Hi,

On 22 May 2018 at 04:46, hubo <hubo@jiedaibao.com> wrote:

> I think I found the cause, happily it is not a PyPy bug (although it is

> PyPy-related). Some generators are calling some cleanup code in the

> "finally" clause, but they are not correctly collected in PyPy until GC

> (because PyPy does not use reference counting). Apparently when they are

> collected by GC module, the "finally" clause is executed in a separated

> thread.

Unlike Java, we're not purposefully running finalizers in its own

threads: instead, if your program has got more than one thread

running, finalizers will run in the one where the GC happens to run.

There isn't much we can do about that.

All these hard GC issues also exist on CPython, but they are just more

hidden. Typically, if there is a cycle of objects which itself holds,

indirectly, a reference to the generator-iterator, then you get the

same effect: the cycle will be broken by the GC module of CPython at a

random point in time and from a random thread, too. So we can argue

that you potentially have a bug even on CPython, in the sense that an

apparently unrelated change somewhere else can make the same bug

appear on CPython.

A bientôt,

Armin.