One final post, now with images/movie to illustrate my points - the first large scale ingame presentation of the game, a helicopter fight on city rooftops in Nepal.
Here's the movie, in case someone hasn't seen it:
http://www.gamersyde.com/stream_uncharted_2_among_thieves_e3_trailer-11587_en.html
[gs]11587_en[/gs]
and some shots from it:
http://www.gamersyde.com/gallery_10873_en.html
So, on to my notes.
This is obviously a huge open area, several square miles. The city is surrounded by mountains, has lots of buildings and so on. But of all those buildings seen from the vantage point in the beginning, you can visit only a few.
You slide down on the cable to the first, get shot at while trying to cross the 'bridge', jump through to a second one; then you move on to a third one and drop down into a hole. You can not go back at all from this point to any of the previous places and you will stay within the buildings from this point on.
(Later through the level you get back to the rooftops to deal with the Mi-24, but those buildings aren't visible from the vantage point at the start - you might be able to see the vantage point itself, as it is a tall tower, but that can be a reduced LOD version).
All the rest of the buildings, the hillsides, bridges, river etc. in the vista are completely noninteractive and uncreachable.
You can zoom in a bit (4x maybe) if you have a scoped rifle, but that's as close as you get to the elements of the vista. A few 512 texture maps are perfectly enough to cover the area and all the shading and lighting can be completely baked into the textures, so no need for normal/spec maps (except maybe the river). This is a huge win in efficiency, creating the illusion of a very large scale background with very limited resources.
Also, these pieces of the background and skybox can be kept in memory during the entire level, so the actual chunks don't have to contain them.
Notice the sequence at 2:00-2:10 where Drake rolls on the ground trying to evade the shots from the copter. This takes precious seconds that the engine can use to stream in the next chunk, because the player can not move around at all.
Also, the player won't notice it because he's trying hard to stay alive... And since he's dropped down from the roof, he probably can't backtrack at all, so the game can also discard the previous chunk at this time.
There's another similar sequence near the end of the movie, where the building collapses and the player is constrained into the room for several seconds, so the engine can once again discard the previous chunk (there's no way to go back) and stream in the next one.
Considering that the game uses memory for the game code and framebuffers, the player character data (model, texture, animation, sounds, weapons), a limited number of NPCs' data; and that a lot of the data is shared between chunks of the same level (skybox/vista, music, more sounds), I'd say that the chunk size probably has to be around 64 megs. 8-10 seconds should be enough to stream in a complete chunk from the HDD, and then the game can copy more data from the BR disc to the HDD while the player passes through the current section.
I'm not entirely sure how frequent 'events' like the ones in this movie are through the game, but it's reasonable to assume that they happen quite often. A single level may be broken down to several dozens of such chunks, so it can use hundreds of megabytes of unique data in total. The entire game is probably at least 10 GB of level data alone, probably even more, whereas GTA4 and AC1/2 has to fit on a single DVD (and GTA4 has unbelievable amounts of voice and music too). Thus there can be about an order of magnitude of difference in texture detail and variance, but the price of that is a very linear, although very exciting and carefully tuned gameplay.