I have a number of map services in ArcGIS Server (10.1, running on 64-bit Windows Server 2008 R2) and only one of them is cached.
Caching is configured on-demand as there's no way I can store the full tile set on disk. Also many of the tiles would be empty as my dataset is quite sparse.
After a period of inactivity (I'm not sure how much) the first hit to the tile service takes a long time to respond. I can't say for sure if this only happens when generating a new tile on demand, when loading an existing tile, or both. The delay itself is annoying but it's also accompanied by an error in the AGS logs:
[Severe] Services containing process crashed. MapServiceName.MapServer instance has crashed
The logs provide no more information, and performance picks up after the initial crash.
The map service's data source is a pair of File GeoDatabases, both are registered with the server, both are in EPSG:4326 and output in EPSG:3857 in line with the ESRI base map. Features are very simple (all polygons, usually four-sided shapes) and don't seem to have any geometry problems. The cache disk is nowhere near full and I haven't previously encountered any permissions issues with any of the AGS accounts.
Can anyone suggest why the service might be crashing, why performance might be slow (if not connected to the crash), and where I might want to take my investigation?