Right now we:

• go through whole map every 100 emits, and if it's over 50KB we try to reduce each key
• every 100 emits, if map is over 100KB, it is dumped to disk in a collection.
  When these threshold are increased you get faster execution.
  Notably a user reported than on a collection with millions of entries and 1KB objects, the incremental dumps actually add up to a massive collection on disk, since there is almost no reduce done ahead of time.