I for example often need to score/model data which doesn't fit in RAM (on my PC) so I use libraries like bigGLM which can use out-of-memory data to build the models. One of the options is SQLite, but you can use an ODBC connection.

Additionally, I can explore slices of the data, which resides only on disk. I don't even need to import it. I can use dplyr (very famous package for aggregations and slicing) which will map the R syntax to SQL which is executed by SQLite.