Things I Love About Go
cambus.net
cambus.net
These are a lot of the benefits that Python had: everybody pretty much has to do things the same way (and there's PEP 8 that establishes non-enforced conventions), so I can start playing with a new codebase without having to figure out their conventions and build system too.
Goroutines and channels are cool because they are simple to use. We have tasks in c# and I could construct channels in countless ways. But go makes them dirt simple.
I don't like the terse c-like naming conventions but that is a manner of taste. Makes it feel more close to the metal and less enterprisey I guess.
I wish there was go#
Printf is all about "formatted printing". What you think the "f" at the end stands for and has stood for since C?
If anything, its the opposite of what you say, ie the "fmt" package name is not measleading but redundant!
for _, item := range slice_of_obj {
item.update()
}
It seems that the 'item' is a copy of, rather than an alias to, the underlying object, so the mutating 'update' call is updating the copy, which is then discarded. (The 'right way' to do this to use the loop index and slice_of_obj[i], but that seems mildly barbaric)
This seemed odd to me, particularly since no error was generated. Can anyone comment if I'm doing anything particularly dumb here? If Go had a concept of 'const', at least it could throw an error when you mutate the copy which is going to be discarded.
Btw code that uses range will always be faster than alternative with for loop and indexing (obj[i]) as range-based code can be optimized by Go compiler not to perform any bound checks.
I think ideally, it would work (as an alias) but if not, I think it should error. Or maybe this is a FAQ, with a good rationale, which I just haven't got to yet.
Whereas slice_of_obj[i] is accessing the element within the slice without a copy. One solution to this problem would be to store pointers in your slice instead of objects.
Some other languages (e.g. perl and I think also java) choose to make the loop var an alias, rather than a copy, to allow mutation:
for my $item (@items) {
$item->frob; # Can happily mutate the item
}
I liked some feature of Ada a lot, it was safe, input, output, in/out parameters for functions. Of course to make itself useful as a lower level language you need to be able to call OS functions, which in unix means C. So then you have to start wrapping... Some parts of Ada were terrible (string handling comes to mind.)
I think its worth a look, but I'm still not sure what problem its trying to solve.
[1] http://existentialtype.wordpress.com/2011/03/15/boolean-blin...
>My problem with Go is that it does not have sum-types and pattern-matching. This means that branching (conditionals) in Go do not gain any type-information. And that means that programmers have to manually keep track of the invariants that hold true in each of their conditionals, and if they get them wrong, they get no help from the compiler.
Example:
if (ptr == NULL) {
... a ...
} else {
... b ...
}
Instead, consider:
data Maybe a = Nothing | Just a
This is known as a "sum" type, because the possible values of the type are the sum of all data constructor values.
We could still use this sum data-type in a boolean-blind way:
if isJust mx then
.. use fromJust mx .. -- UNSAFE!
else
.. can't use content of mx ..
case mx of
Nothing -> ... There is no value of type "a" in our scope
Just x -> ... "x" of type "a" is now available in our scope!
"Maybe" is of course a simple example, but this is also applicable for any sum type at all.
If your language does branching without giving you back any new type information, it means you have to manually track the invariants and conditionals your program is in. If you get it wrong, your program will die at runtime. Instead, the language can almost always provide you some kind of proof that you can pass along of the conditional's truthness, which makes it safe to rely on it.
type Fruit struct {
Name string
Species string
}
// define a method on Fruit structs named Rename
func (f *Fruit) Rename(s string) {
f.Name = s
}
type Candy struct {
Name string
Company string
}
// define a method on Candy structs named Rename
func (c *Candy) Rename(s string) {
c.Name = s
}
// any type that defines a method named Rename is also
// of type Renamable.
type Renamable interface {
Rename(string)
}
// produce something that implements Renamable. Its
// fields will be unknown, but the method Rename will be
// known to exist.
func SomeProducer(...) Renamable {
...
}
x := SomeProducer(...)
// at this point, x is of type Renamable. We do not
// know what fields it has. The only thing we know is
// that it has some method named Rename that takes a
// single string for its only argument and returns
// nothing.
switch t := x.(type) {
case Fruit:
// T is of type Fruit, while x is of type Renamable
fmt.Print(t.Species)
case Candy:
// T is of type Candy, while x is of type Renamable
fmt.Print(t.Company)
default:
// some other type.
}
var x *MyType
...
switch x {
case nil:
// blah
default:
// blah
}
I'm not trying to assert that Go is better than Haskell or that they are the same; I've never programmed Haskell so I'm not qualified to make such a statement. I'm merely addressing the specific "Boolean Blindness" criticism.
I don't really think that criticism is in any way valid, since "Boolean Blindness" isn't a criticism of a language design, it's a criticism of how one might use a particular language. To say that Go committed "Boolean Blindness" is predicated on the following assumptions:
- you can only branch on a boolean
- given a value and a type, the only thing you can do is obtain a boolean indicating whether or not that value is of that type
as far as Go is concerned, those statements are factually inaccurate.
Given some variable `x` of unknown type, and a type `t`, is there some way, in Haskell, to create a boolean `v` such that the value of `v` is `true` when `x` is of type `t` and `false` otherwise? And if so, is there some way to branch on this boolean value `v`? Because if that is true, then how has Haskell not committed the same atrocity of "Boolean Blindness"?
But you do have to establish the provenance of your conditional. You have to keep track of whether you compared against nil yourself.
Whether this is expressed as if(x != nil) or as a boolean-blind switch case, the information is lost and it is up to the user to keep track of the meaning of 'x' under all the conditions it is run.
Of course Haskell is capable of expressing boolean blind and unsafe code. But Haskell, unlike go is also capable of expressing non-blind, safe code.
Go is only able to express boolean blind code (again, it is not about a boolean condition but about whether branching buys you type information). When you compare against nil, it is up to you to keep track of the provenance.
OK, so it's a bit like you're chaining assumptions together, rather than putting them next to one another? This makes Monads, esp. `>>=`, make a bunch more sense to me.
Like, the difference is between chaining/linking two operations (incl. checking what `x` is), by using the result of one check as the input to the other, and basically putting two operations next to one another (not to say that the `if` is meaningless, just that once the `if` is evaluated, the associated block has to take the `if`'s word for it).
The bonus is that Haskell handles the first part, the check, for you if you let it. It's an implicit (well sorta) and immutable precondition to whatever's next, at a logical/linguistic level.
And I guess the reason why booleans in particular are problematic is because you're sorta bolting them on, and they inherently compress information into yes/no, which is otherwise not terribly useful for the block following?
that's true in the case of comparing against nil, which I've already agreed with. I never said that Go is as safe as or safer than Haskell, I said that your assertion that there's "no way to branch and get type information" is a factual inaccuracy that misrepresents the language.
The fact that you haven't commented on type switches in any way shows that you're not really interested in anything other than your own sense of superiority.
Haskell is a better programming language. As a programmer of a programming language that is not Haskell, I am inferior to you.
There.
A type-switch seems to be non-boolean-blind indeed, but it seems to have too much syntactic overhead, as Go code generally uses if branches when a type-safe case is due (e.g: When analyzing return codes).
if x == nil { x.doSomethingCrazy(); // Runtime error? } else { ... }
"Boolean blindness" (bah) is that either the type system should catch this at compile time (because we know x is nil), or the language should make it impossible to phrase (pattern matching to pull the information out, such as with the Maybe type).
Go has many type system oddities (reflection instead of generics, for instance), but the things they _have_ done (auto-implementation of interfaces, for example) are pretty cool.
What's worse is that it appears this was all done out of ignorance, not conscious choice.
As far as I know, these guys have a reputation in industry moreso than in academic circles. It seems entirely believable that they have taken no interest in language design outside of industry, where the state of the art lags a few decades behind that of academia.
Consider all the error-checks in Go -- these are all incorrect from a boolean-blindness perspective, and that would be a lot of code to rewrite if they add proper constructs to Go. Thus, I really believe it is just a mistake -- one that is only possible to make out of ignorance.
if x == nil {
... use x ...
}
if x != nil {
... use x ...
}
match x with
| None -> ... (* use x *)
| Some _ -> () (* do nothing *)
match x with
| None -> () (* do nothing *)
| Some y -> ... (* use y *)
If you go down this road, you end up making the compiler perform a very difficult (probably undecidable) proof. There has been lots of research on this and it's not easy...
So is the distinction here that, basically, you push the distinction between a "nothing" and a "something" down into the language layer? IOW, the `if isJust mx` is the programmer trying to reason about a type, whereas `case mx of` is delegating that reasoning to the compiler/language?
In general I think that's a good principle, but at least in this example I don't quite grok why `if isJust mx` is out-and-out unsafe. Maybe this is just because the (heh) Maybe example is too simple to illustrate the unsafeness.
It does seem more error prone in practice, in terms of engineering & maintenance (read: more likely to lead to bugs). `isJust mx` means you are methodically separating out the verification of what type it is from the access and usage of that type, and since this is all happening at the program layer, you've got to keep the assumptions in your `if` consistent with the assumptions of your type. Conversely the `case of mx` statement leverages the type system to do the work for you.
I'll try to put it another way: the code inside your `if isJust` clause is not semantically related to the clause itself except in the most superficial way. However, with `case mx of`, those clauses are inextricably related since you've (again) pushed the semantics of checking the type and unboxing it down a layer, into Haskell.
Am I anywhere close?
The problem is: What if the condition is changed later? What if you use "fromJust" in the else branch? What if the code is moved around outside of the condition's scope?
The above is meant to illustrate that the "isJust/fromJust" anti-pattern is out-and-out unsafe. You get no compile-time assurances. If you avoid "fromJust" altogether (and equivalent constructs, of course), you have a 100% guarantee that you will not crash from expecting a Just where there's a Nothing (IOW, nil dereference).
Additionally, "Just" and "Nothing" are data constructors (can be viewed as value tags) within the Maybe type, they are not "types" themselves.
Except for these points, I think you are generally describing it correctly.
This kind of thing (avoiding boolean blindness) as well as a few more safety nets in Haskell is the reason lots of Haskell code is said to Just work once it compiles.
So when you say "avoid fromJust," the alternative in practice is to use pattern matching, yes? The example you used above was a case statement. A guard clause is another obvious one, as are additional clauses (?) in a function.
Guard clauses in Haskell are actually boolean conditionals, and are boolean-blind.
Safe pattern-matching can be done via the function definition sugar:
f Nothing = ...
f (Just x) = ...
The various functions that work with a Maybe value make it non-tedious to do things. For example, instead of:
case mx of
Just x -> Just (x + 1)
Nothing -> Nothing
fmap (+1) mx
channels -> rendezvous
Google's implementation makes use of greenlets, while gccgo is using OS threads. Further implementations might decide on others approaches.
As for channels/rendezvous you might be correct, as I only played a bit with Ada.
Go has more in common with Oberon and Component Pascal than Ada
Is this always this way? I'd be interested in a discussion on the pros/cons of this. Some are obvious, but I'm interested in a deep discussion.