print/println should block write final output

[amitmurthy]

As folks start working with more parallel constructs, it will be good to have print/println do a block write of the final string.

julia> @sync begin
           for i in 1:3
               @async begin
                   println("Getting garbled")
               end
           end
       end
Getting garbledGetting garbledGetting garbled

The newline is written separately.

julia> @sync begin
           for i in 1:3
               @async begin
                   println(`Getting garbled`)
               end
           end
       end
```GGGeeettttttiiinnnggg   gggaaarrrbbbllleeeddd```

[amitmurthy]

"block" as in "write as a single block"

[JeffBezanson]

This behavior doesn't seem so great. It's easy to write a string and a newline in a single operation, but we don't want to have to worry about that sort of thing all the time. Maybe some sort of buffering is the answer? @loladiro
Or use buffering for files and sockets, and make TTYs nonblocking?

[Keno]

libuv has the capability to write multiple things at once. It's just that our I/O system doesn't work that way. Making TTY's nonblocking seems fine, since writing to TTYs blocks the whole process anyway.

[JeffBezanson]

But, we have to arrange for the output to be consistent over the different kinds of streams. It would be bad to get different output when redirecting to a file than when testing by printing to the screen.

[Keno]

I don't think you can insist on that. In fact in most situations, you can probably not even depend on the output being the same between julia runs.

[StefanKarpinski]

Line buffering for text streams seems reasonable to me. But of course, we then need a distinction between printing to a text stream and printing to a byte stream, but we've talked about that before already.

[JeffBezanson]

Maybe we can''t insist on identical output in all sorts of scenarios, but getting the expected output at the prompt and then one of Amit's outputs above in a file seems gratuitous to me. The sensitivity to print("a\n") versus print('a','\n') is also a big pitfall.
Even with nondeterminism, we can still pick a model of what sorts of phenomena are possible. Sure, if several tasks print messages I might get them in any order, but hopefully not interleaved character-by-character.

[vtjnash]

I suspect this is showing up now because we are now doing blocking writes, which presumably means there is a a task switch required by each call to the low-level write function.

Another possibility is to introduce locking:

print(s) = 
    lock(STDOUT) do
        print(STDOUT, text)
    end

function lock(fn, stream)
    lock(stream)
    try
        fn()
    finally
        unlock(stream)
    end
end

Where lock/unlock are the traditional recursive binary semaphores.

I'm not sure exactly where this would be put into the hierarchy of show/print.

[JeffBezanson]

No way; I don't want to have to use that lock stuff just to break some output into two print calls.

[amitmurthy]

Regression in 334a2be

[Sisyphuss]

Is the third post in #12600 a related issue?

[amitmurthy]

@Sisyphuss , #12600 is not related and not a bug. for introduces a new scope as documented. global x = i will result in the behavior you are looking for.

[Sisyphuss]

@amitmurthy no I meant the output format: the output "0" is between the print "x:1" and "x:2, x:3".

[KristofferC]

I also get a lot of interleaved printing between Julia and Jupyer in my notebooks.

[amitmurthy]

@KristofferC can you post examples? Currently only println to asynchronous streams is locked. If you are using print statements, you will need to manage it manually - http://docs.julialang.org/en/latest/manual/faq/#asynchronous-io-and-concurrent-synchronous-writes