Each `createManyAndReturn` flush of the export import left about 5 MB of
heap behind until some two dozen calls had piled up. A 300 000-record
synthetic import under a 524 MB heap settled at 53 MB retained; a run of
back-to-back 600-row calls with nothing between them reached 212 MB.
Spot rows now go through the column-array `INSERT … SELECT FROM unnest`
the restore already uses, with `ON CONFLICT DO NOTHING RETURNING`, which
is what `createMany({ skipDuplicates })` emits. The same import holds
11.5 MB and finishes in 35 s instead of 55 s. Ids are minted in the
cuid() shape, defaults are written as the client would write them, and
integration tests compare the returned rows and the resulting table with
`createManyAndReturn` for fresh rows, conflicts on either unique
identity, and duplicates inside one call.
The one-shot hourly rebuild of already-folded heart-rate, HRV and SpO2
days walks one account's whole folded history in one job with the
default fifteen-minute expiry, about fifty statements per day. Like the
drains it now stops between days once 75 % of the expiry has passed or
pg-boss aborts the job, and fails so pg-boss retries it (the boot send
allows three retries); a rebuilt day leaves the candidate set, so the
retry starts at the first day not yet rebuilt.
The off-host backup walked every account in one job with the default
fifteen-minute expiry, and a retry started again from the first account.
On an instance with one large record next to others, an attempt that
ran out of time spent its retry uploading the accounts it had already
done (#1031).
The run now stops between accounts once the job's time budget is spent
and fails, so pg-boss retries it. Every attempt of a job carries the same
creation time; an attempt skips the accounts whose ledger shows a success
since then, and keys its objects on that day, so a retry after midnight
still writes under the run's date. A new job uploads every account again,
and a retry whose accounts were all done passes.
After an Apple Health import the worker folded the rollup table from the
account's first measurement to its last, whatever the import carried, and
so also folded history older than the five-year window that the boot
backfill, its discovery and every read path keep to.
The import now records the oldest and newest instant it wrote
(`measuredSpan` on the job result), and the fold covers that span,
clamped to the fold window. On the seeded account of 1.25 million rows the
full fold took 12 s, the slowest statement 2.7 s against the 60 s
statement limit, so this is about the window rather than time.
Key rotation and the rotation status read every column that was not
flagged as a blob with one `findMany` over the whole table.
`Measurement.notesEncrypted` and `MoodEntry.noteEncrypted` are such
columns, so rotation loaded every measurement of the instance: 1.25
million rows on one large account (#1031), and the rotation status page
did the same on every view.
Every column is now walked with the id-cursor pages the blob columns
already used: 25 rows a page for blobs, 5 000 for everything else. The
maintenance script's own per-column walks page the same way. On the seeded
account with 1 460 notes to rotate, peak heap fell from 272 MB to 153 MB,
and it no longer grows with the table.
`scanned` now counts the rows that hold ciphertext for every column, as
it already did for the blob columns, rather than every row of the table
for some columns and only the ciphertext rows for others. The runbook
says so.
The repair for the inflated cumulative records deleted the suspect rows,
committed, and then re-ran the detector. When the detector failed, which
a timeout on a large account or a restart can cause, the records stayed
deleted, and since discovery looks for the suspect rows the account was
never repaired again (#1031).
The delete and the re-derivation now run in one transaction, with the
detector reading through it, so they commit together or not at all. The
transaction gets five minutes; the detector took 19.8 s on the largest
seeded account.
The personal-record pass, every half hour for every account, read every
live row of each cumulative type into memory to find the best day. With
784 000 active-energy rows on one account (the per-sample rows pile up
when the nightly drain cannot finish, #1031) it ran a 524 MB heap out of
memory; the old code needed 1.1 GB of heap for that account.
It now reads the type in bounded keyset pages and handles one local day
at a time. The canonical-source pick it applies is per day, so the day
totals and the winner are the same; a tie still keeps the earliest day.
Measured on the same account: 189 MB peak heap, the same record. The
warm-up gate counts only up to its threshold instead of counting a whole
history.
`iterateSourcePages` and `iterateDayBuckets` in the consolidation base
take the row type as a parameter, and the day grouping can keep every row.
An account of 1.25 million measurements could not export its data or be
restored from any backup (#1031). Measured in a 1 GB container:
- The full export and the passphrase export built the whole payload and
its JSON string in memory; the app ran out of heap and restarted.
- The weekly backup's JSON is 662 MB, longer than any string V8 can hold,
so `unpackBackupBlob` threw on every read. Restore, download and the
restore preview all answered "could not be decrypted".
- Upload accepted 10 MB, a fraction of a large export.
Every reader and writer of a backup file now streams it:
- `scanBackupJson` reads a backup's bytes once, keeps every section but
the named bulk arrays, and hands their elements out one at a time.
`readStreamedBackup` uses it twice: a first pass that validates each
measurement against the element schema and keeps the rest of the file,
and a second pass that hands the measurements to the writer in batches.
Stored copies and off-host objects are authenticated whole on their
compressed ciphertext before any plaintext is read; only the
decompression is streamed.
- Restore writes the measurements with one `INSERT … FROM unnest` per
batch. Prisma caches a query plan per `createMany` row count, and one
1 000-row plan kept about 145 MB alive, which alone put the restore past
the heap limit. The transaction budget grows with the measurement count.
The awards and ECG sections are handed the ids they reference that were
written, not every written id.
- The export routes write the file into the response as it is produced.
The passphrase export spools its ciphertext to a temporary file because
the HLX1 format puts the tag first; the format is unchanged, so existing
archives open as before.
- Upload accepts the raw body, JSON or gzip, up to the 512 MB request
ceiling, checks it as it arrives and stores it in the same pass; the admin
page compresses the file in the browser before sending it. Multipart stays
for small files.
- Download streams the stored JSON; the preview counts measurements from
the stream. The monthly restore drill and `scripts/restore-backup.ts`
read the off-host object the same way; the script can write `.json.gz`.
On the seeded account in a 1 GB container: export 25 s at 389 MB peak,
passphrase export 25 s, both uploads 19 s, restore of the weekly backup
94 s at 554 MB peak, and export, upload and restore round-trip to exactly
the same rows. The encrypted-column guard now also recognises the raw SQL
statement that writes `DataBackup.data`.
The backup streamed its JSON through gzip and the cipher, but kept the
stored copy as one string and handed that to Postgres in one parameter.
On a seeded account of 1.25 million measurements the copy is 64 MB of
base64, and the pieces, the joined string and the driver's copies of it
grew the process from 309 MB to 946 MB resident. A bare process needed a
heap limit above 400 MB to finish it (the new path finishes under 220 MB).
With the dense retention running beside it in a 1 GB container the
process died of heap exhaustion, and the backup job sat "active" until its
two-hour limit marked it "job timed out" (#1031).
`packBackupBlobInto` hands the stored copy on in pieces of up to 4 MB as
it is produced, and `storeBackupBlob` writes them into a temporary table
on its own transaction and assembles the row with one `string_agg`. The
process never holds the whole copy. The account's previous copy stays in
place until the new one is complete: a failure rolls back to it.
Measured on the seeded account in a 1 GB container, web and worker in one
process: the backup completes in about 30 s with the dense retention
running beside it, where v1.39.0 crashed within 30 s.
The streamed JSON also reports its progress to the job every 50 000 rows.
pg-boss marks a job "job timed out" at its expiry and fires its abort
signal, but does not stop the handler. A drain that could not finish kept
writing after its job was declared dead, and the retry then ran beside it
over the same days.
The nightly drain tick and the per-user dense-retention and daily-mean
jobs now stop between days once 75 % of the job's expiry has passed, or
as soon as pg-boss aborts the job, and complete with `stopped_early`
instead of failing. What they folded is committed. A nightly tick that
stopped early after folding something sends itself again a minute later,
up to 24 times, so a first-time backlog after a large import drains the
same night; a run that folded nothing does not hand on, so a pass that
cannot move does not loop.
The expiries themselves stay at their defaults: with the passes bounded
per day and budgeted, a longer limit would only delay the signal.
The consolidation drains (cumulative, daily mean, dense intra-day
retention, legacy steps) gathered a type's whole history into one array
before folding the first day. On an account of 1.25 million rows the
heart-rate array alone held about 230 MB, and the dense retention peaked
at 460 MB of heap against a 524 MB limit; in a 1 GB container running the
web server and the worker together the process died of heap exhaustion
within a minute and the job was later marked "job timed out" (#1031).
- Pages are now handed on one at a time and each local day is folded as
soon as its last row has been read, so a pass holds one page and one
day.
- The keyset cursor gains a redundant `measuredAt >= cursor` bound.
Without it Postgres could not use the cursor as an index condition and
re-read every earlier row on each page (47 ms and 248 000 filtered rows
for one page halfway through the heart-rate history; 5 ms now).
- The dense fold resolves a day's hourly rows with one lookup and one
insert instead of two lookups and a write per hour. On the seeded
account the pass went from about 290 000 statements to 54 000.
- `shouldStop` lets a caller end a pass between days; the result says
`stoppedEarly`. Days commit one by one and a folded day leaves the
scan, so the next run continues at the first day not reached.
A job that ran into its pg-boss expiry left no trace in the app log. Most
handlers annotate a wide event nobody emits, pg-boss records an expiry only
in its own table, and a worker that dies of heap exhaustion writes one
FATAL line that names no queue. On a large account three queues failed
with "job timed out" for days and the log said nothing about them (#1031).
Every handler now runs inside `observeJob` (through `runJob`), which writes
`job.progress` every two minutes while a handler runs (elapsed time, heap,
RSS and the last progress the pass reported), `job.expired` when the job
passes its expiry while the handler is still running, and `job.finished`
when a long or expired run settles. At worker boot, `worker.boot.jobs_cut_off`
names every job a previous process was still running when it stopped.
The admin card's "job timed out" now says what the limit was and that the
cause is either a run too slow for it or a restart while it ran.
Automatic matching counts a dose taken up to an hour before a slot's
on-time window as on time. A dose pinned to the slot was judged from the
window's start instead, so the same dose read on time unpinned and late
once pinned. A pinned dose now uses the same early grace, and the
confirmation says so plainly: the dose counts as if it had been logged for
the slot, on time or late depending on when it was taken.
Refs #1028
The history view treated every auto-miss anchored before the medication's
creation as the today projector's placeholder and dropped it. That is true
only on the creation day. History that arrives from before the creation, a
restore of an older file for one, can carry real misses, and those vanished
from the history and the rates.
The creation stays the boundary of expected doses: an earlier one, the
first recorded dose or the course start date, would read every day between
it and the creation that no row names as missed, and an import carries
takes and skips only, so those misses would be invented. Before it, every
recorded take, skip and miss claims its slot. Only rows that record nothing
are dropped: a pending row, and an auto-miss within the day before the
creation, which is where the projector's placeholders sit.
Refs #1028
The today projector mints a pending placeholder for every slot of the
current day, so a medication added in the afternoon carries placeholders
for that day's earlier slots. The hourly auto-miss pass stamped them as
forgotten doses a day later, and every reader that counts raw rows saw a
miss on a day before the medication existed: the miss-free streak broke on
the creation day. The dose history, the rates and the Coach already ignore
such a placeholder; the streak read the row as it was.
The pass now stamps a miss only on slots from the medication's creation on.
A placeholder before it stays pending, so the web and the app can still
offer the slot for logging and a dose recorded for it converges onto it. A
migration returns the placeholders an earlier pass already stamped. It
touches only rows that can be nothing else: server-minted, never acted on,
anchored before the creation but within the day before it, which is where
the projector's placeholders sit. An older auto-miss carried in with
restored history keeps its state.
Refs #1028
A portable export left the medication's creation instant out, so a restore
stamped the restore time instead. The creation instant is where a
medication's expected doses start, so the whole restored history then read
as predating the medication: its missed doses dropped out of every
adherence rate. The portable file now carries the creation instant as the
disaster-recovery file already did, and the restore, which already honoured
it, brings it back. A round trip pins both file purposes.
Refs #1028
The adherence calendar counted every off-schedule ledger row as a dose taken
on time. An orphaned skip or auto-miss, a row on an instant that is no slot
of the schedule, carries no taken time, yet coloured its day as taken. Such
a row now counts as taken only with a taken time; an orphaned skip reads as
skipped, and an orphaned auto-miss counts nothing, as it already does in the
rate.
Refs #1028
The confirmation promised that an attributed dose always counts as taken
late. It counts as on time when it falls inside the slot's on-time window,
which also covers a dose taken shortly before the slot's clock time, and
as late otherwise. The copy now says so in every locale, and a test pins
the split it describes.
Refs #1028
A medication added at 16:05 with doses at 09:00, 14:00 and 21:00 showed its
creation day wrongly in "Last 90 days". The 14:00 dose, recorded afterwards
on its own slot, read as an off-schedule take "due" at 21:00, and a 09:00
row with no time appeared as off-schedule too.
The history route minted slots only from the creation instant, so the 14:00
slot never existed for the dose to land on, while the write path, which
mints the whole day, had filed it there. The 09:00 row was the pending
placeholder the today projector mints for every slot of the day, which the
auto-miss pass stamped a day later; with no slot to bind to it surfaced as
an off-schedule row.
The ledger now mints slots over the requested window and takes the
creation instant as the floor of expected slots: a slot before it appears
only when a recorded dose claims it (a take by its window or its pin, or a
skip), and a pending or auto-missed placeholder on it is dropped. The first
archived schedule era also covers the stretch before the creation, so an
edited schedule behaves the same.
Every adherence surface tallies the same ledger with the same floor, so the
detail rates, the calendar and the ledger bar agree on that day. The doctor
report, the Coach and the adherence chips previously read the slots before
a medication existed as missed whenever their window started earlier; they
now read nothing there.
Refs #1028
The live daily fold ranked every row by source before folding, which sorted
the whole window twice. A year of per-minute heart rate (525,600 rows) took
about 0.55 s per series on Postgres 16, with both sorts spilling to disk.
The fold now aggregates per source and bucket in one hashed pass and then
keeps each bucket's ladder-canonical source. Folding the canonical source's
rows is the same as folding that source's group, so the numbers do not
change; the same year reads in about 0.29 s. A new integration case pins the
canonical pick, mean, count and spread for a day with three sources.
A start logged within one period before an existing start folds that start
into itself and soft-deletes it, since the two are one period entered out of
order. Nothing recorded which start was folded, so taking the earlier start
back could not bring the later one back: a mis-tap on the 5th with a real
start on the 10th, then deleted, left the 10th gone and the cycle before it
reopened.
Migration 0345 adds menstrual_cycles.absorbed_into_id. The folded row names
the start that absorbed it; removing that start restores it, re-derives its
own end and hands its days back to it. The column is needed because a
soft-deleted row alone cannot tell a folded start from one the person
deleted, and the audit trail is not state a removal can read. Deleting a
cycle directly now runs the same shared removal as deleting its day-log.
The runtime accepts "+05:30" as a zone id, so isValidTimezone let it through,
but Postgres reads a bare offset in AT TIME ZONE with the POSIX sign (east
of UTC is negative), and the day-grain SQL aggregates would cut every day on
the wrong side of UTC.
isValidTimezone now accepts named zones only: whatever the runtime resolves
to a bare offset is refused, while named zones, their aliases, UTC and the
Etc/GMT family stay accepted. Every write that already validated through it
refuses offsets now, and the screener and mood linked-context request schemas
validate their zone too. A value stored before this change can still sit in
the database, so the SQL call sites that read the stored zone directly (the
dashboard summary, achievements, the doctor report, the tiered AI context)
fall back rather than splice it, as the resolver and the day aggregates
already did.
The aboutMeQuestions capability admits an ai_insights_only receipt, but the
question generator ran its own gate that demanded a Coach receipt, so a
person whose capability read available still got the fallback questions on
an operator-held chain.
The generator now runs the shared wire re-check (aiEgressRefusal) for the
aboutMeQuestions key, so the receipt kinds and the rule for when one is
needed come from the capability table alone.
Migration 0343 moved the operator's Coach answer from module availability
into the Coach switch and folded every live row. Restoring the instance
settings from an older file wrote the dead coach key back into module
availability and left the Coach switch on, so a Coach the operator had off
came back.
The restore now applies the same fold: coach false in the file's module
availability switches the Coach off, and the key is dropped either way.
The visit form's lab and episode pickers fetched with their own page
sizes under the same keys as the Labs list and the illness pages, which
fetch with others. Whichever read landed first served the other: the
picker could see 50 episodes instead of 100, or the Labs list a 200-row
page as the whole list. The picker reads now have keys of their own,
under the lab and illness roots so writes still evict them.
A consent withdrawal purges the stored status notes, briefing, reaction
lines and workout paragraphs in its own transaction, but a provider call
that was already in flight still persisted its reply afterwards, bringing
the text back under a consent that no longer existed.
The consent rule now runs once more after the reply arrives, against the
provider that actually answered and with a fresh receipt read: in
runStatusCompletion (status notes, workout paragraphs, narratives, derived
assessments), before the briefing commit, and before the reaction-line
commit. A refused reply is dropped; the spent tokens stay on the ledger.
A dose or a visit can link documents from its own form, and a
document's sheet can link doses and visits; both ends write the same
table, and the sheet writes the whole set on every tap. Dose and visit
writes left the document reads alone, and the sheet seeded its pickers
from a detail cached for five minutes, so the first tap there after
linking from the dose's form wrote a set without the new link and
removed it.
Dose and visit writes now evict the document root, the sheet reads the
document again on every open, and its pickers wait for that read before
they offer a change.
With no provider set up, a fresh self-host's default, the Coach module
is on but /coach redirects to /insights because the Coach is not
available. The tour read the raw module map, kept its Coach stop, pushed
/coach, was sent back to /insights and pushed /coach again on the
pathname change: a loop.
The tour now reads the same capability-folded map as the navigation, so
the Coach stop is not in the list. And a stop whose page redirects is
dropped after one navigation, in the direction the person was going,
instead of being pushed again.
The daily sweep and the measurement-ingest hook surfaced due reminders into
a Coach conversation whatever the record's coach capability said, and the nag
cap then dismissed them after three ticks although no badge ever showed. The
app-open path already waited on the capability.
surfaceCoachReminders now holds every row while the coach capability is
unavailable (the same answer that keeps the badge dark), and the nag cap
neither charges nor dismisses a held user's surfaced rows. They surface on
the first tick after the Coach is back. The integration fixture gives its
users a provider and consent, and a held case covers the sweep, the nag cap
and the ingest hook.
Every AI surface reads its capability from the ai block on
/api/auth/me, but setting up or removing a provider, signing in to or out
of Codex, saving the fallback chain, the central-Codex opt-in, the
local-OCR switch, a minted consent receipt, the admin's server key and
central Codex, and a guardian's save inside a managed record evicted only
their own reads. The surfaces stayed on the old answer until a reload.
All of these writes, and the ones that already refreshed the payload
(consent, modules, Hide Coach, the operator's switches), now evict one
bundle, aiInputDependentKeys. A structural test holds every client write
to those endpoints to it.
A background job opens one wide event for its whole run, and the per-request
memo cached the operator switches, module availability, module preferences,
provider presence and consent kinds on it. The nightly status passes loop
over every candidate inside that one event, so a switch turned off mid-run
did not stop the remaining users, and the wire re-check read the same cached
answer.
memoizePerRequest takes a freshInBackground option; every input that decides
whether AI work may run sets it, so a background caller reads them on every
call while a request still reads them once.
Reminders are the person's own record, like the stored facts,
conversations and plans. The list and delete routes no longer ask the
Coach module; creating, confirming, resolving or rescheduling one stays
Coach use and is refused by POST and PATCH as before.
Settings -> Coach now always mounts the reminders card (shown with Hide
Coach on only when reminders exist), and Settings -> AI lists them through
the stored-memory block while Settings -> Coach is out of reach. With the
Coach unavailable the card is read-only: every reminder can be read and
deleted, while keep / done / dismiss are not offered.
Plans are the person's own record, like the stored facts and
conversations. The list and delete routes no longer ask the Coach
module, and the plans page turns read-only instead of redirecting to
Insights: every plan can be read and removed, while confirming or moving
one stays Coach use and is refused by the PATCH route as before.
The web no longer reads GET /api/feature-flags: useFeatureFlags and its
query key are gone, and use-ai-capability gains the two readers the Coach
pages and the operator notice need. A structural guard fails when web code
reads the retired projection, and when a module requests an AI route
without being registered with the capability that gates it.
Adds the copy for every new state in all seven locales and drops the
per-card no-provider strings.
Settings, AI says which AI features the server's operator switched off and
shows the one setup hint while no provider is configured. The consent
withdrawal says what it deletes and what stays, and a document-reading
consent on file gets its own withdrawal. Regenerate follows the briefing
capability and the connection test names an operator refusal.
The Coach memory (facts and stored conversations) is never gated on the
Coach: Settings, Coach always shows it, and Settings, AI shows it while
the Coach page is out of reach, whenever rows exist.
Onboarding offers AI only where it can be set up here: the done-screen
panel and the getting-started row go away while the operator has AI off
or inside somebody else's record.
Reading a document, scanning a lab report and turning a medication
description into a schedule follow documentAi, labsOcr and
medicationExtract. Unavailable, the action is not offered. With only the
document-reading consent missing, the place that reads asks for it and
grants exactly that (ai_extraction), nothing wider.
Refusals are read by meta.errorCode: a 403 is no longer always called a
missing consent, and a spent daily budget no longer shows the rate-limit
sentence.
The floating button, the drawer, the launch pills, the nav entry and the
Coach pages read the coach capability from /api/auth/me instead of
combining the operator switch with Hide Coach on their own. A missing
provider or consent now hides the Coach too, instead of opening one that
can only answer with an error.
Stored conversations stay reachable: while the Coach is unavailable the
conversation list turns read-only, opens a transcript in place and keeps
delete. Refusals before the stream opens are read by meta.errorCode and
each code has its own copy.
Eight files mocked `@/lib/feature-flags` with `getAssistantFlags` alone.
The capability loader reads `loadAssistantSwitches`, so wherever it ran
it threw and every capability resolved to `check_failed`. A shared
factory now points both readers at the function the test drives.
Only the config-change chain reached the loader. Its fixture means AI
switched off by the operator; it now supplies the loader's other inputs
and asserts `operator_disabled` on the digest, so a starved loader turns
it red instead of passing on the failure path.
Nine places assigned window.location.href directly: OAuth connects and
sign-in hand-offs that land on API routes answering with a redirect, and
the reload after account deletion. The newer Next lint rule reads each as
an in-app page change. They now go through one helper that says why a
document load is the right tool there.
The Coach correlations tool, the per-metric card and the period
narrative scanned every channel, so a switched-off module still fed
their statistics and the text a model was grounded on. The matrix's
modules option is now required, each caller resolves the record's map,
and the Coach skips the lab pass with labs off. A registry guard holds
every caller to resolving the map itself.
The mask's doc comment now states what it does to the false discovery
correction: fewer channels is a smaller family, so q-values of the
remaining pairs move and a borderline pair can cross the threshold.
Serving stored text and starting model work are separate questions. A
delegate may not start model work on somebody else's record, but reads
the owner's stored briefing, period narrative, status notes, workout
notes and Coach unread signal exactly when the owner would: those
readers now resolve the record's own capability state, not the
viewer's. Generation, warms and enqueues still ask as the viewer, so a
delegate's read of a stale narrative queues nothing. A hidden
model-written narrative is no longer overwritten by a deterministic
warm while AI is unavailable.
The sharing front door reads the digest's briefing lead again, and an
integration case pins both halves.
Refs #975
The push refused outright while the `insights` module was off, which
now means only "AI analysis". The digest it reads already drops the
model-written lead and top signal in that case, so the push goes out
with the deterministic line (the score, or the day's items) and a
digest with nothing to say is still skipped. The `module-off` outcome
and its counter are gone.
Refs #975
Brings in the jobs, stored-text and consent package and the extraction
routes. Conflicts resolved:
- The digest keeps its capability-driven reaction line and check-in,
now through the shared `aiCapabilityForRecord`; the local copy of it
is removed.
- The route inventories classify the extraction routes (documents,
labs OCR, medication text, fenced chat and attachments) as actions;
a capability call counts as the gate for the module inventory.
- With every caller moved, `requireAssistantSurface` and
`AssistantDisabledError` are deleted. The `assistant.disabled.*`
codes stay on the capability refusal, and the guard now asserts the
retired gate is gone.
- `aiCapabilityToServe` answers whether stored model text may be shown
from the record's own state, apart from who is reading.
Refs #975
Marc-André Bombeck