The watchdog's first pass after a deploy records standing incidents as
announced without paging. Their end then sent a recovery note for
something the operator was never told about. Such incidents now carry
the epoch as their announcement time, which no real announcement can
have, and close without a note.
The dose arrival record describes the phone's Apple Health sync. The
measurement batch already leaves it alone for a narrow, single-purpose
credential; the intake batch now does the same, so the record stays
honest should the route ever accept such a credential.
The extra-dose rule split a dose that HealthLog already recorded: a tap
at 08:00 and the same dose arriving from Health at 08:20 became two
doses and consumed stock twice, and a dose edited in Health (a new
record, a new time) became a second dose next to the first. The rule
now follows the calendar and the record: a take on another local day
is another dose; on the same day, a take logged in HealthLog converges
with the incoming record, a new record that names the slot as its
scheduled time is the slot's dose recorded again, and only an
unscheduled record more than a minute from the held one is an extra
dose.
An account whose backfill slice held it made the hourly job wait up to
eight minutes, and the sync after the wait ran as long as it needed:
past the fifteen-minute expiry pg-boss failed the job and retried it
beside the run still going. The pass now spends the job's budget: the
lock wait is capped at what is left, the sync's walks stop when the
budget is spent and leave the rest of the window to the next hour (the
cycle is not stamped), and an account reached after that is skipped.
A stop on the budget is not counted as a failure to retry.
The watchdog sent one page per account, so a provider outage or an
expired app credential paged once for every connected account. A pass
now collects its messages and sends them per provider at its end: more
than three accounts of one provider opening (or closing) an incident in
the same pass become one message naming the count, the reasons with
their earliest start and the accounts, and one recovery note. Every
incident in it counts as announced.
A failure streak reaching the alert threshold paged the operator even
when the watchdog had already announced an open incident for the same
connection. The ledger now checks for that announced incident first,
stamps the streak as announced and sends nothing; the watchdog's
recovery note still ends it.
The watchdog opened an incident for any degraded state that lasted a
day, so an Apple Health account whose phone app stopped running read as
no_new_data for good and paged. Incidents now open for action_required,
for provider errors and for the watchdog's own unexpected silence; stale
or stalled data alone never pages. The first pass after a deploy, with
no watch row yet, records what stands as already announced instead of
paging every long-broken connection at once.
A walk that reads without the edge bound passed over everything earlier
slices wrote on every slice, and once it crossed the edge the skipped
pages were counted against the walk's page ceiling. The skip now stays
counted for the whole walk, runs on the slice's clock and its own
three-minute limit, and saves the page token it stopped at (on a limit,
a rate limit or the slice budget) so the next slice re-reads that one
page and goes on. A saved place counts as progress only when it moves
forward; a token Google refuses, or one that no longer reaches the edge,
sends the walk back to the newest page as before.
Migration 0401 turns healthkit_type_syncs.type into text so the ledger can
hold MEDICATION_DOSE, and adds last_background_at (seeded from rows whose
latest arrival was a background one). POST /api/medications/intake/bulk
takes syncTrigger and records APPLE_HEALTH dose arrivals; the Apple Health
status read serves a MEDICATION_DOSE freshness entry and lastBackgroundAt
on every entry with arrival facts.
One mocked page carries a provider in each state, checked at desktop width
and at 390 px by data attributes, with the menu driven from the keyboard,
the copied diagnostics compared with the shown text, and an axe pass in both
themes. The solid destructive badge replaces the red-on-red wash, which
fell below AA.
The empty note told a connected Oura account to connect Oura. The page
charts the WHOOP and Polar recharge, strain and load readings and the
Oura resilience level; Oura's readiness is the recovery score linked
above. The note now says so.
Oura returns detected periods the person dismissed (type rest) and ones
they deleted in the sleep collection. They were imported as sleep, with
their lowest heart rate, HRV and breathing rate, which added an extra
session to the day. They now map to nothing, and rows an earlier sync
stored for such a period are soft-deleted on the next sync through the
shared write tail.
A nap on the same wake day moved the night's latest instant to the end
of the nap and added its minutes, so the midpoint of a regular night
landed around noon. A few such days scored consistency at 0, and a nap
after the latest night scored timing at 0 as well. The night now carries
its main session (wake instant and asleep minutes), and the midpoint is
read from that alone. Sleep rhythm and chronotype read the same midpoint.
Every connection in Settings renders the same card from the server's health
field: a badge with one word, one sentence per reason, the newest data as a
meta line, a collapsed details block whose text is exactly what the copy
button copies, and an action row with the rare actions in a menu and at most
one primary button. The per-provider banners, the raw provider error line,
the token expiry, the failure counter and the button rows are gone; every
capability they carried is reachable from the sentence, the one button, the
menu or the details. The managed-profile view renders the same card,
read-only. A structural guard keeps the cards off the retired fields.
The guardian's read of a managed record's connections now carries the same
server-resolved health the owner's own envelope does, loaded for the record
and never for the guardian. The legacy state stays on the wire for one
transition release.
Their rings and straps upload only when the phone app is opened, so a day
without opening it is ordinary and must not page the operator. Wearables
whose cloud receives data on its own keep 24 h, Apple Health keeps 48 h,
and scales, cuffs and Strava never alert on silence.
A fixture dated near the day it was written passes until the date slides
out of a window the route counts back from the clock. Five did: a snooze
until a fixed date that becomes the past in January, a prognosis read
against the wall clock over days seeded at a fixed date, sleep rows and a
weight streak read through windows that end now, and a sync token that
would expire on a fixed date. The e2e mood heatmap mock and a timeline
URL are dated from today as well.
A guard scans the integration and e2e trees for near-now absolute dates
within reach of a window parameter, with an allowlist for files whose
windows are fixed on both ends.
1.58 million readings, paged by a generated fake of the list API that
honours the edge and lower-bound filters, with a server error, a timeout and
an hour-long rate limit injected. Pins that the backfill completes over
several slices, imports every reading exactly once without rewriting a row,
lets the hourly sync run between and beside slices, writes one log event per
slice with its requests and edges, and keeps the heap flat across the
history under the production heap ceiling.
A backfill noted progress only once a whole collection was read to the end.
A heart-rate history of years that one attempt could not finish (its budget,
a rate limit it could not wait out, a single failed request among thousands
of pages) started the next attempt from the newest page again, finished
nothing, and failed with "marker not stamped" on every attempt.
- Every spot-reading collection read page by page now records, after each
page it writes, the oldest sample it has written (backfillProgress.cursors).
The next attempt asks Google only for samples at or beyond that edge. The
edge is held back once a page's write fails, so no rows are skipped.
- The backfill runs in slices of five minutes or 500 pages. A slice that
finished a collection or moved an edge continues with the next one; only a
slice that moved nothing fails, and its error names each unfinished
collection and why the slice stopped. The slice ceiling is derived from the
collections and the page ceilings of the resumable ones.
- A rate limit Google will not lift within a minute ends the slice instead of
failing the walk, and the next slice starts five minutes later. Server
errors, timeouts and dropped connections are retried three times.
- The heart-rate stream is read last, so the light collections finish in the
first slice.
- The hourly sync waits up to eight minutes for a running slice instead of
skipping the account for as long as the import runs.
- Each slice annotates googleHealth.backfill.slice with pages, elapsed time,
stop reason, per-collection status and edges, and reports progress on the
job.progress line.
The visit form's lab picker offered the newest 200 results, the condition
picker the newest 100 conditions, the document pickers the newest thousand
documents and the document sheet's visit picker the newest 100 visits.
On a long record everything older could not be linked.
Each picker now opens on the records around the date of the thing being
linked, pages older and newer from there, and searches through the list
route itself. Records already linked always show, from the labels the
record's own read carries and from what was picked in the sheet.
The lab list takes a q filter over analyte and panel, the condition list
from/to/q/offset/sortDir, and the visit list an offset.
An hourly and a daily pass flag a continuous source that delivered on
most of the previous 14 days and has gone quiet, and open one operator
incident per connection when a connection stays degraded or needs action
for more than a day. The incident is announced once through the existing
admin alert path, adopted rather than repeated when the sync ledger
already paged, and closed with one recovery note. Scales, cuffs and
workout logs are watched for errors only, never for silence.
The admission drain ran each full-history import with no wide event in
scope, so every external call, warning and annotation the provider walk
recorded was dropped and a failing backfill left only its error message.
Each admitted run now opens a job.integration_backfill event with its kind,
identity, job id and slice, emitted when the run settles. The Health Connect
import worker gets the same treatment.
The external call list on one event is bounded so a long walk cannot hold
thousands of entries: failed and annotated calls keep the larger share,
the rest is counted in external_calls_omitted.
A HealthKit dose event binds to a slot by window band, and a weekly
schedule's band reaches a day either side of its slot. An extra dose
logged the day after a scheduled one landed on that slot and overwrote it
last-write-wins: the slot's time changed and the extra dose was stored
nowhere. An external dose record whose band slot already carries a
different take is now recorded as its own unscheduled dose.
The dose history now lets the take stored on a slot claim it before an
extra dose stored on its own instant, instead of following the order the
rows were read in, so the extra dose reads as off-schedule.
Marc-André Bombeck