We have observed a pattern (called "multi-plan storms") whereby many concurrent operations may engage in multi-planning for the same problematic query shape. Today it is possible that the sheer magnitude of work done as part of multi-planning these queries prevents the query subsystem from completing any of those plans, caching the winning plan, and speeding up future attempts to plan. This ticket covers introducing a serialization point in the multi-planner, loosely following pseudocode algorithm:

• Ask plan cache for plan given shape
• If “no plan available”, plan and put plan in cache
• Else if “someone is planning that shape already” then “wait for plan on that shape”
• Run