside the python directory, I ran the default python of older version! Based on the stackoverflow link given, I tried to make some documentation. I am attaching the patch!

Hmm, I'm not sure where those explanations belong but I'm not sure should be in the sys module docs (especially as they are quite lengthy, and they also apply to other data such as os.environ). Perhaps the Unicode HOWTO?

New changeset 38f4e468d4b55551e135c67337c18ae142193ba8 by Inada Naoki in branch 'master':
bpo-17110: doc: add note how to get bytes from sys.argv (GH-12602)
https://github.com/python/cpython/commit/38f4e468d4b55551e135c67337c18ae142193ba8

New changeset 5b80cb5584a72044424f2d82d0ae79c720f24c47 by Miss Islington (bot) in branch '3.7':
bpo-17110: doc: add note how to get bytes from sys.argv (GH-12602)
https://github.com/python/cpython/commit/5b80cb5584a72044424f2d82d0ae79c720f24c47

The actual startup code uses Py_DecodeLocale() for converting argv from bytes to unicode. Since which Python version is it guaranteed that Py_DecodeLocale() and os.fsencode() roundtrip?

There is no strict guarantee.

I think ASCII, UTF-8, latin1 with surrogateescape guarantee roundtrip.

Other legacy encodings like cp932 may not roundtrip. But it is not a huge problem because only Windows use them typically.
On Windows:

* wchar_t is used in most case, instead of fsencoding
* fsencoding is now UTF-8 by default

In other words, if you are using legacy encoding on Unix, it may be not roundtripping.

If the encoding supports it, since which Python version do Py_DecodeLocale() and os.fsencode() roundtrip?

The background of my question is that Mercurial goes some extra rounds to determine the correct encoding to emulate what Py_EncodeLocale() would do: https://www.mercurial-scm.org/repo/hg/file/5.4.1/mercurial/pycompat.py#l157 . If os.fsencode() could be used, it would simplify the code. Mercurial supports Python 3.5+.

>
> Manuel Jacob <me@manueljacob.de> added the comment:
>
> If the encoding supports it, since which Python version do
> Py_DecodeLocale() and os.fsencode() roundtrip?
>

Maybe, since Python 3.2. FWIW, fsencode is added by Victor in
https://bugs.python.org/issue8514

> The background of my question is that Mercurial goes some extra rounds to
> determine the correct encoding to emulate what Py_EncodeLocale() would do:
> https://www.mercurial-scm.org/repo/hg/file/5.4.1/mercurial/pycompat.py#l157
> . If os.fsencode() could be used, it would simplify the code. Mercurial
> supports Python 3.5+.
>
>
>

I think it is a right approach.
One of the important use case of os.fsencode is using file path from
sys.argv even if it can not be decoded by filesystem encoding.