We have only tried one very-large Mosaic Dataset in applications (the original imagery size is 2 TB, but using LZW-based storage strategy is 1 TB, at local SAN).
The picture for this try of this latest mosaic dataset in our applications is as follows:
(1). One web application (for safe, minimum 100 users hit daily)
(2). 200 workstations (about 100 engineers/geologists/cartographers use daily)
For web applications, the background is the cached historical images; Onto the top of this background cached imagery, the latest mosaic dataset is programmatically added with scale-dependency.
Inversely, we only use image service(s) of Mosaic Datasets in our applications for most important and latest layers. The cached images are only used for background, and sometimes, also for historical comparision.
Performance comments:
(1). For web applications, it looks OK, at least acceptable; in other word, with this strategy, it is comparable to the cached images;
(2). For workstations, it looks fine, much faster than raster catalog from SDE, but little slower than old image services from standalone ArcGIS Image Server 9.3/10, which are created within image service editor;
PS.,
(1). This background cached images were derived from several-level historical imagery, which is combined with 2000-m image, 200-m image, 15-m images of Landsat 7, 2.5-m images of SPOT 5, 1-m images of IKONOS;
(2). The mosaic dataset (of 0.5-m GeoEye) has two versions of image service: without overviews (updated weekly, even daily sometimes), and with 6-level overviews (rarely updated);
(3). The image service(s), especially, without overview, are programmatically called with scale-dependency into the web applications as separate layer(s).