Test that even when replication lag is high, we don’t see stalls refreshing metadata from config servers. In practice, always attempting to refresh metadata from the primary should eliminate this issue, as targeting a stale secondary should be avoided unless a primary is unavailable.

https://docs.google.com/document/d/1sSGIvuHDlNzFIQEIbUJ-8n0DePJn8ZjtNYRg4jLbOa4/edit?tab=t.0#bookmark=id.1w55t6n7knrw