Currently $wgEnableScaryTranscluding seems to just pull templates from an HTTP
request and cache them for some arbitrary period (probably, not having looked
closely, indefinitely if the page isn't edited, since I bet nothing invalidates
the parser cache). It goes without saying that this is relatively slow and
unreliable; therefore the feature is not enabled on Wikimedia sites. A shared
page database is necessary, and for timely invalidation of caches, the central
repository must also know about the places that are sharing with it.