A rebuild time short enough that the index stops being something to protect.
what made it eleven minutes rather than four hours:
bulk indexing with a 5 MB payload rather than
document-at-a-time
refresh_interval set to -1 during the rebuild
replicas set to 0, restored afterwards
a source query that streams rather than paginating
with OFFSET
and what that changes: the index is no longer backed
up, no longer restored, and a mapping change is a
rebuild rather than a migration.
A derived store that can be rebuilt quickly does not need a backup, which removes an entire operational category. The four settings are each documented and the combination is what matters — dropping the refresh interval alone gets you to forty minutes, which is still long enough that people avoid rebuilding.