We can't include things with unpredictable iteration order (sets, maps) in our
wire test, but we still want to test those things in our GIANT STREAMABLE CLASS
OF DOOM test. So we now have a separate wire format test which just ensures
that we don't spuriously break the encoding of things that should be
predictable.
If we're going to tell you to use an ArrayList instead of toArray(), let's make
that easy. This would be unnecessary if there was an ArrayList constructor that
took Iterable, but there isn't. And DSet is not a Collection.
Every time a client logs on and back off, another DatagramReader thread
is left around. This seems to fix it, but this method should be called
when either the reader or writer shut down..
Also: the didExit methods look funny:
/**
* Callback called by the datagram writer thread when it goes away.
*/
protected synchronized void datagramWriterDidExit ()
{
// clear out our writer reference
_datagramWriter = null;
if (_datagramReader == null) {
closeDatagramChannel();
}
log.debug("Datagram writer thread exited.");
}
Why would it only try to shut things down if there's no reader?
Perhaps the idea here is that it won't shut down the channel until
both the reader and writer exit? It's not documented...
This has been happening for a while on PX and I've been meaning to track
it down. When a peer shuts down an object that's proxied on other peers,
the other peers receive an ObjectDestroyedEvent, and clear the object
from their omgrs. Other listeners may also receive the event and will need
to clear the proxying information, but that causes an unsubscribe to be
queued up, which when it runs will find the DObject already gone, and log
an error.
At first I considered not queuing up the unsubscribe action if the object
is no longer present in the omgr, but in the case that I'm working with
we have 2 or more objects that will all be destroyed simultaneously.
Even if I test to see if the first object is around or not, this same
test would always pass for the second object even though its death event
is next on the queue. The unsubscribe would be enqueued after THAT, so
there's no test to detect that the two objects are linked.
Instead, just suppress this message if we're a peer client. I didn't
want to suppress it altogether because we might want this message logged for
normal client connections. Maybe not.
Perhaps there is a better way. Maybe the PeerNode should add its own
ObjectDeathListener and clear out the proxy information and ensure that
the proxied object is also destoyed, and clients need not unproxy manually
in that case.
RFC.
Consolidate warnings and interval scheduling in doWarning().
Don't let the reboot time drift because of delays on the RunQueue.. each
warning is scheduled based on absolute time from _nextReboot.
doWarning() now takes two args so any subclasses that had overridden
it (unlikley beyond PX) will break...
PendingShutdownObservers will now get the actual milliseconds until reboot.
Basically: after all this, things should essentially work the same except
now there is a protected three-arg version of scheduleReboot that lets you
schedule an exact time, bypassing the normal round-up to the next highest warning time.
Previously a reboot time was only exact if it was further away than
the first warning. Otherwise, the time was rounded upwards to
the next warning time. This is probably still desired for manually
scheduled reboots, but when trying to coordinate reboots between
nodes we need to be able to schedule them exactly even if we're
already in the warning phase.
- Previously if you scheduled a reboot for 5am every day, but the
server came up at 3am, it would schedule the reboot for 5am the next
day. Subtract a day if the current time is before the rebootHour.
- When skipping weekends always push the time out rather than
pulling it in. Questionable, but so was the previous behavior.
If you set it to reboot daily but skip weekends, it would not
skip weekends.
This avoids a needless wait in the run queue for most notifications. Only
notifications that result directly from something one of the communicator
threads did will need to do a loop de loop.
This should close the hole where ClientObserver could be told that the
ClientObject was available *after* some events have come in on the ClientObject
(which they then miss). Now when the subscription response comes in for the
ClientObject, observers are notified immediately, and will be wired up and
ready for any messages that are themselves already in the queue ready to be
dispatched.
Michael Bayne