We currently don't have any metrics that track how big of a problem flipping plan choice is in production. 

peter.volk@mongodb.com's suggestion: Track "the average difference of neighboring plans within the ranked list. E.g. if we have a cost structure that is very dense (only very small differences between plans) then this will indicate that the plans can easily fluctuate if there is a bad sample (depending on the sensitivity of the cost model). Whereas if there are large differences this can be signal that the overall plan choice is more stable."

Link to comment on doc discussing this: https://docs.google.com/document/d/1KXTXOLvktbLNnUNM4-qUEJcTONPIZE68Gm0QWqyu7qM/edit?disco=AAAB90egEj4