Hi Uli,
The practical limitations of the builtin charts engine would rather be based on the time it takes to actually build the columnar cache, the required disk space for it, and the cardinality of the columns (it does not scale very on very-high cardinality columns)
100M rows and low-cardinality columns should definitely be OK.
The cache is automatically dropped when running on managed datasets (ie, datasets built by DSS). For source datasets, we have chosen not to try and autodetect changes in the underlying source, because it would be too expensive, so if the data source changes on an external dataset, you have to click on the "Refresh sample" button.