python file API

[zipher]

For some time now, I've wanted to suggest a better abstraction for the <file> type in Python. It currently uses an antiquated C-style interface for moving around in a file, with methods like tell() and seek(). But after attributes were introduced to Python, it seems it should be re-addressed.

Let file-type have an attribute .pos for position. Now you can get rid ofthe seek() and tell() methods and manipulate the file pointer more easily with standard arithmetic operations.

You've now simplified the API by the removal of two obscure legacy methods and replaced them with a more basic one called "position".

Thoughts?

markj

 

[Dave Angel]

For some time now, I've wanted to suggest a better abstraction for the <file> type in Python. It currently uses an antiquated C-style interface for moving around in a file, with methods like tell() and seek(). But after attributes were introduced to Python, it seems it should be re-addressed.

Let file-type have an attribute .pos for position. Now you can get rid of the seek() and tell() methods and manipulate the file pointer more easily with standard arithmetic operations.

You've now simplified the API by the removal of two obscure legacy methods and replaced them with a more basic one called "position".

Thoughts?

markj

And what approach would you use for positioning relative to
end-of-file? That's currently done with an optional second parameter to
seek() method.

 

[Chris Kaynor]

And what approach would you use for positioning relative to
end-of-file? That's currently done with an optional second parameter to
seek() method.

I'm not advocating for or against the idea, but that could be handled
the same way indexing into lists can index relative to the end:
negative indices.

 

[Chris Angelico]

And what approach would you use for positioning relative to
end-of-file? That's currently done with an optional second parameter to
seek() method.

Presumably the same way you reference a list element relative to
end-of-list: negative numbers. However, this starts to feel like magic
rather than attribute assignment - it's like manipulating the DOM in
JavaScript, you set an attribute and stuff happens. Sure it's legal,
but is it right? Also, it makes bounds checking awkward:

file.pos = 42 # Okay, you're at position 42
file.pos -= 10 # That should put you at position 32
foo = file.pos # Presumably foo is the integer 32
file.pos -= 100 # What should this do?
foo -= 100 # But this sets foo to the integer -68
file.pos = foo # And this would set the file pointer 68 bytes from end-of-file.

I don't see it making sense for "file.pos -= 100" to suddenly put you
near the end of the file; it should either cap and put you at position
0, or do what file.seek(-100,1) would do and throw an exception. But
doing the exact same operation on a saved snapshot of the position and
reassigning it would then have quite different semantics in an unusual
case, while still appearing identical in the normal case.

ChrisA

 

[Dave Angel]

(forwarding to the list)

As size is an oft-useful construct, let it (like .name) be part of the
descriptor. Then


(Or, one could make slices part of the API...)

mark

Well, if one of the goals was to reduce the number of attributes, we're
now back to the original number of them.

 

[zipher]

You raise a valid point: that by abstracting the file pointer into a position attribute you risk "de-coupling" the conceptual link between the underlying file and your abstraction in the python interpreter, but I think the programmer can take responsibility for maintaining the abstraction.

The key possible fault will be whether you can trap (OS-level) exceptions when assigning to the pos attribute beyond the bounds of the actual file on the system...

markj

 

[zipher]

You raise a valid point: that by abstracting the file pointer into a position attribute you risk "de-coupling" the conceptual link between the underlying file and your abstraction in the python interpreter, but I think the programmer can take responsibility for maintaining the abstraction.

The key possible fault will be whether you can trap (OS-level) exceptions when assigning to the pos attribute beyond the bounds of the actual file on the system...

markj

 

[Ian Kelly]

file.pos = 42 # Okay, you're at position 42
file.pos -= 10 # That should put you at position 32
foo = file.pos # Presumably foo is the integer 32
file.pos -= 100 # What should this do?

Since ints are immutable, the language specifies that it should be the
equivalent of "file.pos = file.pos - 100", so it should set the file
pointer to 68 bytes before EOF.

foo -= 100 # But this sets foo to the integer -68
file.pos = foo # And this would set the file pointer 68 bytes from end-of-file.

Which is the same result.

I don't see it making sense for "file.pos -= 100" to suddenly put you
near the end of the file; it should either cap and put you at position
0, or do what file.seek(-100,1) would do and throw an exception.

I agree, but the language doesn't allow those semantics.

Also, what about the use of `f.seek(0, os.SEEK_END)` to seek to EOF?
I'm not certain what the use cases are, but a quick google reveals
that this does happen in real code. If a pos of 0 means BOF, and a
pos of -1 means 1 byte before EOF, then how do you seek to EOF without
knowing the file length?

 

[Chris Angelico]

Since ints are immutable, the language specifies that it should be the
equivalent of "file.pos = file.pos - 100", so it should set the file
pointer to 68 bytes before EOF.

Oh, I forgot that guaranteed equivalency. Well, at least it removes
the ambiguity. I don't like it though.

ChrisA

 

[Mark Lawrence]

For some time now, I've wanted to suggest a better abstraction for the <file> type in Python. It currently uses an antiquated C-style interface for moving around in a file, with methods like tell() and seek(). But after attributes were introduced to Python, it seems it should be re-addressed.

Let file-type have an attribute .pos for position. Now you can get rid of the seek() and tell() methods and manipulate the file pointer more easily with standard arithmetic operations.


You've now simplified the API by the removal of two obscure legacy methods and replaced them with a more basic one called "position".

Thoughts?

markj

This strikes me as being a case of if it ain't broke don't fix it.

 

[Chris Kaynor]

Since ints are immutable, the language specifies that it should be the
equivalent of "file.pos = file.pos - 100", so it should set the file
pointer to 68 bytes before EOF.

There is no reason that it has to be an int object, however. It could
well return a "FilePosition" object which does not allow subtraction
to produce a negative result. Not saying its a good idea... Similarly,
it could be a more complex object with properties on it to determine
whether to seek from beginning or end.

 

[Steven D'Aprano]

Presumably the same way you reference a list element relative to
end-of-list: negative numbers. However, this starts to feel like magic
rather than attribute assignment - it's like manipulating the DOM in
JavaScript, you set an attribute and stuff happens. Sure it's legal, but
is it right? Also, it makes bounds checking awkward:

file.pos = 42 # Okay, you're at position 42
file.pos -= 10 # That should put you at position 32
foo = file.pos # Presumably foo is the integer 32
file.pos -= 100 # What should this do?
foo -= 100 # But this sets foo to the integer -68
file.pos = foo # And this would set the file pointer 68 bytes
from end-of-file.

I don't see it making sense for "file.pos -= 100" to suddenly put you
near the end of the file; it should either cap and put you at position
0, or do what file.seek(-100,1) would do and throw an exception.

I would expect it to throw an exception, like file.seek and like list
indexing.

But
doing the exact same operation on a saved snapshot of the position and
reassigning it would then have quite different semantics in an unusual
case, while still appearing identical in the normal case.

But this applies equally to file.seek and list indexing today. In neither
case can you perform your own index operations outside of the file/list
and expect to get the same result, for the simple and obvious reason that
arithmetic doesn't perform the same bounds checking as actual seeking and
indexing.

 

[Steven D'Aprano]

You raise a valid point: that by abstracting the file pointer into a
position attribute you risk "de-coupling" the conceptual link between
the underlying file and your abstraction in the python interpreter

I don't think this argument holds water. With the ease of writing
attributes, it is more likely that people will perform file position
operations directly on file.pos rather than decoupling it into a
variable. Decoupling is more likely with file.seek, because it is so much
more verbose to use, and you get exactly the same lack of bounds checking:

py> f = open("junk", "w") # make a sample file
py> f.write("abcd\n")
py> f.close()
py> f = open("junk") # now do decoupled seek operations
py> p = f.tell()
py> p += 2000
py> p -= 4000
py> p += 2
py> p += 2000
py> f.seek(p)
py> f.read(1)
'c'


But really, who does such a sequence of arithmetic operations on the file
pointer without intervening reads or writes? We're arguing about
something that almost never happens.

By the way, the implementation of this is probably trivial in Python 2.x.
Untested:

class MyFile(file):
@property
def pos(self):
return self.tell()
@pos.setter
def pos(self, p):
if p < 0:
self.seek(p, 2)
else:
self.seek(p)

You could even use a magic sentinel to mean "see to EOF", say, None.

if p is None:
self.seek(0, 2)

although I don't know if I like that.

 

[Mark Adam]

There are many situations where a little bit of attribute access magic is a
good thing. However, operations that involve the underlying OS and that are
prone to raising exceptions even in bug free code should not be performed
implicitly like this. I find the following a little cryptic:
try:
f.pos = 256
except IOError:
print('Unseekable file')

Well it might be that the coupling between the python interpreter and
the operating system should be more direct and there should be a
special exception class that bypasses the normal overhead in the
CPython implementation so that error can be caught in the code without
breaking syntax. But I don't think I'm ready to argue that point....

markj

 

[Thomas Rachel]

Am 25.09.2012 04:28 schrieb Steven D'Aprano:

By the way, the implementation of this is probably trivial in Python 2.x.
Untested:

class MyFile(file):
@property
def pos(self):
return self.tell()
@pos.setter
def pos(self, p):
if p< 0:
self.seek(p, 2)
else:
self.seek(p)

You could even use a magic sentinel to mean "see to EOF", say, None.

if p is None:
self.seek(0, 2)

although I don't know if I like that.

The whole concept is incomplete at one place: self.seek(10, 2) seeks
beyond EOF, potentially creating a sparse file. This is a thing you
cannot achieve.

But the idea is great. I'd suggest to have another property:

[...]
@pos.setter
def pos(self, p):
self.seek(p)
@property
def eofpos(self): # to be consistent
return self.tell()
@eofpos.setter
def eofpos(self, p):
self.seek(p, 2)

Another option could be a special descriptor which can be used as well
for relative seeking:

class FilePositionDesc(object):
def __init__(self):
pass
def __get__(self, instance, owner):
return FilePosition(self)
def __set__(self, value):
self.seek(value)

class FilePosition(object):
def __init__(self, file):
self.file = file
def __iadd__(self, offset):
self.file.seek(offset, 1)
def __isub__(self, offset):
self.file.seek(-offset, 1)

class MyFile(file):
pos = FilePositionDesc()
[...]

Stop.

This could be handled with a property as well.

Besides, this breaks some other expectations to the pos. So let's
introduce a 3rd property named relpos:

class FilePosition(object):
def __init__(self, file):
self.file = file
self.seekoffset = 0
def __iadd__(self, offset):
self.seekoffset += offset
def __isub__(self, offset):
self.seekoffset -= offset
def __int__(self):
return self.file.tell() + self.seekoffset

class MyFile(file):
@property
def relpos(self):
return FilePosition(self) # from above
@relpos.setter
def relpos(self, ofs):
try:
o = ofs.seekoffset # is it a FilePosition?
except AttributeError:
self.seek(ofs, 1) # no, but ofs can be an int as well
else:
self.seek(o, 1) # yes, it is


Thomas

 

[Thomas Rachel]

Am 25.09.2012 00:37 schrieb Ian Kelly:

Since ints are immutable, the language specifies that it should be the
equivalent of "file.pos = file.pos - 100", so it should set the file
pointer to 68 bytes before EOF.

But this is not a "real int", it has a special use. So I don't think it
is absolutely required to behave like an int.

This reminds me of some special purpose registers in embedded
programming, where bits can only be set by hardware and are cleared by
the application by writing 1 to them.

Or some bit setting registers, like on ATxmega: OUT = 0x10 sets bit 7
and clears all others, OUTSET = 0x10 only sets bit 7, OUTTGL = 0x10
toggles it and OUTCLR = 0x10 clears it.

If this behaviour is documented properly enough, it is quite OK, IMHO.


Thomas

 

[Ulrich Eckhardt]

Am 24.09.2012 23:49, schrieb Dave Angel:

And what approach would you use for positioning relative to
end-of-file? That's currently done with an optional second
parameter to seek() method.

Negative indices.



Uli

 

[Mark Lawrence]

Well it might be that the coupling between the python interpreter and
the operating system should be more direct and there should be a
special exception class that bypasses the normal overhead in the
CPython implementation so that error can be caught in the code without
breaking syntax. But I don't think I'm ready to argue that point....

markj

Something along these lines
http://docs.python.org/dev/whatsnew/3.3.html#pep-3151-reworking-the-os-and-io-exception-hierarchy
?

 

[Steven D'Aprano]

Am 25.09.2012 04:28 schrieb Steven D'Aprano:


The whole concept is incomplete at one place: self.seek(10, 2) seeks
beyond EOF, potentially creating a sparse file. This is a thing you
cannot achieve.

On the contrary, since the pos attribute is just a wrapper around seek,
you can seek beyond EOF easily:

f.pos = None
f.pos += 10

But for anything but the most trivial usage, I would recommend sticking
to the seek method.

The problem with this idea is that the seek method takes up to three
arguments (the file being operated on, the position, and the mode), and
attribute syntax can only take two (the file, the position, e.g.:
file.pos = position). So either there are cases that file.pos cannot
handle (and so we need to keep tell/seek around, which leaves file.pos
redundant), or we need multiple attributes, one for each mode), or we
build a complicated, inconvenient API using special data types instead of
plain integers.

So all up, I'm -1 on trying to replace the tell/seek API, and -0 on
adding a second, redundant API.

Wait, there is another alternative: tuple arguments:

f.pos = (where, whence)

being the equivalent to seek(where, whence). At this point you just save
two characters "f.pos=a,b" vs "f.seek(a,b)" so it simply isn't worth it
for such a trivial benefit.

 

[Dennis Lee Bieber]

Or some bit setting registers, like on ATxmega: OUT = 0x10 sets bit 7
and clears all others, OUTSET = 0x10 only sets bit 7, OUTTGL = 0x10
toggles it and OUTCLR = 0x10 clears it.

If this behaviour is documented properly enough, it is quite OK, IMHO.

I don't think I'd want to work with any device where 0x10 (00010000
binary) modifies bit SEVEN. 0x40, OTOH, would fit my mental impression
of bit 7.

It doesn't even fit my mind if the value is suppose to be the /bit
number/ unless the device considers "bit 7" to be the EIGHTH bit (that
is, the LSB is considered bit 1, not bit 0)