Health signals
Two numbers matter: journal lag (entries written but not yet projected) and checkpoint age. Both are exported as gauges, and an alert fires when lag exceeds fifteen minutes or a checkpoint has not moved during a settlement window.
Replaying a projection
- Pause the projection's queue consumer.
- Truncate the projection and reset its checkpoint to zero.
- Resume the consumer and watch checkpoint age fall.
Replays are idempotent by construction, so the operator does not need to know how far the projection got before it stopped.
Month-end
Close the period by freezing new transfers for the window, waiting for journal lag to reach zero, then running the reconciliation report. A non-zero difference is an incident, not a rounding problem: the journal is the source of truth and the report is a projection of it.