Good question, and to answer this well we should probably do a blog post or something. In the meantime you could dig into the code since the clients are all open source. :-)
But basically, sync is split into two halves: writes and listens. Clients store pending writes locally until they're flushed to the backend (which could be a long time if the app is running offline). While online, listen results are streamed from the backend and persisted in a local client cache so that the results will also be visible while offline (and any pending writes are merged into this offline view). When a client comes back online, it flushes its pending writes to the backend which are executed in a last-write-wins manner (see my answer above to ibdknox for more details on this). To resume listens, the client can use a "resume token" which allows the backend to quickly get the client back up-to-date without needing to re-send already retrieved results (there are some nuances here depending on how old the resume token is, etc.).
