A patient asked why she got two reminder texts a day when everyone else got one. Fair question. She had two open episodes in my Medicare RTM engine, and I had written the reminder logic to run once per open episode. Two episodes, two clocks, twice the nagging. The scheduler fired exactly on time. It always does. A few months later, in a different industry, I caught myself starting to write the same bug again. That was when I finally saw what the bug was.
The scheduler is innocent
When a cadence goes wrong, the scheduler is the first place everyone looks. I looked there too. I read the cron expression. I checked the timezone. I hunted for drift. Everything was fine, because cron is almost always fine. The scheduler is the most audited hundred lines in any system I run. The bug was somewhere I had never thought to look, because I had never noticed I was deciding anything there. I had let the episode own the clock.
The fix was one sentence long: remind once per patient, not once per open episode. A patient does not experience episodes. She experiences her phone buzzing at dinner. The moment I said the fix out loud I could hear how obvious it was, and that is exactly why it survived review. It did not look like a decision. It looked like plumbing.
