重读 eXplode 知道的事情

The reader should keep in mind that conceptually what EXPLODE does is very simple. If we assume infinite resources and ignore some details, the following would approximate its implementation: 1. Create a clean initial state (§3.3) and invoke the client’s mutate on it. 2. At every choose(N) call, fork N children. 3. Onclient request, generate all crash disks and run the client check method on them. 4. When mutate returns, re-invoke it

这个简化是非常不讲道理的。作为不需要复现这种技术的人来说，当然可以假设你在每一个分支的地方都去作 Fork，但是。。。这始终不是很对。

对于想要模仿这种做法的人来说，完全找不到任何方法可以忽略过度使用 fork() 带来的进程数量爆炸。问题不光处在系统性能资源上。甚至可以说，连进程数量上限都能摸到

当然作者的说法是很狡猾的，“你们大可以把这件事情看成是这样”。这句话隐含的意思就是实际情况和他所说的相去甚远。

当然，作者还是有所补充的：

As a result, EX PLODE unsurprisingly looks like a primitive operating system: it has a queue of saved processes, a scheduler that picks which of these jobs to run, and time slices

所以总的来说，做法还是状态存取和调度器。

接下来有点费解。

exhaust all possible combinations of choices within a single mutate call before doing another.

也就是原地把所有的可能性都走一步，再往下走。有一种广搜的感觉，而不是：我尝试这个分支，然后从这里继续往后探索。

eXplode 的 ckpt 方法是确定性回放。也就是 stateless。

EXPLODE uses computation rather than copying to recreate states.