The memory usage limit for the aggregation pipeline is 100 MB.

For a sharded cluster, this happens in two stages: a distributed stage on the various shards, and a final stage on the primary shard for the database.

We should mention that the 100 MB limit is a per-server limit, so for the first (distributed) stage, it can be more. Also, I haven't checked this, so please verify it with someone in the know if you do plan to publish it; I'm only 80% sure.