Scheduled jobs platform

Design a scheduled-jobs / cron platform that lets ~10K internal services register recurring jobs (cron expressions) and one-off delayed tasks. Scale: ~2M scheduled executions/day, jobs range from 100ms to 30min, some are critical (billing runs) and must fire exactly on time. Requirements: a job must fire even if a scheduler node crashes, must NOT fire twice (or must be idempotent), late jobs (missed window during an outage) need a defined catch-up policy, and the system must scale horizontally without two schedulers double-firing the same job.