Precondition for the merge-join in backend_consistency (step 6 /
option A of docs/plan/derived-blobs.md), landed separately because it
is independently useful and carries the risk.
Two contract changes on BlobStorageBackend::list_blob_hashes:
1. Entries MUST be in ascending hash order. Every shipped backend
already did this — local sorts within each shard and walks 00..ff,
and since the shard IS the hash prefix that is globally sorted; S3
and Azure list lexicographically by key and blobs/<xx>/<hash> sorts
identically to <hash>. It was accidental, and a future backend
enumerating in any other order would have silently made the
merge-join emit bogus blob_missing_from_backend findings at
data_loss severity.
2. The cursor is the last hash returned, not an opaque backend token.
This is what lets a caller resume from a checkpoint it already
holds — the merge-join keeps one cursor for both the DB walk and
the backend walk instead of a compound one, which in turn means
blobs_consistency's existing cursor format survives and no paused
run is stranded.
Local already derived its position from a hash; it now emits the bare
hash instead of "<shard>/<hash>", and still accepts both legacy forms
so a run paused across this deploy resumes. The bare-shard form works
through the same path unchanged, since "3f" sorts before every 64-char
hash beginning "3f".
S3 moves from continuation_token to StartAfter, which supports this
natively. One non-obvious case handled: a page can contain only
non-canonical keys (.tmp spool files, .corrupt sidecars), which are
filtered into `unknowns`, leaving `blobs` empty — a naive
blobs.last() would return no cursor and silently end enumeration while
is_truncated said otherwise, making an audit job under-report. It now
falls back to the last key seen; StartAfter is a string comparison, so
a non-hash resume point is fine. "Cursor is a hash" constrains what
callers may synthesise, not what backends may return.
Azure is unaffected — it does not implement list_blob_hashes (TODO,
inherits the NotSupported default).
Adds the first test for enumeration at all: ordering across shards with
deliberately out-of-order inserts, complete paged traversal, and
resume from a caller-synthesised cursor.
NOT verified against real S3 — no bucket available here. The local path
is covered by the new test; the StartAfter change is reasoned from the
API contract and needs exercising against a real bucket before it is
relied on.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three benchmark-gated optimizations from the ROUND25 backlog (benches/ROUND26.md),
each with a BEFORE/AFTER gate that rolls back if AFTER does not beat BEFORE:
- P1 drive_pg_repository policy reads: decode d.policies through
sqlx::types::Json<DrivePolicies> (one from_slice over the raw JSONB bytes)
instead of a throwaway serde_json::Value DOM + DrivePolicies::from_value —
6 -> 0 allocs/read, 2.77x wall. A shared policies_from_row helper preserves
the lenient unwrap_or_default fallback (malformed bag -> all-false).
- D1 CachedBlobBackend: pre-create the 256 {00..ff} shard dirs at initialize()
(mirroring LocalBlobBackend, reusing HEX_PREFIXES) and drop the redundant
per-write create_dir_all on already-existing shards — ~45us + a blocking-pool
dispatch removed per cache write on cached-remote deployments.
- G1 delta-upload have/need hash sets (distinct_hashes, authorize_chunk_download):
SipHash -> foldhash::quality::RandomState — a fast hasher that stays
DoS-resistant via a per-instance random seed, the required property for the
attacker-controlled 64-hex client hashes — 2.37x wall on a 40k-hash
negotiation. foldhash was already in the lockfile transitively (hashbrown).
Tested and REVERTED (kept as-is): moving the moka eviction unlink off the reactor
via spawn_blocking. The benchmark refuted it — on the local cache dir the
spawn_blocking dispatch (~20us) costs more than the inline unlink (~7us) it would
replace. See ROUND26.md §D2.
Adds bench_round26_{micro,diskio,hasher} (counting allocator / async wall / wall).
Verified: cargo fmt clean, cargo clippy --features bench -D warnings clean,
cargo test --lib --features bench = 529 passed / 0 failed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L8gs91AhmazoxMsDcNk3KT
This reverts ee51b32. The create_new change showed no measurable throughput
benefit — three 9-rep interleaved runs on the same ext4 device swung −12%..+21%
at the 256 KiB CDC size (a negative stat on a warm dentry cache is ~µs, below
the shared-disk noise floor). Applying the same "no change without a measured
win" bar used for the pool-sizing revert: the idiom/TOCTOU angle is real but the
race is already prevented upstream by the PG pin-or-classify serialisation, so
it's defence-in-depth only — not enough to keep an unmeasured change. Reverts
the production change, the bench, and its doc together.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JG5yYZ9s868mJwqT2Qz7ez
Replace the try_exists(stat) + File::create(O_CREAT|O_TRUNC) pair with a single
OpenOptions::create_new (O_CREAT|O_EXCL), treating AlreadyExists as the existing
idempotent skip. One metadata syscall per new chunk instead of two (each a
spawn_blocking round-trip), and O_EXCL closes the check-then-create TOCTOU the
old pair left open (a racing writer could be truncated).
Honest measurement caveat (benches/BLOB-WRITE.md): the wall-clock throughput
effect is BELOW the noise floor of the test environment — three 9-rep
interleaved runs on the same ext4 device swing −12%..+21% at the 256 KiB CDC
size, because a negative stat on a warm dentry cache is ~µs, dwarfed by the
chunk's create+write+flush. So this is justified as a code-quality / correctness
change (canonical idiom, strictly fewer syscalls, closes a TOCTOU, zero
downside), NOT as a benchmarked perf win.
The sibling idea — reusing the File handle for the fsync sweep — is deliberately
NOT done: sync_blobs is a single end-of-stream sweep over all the upload's new
hashes, so retaining handles would hold thousands of FDs open (>ulimit) on a
large upload. The re-open sweep is a deliberate FD-frugal design.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JG5yYZ9s868mJwqT2Qz7ez
The local backend inherited the trait's conservative read_prefetch() = 1
(strictly sequential chunk reassembly), while S3/Azure already use 8. The
prior rationale was that concurrent opens over scattered content-addressed
chunk files turn one sequential read into competing random I/O ("slower
cold"). Benchmarked that assumption with examples/bench_blob_prefetch
(sweeps the buffered(N) depth over a real LocalBlobBackend under disk-bound
vs network-bound consumers and warm vs cold page cache).
Result on SSD-class storage (median MB/s vs N=1):
warm disk-bound N=2 +11.8% N=8 +3.9% N=16 -4.4%
cold disk-bound N=2 +7.2% (no cold regression on SSD)
network-bound (throttled) ~0% at any N — the socket, not the disk, caps it
So N=8 is wrong for local (leaves gain on the table, risks HDD seek thrash)
and the network-bound win the analysis assumed doesn't materialize: buffered()
here overlaps the per-chunk File::open (cheap on local disk), not the data
read. N=2 captures most of the disk-bound gain — which covers localhost/LAN
downloads AND the internal blob reads that drain as fast as the disk delivers
(thumbnail render, transcode, ZIP export, content extraction) — at the lowest
fan-out. Env-tunable via OXICLOUD_LOCAL_READ_PREFETCH (set 1 on seek-bound
HDDs to restore the old behaviour; raise on fast NVMe). Signature unchanged,
so all ~16 LocalBlobBackend::new call sites are untouched.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JG5yYZ9s868mJwqT2Qz7ez
Every upload surface previously wrote each byte to disk twice: the HTTP
body was spooled to a temp file (or assembled from chunk parts), then
mmap-re-read for FastCDC analysis, and finally the new chunks were
written to the blob backend. CDC could not start until the last byte
arrived, so large uploads paid receive + reread + rewrite latency.
The dedup engine now chunks, hashes and settles the stream WHILE it
arrives (fastcdc AsyncStreamCDC + incremental BLAKE3):
- Each batch of distinct chunks is pinned-or-classified by ONE
`UPDATE … RETURNING` (no check-then-bump TOCTOU; pinned chunks can't
be reclaimed mid-upload), and only chunks the store doesn't have are
written — a full dedup hit performs zero content writes.
- Durability before visibility is preserved: one batched fsync sweep,
then one batched INSERT, then the manifest. Identical concurrent
uploads are resolved at the manifest INSERT via ON CONFLICT (the
loser releases its references and becomes a dedup hit).
- A drop guard rolls back pins and surfaces written-but-unregistered
chunks to GC if the request future is cancelled mid-stream.
- MIME sniffing now peeks the first bytes in-flight; client-requested
MD5/SHA-256 checksums are computed by a stream tee — the post-upload
re-read of the assembled file is gone.
All surfaces converge on the new interfaces::upload_ingest helper:
REST multipart, WebDAV PUT, NextCloud PUT, WOPI PutFile, the dedup
endpoint, and both chunked-upload completions (which now stream their
ordered parts straight into the store instead of writing an assembled
file — chunk parts persist until finalize, so completion is genuinely
retryable). The legacy blob re-chunk migration streams from the
backend with no spool file either.
Legacy removed: store_from_file + mmap CDC analysers + temp-path
plumbing through every port (pre_computed_hash, save_file_from_temp,
update_file_content_from_temp), upload_spool + assembled-file
assembly in both chunked services, create_file/update_file byte-slice
variants (no callers), common::temp, the OXICLOUD_UPLOAD_TMPDIR
config, and the memmap2 dependency.
Verified end-to-end against PostgreSQL 16: 8 MB upload (26 chunks),
identical re-upload (dedup hit, zero writes), 3-byte edit re-upload
(26 chunks, 1 written), byte-identical downloads, Range across chunk
boundaries, concurrent identical-upload race (manifest ref 2), and
trash-empty reclaiming exactly the unshared chunk while the shared 25
survive for the edited file. The empty/sub-8KB multipart path found a
post-EOF re-poll panic in the MIME peek (fixed with fuse + regression
test).
https://claude.ai/code/session_01WdNenpnujNR2sc32XVvwfS
Storing a new file through CDC dedup issued sync_all + a parent-dir
fsync for every ~256 KB chunk (~8,200 fsyncs for a 1 GB upload), plus
one PG INSERT round-trip per chunk. The actual durability boundary is
the manifest INSERT: chunks only need to be durable before any PG row
references them, not one by one.
- BlobStorageBackend grows put_blob_from_bytes_unsynced + sync_blobs
with conservative defaults (unsynced delegates to the synced write,
sync_blobs is a no-op) so backends that don't opt in keep the
per-write durability semantics. Remote stores are durable on PUT.
- LocalBlobBackend writes chunks without fsync and implements
sync_blobs as a parallel sweep: every listed blob file (hard
requirement) plus each distinct prefix directory exactly once
(best-effort, same tier as fsync_parent_dir).
- DedupService::store_chunks writes new chunks unsynced, runs one
sync_blobs sweep, then registers all new chunks in ONE batched
UNNEST INSERT - durability before visibility, and the per-chunk PG
round-trips collapse into one.
- Encrypted/Migration decorators forward both methods so the
optimization survives encrypted-local and live-migration stacks.
https://claude.ai/code/session_013Bk4BMQEvR9QxCU7QXLRwv
Edouard Vanbelle