Normally, when a transfer times out, it's because it's stuck deep down in I/O wait, meaning SIGKILL
is not going to be handled in a timely manner, and, so, trying to .communicate() with it again is just futile.
With the current code, after a timeout, the worker just gets stuck in the same wait it tried to get out of with the kill, as if there were no timeout given.
So, let's just send it a kill and then abandon it.
Related to #175 (abandonning subprocesses may lead to zombies), but this doesn't really address the items brought up in that issue, which are higher-level ideas.
Normally, when a transfer times out, it's because it's stuck deep down in I/O wait, meaning SIGKILL is not going to be handled in a timely manner, and, so, trying to
.communicate()
with it again is just futile.With the current code, after a timeout, the worker just gets stuck in the same wait it tried to get out of with the kill, as if there were no timeout given.
So, let's just send it a kill and then abandon it.
Related to #175 (abandonning subprocesses may lead to zombies), but this doesn't really address the items brought up in that issue, which are higher-level ideas.
I have deployed this to cedar.