With few packs a request per pack is fewer requests than 256 listings. On
FTP every listing takes a data connection; the FTP server of the integration
tests has ten passive ports, and 256 listings in a row slowed it down until a
read failed ("Not found: snapshots/...").
A Hetzner Storage Box drops an SFTP connection that carried a lot of data
and then stayed quiet for a few minutes; files_external keeps using the dead
client, on which every file looks missing. StorageBackend now checks the
connection after a minute without requests, and replaces a connection that
is dead after a failed request (once per request; put, move and delete are
not repeated). SftpReadStream closes a connection it left in an unknown state.
The verify cache refresh took more than 6 minutes on PostgreSQL for 27,547
packs, one autocommitted insert per pack: putMany()/replaceAll() write in one
transaction. After a verify the cache holds the index files as they are, so a
blob in two packs stays cached for both.
Review notes on the open pack stream: it stayed open while a backup run went on to upload
(e.g. the snapshot after reading the file list back), which SFTP cannot do on one connection.
Reading moves to PackReader, and Repository reaches the location through PackReaderGuard,
which closes the stream before every other request. A pack stream that cannot seek is read up
to the offset. Deep verify no longer reports an \Error during decoding as damage.
Restores and deep verifies requested every blob separately: a new file handle and a seek on
the location each time. They read the blobs of a snapshot in the order they were written, so
the stream of the pack read last now stays open and is read on; a jump backwards, to another
pack, or more than 8 MiB ahead opens the pack again at the offset, as before. On a Hetzner
Storage Box over SMB, restoring 86 MB in 202 files went from 0.78 MB/s to 4.2 MB/s.
Opening a repository lists index/, which holds one file per pack.
StorageBackend::list() checked every entry with is_dir(), and on a
remote storage each check is a round trip. Over SFTP to a Hetzner Storage
Box (33 ms) with a 60 GiB repository, occ backup:run (30 s steps) spent
nearly every step on these requests and uploaded 45 KiB/s; with this
change it uploads 5.8 MB/s. The cost grows with the number of packs.
index/, snapshots/, catalog/ and trash/info/ only ever hold objects:
list them with the new IBackend::listFiles(), one listing without
per-entry checks.
A database dump that fills a whole pack (32 MiB) uploads that pack while
the dump transaction is still open, and the pack's index cache row was
inserted on the same connection. PostgreSQL refuses that (25006, read-only
transaction) and the run failed; on other databases the row was rolled
back with the dump transaction. The index now holds those cache writes
back until the transaction is over.
Catch \Exception (storage adapters throw their own exception types) instead
of \Throwable, so an \Error such as a TypeError is not retried and reported
as an unreadable blob.
An unreadable blob may still be a truncated pack, so don't rule out damage
in the verify message; README: Hetzner needs WebDAV/SMB and external
reachability enabled for the sub-account.
verifyBlob() caught every exception from loading a blob and counted it as
a failed (corrupted or tampered) blob, so a transient network error during
a deep verify over WebDAV looked like data corruption. Read the range
separately, retry it, and report a read that keeps failing as unreadable
with its error; only a blob that was read but does not decrypt or match
its id counts as failed.
forget() and untrash() uploaded a remote read stream directly, which the
new size check could not judge on SFTP, so cleanup failed there. They now
copy through a local buffer, and the size check only trusts local streams
(php://temp, php://memory, plain files).
The disaster recovery test calls backup:restore:full with --no-interaction:
docker exec without a terminal still counts as interactive, so the new
confirmation question read an empty answer and cancelled.
Storage and repository:
- Listing skipped every name containing ".part-", directories too, so the
exports of a user like "john.part-time" were invisible to prune on S3,
SFTP, FTP and SMB and their data was deleted. Only real temporary uploads
("<name>.part-<8 hex>") are skipped now.
- prune refuses to run while a recorded per-user export's manifest is
missing, instead of freeing that export's data.
- forget() and untrash() copy first, switch the catalog, and delete the
originals last, so a crash can no longer leave the catalog pointing at a
missing snapshot or database dump.
- Writes are checked against the expected size, so a short upload is an
error instead of a truncated pack; raw restores check rename() and touch().
- A catalog write refuses to overwrite a generation another process just
wrote, and the catalog lock lasts 300 s, so opening a large repository
inside it no longer lets the lease run out.
Backup runs and scheduling:
- A file that keeps changing is read again at most three times; then the
last read is kept and reported, instead of restarting forever and
blocking cleanup.
- Unreadable files and folders are recorded, logged, shown by occ
backup:run, stored in the snapshot and notified to admins, instead of
being skipped silently.
- After a failed scheduled backup, at most three attempts a day, an hour
apart, instead of a new run and notification every five minutes.
- The job no longer spins on a run whose lease another process holds, and
the weekly cleanup catches up after a Sunday spent backing up.
- Notifications for the per-user export retention events.
Full restore, raw directory restore, per-user export and the safety export
held a fixed-TTL lease on the location that nothing refreshed. A restore or
export running longer than that TTL let a cleanup take over and delete data
still in use. They now refresh the lease while they work (per blob, per row
batch, per written chunk), throttled to once a minute.
Also:
- backup:prune still applies snapshot retention when user-export retention
fails for a location.
- The recovery kit carries the deletion delay, so a disaster recovery keeps
the original protection instead of falling back to 7 days. Older kits
still import with the default.
- CI checkouts no longer persist the token in the git config.
- recoveryKit() has its real return type; the stale baseline entry is gone.
A full backup opened its repository at the start of a step and wrote the
catalog at the end of it, without the catalog lock. A per-user export or
restore that wrote the catalog in between could lose its entry, or the
snapshot could lose its own.
The final step now only stores the snapshot object and moves the run to a
new 'commit' phase. BackupService then records the snapshot under the
catalog lock, from a repository opened inside that lock; if the lock is
busy, the next step retries the commit. Repository::writeSnapshot() keeps
its old behaviour as putSnapshot() plus recordSnapshot().
Catalog: forgetting a user export or purging trash now writes the catalog
before deleting the objects, so a crash leaves at most an unreferenced file
instead of a catalog entry pointing at nothing. Per-user backup and the
safety export reopen the repository inside the catalog lock, so they build
on the latest catalog generation.
Commands: backup:user:backup and backup:user:restore resolve the location
before asking to install user_migration, and a successful install always
stops with a rerun instruction. backup:retention rejects a malformed
--user-last before changing any policy. backup:key:import fails clearly
when the passphrase file cannot be read.
backup:verify accepts a user export path (users/<uid>/<export>): every blob its manifest references
is checked the same way as for a snapshot, shallow or --deep.
The SMB backend is refused at backup:target:add when the native smbclient PECL extension is not
loaded, instead of registering and then failing the first backup with "Seek failed": with only the
smbclient binary files_external falls back to a shell stream that cannot seek.
The per-user integration job now runs on Nextcloud 32-35, not just 35: it calls user_migration's
internal UserMigrationService, whose API may change between versions.
Repository::verify(): every blob a snapshot's files and database dump reference must be in a
freshly re-downloaded, re-authenticated index (never trusted from the local cache, same reasoning
as prune()) and have its pack still present on the location -- shallow, cheap, no blob content
read, catches a deleted pack, a tampered/rolled-back index, or a database manifest that went
missing (reported as a failure, not silently skipped). With $deep, each blob is also downloaded
and decrypted, same as loadBlob() already does on every real read (AEAD authentication tag,
content hash against its blob id), done here deliberately for everything in the snapshot rather
than waiting for it to come up during an actual restore. A blob referenced by more than one file
is only checked once; pack presence is cached per pack, not re-checked per blob.
occ backup:verify <target> <snapshot> [--deep], meant to run periodically (cron, a sample of
recent snapshots), per PLAN.md's "Verificatie: periodiek checksums controleren, steekproef-restore".
The lease covering the audit is refreshed roughly every half of its TTL while blobs are checked,
rather than relying on a single long-lived lease like the other new phase-2 commands do.
Verified against a real S3-backed 1.4 GiB / 603-blob snapshot (shallow: instant; --deep: ~28s,
fully decrypted) in addition to the new unit tests and the roundtrip.sh addition (shallow and deep
on a snapshot with real, varied file sizes, now covered on every CI matrix combination).
Local review (Codex) on the first pass found three real issues, all fixed: a missing database
manifest was silently skipped instead of reported; shallow verification trusted the (possibly
stale or unauthenticated) local index cache instead of re-checking the authenticated index on the
location; and the lease had no refresh, so a --deep audit of a very large snapshot could outrun
its TTL and run unprotected against a concurrent prune.
New commands, per PLAN.md 3.7:
- occ backup:key:import: adopt the master key from a recovery kit on a fresh installation
(instead of occ backup:key:init generating a new one), so occ backup:target:add opens the
existing repository on the location rather than creating an empty one.
- occ backup:restore:full: replace the database (consistent, one transaction), restore the data
directory in place, and merge config.php -- refuses up front and lists what's missing when the
Nextcloud major version or any app enabled at backup time doesn't match here (their tables must
exist before the database restore can use them), rather than guessing or partially applying.
Repository::restore() gained an optional $stripPrefix (write "data/<uid>/..." straight into the
real data directory, not a "data/" subdirectory of it) -- restoring "data/" and "config/" this way
needs no temporary copy of what can be the largest part of an instance.
occ backup:target:add now backfills the local snapshot cache (ngb_snapshots, what occ backup:list
actually reads) when it adopts an existing repository: without this, a freshly connected location
showed no restore points at all, even though the repository itself had them. One unreadable
snapshot is skipped rather than aborting the backfill or the target add itself.
Three self-referential problems found and fixed by testing this against two real, separate
Nextcloud instances sharing only the backup location (tests/integration/full-restore-roundtrip.sh
covers this; the per-process nature of these bugs meant they were invisible from inside a single
occ invocation, including the one used to write the fixes, which is exactly why a second, fresh
process re-running occ backup:status etc. was necessary to catch them):
- The full database dump includes ng_backup's own tables and appconfig rows; restoring them
verbatim would overwrite the key/target that is actively performing the restore, on THIS
installation, with the backed-up instance's. Both are now preserved: ng_backup's own tables
(and, on PostgreSQL, their sequences -- excluding a table does not exclude its sequence) are
excluded from what gets restored, and its appconfig rows are captured before the database
restore and written back after, in their own transaction.
- config.php's 'secret' is merged in from the backup (needed so already-encrypted content in the
restored database, e.g. share tokens, stays decryptable), but Target::options, the master key,
the wrapped key and the recovery kit code are all encrypted with that same instance secret
(OCP's ICrypto, or Nextcloud AppConfig's own "sensitive" flag underneath), not with ng_backup's
own key. Changing it without re-encrypting them first left all four permanently undecryptable
the moment any later occ command ran. Fixed with two helpers
(TargetService::reencryptOptionsAround(), KeyService::reencryptForSecretRotation()) that decrypt
before and re-encrypt after whatever callback actually changes the secret; config.php is applied
dead last in restoreFull(), after every other repository access. Not crash-safe (a process kill
mid-rotation would leave a key missing, not just stale) -- disaster recovery is run once by
hand, right after occ maintenance:install, a window too narrow yet to justify a two-phase write.
- A subtler variant of the same bug: Nextcloud's own AppConfig additionally encrypts a "sensitive"
value itself, tied to the same secret, underneath our own encryption. Overwriting an existing
sensitive key makes it read the old value first to compare, which then also fails against the
new secret. Deleting the key before setting the new value avoids that read entirely.
Roadmap (PLAN.md 9): phase 3 (in-place full restore, user self-service) moved out to v1.3.0; phase
2 continues toward v1.2.0 with FTP/SMB, a verification/checksum-audit command, user-export
retention, and deciding whether per-user export joins the automatic schedule.
Wires the existing RepositoryExportDestination/RepositoryImportSource (see
lib/UserMigration/) into user_migration's own UserMigrationService, resolved on demand
through a new UserMigrationServiceLocator since the app is an optional dependency
(PLAN.md 3.8) -- without it, only this feature is unavailable, the full backup still
works.
New: UserBackupService::backupUser() (occ backup:backup:user) streams a user_migration
export straight into the repository, under the same shared location lock as a regular
backup. UserRestoreService::restoreUser() (occ backup:restore:user) never imports over a
live account: --mode=replace safety-exports the current state first, then deletes and
recreates the account from the backup under the same uid; --mode=as-backup restores into
a new "<uid>-bak" account instead, since Nextcloud cannot rename a uid.
The catalog now also tracks user export manifests (Repository::recordUserExport/
listUserExports/forgetUserExport/isRecordedUserExport), rollback-protected like a
snapshot: listUserExports() verifies each one is still present on the location, and
restore only ever reads a manifest the verified catalog actually recorded, never an
arbitrary caller-supplied path. prune()'s existing raw users/ scan for blob retention is
unchanged. A new short exclusive "catalog" lock (PruneService::catalogLockKey())
serializes the catalog write itself, so two concurrent per-user backups (or one
alongside a full backup) can no longer race to write the same next generation and
silently drop one of them -- the broader per-target shared lock only ever protected
against a concurrent cleanup, not against other shared holders.
Two real bugs found by testing this against a live Nextcloud instance, not by unit
tests (RepositoryExportDestination/RepositoryImportSource implement OCP interfaces that
plain PHPUnit can't load, see tests/Unit/Repository/UserExportTest.php's docblock; the
new tests/integration/user-roundtrip.sh covers the real round trip instead):
- RepositoryImportSource::copyToFolder() double-closed the stream it handed to
putContent()/newFile() (both already close it themselves), which throws a TypeError on
PHP 8 and broke every restored file with content.
- A replace-mode restore failed on NotPermittedException: IRootFolder::getUserFolder()
caches a Folder per uid for the process lifetime, so after delete+recreate of the same
uid the import still got the deleted account's (now storage-less) folder. Fixed by
dispatching FilesystemTornDownEvent, which Root already listens for to clear exactly
that cache -- no internal OC\ classes needed.
Local review (CodeRabbit + Codex) against this diff found three more real issues, all
fixed: the catalog write race above, a failed as-backup import leaving a broken stub
account behind (now deleted on failure), and an export stream leak on a read/backend
failure in RepositoryExportDestination::copyFolder() (now closed in a finally).
Deliberately left out: wiring per-user export into the automatic backup schedule
(on-demand via occ only, for now) and retention/pruning of old user exports -- both open
design questions, not yet decided. The lease covering a single export/import still has a
fixed TTL (3600s) with no mid-call refresh, since UserMigrationService exposes no
progress; a resumable, step-based per-user backup (like the main BackupService run)
would remove that cap but is a bigger change than this pass.
RepositoryImportSource::copyToFolder() called putContent() on the generic
return type of Node::get(), which doesn't declare that method -- only
File does. Existing-file overwrite during a per-user restore would have
fatally errored at runtime; added an instanceof check and a resource leak
fix (finally-close) while there. Also guarded getFileAsStream() against
fopen() returning false.
Everything else Psalm (strict level) currently flags is mechanical
(missing `final` on leaf classes -- applied everywhere safe) or pre-existing
array-shape/type gaps, captured in psalm-baseline.xml so CI fails only on
new issues from here on, not the whole backlog at once.
Added tests/integration/targets-matrix.sh, which runs the full backup/restore
round trip through files_external's AmazonS3 and SFTP backend classes (S3
via a disposable s3mock container, SFTP via atmoz/sftp), since the existing
matrix only ever exercised the local backend.
This caught a real bug: moving a snapshot to the trash read it from the
location and streamed it straight back into a write on the same connection,
which SFTP can't do reliably (phpseclib fails to interleave a read and a
write on one connection). Backends now expose a native move() and the
id-bound snapshot object is moved with that instead of a read+write copy;
the path-bound database dump still goes through decrypt/re-encrypt, since
its AEAD authentication is tied to its path.
S3 has no bucket quota, so NG Backup enforces one: before a pack is
uploaded, the stored size plus the pack must stay under the location's
limit, otherwise the backup stops with an error (and administrators are
notified through the failed run). Repacking during prune is exempt, as it
is how space is freed. occ backup:target:limit NAME [SIZE|none].
Also ignore credential files in the repository.
Every change to the set of snapshots or trash entries writes an
encrypted catalog generation bound to the repository and its position,
with the hash of the previous generation. The server anchors the newest
generation it has seen in its app config; on open, the location must
show at least that generation and chain back to it, and every snapshot
listed must exist. A rolled-back or tampered location is refused (the
run fails and administrators are notified). Opening with the passphrase
on a new server accepts the newest valid chain.
occ backup:target:trust-current accepts a legitimately older state
(e.g. restored from the provider's snapshot) after typing the location
name; audited and announced to all administrators.
Tests: chain advances, rollback, hidden snapshot, swapped generation,
disaster recovery without anchor. Tested on NC34 by deleting the newest
catalog generation on the SFTP location.
- restore runs from the web are created under the location's exclusive
lock, like backup runs
- raw restore writes to an unpredictable temporary name opened with
exclusive create (never follows a planted symlink) and validates the
path prefix; the restore service rejects non-canonical paths instead
of trimming them (occ normalises user input before calling it)
- restore jobs report progress every minute, so a long restore is not
taken for a dead one
- README: Object Lock turns deletions into delete markers
- one locking protocol per location: backup steps and restores hold a
shared lock, prune and untrash an exclusive one and refuse while any
run (backup, restore, export) exists; starting a backup checks and
creates the run under the exclusive lock, closing the race with prune
- snapshot entries are validated when read (relative path without '.',
'..', NUL, backslash or control characters); raw restore also checks
that the resolved directory stays below the target; the restore API
and service accept canonical paths only
- recovery kit confirmation requires a recent password confirmation and
the confirmation code printed in the current kit file (new code per
kit version), proving the kit was actually downloaded
- prune decides on the authenticated index files on the location and
refreshes the local index cache from them
- restore job marks the run failed on any error; restore runs without
progress for 6 hours are failed so they cannot block cleanup forever
- tests for safe paths and a malicious snapshot path
- removed snapshots go to a trash on the location and their data is only
freed after the deletion delay, which is chosen when the key is created
(default 7 days, 1-365) and cannot be changed afterwards; any admin can
bring them back (occ backup:trash, backup:trash:restore, web UI)
- up to three passphrase slots open the same master key (like LUKS), so
one person leaving or dying does not lock the company out; slot changes
are written to every location, announced to all admins and require a
new recovery kit confirmation, while backups keep running
- occ backup:key:slots, backup:key:slot:add|replace|remove; key init takes
--label and --delete-delay-days
- web UI: deletion delay and first slot at key creation, slot management,
trash per location with restore; all behind password confirmation
- README: key, slots (safe / CTO / head of IT), and protection against
ransomware: object lock in compliance mode (e.g. 90 days), credentials
without delete rights, provider snapshots, off-site copy
- RetentionPolicy (last/daily/weekly/monthly) and PruneService: forget
snapshots the policy drops, delete packs nothing uses and repack mostly
unused ones by copying encrypted blobs as-is; new packs are indexed
before old ones are removed
- a location lock keeps prune and backups apart; append-only locations
are never pruned (cleanup via the location's lifecycle rules)
- weekly prune from the background job; occ backup:prune [--dry-run],
backup:retention, backup:target:append-only
- tests: forget+prune frees space and the remaining snapshot restores
identically, repack of mixed packs, dry-run preview, retention buckets
Opening a repository lists index/ once and downloads only index files
that are not cached; packs written by this server are cached on write,
stale cache rows are dropped. Each resumable step therefore no longer
downloads every index file from the location.
TreeWalker walks one or more roots in a stable segment-wise order with
exclusions and resumes after a cursor; ParentCursor merge-joins the
previous snapshot in the same order. BackupRun no longer keeps a file
list or the parent's entries in memory or in its state.
- migration: ngb_targets, ngb_runs, ngb_snapshots
- KeyService: master key stored encrypted with the instance secret
(unattended backups), passphrase-wrapped copy in every repository and
in the recovery kit, recorded admin confirmation per key fingerprint
- TargetService: locations built from files_external backends without a
mount; files_external is loaded through IAppManager::loadApp() and its
enabled state is left alone; credentials encrypted with ICrypto;
existing repositories are adopted only if made with this key, and a
repository cannot be registered twice
- Repository: initWithKey/openWithKey with an encrypted key check
- occ: backup:key:init|kit|confirm, backup:target:backends|add|list|remove
Dev tooling via composer (phpunit, psalm, nextcloud/ocp); vendor/ is
mounted from outside the repository by a helper script. 28 unit tests
ported from the phase-0 spikes (round trips, corruption, wrong keys,
dedup, incremental runs, partial restore, tampering, object binding,
time-boxed steps, file changed between steps).
- detect files changing during reading or between resumable steps and
re-read them (size/mtime/ctime/inode), spike F overwrites mid-run
- drop pending blob ids when a pack or index upload fails
- dump and restore PostgreSQL sequence state exactly, bound parameters,
validated catalog names; spike C checks all sequences exactly
- keep invalid UTF-8 byte-exact (tagged base64)
- refuse large tables without a primary key
- StorageBackend: check write/flush/close results, validate base path
- user_migration import/export: root paths, intermediate folders,
partial writes, read errors
- spikes: never touch an existing account, own fixtures, 512M note
BackupRun is a resumable state machine: each step works until a deadline,
uploads its pack and tree segment, and returns the state; large files are
resumed at the next blob. Snapshot is written last.
Spike F: 200 MiB file resumed across steps in separate processes; with 6
random kill -9 the run completes and restores bit-identically.
spikes/RESULTS.md summarises phase 0 and the findings for the plan.
RepositoryExportDestination / RepositoryImportSource implement the
public OCP\UserMigration interfaces on top of the repository, so
user_migration's migrators (calendar, contacts, trashbin, account,
files) export straight into the encrypted repository.
Spike E (NC34): export 0.8s, unchanged second export stores nothing;
after deleting the user, import restores account, files (bit-identical),
calendar event and contact. Findings: calendars/address books come back
as 'migrated-*' with new object URIs; importing over an existing user
duplicates contacts and adds a migrated birthday calendar; creating the
user on import needs >128 MB (password_policy word list).
ExternalStorageFactory builds storages with files_external's own
backends and auth mechanisms from an unsaved StorageConfig (as its
ConfigAdapter does), and StorageBackend adapts any IStorage to IBackend
(writeStream, seekable range reads, temp name + rename except on S3).
IBackend::put() now owns and always closes the stream, because
files_external's S3 writeStream closes it.
Spike D (NC34): S3 (SeaweedFS) and SFTP: 120 MiB backup ~6s, unchanged
run uploads nothing, restore bit-identical, no files_external mount
created, no temp files (max 12 KiB in /tmp). Works with files_external
disabled when its classes are loaded. Backup peak memory 85-89 MiB on
these targets (to investigate).
DbDumper reads every prefixed table through the Nextcloud connection in
one REPEATABLE READ (read-only) transaction, with keyset pagination on
the primary key, and streams each table as JSON lines into its own blob
stream (BlobWriter), so unchanged tables are deduplicated. Restore
replaces table contents in one transaction and resets PostgreSQL
sequences.
Spike C (NC34, PostgreSQL, standard image without pg_dump/pgsql ext):
300k-row table with bytea: dump 6.3s / 50 MiB peak, 94 MiB JSON -> 20 MiB
stored; unchanged second dump stores nothing; restore makes every table
checksum-identical and all 110 sequences continue above the maximum.
Content-addressed repository on an IBackend (local implementation):
4 MiB blobs, deflate when it helps, each blob encrypted with its id as
associated data, ~32 MiB packs with an encrypted header, one index file
per pack, snapshots as encrypted JSON lines written last.
Spike B: 390 MiB tree, peak 54 MiB; unchanged run reads nothing; a 1 MiB
change in a 300 MiB file uploads one blob; full and partial restore are
bit-identical; tampering is detected.
dillard