queue up an invoker unit which would go off and do some database stuff and then
by the time it came back and was ready to publish its results to a distributed
object, the object in question would have been destroyed for any of a variety
of fairly natural reasons (client disconnected or logged off, game was
abandoned, dog ate homework).
One "solution" to this problem would be to litter our games' code with
thousands of calls to isActive() in the handleResult() methods of our invoker
units. We've done a bit of that in Yohoho but I've resisted starting down that
path in our other games.
Another solution would be to create an Invoker.Unit wrapper that takes a
reference to the distributed object (or objects) that it will be modifying and
have the common unit code check that the object(s) in question are still alive
at the end of the asynchronous operation and not call handleResult() if they
are not. This has numerous problems: what do you do if one object is alive but
not another, how do you incorporate this functionality in with the numerous
other Invoker.Unit derivations we have that simplify our lives in other ways
(without getting crazy and starting to use something like AOP), do you silently
abort the operation or log something?
So instead, I've come around to the idea that this is simply a dirty fact of
life in asynchronous programming and the fact that we can accept modifications
to distributed state after the distribted object in question is dead is a good
thing. We used to log a warning every time this happened and freak out even
more substantially if one tried to start a transaction on a dead object. Now we
will simply log an informational message (I don't think this sort of thing
should be silently ignored because there are some cases where it is an
indication of incorrect code, those are simply more rare). We will also allow a
transaction to be started on a dead object and when the transaction is
committed, all the events involved will be dropped just like a single
modification would have been dropped on that object.
This allows the most sensible thing to happen which is any results that are
published to still live objects will actually be published and results
published to dead objects will be dropped without making a big fuss. Since a
dead object by definition cannot have subscribers, no one could possibly have
cared about the dropped events anyway.
Also widened.
git-svn-id: svn+ssh://src.earth.threerings.net/narya/trunk@4545 542714f4-19e9-0310-aa3c-eee0fc999fb1
Michael Bayne
03dab36e3e