Google Health synchronization

Google Health is a read-only local mirror. One serialized worker runs fixed-rate cycles, normally every 30 minutes. Busy ticks are skipped. Each cycle has a new sync ID and a deadline one minute before the next nominal tick. Manual sync, reload, and retry use the same guard, budgets, request limiter, and durable cooldown.

Scheduling deadlines stop new requests. Complete batches save under the separate database timeout, even across a stream or cycle deadline. Records, checkpoints and historical rotation commit atomically before yielding; batch sizes are unchanged. Historical finalization is a separately admitted transaction. A cycle may overrun while its current batch finishes; busy ticks still create no new job.

Both lanes have reserved time and may borrow unused capacity. Pending full and reload work has a positive reservation inside recent time. Route exports have a bounded turn inside that reservation. Fast-first work does not create extra jobs or continuation chains. A completed full pass records its next due time; another pass starts at its new cutoff rather than the old pass's end.

Temporary provider/network errors and transient database failures retain WAITING work. From the fourth failure, the normal sync interval doubles per failure with up to 20% added jitter, capped at work-retry-max-delay (six hours). Validation and other non-recoverable failures retain the four-attempt skipped/parked limit. Routes use the same policy; a failed route without a retry time needs explicit retry. Quota cooldowns, budget yields and cancellation consume no work attempt. Full recent passes advance past temporary cells in slow retry and finish with errors, retaining their waiting gaps. Fast refreshes and later full passes retry those gaps when eligible, so one failing stream cannot block new shared passes. Reload parents inherit the cell's failure state and retry time without counting another failure while the cell waits. Committed historical pages survive retries. Unrecoverable pagination-token errors retain their immediate parking after token recovery fails. Successful work clears its active error and retry time; earlier failed-attempt audits remain history until the local mirror is deleted.

Physical instants, civil components, and calendar dates retain their native query keys. Civil queries cover UTC−12 through UTC+14 at both ends. Date queries cover whole native days, including the upper civil date. Requests and checkpoints use microsecond precision. Exercise ownership uses civil start; sleep ownership uses UTC end day. Sleep charts still group whole sessions by their start day in the viewing time zone. A rejected upper envelope is a failure, never an empty success.

Named points retain Google identity, including heart-rate samples. Nameless raw identity includes canonical observations and source metadata; differing same-time observations survive. Nameless reconciled and rollup keys use native intervals. Identical duplicates collapse; conflicting duplicates invalidate a complete query. Historical pages are normalized once before their guarded transaction resolves sources and writes bounded multi-row INSERT … ON CONFLICT batches. Unchanged rows are not rewritten. Automatic history performs no membership stamping, pre-write changed-count scan, missing-record sweep, or completion-time mirror scan. Write results supply changed counts; stamp and deletion counts stay zero.

Named heart-rate records use a partial unique index on connection and identity; timestamp corrections still update one row. Nameless identities include their physical time and content, so their partial unique index starts with connection and sample time before identity. Recent and maintenance lookups include that time. Record keys and logical uniqueness stay unchanged. The time prefix avoids scattered reads across the whole history during an import. Mirror fingerprint indexes are no longer maintained; lookups use identity and native time.

Content-independent identities retain a bounded identity/fingerprint map in history_child.query_fingerprints, committed with the page checkpoint. Resumed queries restore it; restarted, revoked, and completed queries discard it. Nameless raw identities already include content and need no separate conflict map.

Historical data is treated as fixed. One successful history_audit record per account, stream, representation, calendar policy, key policy, and native day is the completion authority. All required queries and pages must succeed before the audit and work completion commit together. Valid returned counts distinguish DATA from EMPTY; an unchanged existing record still counts as data. Measurement gaps are valid. An empty result does not delete local rows. Bounded empty-range queries still produce a separate completion record for each day.

Each cycle keeps a search position per stream and reuses its selected day when creating work. The position moves backward only after that selected work and its day audits commit. Completing older work resumed from another cycle does not move it past newer gaps. A completed stream needs one gap search per cycle; saved retries remain eligible after the search is exhausted. Positions reset every cycle, and ownership guards still reject work revoked by reload. Audits remain the completion authority. The cycle log's historyGapSearches counts these gap queries, separately from provider requests.

Automatic sync skips successful days permanently, including after restart or a wider recent window. Explicit range/all reload supersedes matching marks and revokes overlapping work without pre-deleting records. A durable RELOAD_REQUESTED audit records the requested range; new historical work sets replace_missing and retains maintenance stamping and replacement, including recent-to-history handoff. Request audits do not count as completion. Unrequested parked portions retain their state under new owners. Retry releases incomplete work without reopening successful days. Reset clears mirror, audits, work, failures, passes, ownership, and caches. Empty audits and retained owners protect account identity.

Route ownership is separate from mirror points. Valid exports use atomic file publication and keep their indexing associations. Recent replacement and explicit reload update absent exercise presence while retaining its file, owner, and track. Track replacement also updates absent owners. Adding location access exports stored missing routes without reopening history. Automatic synchronization never removes published route files; file deletion is confined to confirmed administration actions.

Status separates cycle counts from cumulative pass/traversal counts. It reports fast bounds and completion, frozen full positions and due time, skipped/handed-over chunks, parked ranges, and contiguous history finalized back to the first gap. The Processing screen separates Recent, History, and Routes, with searchable data-type disclosures and persistent expansion across cycle transitions. Growing cycle totals use indeterminate activity; incomplete or paused work cannot appear as a completed refresh. Mirror totals and history coverage carry snapshot times. Coverage snapshots are bounded and cached; status polling does not reload all audits each time.

Shared client count timelines animate confirmed count changes, settle completed cycles, and reset for each new sync ID. Active Data & Records, Processing, and service views share one status request every eight seconds. Reopening a view reuses the same in-memory snapshot and animation timeline.

Overview opens with only the grouped chart query. It reads any existing shared status without requesting or polling it. Expanding Latest values loads the per-type date bounds and latest measurements; successful results are reused until the screen closes or a shared sync update makes them stale. Closing the section discards pending results and prevents follow-up requests. Overview does not fetch catalogue counts for its footer.

Health screen loaders track their own requests. Shared sync updates refresh Overview charts and expanded latest values. Data & Records refreshes records, summaries, and sources without invalidating an in-flight status retry. Closing a screen invalidates all of its pending results.

Record list and detail responses expose an optional typed measurement with a required numeric value and unit, plus optional minimum and maximum. Values use the chart units. The server derives valueSummary from that same measurement and preserves numeric zero. Unsupported or unavailable measurements omit the typed value. OpenAPI defines the response and generated client types. List queries project only measurement operands for the selected page. The full provider payload remains available in record details for diagnostics.

The heart-rate covering-index migration drops a leftover index before retrying its concurrent build. A recorded successful original migration keeps its index; its original checksum remains valid.

There is no legacy backfill or identity migration. Successful audit marks stay authoritative. An incomplete checkpoint without query-validation state restarts its subdivision under new ownership without deleting committed records. Legacy route columns remain inactive.

The implementation plan contains the acceptance gate. Real-account density, upper-envelope acceptance, first-seen delays, and comparative request/byte costs still need measurements; synthetic checks do not establish a production freshness or speed guarantee. See authorization and the user guide.

Chart grouping

The grouped metrics endpoint supports AUTO, SECOND, DAY, WEEK, MONTH, and YEAR. SECOND accepts one local calendar day using the same inclusive start date, exclusive end date, and viewing time zone as the calendar groupings. This includes 23-hour and 25-hour daylight-saving days. The overview labels SECOND as Intraday.

Heart rate reuses the repository's time buckets with a one-second interval. Timestamped weight, body fat, body temperature, and blood glucose readings are grouped by second after conversion to display units. Both retain average, minimum, and maximum in detailedPoints, with the selected day's summary in buckets. Daily values are not substituted for missing samples. The same chart component draws the time series and breaks a line where the gap between readings exceeds three times the series' median reading spacing, so devices that record every few seconds or minutes still draw a line. Readings with no neighbor within that spacing are drawn as markers.

The overview owns one time window and one marked time for all intraday graphs. The time window control, drag selection, and pinch update that shared window without another metrics request; each graph applies it to its axis without rebuilding its series. Selecting a point marks that time on every graph, and each graph shows its nearest reading within the gap spacing. The window and mark are kept when expanding a graph and reset when a different day or grouping loads. A daily measurement graph offers the selected day as an intraday view.

Setting constraints

The settings are listed in Google Health setup. The fast window must be at least the sync interval and no longer than the physical recent window. Duration settings must be positive. recent-window covers at least two whole days, cycle-deadline-margin stays smaller than the interval, and history-budget-fraction stays strictly between zero and one. Explicit history-start-date values are preserved. Request spacing is one second across pages, probes, retries, and route exports. Provider cooldowns persist across restarts. Civil offsets are limited to UTC−12 through UTC+14; the earlier parser's ±18-hour range is no longer accepted.

cursor-overlap, initial-window, and daily-reconciliation-history are retired. Daily/monthly reconciliation schedules and automatic continuations are removed. The default history floor changed from 1970 to 2010; it is a configured limit, not Google's earliest available date. The current layout targets fresh mirrors; existing checkpoints and route state from earlier development builds are not migrated or silently discarded.