The incident started as a latency spike and ended as a deadlock investigation. The hard part was not finding one broken query. It was recognizing that several otherwise acceptable queries all converged on the same hot records during peak traffic.