[Haskell-beginners] lazy IO in readFile
Stephen Blackheath [to Haskell-Beginners]
mutilating.cauliflowers.stephen at blacksapphire.com
Sat May 8 07:38:25 EDT 2010
On 08/05/10 15:59, Andrew Sackville-West wrote:
> On Sat, May 08, 2010 at 03:41:43PM +1200, Stephen Blackheath [to Haskell-Beginners] wrote:
>> In Haskell, lazy I/O is a form of cheating, because Haskell functions
>> are supposed to have no side effects, and lazy I/O is a side effect. At
>> first, cheating seems attractive, but it takes a bit of experience to
>> really understand why cheating really is not a good idea, and that
>> Haskell is so powerful that it doesn't matter that you shouldn't
> So, are you saying that using something like readFile is cheating? Or
> just that lazy IO itself is cheating?
I'm saying that 'readFile' is lazy I/O, lazy I/O is cheating, and
cheating is bad. I want you to understand what I mean by "bad" rather
than taking it on authority. The thing is, Haskell is all about giving
you safety. Haskell gives you certain advantages that come only from
going all the way, saying that pure really means pure. Cheating means
that you can't make this assumption, and the rot can spread more quickly
than you expect. This is fine for small programs, but can be a road to
ruin in large programs. So, why cheat when Haskell makes it so easy for
you not to, and rewards you so handsomely for your good behaviour?
You may ask, if that's true, then why is 'readFile' in the standard
libraries? My own opinion on this subject is that even though the
designers of Haskell had incredible foresight, they didn't have our
luxury of experience programming in the language they were designing.
Lazy I/O is _very_ convenient, after all.
All this is my opinion, but I honestly believe that the great majority
of hardcore Haskellers agree with me that cheating is bad.
>> So the short answer is, no - there is no way to force the file returned
>> by readFile to close.
> I figured as much. I'm not completely unhappy with my solution since
> it irks me to write out an empty list anyway. And it's really a simple
> little project for my personal use...
I will sometimes use lazy I/O if the program is simple enough, but it
seems like you're already getting into a situation where it's causing
It's a difficult one, because, like I said, lazy I/O really is useful.
My own approach is this: I ask myself, "Is it *really* referentially
transparent?" If you can say "yes" to this question, then it's
logically equivalent to not cheating (but you must still feel guilty).
In the case of 'readFile', this becomes, "Can I assume that for the life
of the program, the file's contents will never change?" (Haskell abides
in a Zen-like 'eternal now'.) For a program that reads a config file
once on startup, the answer might be "yes", and cheating may not
introduce any risks.
Perhaps I can give you some more insight by saying this: Laziness adds
complexity to the reasoning necessary to understand how your program
will execute. Purity means that this complexity is neutralized and
becomes the compiler's problem, not yours. You have to remember that
while the tiniest little piece of information calculated from the
contents of your 'readFile' remains unevaluated, the file will not
close. That's very difficult to reason about.
Another thing to consider is, if the code you're implementing relies on
lazy I/O, might you want to use it in a big program? If so, surely it
would be better to do it in a more general way to begin with. One of
the things monads are especially good for is replacing lazy I/O.
You might object, "Lazy I/O is so incredibly brilliant, but you are
telling me I can't use it! That really ruins Haskell for me if one of
its most amazing features is not allowed!"
I know this seems very unfair. But my reply is, "Purity really is
*that* good. It's even worth giving up lazy I/O for - that's how good it
>> I'd recommend using withFile and hGetLine, like this:
>> withFile "testfile" ReadMode $ \h -> do
>> l <- hGetLine h
> and using this to read through the entire file and then closing it?
> (Don't answer that, I'll do the reading). hmm... a little thought
> suggests that laziness will still get me unless I put some strictness
> in somewhere. I'm still left with a case where the history list is
> never completely evaluated forcing the reading of EOF. I will apply
> some thought to it and see what happens.
I can't resist a couple of comments: Note that withFile closes the
handle for you explicitly. It's completely safe in that respect (unless
you pass 'h' as a return value from withFile - which is obviously a bad
idea - but if you want the type system to make this impossible, this
*can* be achieved!). You know exactly when it's being closed. With
withFile/hGetLine, laziness can't get you, except in the usual way (that
is, as it relates to memory and CPU usage).
A parting thought: One of the great things about Haskell (compared with
imperative programming) is that there are dozens of things you don't
have to reason about any more, so you can concentrate on solving your
problem. Do you see what I'm saying? Why even bother reasoning about
whether laziness can get you? Just make everything pure and you don't
have to. *
(* I don't want to mislead you and make you think Haskell is something
it's not. Therefore I need to add here that you *do* need to reason
about the *space and CPU usage* of your code in the presence of
laziness. IMO this is the only serious cost of using Haskell - the rest
>> On 08/05/10 14:47, Andrew Sackville-West wrote:
>>> I'm trying to suss out the best (heh, maybe most idiomatic?) way to
>>> handle laziness in a particular file operation. I'm reading a file
>>> which contains a list of rss feed items that have been seen
>>> previously. I use this list to eliminate feed items I've seen before
>>> and thus filter a list of new items. (it's a program to email me feed
>>> items from a couple of low frequency feeds).
>>> So, the way I do this is to open the history file with readFile, read
>>> it into a list and then use that as a parameter to a filter
>>> function. Instead of getting confusing, here is some simple code that
>>> gets at the nut of the problem:
>>> import Control.Monad
>>> isNewItem :: [String] -> String -> Bool
>>> isNewItem  = \_ -> True
>>> isNewItem ts = \x -> not (any (== x) ts)
>>> filterItems :: [String] -> [String] -> [String]
>>> filterItems old is = filter (isNewItem old) is
>>> getOldData :: IO [String]
>>> getOldData = catch (liftM lines $ readFile "testfile") (\_ -> return )
>>> main = do
>>> let testData = ["a", "b", "c", "d"] :: [String]
>>> currItems <- getOldData
>>> let newItems = filterItems currItems $ testData
>>> print newItems -- this is important, it mimics another IO action I'm
>>> -- doing in the real code...
>>> appendFile "testfile" . unlines $ newItems
>>> Please ignore, for the moment, whatever *other* problems (idiomatic or
>>> just idiotic) that may exist above and focus on the IO problem.
>>> This code works fine *if* the file "testfile" has in it some subset of the
>>> testData list. If it has the complete set, it fails with a "resource
>>> busy" exception.
>>> Okay, I can more or less understand what's going on here. Each letter
>>> in the testData list gets compared to the contents of the file, but
>>> because they are *all* found, the readFile call never has to try and
>>> fail to read the last line of the file. Thus the file handle is kept
>>> open lazily waiting around not having reached EOF. Fair enough.
>>> But what is the best solution? One obvious one, and the one I'm using
>>> now, is to move the appendFile call into a function with guards to
>>> prevent trying to append an empty list to the end of the file. This
>>> solves the problem not by forcing the read on EOF, but by not
>>> bothering to open the file for appending:
>>> writeHistory  = return ()
>>> writeHistory ni = appendFile "testfile" . unlines $ ni
>>> And this makes some sense. It's silly to try to write nothing to a
>>> But it also rubs me the wrong way. It's not solving the problem
>>> directly -- closing that file handle. So there's my question, how can
>>> I close that thing? Is there some way to force it? Do I need to rework
>>> the reading to read one line ahead of whatever I'm testing against
>>> (thereby forcing the read of EOF and closing the file)?
More information about the Beginners