Currently Daffodil requires that all incoming data fit in one java.nio.ByteBuffer. A separate issue (DFDL-881) is about allowing > 4GB files, but data sizes would still be limited by available address space.

A streaming approach has great advantages. It requires that the input can be streamed in (e.g., from a java.io.InputStream), but also requires that the DFDL Infoset can be streamed out. (Think SAX parser 'events' coming out). This is complicated by the DFDL notion of points of uncertainty. E.g., until a choice branch has been resolved none of the elements on any branch can be emitted since "backtracking" may invalidate them.