Is the limit per "table/dataframe" or for all tables/dataframes put together?

The driver collects the data from all executors (which are having the respective table or dataframe) and distributes to all executors. When will the memory be released in both driver and executor? Or does it hold on to this memory through out the pipeline/application?