[Jython-dev] Nobody uses SystemRestart

[Jeff A. <ja...@fa...>]

I know this because it hasn't worked since 2014 and no-one has 
complained. Should we keep it?

First, some archaeology ... Restarting was introduced by Leo Soto in 
2008 (rationale here: https://codereview.appspot.com/2810) and is still 
marked as "experimental". Along the way it has had some work including 
the introduction of shutdownInterpreter() here: 
https://hg.python.org/jython/rev/94b5f009fb77#l3.36 and the accompanying 
socket._closeActiveSockets() that it calls. The motivation is to restart 
Jython when used in a web-server without incurring the application 
start-up time. Fast-forward to 2014 and _closeActiveSockets gets blown 
away (https://hg.python.org/jython/rev/107fe4a4c96b). From here on, 
raising SystemRestart gets you an NPE, but nobody notices that. Here it 
is in 2.7.0:

PS startup2> jython test_restart.py
Exception in thread "MainThread" java.lang.NullPointerException
         at org.python.util.jython.shutdownInterpreter(jython.java:439)
         at org.python.util.jython.run(jython.java:372)
         at org.python.util.jython.main(jython.java:142)

Restart gets a mention here: http://bugs.jython.org/issue1475 so it was 
used in 2.5.1 ...

I can keep it (really, restore it) but I'm not sure what to do to 
reproduce the missing _closeActiveSockets() or whether it matters. Just 
for now I've made the absence non-fatal while I work on the rest of 
http://bugs.jython.org/issue2686.

I think the idea of this has a couple of flaws we cannot easily get around.

  * It is not compatible with the idea of multiple interpreters
    (multiple sys modules). A restart, invoked by one interpreter, stops
    *all* Jython threads, cleans up *one* interpreter, and restarts the
    *whole* application.
  * Static initialisation is not reset unless you know to do it. Some
    static initialisation (the type system, loaded classes) is
    intentionally preserved because it drives the start-up time we're
    trying to save. But others (initially empty lists in sys, Options
    flags, the registry) can trip up the interpreter on a second pass.
    Every such case I find has an easy corrective: countless others are
    waiting to pounce.

Any thoughts on this experiment?

Jeff

-- 
Jeff Allen