A possible solution I could think of currently:

1. Always choose to use HashJoin when there is no statistical information indicating that both tables are large.
2. Memory tracking while building hashtable for building side.
3. When the hash-builder fails to grow its memory
   3.1. sort and spill the in-memory hashtable into spill0, free memory.
   3.2. buffer and sort the incoming records for the buffer table until it's exhausted, do a sort.
   3.3. buffer and sort the records for the streaming side until it's finished, do a sort.
   3.4 MergeJoin the two sides.