I can take a shot at these. #1 is difficult without seeing the problem. If you're in a Web app, the map's not going to do a dynamic draw when you zoom in. Instead, it should create the tiles on demand. This can take longer than a normal dynamic draw because the server draws a large chunk of tiles at once. If you waited a while and you're sure nothing's going to draw, the next thing might be to check your data connections, perhaps looking through the log files for broken layers. Some folks deploy a tile cache, then remove the source data from the production machine; however, removing the source data is going to cause a problem if on-demand caching is enabled. You would see...nothing.

For question 2, there's often not a direct 4:1 ratio in tiles between scale levels due to the shape of the area being cached. As you reach larger scales the created tiles follow the area of interest boundary in a more fine-grained way.

With regards to the caching time, at larger scales more layers tend to be visible and layers tend to have more complex symbology. This makes the tile creation time for large scales disproportionately long.

If you don't have cache on demand turned on, it won't draw the layer dynamically in a Web application. This model of "you either get a tile or you don't" is by design, and it helps the layer maintain its scalability. Bing Maps and Google Maps work this way when you pan to some area where there's no data. You just see a generic tile that says "Data not available". There's no way MS and Google are going to let potentially millions of people make dynamic requests against their servers.

And yes, you're right, you can put a "dummy" MXD on the production server and still see the tiles. You need some feature class in there (turned off) to define the map extent, but that's about it. This saves you the need to put all your data on the production server. It doesn't work if you're doing queries or cache on demand, though.

Slightly off the topic: I have a problem that isn't getting addressed in the forums and I don't know where else to turn. I have a service using the Google tiling scheme for most of the northern hemisphere. When I cache, the small scales cut off some of the data with each increase in scale, until parts of AK and HI are missing by 1:4,000,000 or so. It replaces the map with white space; I'm using a "map data not available" tile that appears where there's no map data, but it doesn't show up in these cut-off areas. Any clue???

I was going to reply back and say what worked, for anyone else viewing this, but sadly I still haven't fixed it. I may have it narrowed down to a data problem. I've been projecting our NAD83 source data to Web Merc, so the service can ultimately be mashed up with Google Maps etc. The source mxd looks OK, but adding the uncached service to a new ArcMap session give the "inconsistent extents!" error, which I think is causing the problem with the cache later on. The extents get changed when I project the data, since they go from being in long/lat to meters, so that they're so big they don't make any sense. Ideas, Sterling or anyone else?

How do you mean? Sliverlight, Flex, JScript are client-side technologies the rest is server-side. Cost comparison comes down to your/your team skills