subtle side effect of generator/generator expression

[bonono]

Hi,

I initially thought that generator/generator expression is cool(sort of
like the lazy evaluation in Haskell) until I notice this side effect.

a=(x for x in range(2))
list(a) [1,2]list(a)

[]

Would this make generator/generator expression's usage pretty limited ?
As when the program/system goes beyond a single module, this behaviour
can cause subtle bugs ?

[Fredrik Lundh]

bo****@gmail.co m wrote:

I initially thought that generator/generator expression is cool (sort of
like the lazy evaluation in Haskell) until I notice this side effect.

a=(x for x in range(2))
list(a) [1,2]list(a) []

Would this make generator/generator expression's usage pretty limited ?
nope.
As when the program/system goes beyond a single module, this behaviour
can cause subtle bugs ?

sure, in the same way as

f = open(filename)
f.read() 'hello world\n' f.read() # oops!

''

causes subtle bugs (that is, almost never)

</F>

[bonono]

That is exactly what I meant, in fact. These IO thing are expected to
have side effects so they are not subtle. Generator on the other hand,
is sort of "clever iteratables".

Now that I notice that, Of course I can be sure I would be careful. But
what about the following situation :

I import some function from some third party module which said, "oh, my
function returns a iteratable". Then I can only list(o) if I want to
pass it around or I consume it in only one and only one place. turning
it into a list completely negate what generator is intended
for(generate as much as you need) and using it in only one and only one
place(like your fread example) would IMO means its usage is pretty
limited.

Beside, I can do a fseek to rewind a file handle but not a generator
object.

Fredrik Lundh wrote:

As when the program/system goes beyond a single module, this behaviour
can cause subtle bugs ?

sure, in the same way as

>>> f = open(filename)
>>> f.read() 'hello world\n' >>> f.read() # oops!

''

causes subtle bugs (that is, almost never)

</F>

[Diez B. Roggisch]

bo****@gmail.co m wrote:

That is exactly what I meant, in fact. These IO thing are expected to
have side effects so they are not subtle. Generator on the other hand,
is sort of "clever iteratables".

Now that I notice that, Of course I can be sure I would be careful. But
what about the following situation :

I import some function from some third party module which said, "oh, my
function returns a iteratable". Then I can only list(o) if I want to
pass it around or I consume it in only one and only one place. turning
it into a list completely negate what generator is intended
for(generate as much as you need) and using it in only one and only one
place(like your fread example) would IMO means its usage is pretty
limited.

Beside, I can do a fseek to rewind a file handle but not a generator
object.

Its a question of protocol. A iterator only guarantes that you can fetch
items from it until it is possibly exhausted. In the same way, a stream
may only guarantee you that you can read from it. If you need more, you
have tgo use a different protocol. Lists are one that allows you to
randomly access it. Files allow to seek, in addition to stream semantics.

And given that python is an imperative language with side-effects, you
don't know what sideeffects a generator can cause - which makes a
"rewind" or copy-semantics to make it work several time impossible (in
the general case, that is). So it has to focus on what it _can_ guarantee.

Haskell and other FP languages canwork differently here, as sideeffects
are made excplicit using e.g. monads or other means - which allows for
re-use of a generator-like construct. But there is no way to make this
work in pythons paradigm.

Diez

[Fredrik Lundh]

Diez B. Roggisch wrote:

Files allow to seek, in addition to stream semantics.

Some files. Not all files support seek operations. Some only support
forward seek.

</F>

[Simon Percivall]

If you find that you want to iterate over an iterable multiple times,
have a look at the solution that the tee() function in the itertools
module provides (http://docs.python.org/lib/itertools-functions.html).
(Have a look at the rest of the itertools module as well, for that
matter.)

[bonono]

True. That is why I have now reverted back to use list whenever
possible. As while list can also be modified(say in a multi-thread
situation), at least if I don't do the update(coding policy, practice
or whatever), they are sort of "guaranteed ".

I would only use generator as IO monad in Haskell, i.e. don't use it
unless I absolutely need to.

Diez B. Roggisch wrote:

Its a question of protocol. A iterator only guarantes that you can fetch
items from it until it is possibly exhausted. In the same way, a stream
may only guarantee you that you can read from it. If you need more, you
have tgo use a different protocol. Lists are one that allows you to
randomly access it. Files allow to seek, in addition to stream semantics.

And given that python is an imperative language with side-effects, you
don't know what sideeffects a generator can cause - which makes a
"rewind" or copy-semantics to make it work several time impossible (in
the general case, that is). So it has to focus on what it _can_ guarantee.

Haskell and other FP languages canwork differently here, as sideeffects
are made excplicit using e.g. monads or other means - which allows for
re-use of a generator-like construct. But there is no way to make this
work in pythons paradigm.

Diez

[Steven D'Aprano]

On Sun, 16 Oct 2005 15:52:54 +0200, Fredrik Lundh wrote:

bo****@gmail.co m wrote:

I initially thought that generator/generator expression is cool (sort of
like the lazy evaluation in Haskell) until I notice this side effect.

>>>a=(x for x in range(2))
>>>list(a)

[1,2]

>>>list(a)

[]

Would this make generator/generator expression's usage pretty limited ?

nope.

In fairness, it is a pretty big Gotcha.

As when the program/system goes beyond a single module, this behaviour
can cause subtle bugs ?

sure, in the same way as

>>> f = open(filename)
>>> f.read() 'hello world\n' >>> f.read() # oops!

''

causes subtle bugs (that is, almost never)

Are you saying that the bugs it causes aren't subtle? *wink*
--
Steven

[bonono]

thanks. I was looking for scanl in itertools but can't find it so I
implement my own then run into some subtle bugs which first made me
think my scanl is the problem. Then notice my wrong perception about
generator(and iterable in general, though the built-in iterables like
list, dict don't seem to have this characteristic) .

Simon Percivall wrote:

If you find that you want to iterate over an iterable multiple times,
have a look at the solution that the tee() function in the itertools
module provides (http://docs.python.org/lib/itertools-functions.html).
(Have a look at the rest of the itertools module as well, for that
matter.)

[vivainio]

> Are you saying that the bugs it causes aren't subtle? *wink*

Exactly. Destructive generator problems are caught almost immediately.