I'm guessing that it's the order in which that doc was (first? most recently?) inserted in the index, but would like to confirm that.
Context: I'm looking for a cheap default/secondary sort which will behave consistently for the same data even for separate copies of that data in separate indices each of which is populated by multiple loaders running in parallel. (This is mostly to keep things well-defined for regression testing.) Which I think means I want _uid rather than the faster _doc.