[Opensim-dev] Hey all, Grid Comms, border crossings, Request for comments, red regions and tolerance to network failure.

[Teravus Ovares]

Hi everyone

People are practically begging for more reliable grid comms.   So,
naturally, I want to give it to them.    I really want to do it with
the least amount of work/turnaround time as possible.   We've all got
only so much time to work on something.

So, at least with border crossings, I've observed a few phenomena

1. Fly off into infinity
2. Get put arbitrarily in a spot in the region you were previously in
and get stuck there

Item one is almost exclusively from child agent disconnects (where a
region that you're a child agent in says, 'close was called' but you
didn't actually want it to).  Eventually the region will turn red on
your minimap.

Item number two seems to be corruption of some kind.  It could be a
race condition or lack of synchronization with the client.

Fact: When a region disconnects on your minimap in llsimulators, the
region will appear again.

If you watch the console you'll see that the region that you're a root
agent in sends the 'disable simulator'  message and then the 'enable
simulator' message.

This implies that when you get disconnected as a child agent from the
region you're a child in, 'that region' mentions the fact that 'it
lost you' to the region you're a root agent in.  Then the region
you're a root agent in sends the 'disable simulator message' and then
sends the 'enable simulator' message to reconnect you to that region.

Fact: When a region goes red on your minimap with opensim, the region
will never be available again until you relog and when you try to
cross the border, you will fly off into infinity.   Even if the region
hasn't turned red *yet*, you will still fly off into infinity.

So, MW suggested I post a request for comments on this topic since it
has some potential to affect everyone.

Here's what I propose

When sending a child agent notification, the current region is also
included so that, when a child agent disconnects, the region can tell
the neighbor that the child is a root agent in to re-inform the agent
of the neighbor.

Add a routine that the clientview.close method calls in the scene
which checks for a child agent and sends a 'I lost em' message to the
region the agent came from.  From there the simulator with the root
agent in it can make the choice to send the child the reconnect
messages or not.

This will ensure that a child agent connection is established.

Ensuring that a child agent connection is established will make border
crossings more successful and less of Item one from above will occur.

I'm not looking to re-do the protocol entirely.  I want to make it
more tolerant to the inevitable network failure.  I'm not looking to
change much.

I also don't really want to change any of the existing protocol stuff.
  We need a child agent data update anyway..   and it's half
implemented already in the OSG1 code(for a month now).

The child agent data update contains things like..  child throttles..
 so the neighbor sims don't flood the avatar, where the avatar is in
world and what their draw distance is so that we can enable sending
prim to agents in other regions.

I'm not looking to make OSG2 here.  I'm looking to make OSG1 work better.

With that..       comments anyone?