Example: the content type Company declares two indexing fields, "ceo.name" and "city.state.title", with "ceo", "city" and "state" content metadatas of respective types "Person", "City" and "State".
If a person content "name" is modified, contents (companies) referencing it should be found and reindexed (in the best-case scenario, only the corresponding indexing field should be reindexed).
The same goes for the state title: when a State is modified (in case its title changes) OR when a City is modified (if its state is changed), companies referencing the "old" city or state should be reindexed.

Problem: searching the contents referencing the modified one may be expensive. For instance, many content types will reference City as an indexing field, and searching for referencing contents will require executing as much queries. If this is done synchronously in an observer every time a content is modified, saving a content will take minutes.