Authoritative shards cause all the CSRs to be cleared at live recovery rollback (when the process doesn't restart). This is a regression from FCV 8, where we left CSRs intact on live rollback.

We actually think that even in authoritative shards, clearing the CSR is not really needed. The protocol ensures that CSRs are updated under a critical section, and once a critical section is taken, DDLs always retry until successful. The steps they try are also idempotent. In other words, the system will likely come out clean after rollback recovery with the correct CSR installed, regardless of whether we clear them at rollback.

SERVER-130703 introduced a server parameter that controls whether or not to reset the CSRs at rollback recovery.

We will use this ticket to prove and test that it is safe and correct to leave the CSRs intact at rollback. Then, if needed in production and a customer is impacted, we can offer them the server parameter as mitigation.