question about bandwith and datatransfer

The server consuming way beyond 4GB of RAM does not mean that the game state is just as big. You are most likely going to have a lot of redundant information and bloated data structures in RAM which are required to speed up the simulation, but don't add to the condensed game state.

Yes, much more like this. Data comes out of the sim in the form of curves and get stored inside the history server. The history server then send the relevant info to each client. If you want to do a replay you just setup a history server without the sim part. If you want to load game you inject the state back into the sim.

Of course, that's a primary design goal.

I infer from the sheer number of posts in this thread that this topic is one that Neutrino is exceptionally passionate about.

It's like everything you guys do gives me even more faith in how good of a job you are going to do. I was thoroughly convinced thirty seconds into the KS video, so I think we'll have to invent a new scale to define just how much Uber is going in the right direction.

Not just me. Everyone at Uber was amazed to see this stuff work.

If anything makes me excited, it's statements like this.

HELL YEAH THERE MAY BE A NEW ERA OF RTS LIVESTREAMING INCOMING.

sry fir screaming but i am pumped right now.

It is already possible for FA and RUSE (and I guess wargames), and I think it's possible for starcraft 2 (it's just a matter of sending the datas of a replay through the net while it's written).

a few questions about this :

is the replay file candidate for high compression ?
is the replay engine able to uncompress and feed the core engine in real time with required data ?

The advantage of this system is it lends to flexibility; your replay and save files are identical and you can load from any point in your replay/save. Possibly more depending on details.