Hi, I need share operator. The <implementations> s...

# coroutines

[Jan Skrasek]

Hi, I need share operator. The implementations seem to be quite complicated so I've written this.
Do you have any idea why it shouldn't work correctly?

import kotlinx.coroutines.*
import kotlinx.coroutines.channels.*
import kotlinx.coroutines.flow.*
import java.util.concurrent.atomic.*

fun <T> Flow<T>.share(): Flow<T> {
  val channel = ConflatedBroadcastChannel<T>()
  val counter = AtomicInteger()
  var job: Job? = null

  return channel
    .asFlow()
    .onStart {
      if (counter.incrementAndGet() == 1) {
        job = GlobalScope.launch {
          this@share
            .catch {
              channel.close(it)
            }
            .collect {
              channel.offer(it)
            }
        }
      }
    }
    .onCompletion {
      if (counter.decrementAndGet() == 0) {
        job?.cancelAndJoin()
        job = null
      }
    }
}

[elizarov]

You have two pieces of state here: counter and job and there's a race in how you check/update them. E.g. you can end up in an invalid state with counter > 0 && job == null

[Jan Skrasek]

Thanks!

[bezrukov]

Also there is an issue with Conflated channel - it seems unexpected that share() operator will conflate emission by default. You probably need to make it

fun <T> Flow<T>.share(capacity: Int = CONFLATED/BUFFERED): Flow<T> { 
    val channel = BroadcastChannel<T>(capacity)
.....

[Jan Skrasek]

Isn't rendevouz more clean option? share() operator will do the sharing but not replay?

[bezrukov]

there is no rendevouz broadcast channel - only conflated and buffered. Closest to rendevous is buffered with capacity = 1

[bezrukov]

buffered broadcast channel won't replay last element

[Jan Skrasek]

Why not? It will replay the last element if there is an open subscription, won't it?

[bezrukov]

with Conflated will, with Buffered won't -
Conflated: https://pl.kotl.in/tOI0C_NCA
Buffered: https://pl.kotl.in/NtvlYT9gV

You can check ArrayBroadcastChannel (it's under the hood when you choose BUFFERED) docs:

* Note, that elements that are sent to this channel while there are no subscribers are immediately lost.

👆 it applies to the second subscriber as well - if first subscriber consumed element, it lost