Why Go Is Not Good (2014)

File I/O or everything else treated as syscall by Go runtime might turn you program into 10k-os-threads-monster. Scheduler will be creating new OS threads to replace those locked on syscalls until thread limit is reached and whole program crashes. Only way to prevent it is to restrict your syscall layer into fixed-size goroutine pool.

I had an interesting case recently - my app serves some data from tons of files laying in NAS, accessing it by NFS mount and one day NAS hunged completely, every I/O call to it was lasting forever. Even 'ls /mount-point-of-nas' was just doing nothing forever until Ctrl-C. In my case I've applied poweroff-poweron cycle to NAS, and everything went right in minutes, just as NAS booted. And after it I wondered, what if my server was written in Go, instead of Erlang...

And, BTW, you can never be sure, that underlying libraries of your code are safe to use.

> It has the good parts of both cooperative and preemptive scheduling, but yes, it's certainly cooperative.

For me, the best part of cooperative scheduling is that you can work entirely without locking shared data structures, because you get "transactions" for free. This means it's rather difficult to get data races, deadlocks, etc. Go's scheduler certainly does not give you that, trades it for spreading work over different cpus.

So it has problems:

1) necessity of locking, using IPC mechanisms, ... (like preemptive schedulers, and let's face facts here : channels aren't enough in real world apps)

2) everything gets blocked by large calculations (like cooperative schedulers)

3) more generally, easy to crash by misbehaving thread due to unrestricted access shared resources (not just cpu) (like cooperative schedulers)

And advantages:

1) Actually uses multiple cpu's/cores/... (like preemptive schedulers)

2) integrated event loop that scales (like cooperative schedulers)

If you want to see a programming language with a "scheduler" that doesn't have the bad parts of cooperative schedulers, check out Erlang. If you attempt to crash erlang with infinite loops, bad memory allocation, ... (on a properly configured system) that just won't work, the offending threads/"goroutines" crash leaving the rest of your program running fine. The offending threads will restart if you configure them to do so (which is really easy).

The same can be achieved, with much more work, on the JVM, or, also with much more work, with python's "multiprocessing" library, part of the standard library.

> > On the linux scheduler with proper ulimits it's bloody hard to crash the system, for instance, forkbombs, memory bombs, ... won't do it. I hope we'll get a language where you can do that too. > I don't understand the analogy. It is not clear what "crash" means here, and it is not clear how it would apply to a runtime environment. All that stuff, forkbombs, etc, means that you can configure the system so arbitrary code can't affect the system in those particular ways.

Crash means that the system/"program" doesn't respond (in a useful manner) anymore.

There are plenty of bothersome things about "immutable only" too.

I've employed both approaches in earnest. Each have their own set of trade offs.

> Of course you can't do that in Go. Not only you don't get static guarantees, you can't even write a generic map with atomic access. You have to remember to use RWMutex every single time. No generics.

I'm quite aware of Go's limitations, thanks.

I find it amusing that you've dismissed an entire category of practice to statically eliminating data races at barely a glance. Irony, it seems. The very thing that people lament about Gophers is precisely the behavior you've demonstrated here! (Quite literally in fact. How many gophers have you heard say something like "generics is bothersome"?)

You've been polite, but snobbery is vexing, no matter where it comes from.

If one adopted such a scheme in Go (which would also be fast to compile), you'd end up sacrificing a great deal of runtime performance, because everything generic would be forced behind a box. Such a thing is no big deal in a language like Javascript, but for a language like Go that aims to be fastish, it's probably a non-starter.

Finding a compiler that monomorphizes generics and is also as fast as Go is probably a much harder challenge. I've had some folks claim that one of D's compilers is pretty fast, but I haven't seen any good benchmarks to support that claim. Many of the other monomorphizing implementations I've tried (Ocaml, ML, Haskell, Rust) are all pretty slow. At least, much slower than Go's compiler in my own experience.

Those sound like pretty compelling use cases to me, and also ones that seem to be well suited for Rust. You might also consider looking at Servo; I bet its engineers could list a myriad number of reasons why immutable-only data structures are insufficient.

I note that performance is not the only trade off worth examining (to be fair, I think you acknowledged this). Another aspect of the trade off is abstraction, albeit this is fuzzier. Mutation can be more natural to a lot of folks. My pet theory is that we've built up a defense mechanism against mutation because it's the source of so many bugs; but Rust's static guarantees are worth consideration here. They remove many of the problems normally ascribed to mutability. For example, Rust not only prevents data races, but it also prevents aliasing mutable pointers to data at compile time, which defeats another class of bugs not related to concurrency at all.

My main point of contention with your comments is that you think you've stumbled on to the "right" way of doing something. In my opinion, that's nonsense. What's the point, even, to declare such a thing? Instead, focus on what the trade offs are, then make a decision based on the constraints you've imposed in any given situation. (Valid constraints absolutely include "immutable data structures are easier for me to reason about intuitively.")

> In projects where I can afford a GC, I'd always take the GCed option (in my case an overwhelming majority). Same for immutable data structures (use whenever they can be afforded). Are those bad heuristics? (Afforded here refers to performance/memory constraints only)

They don't seem like bad heuristics to me. They don't really correspond to my own heuristics, depending on what problem I'm trying to solve. (I once chose a language for a project based purely on the fact that I wanted to target non-programmers.)