The join optimizer produces new kinds of SBE plans which may have different data access patterns than plans generated in the past. We should investigate whether the SBE engine's yielding behavior is sufficient, or whether it is possible for the join optimizer to generate a plan that (depending on the shape of the data) may not yield for seconds or minutes. My recollection is that SBE plans currently trigger yielding at the leave stages doing data access (scan or ixscan), but perhaps it is possible to generate a join plan which spends lots of time processing intermediate results without ever accessing the base collections.