For review: PEP 343: Anonymous Block Redux and Generator Enhancements

[Guido van Rossum]

After many rounds of discussion on python-dev, I'm inviting public
comments for PEP 343. Rather than posting the entire PEP text here,
I'm inviting everyone to read it on line
(http://www.python.org/peps/pep-0343.html) and then post comments on a
Wiki page I've created for this purpose
(http://wiki.python.org/moin/WithStatement).

I think this is a good one; I hope people agree. Its acceptance will
obsolete about 4 other PEPs! (A sign that it fulfills a need and that
the proposed solution is powerful.)

 

[Nicolas Fleury]

After many rounds of discussion on python-dev, I'm inviting public
comments for PEP 343. Rather than posting the entire PEP text here,
I'm inviting everyone to read it on line
(http://www.python.org/peps/pep-0343.html) and then post comments on a
Wiki page I've created for this purpose
(http://wiki.python.org/moin/WithStatement).

I think this is a good one; I hope people agree. Its acceptance will
obsolete about 4 other PEPs! (A sign that it fulfills a need and that
the proposed solution is powerful.)

I like the PEP very much; I guess most C++ programmers are missing that
capability in Python. (I was following the discussion on python-dev,
and I'm pleased and surprised how good the result/compromise is).

What about making the ':' optional (and end implicitly at end of current
block) to avoid over-indentation?

def foo():
with locking(someMutex)
with opening(readFilename) as input
with opening(writeFilename) as output
...

would be equivalent to:

def foo():
with locking(someMutex)
with opening(readFilename) as input
with opening(writeFilename) as output
...

Regards,

Nicolas

 

[Andrew Dalke]

What about making the ':' optional (and end implicitly at end of current
block) to avoid over-indentation?

def foo():
with locking(someMutex)
with opening(readFilename) as input
with opening(writeFilename) as output
...

would be equivalent to:

def foo():
with locking(someMutex)
with opening(readFilename) as input
with opening(writeFilename) as output
...

Nothing in Python ends at the end of the current block.
They only end with the scope exits. The order of deletion
is not defined, and you would change that as well.

Your approach wouldn't allow the following

with locking(mutex):
increment_counter()

x = counter()

with locking(mutex):
decrement_counter()


except by making a new block, as

if 1:
locking(mutex)

x = counter()

if 1:
locking(mutex)


If the number of blocks is a problem it wouldn't be that
hard to do

with multi( locking(someMutex),
opening(readFilename),
opening(writeFilename) ) as _, input, output:
...

Untested sketch of an implementation


class multi(object):
def __init__(self, *args):
self.args = args
def __enter__(self):
results = []
for i, arg in enumerate(self.args):
try:
results.append(arg.__enter__())
except:
# back up through the already __entered__ args
exc = sys.exc_info()
for j in range(i-1, -1, -1):
try:
self.args[j].__exit__(*exc)
except:
# Need to get the new exception, to match the PEP behavior
exc = sys.exc_info()
raise exc[0], exc[1], exc[2]
return results

def __exit__(self, type, value, traceback):
for arg in self.args[::-1]:
try:
arg.__exit__(type, value, traceback)
except:
type, value, traceback = sys.exc_info()

Andrew
(e-mail address removed)

 

[Nicolas Fleury]

Nothing in Python ends at the end of the current block.
They only end with the scope exits. The order of deletion
is not defined, and you would change that as well.

There's no change in order of deletion, it's just about defining the
order of calls to __exit__, and they are exactly the same. As far as I
know, PEP343 has nothing to do with order of deletion, which is still
implementation-dependant. It's not a constructor/destructor thing like
in C++ RAII, but __enter__/__exit__.

But yes, it creates a precedent by creating a statement affecting the
end of the current indentation block. But that's what this PEP is all
about...

Your approach wouldn't allow the following

No, I said making the ':' *optional*. I totally agree supporting ':' is
useful.

If the number of blocks is a problem it wouldn't be that
hard to do

with multi( locking(someMutex),
opening(readFilename),
opening(writeFilename) ) as _, input, output:
...

True. But does it look as good? Particularly the _ part?

Regards,
Nicolas

 

[Ilpo =?iso-8859-1?Q?Nyyss=F6nen?=]

What about making the ':' optional (and end implicitly at end of current
block) to avoid over-indentation?

def foo():
with locking(someMutex)
with opening(readFilename) as input
with opening(writeFilename) as output
...

How about this instead:

with locking(mutex), opening(readfile) as input:
...

So there could be more than one expression in one with.

would be equivalent to:

def foo():
with locking(someMutex)
with opening(readFilename) as input
with opening(writeFilename) as output
...

The thing is that with normal try-finally block, you can add more
things to it easily. This kind of with thing does not allow adding
more stuff to it in any other way than adding more indentation.

Anyway, I like the idea of the PEP too.

 

[Andrew Dalke]

There's no change in order of deletion, it's just about defining the
order of calls to __exit__, and they are exactly the same.

BTW, my own understanding of this is proposal is still slight.
I realize a bit better that I'm not explaining myself correctly.

As far as I
know, PEP343 has nothing to do with order of deletion, which is still
implementation-dependant. It's not a constructor/destructor thing like
in C++ RAII, but __enter__/__exit__.

I'm mixing (because of my lack of full comprehension) RAII with
your proposal.

What I meant to say was in the PEP

with locking(someMutex)
with opening(readFilename) as input
with opening(writeFilename) as output
...

it's very well defined when the __exit__() methods are
called and in which order. If it's


with locking(someMutex)
with opening(readFilename) as input
with opening(writeFilename) as output

with the __exit__()s called at the end of the scope (as if it
were a __del__, which it isn't) then the implementation could
still get the __exit__ order correct, by being careful. Though
there would be no way to catch an exception raised in an __exit__.
I think.

No, I said making the ':' *optional*. I totally agree supporting ':' is
useful.

Ahh, I think I understand. You want both

with abc:
with cde:
pass

and

with abc
with def

and to have the second form act somewhat like RAII in that
the __exit__() for that case is called when the scope ends.


Hmm. My first thought is I don't like it because I'm a stodgy
old traditionalist and don't like the ambiguity of having to look
multiple tokens ahead to figure out which form is which.

I can see that it would work. Umm, though it's tricky. Consider

with abc

with defg:
with ghi
with jkl:
1/0



The implementation would need to track all the with/as forms
in a block so they can be __exit__()ed as appropriate. In this
case ghi.__exit() is called after jkl.__exit__() and
before defg.__exit__

The PEP gives an easy-to-understand mapping from the proposed
change to how it could be implemented by hand in the existing
Python. Can you do the same?

True. But does it look as good? Particularly the _ part?

I have not idea if the problem you propose (multiple with/as
blocks) will even exist so I can't comment on which solution
looks good. It may not be a problem in real code, so not needing
any solution.

Andrew
(e-mail address removed)

 

[Robin Becker]

Ilpo Nyyssönen wrote:
........

with locking(mutex), opening(readfile) as input:
...

.....

with EXPR as x:
BLOCK

EXPR can be a tuple so the above would be ambiguous.

 

[Kent Johnson]

with EXPR as x:
BLOCK

EXPR can be a tuple so the above would be ambiguous.

I don't think EXPR can be a tuple; the result of evaluating EXPR must have __enter__() and __exit__() methods. *x* can be a tuple.

Kent

 

[Robin Becker]

I don't think EXPR can be a tuple; the result of evaluating EXPR must
have __enter__() and __exit__() methods. *x* can be a tuple.

Kent

Well perhaps this would fly then. I reread the PEP and it says EXPR is
arbitrary, but cannot be an expression list so maybe this was already
considered and rejected.

 

[oren.tirosh]

How about this instead:

with locking(mutex), opening(readfile) as input:
...

+1, and add PEP-328-like parentheses for multiline.

Oren

 

[Steven Bethard]

How about this instead:

with locking(mutex), opening(readfile) as input:
...

I don't like the ambiguity this proposal introduces. What is input
bound to? The return value of locking(mutex).__enter__() or the return
value of opening(readfile).__enter__()? Seems ambiguous to me. And is
the file opened with the mutex held, or not? Sure, all of these
questions can be answered with an arbitrary decision. But the point is
that, whatever decision you make, I now have to *memorize* that decision.

Note that if I wrote:

with locking(mutex):
with opening(readfile) as input:
...

it's clear that input is the return value of
opening(readfile).__enter__(), and that the mutex is held while the file
is opened. I don't need to memorize these things; they are explicit in
the syntax.

I can see making the with-statement proposal more complex if you had
some very good motivating examples for wanting the multiple-expressions
extension. But you have yet to provide a real-world use case. Go
search your codebase, and find some examples of where you would actually
use this. For the complexity that you want to add to the
with-statement, you need to show that there's a *large* advantage to a
*variety* of use cases in *real-world* code.

STeVe

 

[Nicolas Fleury]

The implementation would need to track all the with/as forms
in a block so they can be __exit__()ed as appropriate. In this
case ghi.__exit() is called after jkl.__exit__() and
before defg.__exit__

The PEP gives an easy-to-understand mapping from the proposed
change to how it could be implemented by hand in the existing
Python. Can you do the same?

I think it is simple and that the implementation is as much
straight-forward. Think about it, it just means that:

> with abc
>
> with defg:
> with ghi
> with jkl:
> 1/0

is equivalent to:

> with abc:
>
> with defg:
> with ghi:
> with jkl:
> 1/0

That's it. Nothing more complex. It's only about syntax.

I have not idea if the problem you propose (multiple with/as
blocks) will even exist so I can't comment on which solution
looks good. It may not be a problem in real code, so not needing
any solution.

Good point. As a C++ programmer, I use RAII a lot. However, there's
situations in C++ that don't apply, like allocating dynamic memory in a
scope. However, I still expect some Python programmers to use it enough
to ask for that syntax to be added, in the same way operators like +=
have been added.

But there's another point that has nothing to do with how many "with"
statements you have in a function. In C++, very very rarely I've seen
something like:

void foo() {
{ // (define a scope for the Lock object)
Lock locking(myMutex);
...
}
...
{
Lock locking(myMutex);
...
}
}

So I come to another conclusion: the indentation syntax will most of the
time result in a waste of space. Typically a programmer would want its
with-block to end at the end of the current block.

So basically, there's two 10-90% points, one in favor of my proposal,
one against:

- Most of the time, you don't have a lot of with-statements in a single
function, so not so much indentation.
- Most of the time, a with-statement ends at the end of current block,
so indentation-syntax most of the time result in a waste of space.

The way to see my proposal is not "to be used when you have multiple
with-blocks" but instead "never use the ':' syntax, unless necessary".
The day some code need it, it's very easy to add a ':' and indent some
code with our favorite editor.

Regards,
Nicolas

 

[Andrew Dalke]

I don't like the ambiguity this proposal introduces. What is input
bound to?

It would use the same logic as the import statement, which already
supports an 'as' like this

But the point is
that, whatever decision you make, I now have to *memorize* that decision.

It's the same rule so the rule would be "ahh, uses the 'as' form".

Andrew
(e-mail address removed)

 

[Andrew Dalke]

I think it is simple and that the implementation is as much
straight-forward. Think about it, it just means that:

Okay, I think I understand now.

Consider the following

server = open_server_connection()
with abc(server)
with server.lock()
do_something(server)

server.close()

it would be translated to

server = open_server_connection()
with abc(server):
with server.lock()
do_something(server)
server.close()

when I meant for the first code example to be implemented
like this

server = open_server_connection()
with abc(server):
with server.lock()
do_something(server)

server.close()


(It should probably use the with-block to handle the server open
and close, but that's due to my lack of imagination in coming up
with a decent example.)

Because of the implicit indentation it isn't easy to see that
the "server.close()" is in an inner block and not at the outer
one that it appears to be in. To understand the true scoping
a reader would need to scan the code for 'with' lines, rather
than just looking at the layout.

Good point. As a C++ programmer, I use RAII a lot.

And I've used it a few times in Python, before I found
out it wasn't a guaranteed behavior by the language.

So I come to another conclusion: the indentation syntax will most of the
time result in a waste of space. Typically a programmer would want its
with-block to end at the end of the current block.

A test for how often this is needed would be to look in existing
code for the number of try/finally blocks. I have seen and
written some gnarly deeply stacked blocks but not often - once
a year?

That's not to say it's a good indicator. A lot of existing code
looks like this

def get_first_line(filename):
f = open(filename)
return f.readline()

depending on the gc to clean up the code. A more ... not
correct, but at least finicky ... implementation could be

def get_first_line(filename):
f = open(filename)
try:
return f.readline()
finally:
f.close()

Almost no one does that. With the PEP perhaps the idiomatic
code would be

def get_first_line(filename):
with open(filename) as f:
return f.readline()


(Add __enter__/__exit__ semantics to the file object? Make
a new 'opening' function? Don't know.)

What I mean by all of this is that the new PEP may encourage
more people to use indented blocks, in a way that can't be
inferred by simply looking at existing code. In that case
your proposal, or the one written

with abc, defg(mutex) as D, server.lock() as L:
..

may be needed.

Andrew
(e-mail address removed)

 

[Steven Bethard]

It would use the same logic as the import statement, which already
supports an 'as' like this

Ahh, so if I wanted the locking one I would write:

with locking(mutex) as lock, opening(readfile) as input:
...

There was another proposal that wrote this as:

with locking(mutex), opening(readfile) as lock, input:
...

which is what was confusing me. Mirroring the 'as' from the import
statement seems reasonable.


But it doesn't address my other concern, namely, is

with locking(mutex), opening(readfile) as input:
...

equivalent to the nested with-statements, e.g.:

_locking = locking(mutex)
_exc1 = (None, None, None)
_locking.__enter__()
try:
try:
_opening = opening(readfile)
_exc2 = (None, None, None)
input = _opening.__enter__()
try:
try:
...
except:
_exc2 = sys.exc_info()
raise
finally:
_opening.__exit__(*exc)
except:
_exc1 = sys.exc_info()
raise
finally:
_locking.__exit__(*exc)

Or is it equivalent to something different, perhaps:

_locking = locking(mutex)
_opening = opening(readfile)
_exc = (None, None, None)
_locking.__enter__()
input = _opening.__enter__()
try:
try:
...
except:
_exc = sys.exc_info()
raise
finally:
_opening.__exit__(*exc)
_locking.__exit__(*exc)

Or maybe:

_locking = locking(mutex)
_opening = opening(readfile)
_exc = (None, None, None)
_locking.__enter__()
input = _opening.__enter__()
try:
try:
...
except:
_exc = sys.exc_info()
raise
finally:
# same order as __enter__ calls this time!!
_locking.__exit__(*exc)
_opening.__exit__(*exc)

All I'm saying is that any of these are possible given the syntax. And
I don't see a precedent in Python for preferring one over another.
(Though perhaps there is one somewhere that I'm missing...)

And if it *is* just equivalent to the nested with-statements, how often
will this actually be useful? Is it a common occurrence to need
multiple with-statements? Is the benefit of saving a level of
indentation going to outweigh the complexity added by complicating the
with-statement?

STeVe

 

[Andrew Dalke]

Ahh, so if I wanted the locking one I would write:

with locking(mutex) as lock, opening(readfile) as input:
...

That would make sense to me.

There was another proposal that wrote this as:

with locking(mutex), opening(readfile) as lock, input:
...

which is what was confusing me. Mirroring the 'as' from the import
statement seems reasonable.

Ahh, you're right. That was an earlier proposal.

But it doesn't address my other concern, namely, is

with locking(mutex), opening(readfile) as input:
...

equivalent to the nested with-statements, e.g.:

I would think it's the same as

with locking(mutex):
with opening(readfile) as input:
...

which appears to map to the first of your alternatives

Or is it equivalent to something different, perhaps:

_locking = locking(mutex)
_opening = opening(readfile)
_exc = (None, None, None)
_locking.__enter__()
input = _opening.__enter__()
try:
try:
...
except:
_exc = sys.exc_info()
raise
finally:
_opening.__exit__(*exc)
_locking.__exit__(*exc)

That wouldn't work; consider if _opening.__enter__() raised
an exception. The _locking.__exit__() would never be called,
which is not what anyone would expect from the intent of
this PEP.

Or maybe:

_locking = locking(mutex)
_opening = opening(readfile)
_exc = (None, None, None)
_locking.__enter__()
input = _opening.__enter__()

Same problem here

finally:
# same order as __enter__ calls this time!!
_locking.__exit__(*exc)
_opening.__exit__(*exc)

and the order would be wrong since consider multiple
statements as

with server.opening() as connection, connection.lock(column) as C:
C.replace("X", "Y")

The inner with depends on the outer and must be closed
in inverted order.

And if it *is* just equivalent to the nested with-statements, how often
will this actually be useful? Is it a common occurrence to need
multiple with-statements? Is the benefit of saving a level of
indentation going to outweigh the complexity added by complicating the
with-statement?

Agreed.

Andrew
(e-mail address removed)

 

[Nicolas Fleury]

How about this instead:

with locking(mutex), opening(readfile) as input:
...

So there could be more than one expression in one with.

I prefer the optional-indentation syntax. The reason is simple (see my
discussion with Andrew), most of the time the indentation is useless,
even if you don't have multiple with-statements. So in my day-to-day
work, I would prefer to write:

def getFirstLine(filename):
with opening(filename) as file
return file.readline()

than:

def getFirstLine(filename):
with opening(filename) as file:
return file.readline()

But I agree that in the case of only one with-statement, that's no big deal.

Also, if multiple with-statements are separated by other indentating
statements, your proposal doesn't help:

with locking(lock)
if condition:
with opening(filename) as file
for line in file:
...

would still be needed to be written:

with locking(lock):
if condition:
with opening(filename) as file:
for line in file:
...

Regards,
Nicolas

 

[Nicolas Fleury]

Consider the following

server = open_server_connection()
with abc(server)
with server.lock()
do_something(server)

server.close()

it would be translated to

server = open_server_connection()
with abc(server):
with server.lock()
do_something(server)
server.close()

when I meant for the first code example to be implemented
like this

server = open_server_connection()
with abc(server):
with server.lock()
do_something(server)

server.close()

That's an interesting point. But I'm not sure if it is a realistic
error. It would be like using a with-statement without knowing what it
does. Also, and it seems we agree, these cases are very rare.

(It should probably use the with-block to handle the server open
and close, but that's due to my lack of imagination in coming up
with a decent example.)

Because of the implicit indentation it isn't easy to see that
the "server.close()" is in an inner block and not at the outer
one that it appears to be in. To understand the true scoping
a reader would need to scan the code for 'with' lines, rather
than just looking at the layout.

But with both syntaxes you only look at the indentation layout, no? If
you want to know the "scope" of something as you said, it means you have
already scan its "with" statement. You only need after that to look at
the indentation layout, as with indentation syntax.

It's however less obvious at first look that with-statements implies
some scope, but I feel that Python newcomers usually do the opposite
error instead, thinking their variables have a defined scope as in some
other languages.

A test for how often this is needed would be to look in existing
code for the number of try/finally blocks. I have seen and
written some gnarly deeply stacked blocks but not often - once
a year?

That's not to say it's a good indicator. A lot of existing code
looks like this

I agree. It's hard to find a good indicator. In my case I use both my
Python code (with try/finally typically ending a function) and my C++
code (with typically no {} block created for RAII object scope). So, my
conclusion, most of the time the indentation will be useless.

What I mean by all of this is that the new PEP may encourage
more people to use indented blocks, in a way that can't be
inferred by simply looking at existing code. In that case
your proposal, or the one written

Totally agree. The with-statement will open the door to new programming
patterns in Python and it's hard to tell from status quo how much it
will be used.

Regards,
Nicolas

 

[Steven Bethard]

I prefer the optional-indentation syntax. The reason is simple (see my
discussion with Andrew), most of the time the indentation is useless,
even if you don't have multiple with-statements. So in my day-to-day
work, I would prefer to write:

def getFirstLine(filename):
with opening(filename) as file
return file.readline()

than:

def getFirstLine(filename):
with opening(filename) as file:
return file.readline()

One of the beauties of PEP 343 is that it can be explained simply in
terms of current Python syntax. The expression:

with EXPR as VAR:
BLOCK

is equivalent to:

abc = EXPR
exc = (None, None, None)
VAR = abc.__enter__()
try:
try:
BLOCK
except:
exc = sys.exc_info()
raise
finally:
abc.__exit__(*exc)

Can you do the same thing for your proposal? As I understand it you
want some sort of implicitly-defined BLOCK that starts the line after
the with statement and runs to the end of the current block...

Do you think the benefit of less indentation outweighs the added
complexity of explaining the construct? I still can't think of a good
way of explaining the semantics you want. If you could provide an
explanation that's simple and as explicit as Guido's explanation in PEP
343, I think that would help your case a lot.

STeVe

P.S. I think it's a great sign that people are mainly squabbling about
syntax here. More likely than not, Guido's already made his mind up
about the syntax. So, since no one seems to have any real problems with
the semantics, I'm hopeful that we'll get a direct implementation of PEP
343 in the next Python release. =)

 

[Nicolas Fleury]

Can you do the same thing for your proposal? As I understand it you
want some sort of implicitly-defined BLOCK that starts the line after
the with statement and runs to the end of the current block...

Yes. I totally agree with the syntax in the PEP, it provides a
necessary feature. I'm just suggesting to make the indentation
*optional*, because most of the time it is not necessary.

Do you think the benefit of less indentation outweighs the added
complexity of explaining the construct? I still can't think of a good
way of explaining the semantics you want. If you could provide an
explanation that's simple and as explicit as Guido's explanation in PEP
343, I think that would help your case a lot.

Since the current syntax would be there, the no-indentation syntax can
be explained in terms of the indentation syntax:

"""
To avoid over-indentation, a with-statement can avoid defining a new
indentation block. In that case, the end of the with block is the end
of the current indentation block.

with EXPR as VAR
REST OF BLOCK

is equivalent to

with EXPR as VAR:
BLOCK
"""

What do you think? I fail to see the complexity...

P.S. I think it's a great sign that people are mainly squabbling about
syntax here. More likely than not, Guido's already made his mind up
about the syntax. So, since no one seems to have any real problems with
the semantics, I'm hopeful that we'll get a direct implementation of PEP
343 in the next Python release. =)

It's important to note that nobody is against the PEP syntax. We are
only talking about adding things to it (optional indentation, or
multiple-as as with import).

All these changes can also be made later, so no proposition should slow
down the implementation of the PEP.

Regards,
Nicolas