Repository navigation
Data policy: decide when "no migration" ends, and plan history retention for health trends #123
Description
Update (2026-10-07): the owner confirmed there are no users yet, so the pre-1.0 "no migration" policy is fine as it stands. Two things remain open: when the policy ends, and how much history the app keeps.
The decision
- AGENTS.md: "Pre-1.0: no back-compat or migration code … change formats freely and document the reset".
- ADR-0002: "rows written by an older format are dropped on load, never migrated."
- Draft Replace the history CSV with an append-only raw ChargeState log #94 replaces the CSV without importing it ("Pre-store CSV history is not migrated").
That's reasonable while the only install is the owner's.
What to decide now, while it's cheap
- The trigger that ends the policy. The rolling preview, the Obtainium instructions on the README and website, and five tagged betas mean anyone can install today. Pick the event — the first non-preview tag, a store listing, or the first known user — after which on-device formats carry a version, and every format change ships an importer or at least an in-app "history was reset" notice.
- Retention. With Replace the history CSV with an append-only raw ChargeState log #94, history is capped at 20,000 records per vehicle, and every 10 s poll is logged with no deduplication (ADR-0006). That covers roughly 2–7 days of awake time, compared with about two weeks before Replace the history CSV with an append-only raw ChargeState log #94. Degradation trends need months to years.
Proposal
- Record the cut-off trigger in AGENTS.md and ADR-0004.
- Define retention: keep raw records for N days, then downsample to per-session and per-day summaries kept indefinitely, and have the health estimators read the summaries.
- Once the cut-off passes: add a golden fixture file for each shipped format, plus a test that the current reader accepts them.
Related: #94, #112 (a slower parked cadence also stretches retention).
Activity
[-]Data policy: "no migration pre-1.0" vs a publicly distributed preview; users silently lose history and pairing[/-][+]Data policy: decide when "no migration" ends, and plan history retention for health trends[/+]dzid26 commentedon Nov 1, 2026
Closing as decided: both questions in this issue have an answer on
main.AGENTS.md: the first normal release tag. Prereleases do not count. From that tag every on-device format is versioned, each change ships an importer or an in-app reset notice, and each shipped format gets a golden fixture.events.dbexpires after 90 days,derived.dbholds the thinned long-term samples.RoomBackedRecordStoreTestproves the policy: "phone-side events expire after 90 days and car records do not" and "the car cap drops the oldest of the time span across series and leaves the events file alone".Still open and tracked in ADR-0014, not here:
owner.db, per-day rollups, and pointing the health estimators atderived.dbfor periods the raw window no longer covers.