Keying the job on (account_alias, report_date) instead of the execution timestamp is the right fix, since it makes the schedule idempotent no matter how often the clock fires. A complementary option is to schedule in UTC and derive the local business date inside the job, which avoids ambiguous local times altogether. Do you also record a status column such as pending, running, sent so a crash mid-send can be resumed without posting the Slack message twice?
