(filing based on discussions with Igor)

The problem is: if one uses mrr_cost_based=ON, it is possible for the following
to occur:

1. The optimizer decides to use DS-MRR (or whatever MRR implementation the engine has)
2. Execution starts
   1. The join buffer is filled, it turns out that it has fewer records then expected
   2. DS-MRR scan is used (according to the decision), but use of DS-MRR only adds overhead.

The idea to address this is as follows: after BKA code has filled the join buffer, we should note the amount of data we have in the join buffer and check again whether to use DS-MRR.

Tasks to be done

• Fix the current DS-MRR cost choice formulas (or decide to use current ones)
• Implement "the idea" described above.

Unresolved questions

• DS-MRR cost formulas (use current? take 5.6's? something else?)
• Do we make the late-decision every time we've filled the join buffer? or just the first time?