A restore of an account with 1.9 million readings, most of them deleted
by the nightly consolidation and still awaiting the purge, was cancelled
in its clearing step: `DELETE FROM measurements` ran past the 60 s
statement timeout every pooled connection carries, and the whole restore
rolled back with a message that named no cause (#1031).
Most of that statement was not the delete. Two columns point into
`measurements` with ON DELETE SET NULL and had no index, so Postgres
scanned `personal_records` and `ecg_recordings` once per deleted reading:
38 of 40 s on a seeded account with 300 personal records. Migration 0362
indexes both, which also speeds up the tombstone purge and account
deletion.
The restore transaction now sets its own statement and idle-in-transaction
limits to the transaction budget with SET LOCAL, and the budget counts
the rows the account holds as well as the rows the file carries. A
rolled-back restore names what the database refused it over (timeout,
lock, constraint, storage, connection, other) on the job row, the
console, the audit row and the wide event.
Exporting the whole record, making a clinician share link, minting a token,
and the admin backup download, upload and restore, data wipe and password
reset ran on a live session alone. The plaintext full backup did, while the
encrypted one beside it asked for a second factor. Whoever held a stolen
session could take the record, or leave behind a credential that outlives
the session.
These now go through one gate, requireRecentProof (assertRecentCookieProof
for the routes that resolve their caller themselves). A browser session
passes after a sign-in or re-proof within five minutes: a second factor on
an account that has one, otherwise the password or a passkey. The refusal is
401 auth.reproof.required with meta.methods, and the web opens the re-proof
dialog, posts the proof to the new POST /api/auth/reproof and retries. A
password proof stamps the new Session.reproofAt (migration 0359), never
mfaVerifiedAt, so it cannot stand in for a second factor anywhere.
On Bearer, the full backup asks for an X-Step-Up elevation; the encrypted
export asks for a second-factor elevation on an account that has one. Share
links and connector tokens keep taking the token alone, because the shipped
app calls them without an elevation and reads a 401 there as a dead
session; the guard test names them.
Key rotation keeps its stricter second-factor gate.
A backup carries every encrypted column as the stored ciphertext, and a
restore writes it back verbatim. Key rotation re-seals the stored copy's
envelope but cannot reach inside it, so a copy taken before a rotation
still needs the retired key. Without it the restore went through and
left rows no reader can open.
The restore, the restore preview, the admin upload and the monthly
restore drill now collect the key ids of every inner ciphertext value,
hold them against ENCRYPTION_KEYS and decrypt one value per key. A
missing key, or an id configured with different key material, is
refused before anything is deleted, with the key id named. The restore
status and the preview say which key to put back.
The restore also checks, inside its transaction, that it added no
reference from the account's rows to rows another account owns. The
check is derived from the schema's foreign keys rather than written per
section, so the tag and category upserts keyed on a global key are
covered along with every link table. The illness parent update is
scoped to the owner as well.
The off-host listing now follows every page: the first page held only
the oldest dates, so the drill could read an old object and report the
chain as stale.
Key rotation re-sealed a single-value backup in place, which holds the
stored string, its bytes, the plaintext and the new ciphertext at once; a
copy can be over a hundred megabytes and the in-app rotation walked 25 of
them per page, enough to run a 1 GB container out of memory. Rotation now
converts each such copy into pieces under the active key instead, one row
at a time. The ~hlgcm1. stream of v1.38.6 to v1.39.1 is read a megabyte
at a time with its tag checked at the end, inside one transaction, so a
copy that fails the check is left untouched; the copy keeps its date. The
scan reads only the first characters of each value. Measured on a copy of
a 2.6 million-reading account (126 MB stored): converted in 25 s with the
heap flat at 57 MB under a 200 MB limit, where re-sealing in place ran out
of heap.
The weekly and manual backup leave an account alone while a restore of
its current copy is queued or running, and count it as skipped, not
failed. A reader whose copy is replaced mid-read answers 409
backup_changed, "replaced by a newer one while it was being read", rather
than the undecryptable error; a restore rolls back with the same code.
Pieces are sealed with a fixed associated-data label, so a value the same
key sealed for another purpose never opens as a piece; rotation re-seals
under the same label. A row holding both a single value and pieces, which
only an older release writing after a downgrade produces, is refused with
a message saying so, and the next weekly run replaces it. Rotation treats a
row deleted between its read and its write as gone rather than an error.
A stored backup was one value in data_backups.data. v1.39.1 wrote it a
piece at a time but still joined the pieces into that one value, and one
value has to pass through the app whole, so a copy was capped at a fifth of
the heap: 105 MB in the default 1 GB container. An account with 1.75
million readings is past that and could not be backed up at all (#1031).
A copy is now kept as ordered pieces of about a megabyte in the new
data_backup_chunks table (migration 0352). Each piece is sealed on its own
with AES-256-GCM, and its sealed header names the copy (a random stream id
kept on the row), its position and whether it is the last. A reader checks
every piece before it releases a byte, so a piece that is missing, moved,
altered, taken from another copy, or a copy cut short is refused with
backup.payload.undecryptable before a restore deletes anything, and the
second read during the restore checks each piece again.
The write is one transaction, so the previous copy stays whole and readable
until the new one commits. Pieces are written through the model API: a raw
query with a large parameter is kept alive by Prisma's query plan cache,
which left 160 MB of heap behind per hundred one-megabyte inserts.
Restore, preview, download, the restore job and the admin list read both
forms; copies written before this change keep their single value and still
restore. The heap-derived limit is gone. The one limit left is on storage,
BACKUP_MAX_STORED_MB (default 2048), and its message names the setting. Key
rotation covers the pieces, the wipe and account deletion take them with
their backup, and the backup plan lists the table as not carried.
Measured on an account of 2.6 million readings in a 1 GB container with web
and worker in one process: backup 87 s, 97 MiB in 96 pieces, preview 20 s,
download 25 s, restore 264 s with every reading back exactly, peak 567 MiB,
no restart.
A restore commits its transaction and then rebuilds the chart tiers. A
worker that stopped during that rebuild left the job running with a stale
heartbeat, and the boot sweep and the admission check treated it like any
interrupted job: the sweep queued it again, so the second run deleted and
re-wrote the account and lost whatever was written in between, and a second
interruption closed it with "rolled back, nothing was changed", which was
false.
The job now records `committedAt` (added to migration 0347, which has not
shipped) and awaits that write before the rebuild starts. The sweep and the
admission check close a committed job as `failed_after_commit`, never queue
it again, and evict the account's caches. The rebuild itself is left to the
nightly and boot passes that already rebuild missing tiers.
The module comment now says what an interrupted restore does (restarted
once, then failed) and what happens to other writes while a restore runs:
new rows are written at once, updates to rows the restore already deleted
wait on their lock until it commits or rolls back.
Restoring a large account was one request: 94 s for 1.25 million readings.
A reverse proxy that cuts a request at 60 s (nginx's default) showed the
operator an error for a restore that went on to finish on the server.
The restore request now checks the typed confirmation, that the backup
exists and that its stored copy opens, admits one job for the account and
answers 202 with its id. The `backup-restore` queue runs the restore and
records its phase, counts and outcome on a `backup_restore_jobs` row
(migration 0347), which `GET /api/admin/backups/restores` and
`GET /api/admin/backups/restores/{id}` read back.
What a restore checks and writes is unchanged. The body moved from the
route into `restoreBackup` (src/lib/export/restore-backup.ts): every
refusal still happens before the first delete and the replacement is still
one transaction, so a failure part-way leaves the account as it was. While
it runs the account keeps showing its current data, and other writes are
neither refused nor held, as before.
- One restore per account: a partial unique index admits one queued or
running job, and a second request gets 409 `backup.restore.active` naming
the running job. A per-account advisory lock inside the transaction holds
even if a worker thought gone is still writing.
- Time budget: the job's expiry is two hours and a file whose transaction
limit would run past three quarters of it is refused before anything is
deleted (`jobDeadline`, beside `jobBudget`).
- Resuming: the job writes a heartbeat every two seconds. A boot sweep
queues a job whose worker stopped again from the start, at most twice,
then fails it as interrupted. The admission check releases an abandoned
job so it never holds the account's slot.
- The staged file is the stored `DataBackup` row, already encrypted at
rest. It stays whatever the outcome, so a retry needs no new upload. The
job records a digest of the copy and refuses one that was replaced while
it waited.
- Failures carry a stable code and a sentence that says what to do.
- The console follows the job: phase and counts while it runs, the outcome
after, and after a reload it finds the job again from the server. A poll
that fails on the network is never shown as a failed restore. A web
process separate from the worker evicts the account's caches when it
first reports the restore finished. The success line goes through the
outcome module, so an empty file is not drawn as a tick.
The model is registered in the backup plan (not carried) and the wipe plan
(wiped with the account); the queue in the maintenance registrar, the
subsystem surface and the job-fact allowlist; the new code in the error
catalogue; the routes in the OpenAPI contract. The tests that pin restore
semantics now drive the real route and the real job through one helper, and
`backup-restore-job.test.ts` pins the asynchronous contract itself.
A `DataBackup` row that fails to decrypt answered 500 on the restore, the
download and the preview. It is not a server fault: either the key that wrote
the copy is no longer in `ENCRYPTION_KEYS` — a rotation that dropped the legacy
entry before the script reported zero rows, the case the rotation runbook warns
about — or the stored bytes are not the ones that were written. Every other
bad-input arm on those routes already answers 4xx, and the 500 also paged the
error reporter as if the process had broken.
All three now answer 422 with `meta.errorCode = backup.payload.undecryptable`
and one shared sentence, so the three routes cannot drift into three wordings
for one condition. The download's second 500 stays: a payload this instance
wrote and cannot parse really is a fault here.
The two routes that open a stored copy move from the unpublished list into the
contract, because the refusal is now a promise an operator can look up rather
than whatever the handler happened to answer. The restore publishes its
`Idempotency-Key` header with them.
Marc-André Bombeck