Dropped readings left silent gaps that looked like flat readings downstream.
Conversation
Turn 1 — User
A sensor dropped out for an hour and the dashboard drew a flat line through it.
Turn 2 — Assistant
There was no distinction between a missing reading and a repeated one. Gaps are now detected against the expected interval and recorded explicitly, so downstream can draw a break instead of interpolating.
Backfill re-requests a bounded window when a device reconnects, and refuses beyond that rather than pulling unbounded history.
Turn 3 — User
How does it know the expected interval?
Turn 4 — Assistant
Declared per device rather than inferred. Inference was the original approach and it adapted to the outage, which is exactly when you need it not to.
Subjects
- [[Time Series]]
- [[Observability]]
- [[Python]]