Go channels are bad

Novices to the language have a tendency to overuse channels. Here is Andrew Gerrand addressing precisely this point two years ago at GopherCon: https://www.youtube.com/watch?v=u-kkf76TDHE&t=815

Neither the referenced performance characteristics via Tyler Treat, nor the FUD surrounding channel-based program architecture, invalidate channels generally. One does have to think carefully about ownership hierarchies: only one goroutine gets to close the channel. And if it's in a hot loop, a channel will always perform worse than a mutex: channels use mutexes internally. But plenty of problems are solved very elegantly with channel-based CSP-style message passing.

It's unfortunate that articles like this are written and gain traction. The explicit instruction to [new] Go programmers is that they should avoid channels, even that they are badly implemented, and both of those things are false.

Wait, why would you say that?

In general, if "orchestrating concurrency" involves guarding access to shared mutable state, then you can't orchestrate it at the callers site. It would be a massive encapsulation leak, because synchronization is not composable, requires special knowledge, plus you don't necessarily know how to synchronize unless you have knowledge about internals. Furthermore, because it is after the fact, your only choice of handling it is by means of mutexes, which has really terrible performance characteristics. Even if you could do ordering by other means, you end up paying the price of LOCK XCHG or whatever mutexes translate to, not to mention that you'll have problems if you want (soft) real-time behavior, because now you can end up with both dead-locks and live-locks.

And this brings us to another problem. If you end up doing such synchronization in Go, then Go's M:N multi-threading ends up doing more harm than good, because if you need such synchronization, you also need to fine tune your thread-pools and at this point 1:1 would be better. On top of 1:1 platforms you can build M:N solutions, but it doesn't work well in the other direction.

> Novices to the language have a tendency to overuse channels

Novices to software development have a tendency to overuse mutexes as well.

(I decided to make a new comment rather than edit my existing comment)

Shared mutable state is generally behind an API boundary. I'm talking about the exported method set of that API. That is,

    func (f *Foo) Update(i int) (int, error)             // yes
    func (f *Foo) Update(i int) (<-chan int, error)      // no
    func (f *Foo) Update(i int, result chan<- int) error // no

What this generally means in Go is that it is an antipattern for your library to provide something like "a method that makes an HTTP request in a goroutine". In Go, you should simply provide code that "makes an HTTP request", and it's up to the user to decide whether they want to run that in a goroutine.

The rest of what you're talking about is a completely different issue.

Channels are smelly in an API. IIRC in the entire standard library there's less than 10 functions/methods that return a channel. But the use case does occasionally arise.

Just to add to this, I found the blog post he mentions[1] towards the bottom that supports this conclusion a really good read.

[1] http://journal.stuffwithstuff.com/2015/02/01/what-color-is-y...

If it truly is the accepted best practice for novices to avoid channels, then that PR campaign has been tried and found lacking. EDIT: whoops, read parent wrong.

and all this turns out to be a really amazing system of paradigms when designing apps to work in the real world

Without exception, hitting one of these corner cases exposes an error in design, from Go's perspective on CSP. You can disagree with that perspective on a subjective basis ("hard to understand") -- but you can't lift that opinion to objective fact, and you certainly can't claim these artifacts of design as evidence of incompetence or neglect.

In addition, making bad design decisions that you think are good is actually one of the best types of evidence for incompetence (though not neglect, in this case).

I don't personally have enough data to have a strong opinion on where Go channels falls here, but I don't think any of your arguments here have any bearing on the idea that Go's channel implementation is bad.

> I don't personally have enough data to have a strong opinion on where Go channels falls here, but I don't think any of your arguments here have any bearing on the idea that Go's channel implementation is bad.

If sagichmal is correct, zeeboo is trying to use channels in a way that they were explicitly not designed to be used. That makes zeeboo's criticism very likely to be invalid. (It is the one who uses them as they were designed to be used who knows what the actual problems with the design are.)

The argument could be made that go's channels SHOULD be able to handle zeeboo's use case and the fact that they weren't designed to be able to handle it makes them bad.

Nil channels blocking is definitely a deliberate design decision and has valid use cases. I use them frequently when I have a channel based design. It also isn't the thing that most people first expect since they have the opposite analog for using anything else that is nil: panics. The article which I assume you read makes only this point.

I never attempted to lift statements that are obviously opinion based (anything that has a judgement of something good or bad) as objective fact.

Here's a proposal I worked on with a coworker to make channels better that might give you more of an idea of why I'm suggesting that the current design has flaws: https://github.com/golang/go/issues/14601

Given how much weight channels are in the language specification and memory model, it would be nice if they were more generally applicable and easier to use for more concurrency situations.

It's as if someone created a gun that fired backwards and I said "hey, it might be better if the gun fired forwards. we'd be able to use it in more situations." and people responded with "you shouldn't use a gun that fires backwards if you want to fire forwards." I totally agree, but it's missing the point.

  -module(someone_was_wrong_on_the_internet).
  -export([init/0, fizzbuzzer/1]).
  
  -spec init() -> list(pos_integer() | binary()).
  init() ->
      List = [-1, 0, 0.1, 1, 2, 3, 5, 15, 2, 3, 5, 15, 1],
      [fizzbuzzer(Result) || Result <- List].
  
  -spec fizzbuzzer(pos_integer()) -> pos_integer() | binary().
  fizzbuzzer(Number) when Number rem 15 =:= 0 ->
      <<"FizzBuzz">>;
  fizzbuzzer(Number) when Number rem 5 =:= 0 ->
      <<"Buzz">>;
  fizzbuzzer(Number) when Number rem 3 =:= 0 ->
      <<"Fizz">>;
  fizzbuzzer(Number) ->
      Number.

The code itself is a valid program that runs, but it produces incorrect output, because 0 rem 15 =:= 0, so you get <<"FizzBuzz">> where you'd expect to get a 0 in the list. By running Dialyzer in my build chain I can catch that my implementation doesn't match my constraints at compile-time. In a way that I otherwise would have only found at runtime.

Though while creating this little pointless example one thing I'm not super clear on is why Dialyzer fails to notice that my return type from

  fizzbuzzer(Number) ->
      Number.

  fizzbuzzer(Number) ->
      -Number.

But channels use locks internally. The choice of channels vs. mutexes is one of design, not implementation. Also, mutexes are blocking; LOCK XCHG isn't. Sure, mutexes also use LOCK XCHG (but so do channels, and nearly all concurrent data sctructures), but they also block (as do channels).

> your only choice of handling it is by means of mutexes, which has really terrible performance characteristics

That's just not true. There is a way to translate any locking algorithm to a non-blocing one (in fact, wait-free, which is the "strongest" non-blocking guarantee), yet only only a handful of wait-free algorithms are used in practice. Why? Because it's hard to make them more efficient than locks in the general case.

> not to mention that you'll have problems if you want (soft) real-time behavior, because now you can end up with both dead-locks and live-locks.

Again, channels are blocking data structures.

> If you end up doing such synchronization in Go, then Go's M:N multi-threading ends up doing more harm than good, because if you need such synchronization, you also need to fine tune your thread-pools and at this point 1:1 would be better.

I don't know where you're getting that. AFAIK, Go's mutexes don't block the underlying thread; only the goroutine.

The question of which concurrency mechanism should be used is a difficult one (and in general, more than one is necessary; even Erlang has shared, synchronized mutable state with its ETS tables), but you are very misinformed about how concurrency constructs are built and about their performance behavior.

I choose wording carefully for a reason