You run six instances of a service and need a nightly job to run exactly once. How do you arrange that?

Do not rely on every instance running a local cron. Use a leader lock, such as a row or key acquired with a lease and a time to live, so only the holder runs the job, or delegate scheduling to a platform level cron that enqueues one job for a worker to pick up. Either way, make the job idempotent and observable, because a lease can expire mid run and cause an overlap.

The naive version runs six times, and the version where one special instance is designated stops working the day that instance is unhealthy. So I either use a lock or take scheduling out of the application entirely. With a lock, each instance tries to acquire a lease at the scheduled time, typically an insert or a conditional update on a jobs table with an expiry, and only the winner runs, renewing the lease while it works. The subtlety is that a lease can expire during a long run, and then a second instance starts, so the job still has to be safe to run twice. My preference at any scale is the second approach: platform level scheduling that enqueues a single message, and ordinary workers consume it. That gives retries, visibility and history for free, and the application code becomes just a handler. Whichever way, I record start and finish times and alert when a job has not completed within its expected window, because a job that quietly stops running is the failure people notice weeks later.