Java8并行流是否对序列使用相同的线程

假设我们有这样的东西：

LongStream.range(0, 10).parallel()
.filter(l -> {
  System.out.format("filter: %s [%s]\n", l, Thread.currentThread().getName());
  return l % 2 == 0;
})
.map(l -> {
  System.out.format("map:    %s [%s]\n", l, Thread.currentThread().getName());
  return l;
});

filter: 8 [ForkJoinPool.commonPool-worker-2]
filter: 9 [ForkJoinPool.commonPool-worker-7]
filter: 0 [ForkJoinPool.commonPool-worker-6]
filter: 1 [ForkJoinPool.commonPool-worker-3]
filter: 4 [ForkJoinPool.commonPool-worker-5]
filter: 2 [ForkJoinPool.commonPool-worker-1]
filter: 6 [main]
filter: 7 [ForkJoinPool.commonPool-worker-4]
filter: 3 [ForkJoinPool.commonPool-worker-6]
filter: 5 [ForkJoinPool.commonPool-worker-2]
map:    0 [ForkJoinPool.commonPool-worker-6]
map:    2 [ForkJoinPool.commonPool-worker-2]
map:    8 [ForkJoinPool.commonPool-worker-4]
map:    6 [main]
map:    4 [ForkJoinPool.commonPool-worker-6]

如果运行此程序，输出将如下所示：

filter: 6 [main]
map:    6 [main]
filter: 5 [main]
filter: 4 [ForkJoinPool.commonPool-worker-2]
map:    4 [ForkJoinPool.commonPool-worker-2]
filter: 1 [ForkJoinPool.commonPool-worker-3]
filter: 2 [ForkJoinPool.commonPool-worker-1]
filter: 0 [ForkJoinPool.commonPool-worker-3]
filter: 3 [ForkJoinPool.commonPool-worker-2]
filter: 8 [main]
filter: 7 [ForkJoinPool.commonPool-worker-2]
filter: 9 [ForkJoinPool.commonPool-worker-2]
map:    0 [ForkJoinPool.commonPool-worker-3]
map:    2 [ForkJoinPool.commonPool-worker-1]
map:    8 [main]`

正如我们所看到的，每个long的每个任务序列都是由同一个线程执行的。这是我们可以依靠的，还是仅仅是巧合？线程能否在执行期间“共享”任务？

来自“副作用”部分：

如果行为参数确实有副作用，除非明确说明，否则不能保证这些副作用对其他线程的可见性，也不能保证在同一线程中执行相同流管道中“相同”元素上的不同操作

---

这不是巧合，而是流API目前在OracleJDK/OpenJDK中的实现方式：无状态操作（如

filter

、

map

、

peek

和

flatMap

）被融合到一个操作中，在单个线程中顺序执行步骤。但是，引入一些有状态操作可能会改变情况。例如，让我们添加一个

限制

：

LongStream.range(0, 10).parallel()
.filter(l -> {
  System.out.format("filter: %s [%s]\n", l, Thread.currentThread().getName());
  return l % 2 == 0;
})
.limit(10)
.map(l -> {
  System.out.format("map:    %s [%s]\n", l, Thread.currentThread().getName());
  return l;
})
.forEach(x -> {});

现在，limit引入了一个屏障，将管道分为两部分。结果如下：

LongStream.range(0, 10).parallel()
.filter(l -> {
  System.out.format("filter: %s [%s]\n", l, Thread.currentThread().getName());
  return l % 2 == 0;
})
.map(l -> {
  System.out.format("map:    %s [%s]\n", l, Thread.currentThread().getName());
  return l;
});

filter: 8 [ForkJoinPool.commonPool-worker-2]
filter: 9 [ForkJoinPool.commonPool-worker-7]
filter: 0 [ForkJoinPool.commonPool-worker-6]
filter: 1 [ForkJoinPool.commonPool-worker-3]
filter: 4 [ForkJoinPool.commonPool-worker-5]
filter: 2 [ForkJoinPool.commonPool-worker-1]
filter: 6 [main]
filter: 7 [ForkJoinPool.commonPool-worker-4]
filter: 3 [ForkJoinPool.commonPool-worker-6]
filter: 5 [ForkJoinPool.commonPool-worker-2]
map:    0 [ForkJoinPool.commonPool-worker-6]
map:    2 [ForkJoinPool.commonPool-worker-2]
map:    8 [ForkJoinPool.commonPool-worker-4]
map:    6 [main]
map:    4 [ForkJoinPool.commonPool-worker-6]

请参见元素#2在FJP-1线程中过滤，但在FJP-2线程中映射

请注意，正如@Misha正确引用的，即使对于无状态操作，也不能保证使用相同的线程。未来或替代的流API实现可能会改变这种行为（例如，使用生产者-消费者方法）