[lenkite]

If results[i] must be printed only AFTER print of results[0]...results[i-1], then you effectively need to wait for max of (time to compute results[0]...results[i]), since even if results[i] is computed earlier you can't print it out if results[0]..results[i-1] are not available. If result[0] takes the longest compute time, then you will definitely need to wait till the end.

Frankly, a simple sequential for loop seems to the simplest solution here :)

Anyways, I think this: https://go.dev/play/p/lFBpzUVVzUj satisfies all the constraints. Only look at the output with the "(main)" prefix, the other prints are for elucidation.

        package main

        import (
         "fmt"
         "math/rand"
         "time"
        )

        type Result struct {
         value    int
         computed bool
         consumed bool
        }

        func main() {
         args := []int{5, 2, 4, 1, 8}
         results := make([]Result, len(args))
         signal := make(chan bool)
         signalCount := 0

         slowSquare := func(arg int, index int) {
          randomMilliseconds := rand.Intn(1000)
          blockDuration := time.Duration(randomMilliseconds) * time.Millisecond
          <-time.After(blockDuration)
          square := arg * arg
          results[index] = Result{value: square, computed: true}
          fmt.Printf("(#%d)   Squared %d, index=%d, result=%2d, duration=%s, sending signal. \n", index, arg, index, square, blockDuration)
          signal <- true
         }

         for i, x := range args {
          go slowSquare(x, i)
         }

         for {
          if signalCount == len(results) {
           break
          }

          <-signal
          signalCount++

          for i := 0; i < len(results); i++ {
           if !results[i].computed {
            break
           }
           if !results[i].consumed {
            fmt.Printf("(main) Squared %d, index=%d, result=%2d\n", args[i], i, results[i].value)
            results[i].consumed = true
           }
          }
         }

        }

Example Output

        (#0)   Squared 5, index=0, result=25, duration=8ms, sending signal. 
        (main) Squared 5, index=0, result=25
        (#4)   Squared 8, index=4, result=64, duration=210ms, sending signal. 
        (#1)   Squared 2, index=1, result= 4, duration=777ms, sending signal. 
        (main) Squared 2, index=1, result= 4
        (#3)   Squared 1, index=3, result= 1, duration=867ms, sending signal. 
        (#2)   Squared 4, index=2, result=16, duration=924ms, sending signal. 
        (main) Squared 4, index=2, result=16
        (main) Squared 1, index=3, result= 1
        (main) Squared 8, index=4, result=64

[Thorrez]

That code has a race condition. One goroutine can modify an element of results at the same time a different goroutine is reading it. Running with `go run -race` detects the race.

This can be fixed with a mutex.

Now that we've got working code (with the mutex), we have to ask: have we proved Chris wrong? I don't think so. Chris never said it's impossible to implement this in Go. Chris just said that implementing it in Go is uglier than using an array of promises in some other language. And I think this code is uglier than an array of promises.

Although I think I disagree with Chris. He says that an array of channels is significantly uglier than an array of promises. I don't think so. I think an array of channels is fine (and is easier to understand than the signal, signalCount, computed, consumed, +mutex code).

    package main

    import (
     "fmt"
     "math/rand"
     "time"
    )


    func main() {
     args := []int{5, 2, 4, 1, 8}
     results := make([]chan int, len(args))

     slowSquare := func(arg int, index int) {
      randomMilliseconds := rand.Intn(1000)
      blockDuration := time.Duration(randomMilliseconds) * time.Millisecond
      <-time.After(blockDuration)
      square := arg * arg
      fmt.Printf("(#%d)   Squared %d, index=%d, result=%2d, duration=%s, sending signal. \n", index, arg, index, square, blockDuration)
      results[index] <- square
     }

     for i, x := range args {
      results[i] = make(chan int, 1)
      go slowSquare(x, i)
     }

     for i := 0; i < len(results); i++ {
      fmt.Printf("(main) Squared %d, index=%d, result=%2d\n", args[i], i, <-results[i])
     }
    }

[lenkite]

Yes, I didn't quite get why he wanted a solution without multiple channels..it is clearly the best way. The only other option without multiple channels is to use one of the atomic types: `sync/Atomic.Int32` (to avoid the mutex) and then a single channel for signalling.