使用重播（selectorFoo）但不发布（selectorFoo）时的OOM

Question

This crashes with OOM:

Flowable.range(1, 5000)
        .map(__ -> new byte[1024 * 1024])
        .replay(
          fb ->
            fb.take(1)
              .concatMap(__ -> fb)
          ,1
        )
        .count()
        .toFlowable()
        .blockingSubscribe(c -> System.out.println("args = [" + c + "]"));

This is, I think because replay is holding to the emissions from upstream, even though I would have thought that the 1 buffer sizse hint would make it not to.... what am I missing?

This doesn't crash:

Flowable.range(1, 5000)
        .map(__ -> new byte[1024 * 1024])
        .publish(
          fb ->
            fb.take(1)
              .concatMap(first -> fb.startWith(first))
          ,1
        )
        .count()
        .toFlowable()
        .blockingSubscribe(c -> System.out.println("args = [" + c + "]"));

But I am not sure if I am guaranteed that I will get ALL the emissions from upstream like that...

Answer 1

I've investigated this and found the cause of the issue: a bug in replay in RxJava 2.

What happens is that replay holds references to 2 subscribers, one for take and the other for the concatMap 's inner consumer in a local variable, thus there is a GC root from the main thread to the defunct take still referencing the very first item. Since the bounded replay uses a linked list, this very first item then keeps referencing the newer and newer items via its "next" links and ends up exhausting the memory.

publish doesn't keep references to old values thus this is not an issue there.