See this comment- we should add a load test for this scenario & see if its easy to reproduce or if the proposed mechanism of 5 retries is sufficient.

If it proves to be a problem, some ideas for addressing it:

• Don't awaken all threads, only awaken some at a time (e.g. number of free tokens, half the sleeping threads). This may have a performance downside for the (hopefully common) case where we have a plan cache activated & all sleeping threads can now short-cut to this plan
• Use a query knob for query retries so that there is a mitigation available if customers run into this issue