During the evaluation of the new cost model in SPM-3658, we found that IXSCANs on multikey indexes are typically underestimated. There is currently no special case for multikey indexes in the costing logic, nor in the calibration scripts.

One of the possible reasons for the underestimation is that we don't account for the deduplication which needs to happen for multikey IXSCANs (as exposed by dupsTested and dupsDropped in the explain output).

Since this can lead to poor index choices, it's worth investigating whether we can calibrate for the deduplication work.