Five benchmark-gated optimizations from a fresh six-way audit (benches/ROUND25.md),
each with a BEFORE/AFTER gate that rolls back if AFTER does not beat BEFORE:
- M1 EncryptedBlobBackend::decrypt_bytes: replace split_off (a fresh Vec + full
ciphertext memcpy on every decrypted chunk, contradicting its own "in place"
doc) with in-place detached decrypt + a zero-copy Bytes::slice past the nonce.
Peak RAM per read drops from ~2x to ~1x the payload (-262KB/op at 256KiB;
scales with blob size). Plaintext byte-identical; tamper/wrong-key tests pass.
- M2 delta commit: move-unzip the owned chunk list instead of a third
per-occurrence hash clone (-4000 allocs on a 4000-chunk commit).
- M3 folder ZIP download: drop the dead Query<HashMap> extractor it never read
(byte-identical response; 5->0 allocs/request).
- Q1 public-playlist listing: fold the per-playlist COUNT(*) N+1 into one
LEFT JOIN ... GROUP BY via a new inherent repo method (101 -> 1 round-trips,
36x wall on a 100-playlist page).
- Q2 contact REST listings (paginated/search/by-group): stop over-fetching the
multi-KB vcard TEXT the ContactDto discards, via a shared lite row mapper and
narrowed SELECTs (6.4x wall on 1000 contacts with 8KiB vcards). The whole-book
vCard export and CardDAV sync paths keep the column.
Adds bench_round25_micro (counting allocator tracking count+bytes) and
bench_round25_queries (live Postgres), both with equivalence gates and a
rollback exit(1). Verified: cargo fmt clean, cargo clippy -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
Benchmark-gated (benches/ROUND20.md), same rule as rounds 2-19: every change
ships with a BEFORE/AFTER counting-allocator micro-benchmark and a byte-value
equivalence gate; a non-winning AFTER is rolled back (never applied). The
rollback rule is encoded in the harness (GATE FAIL exit). All 8 sections pass.
Reproduce: cargo run --release --features bench --example bench_round20_micro
- A1 CalendarEvent iCal parse: replace the throwaway per-property
HashMap<String,Vec<String>> (DTSTART/DTEND/RECURRENCE-ID) with a direct
VALUE=DATE scan; prop_with_params kept #[cfg(test)] (6->2 allocs/event, 4.2x)
- A2 UserDto::from: add User::into_parts and MOVE image (<=512 KiB data URI)
+ ui_preferences JSON instead of cloning on every /api/auth/me (27->14 allocs)
- A3 parse_vcard: drop the per-line to_ascii_uppercase copy + the lines Vec;
promote ascii_ci_contains to common::text and share it (8->1 allocs/contact)
- A4 Calendar/AddressBook DTO: into_parts move incl. custom_properties map (18->10)
- I1 file-listing repos: collect::<Result<Vec>>() size-hints to 0 and grows from
capacity 0; pre-size with Vec::with_capacity (8->1 container reallocs, 4 sites)
- I4 plaintext_stream: lazy emit iterator instead of eager Vec collect (43x wall)
- C1 NC write_etag_element: borrowed pre-escaped quote events, no owned quoted
String/escape re-alloc; byte-identical output (3->0 allocs/PROPFIND row)
- C3 NC favorites REPORT: map.remove() move instead of get().clone() (~7 allocs/fav)
Deferred (documented in ROUND20.md): NC oc:id/trashbin buffer reuse, I1 sibling
CardDAV/CalDAV listing paths, Contact JSONB Json<Vec<_>> decode, dedup
settle_batch &str bind, and a fast DoS-resistant hasher for hot trusted-key maps
(needs a dependency decision).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JsJjcVX9RoN96DMa35Wqzd
Backend (each change benchmark-gated with BEFORE replicas + equivalence
gates; see examples/bench_round11_micro.rs, bench_round11_queries.rs,
bench_log_writer.rs and benches/ROUND11.md — final numbers land in the
follow-up doc commit):
- StoragePath re-representation: single canonical joined String, segments
derived on demand; File/Folder drop the duplicated path_string field
(4000→1000 allocs per 500-row listing page)
- Display classifier fusion: classify_display shares one stack-lowered
extension across the three decision trees; call sites in FileDto,
folder/favorites/recent handlers, trash, path-resolver (+ interning
where Arc::from was still used)
- /status.php and /openapi.json memoized into OnceLock<Bytes> (openapi
rebuilt a 171 KiB spec per request: 2.8 ms → 18 ns)
- NC upload-session PROPFIND: write! + pre-sized body + stack RFC2822
dates (2.3-2.6x, 2582→772 allocs at 256 chunks)
- REST download: dead FileDto clone removed (capture mime/size + move)
- CalendarEventDto/TrashedItem into_parts moves (11 KiB ical_data memcpy
gone per CalDAV row); CardDAV getlastmodified stack render
- 4xx path: borrowed ErrorResponse serialize, ErrorKind::as_str,
not_found/already_exists clone kill
- vCard emit via write!; search page moved out with into_iter skip/take;
content-hit UUIDs parsed once; group last-user check via HashSet
- RateLimiter: lock-free get + insert (and_upsert_with variant REJECTED
by benchmark); CSRF token borrow-compare + borrowed cookie extraction
- Thumbnail/preview ETags built from as_str (Debug-identical bytes)
- Encrypted backend: encrypt_in_place_detached single-buffer write path,
chunk-sized reserve in collect_stream; retry labels made lazy
- PG: deferred upload registration 3→1 round-trips (persist_file CTE
template); direct_grant_cache for Calendar/AddressBook/Playlist authz
(single-flight + set_role/clear_role invalidation); expand_user
tokio::join!; geo clusters min(uuid)::text; recluster face assignment
batched into one UNNEST update
- People recluster cosine: norms precomputed once (bit-identical gate)
- NC capabilities poll logs demoted to debug; tracing-appender dep added
for the log-writer benchmark
Frontend:
- ResourceList.selectedEntries O(N)-per-toggle → id-index projection
O(k log k); favorites/recent consume the batchToolbar snippet param and
drop their duplicate filter + dead selectedIds mirror
- Recent: star state via new favoriteIds prop — a star click no longer
rebuilds all N entries
- admin timeAgo >30d fallback uses the cached Intl.DateTimeFormat
- vitest gates in src/lib/components/round11.bench.test.ts
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ABhTEHuGujvwoodh67Kga7
read_blob_stream / read_blob_range_stream reassembled a CDC file by fetching
its chunks with `buffered(1)` — strictly sequential, so the next chunk's
backend fetch (a file `open` locally; a full request round-trip on S3/Azure)
only started after the current chunk was fully drained.
A benchmark of the exact pipeline (stream::iter(chunks).map(get).buffered(K)
.try_flatten()) showed a blind `buffered(4)` is the WRONG fix: on a local
disk it is neutral on a warm page cache and ~37% SLOWER cold, because
concurrent opens turn one sequential read into several competing random-I/O
streams over content-addressed (scattered) chunk files. The win is entirely
on remote backends, where per-chunk request latency dominates and overlapping
fetches hide it (≈ linear in K).
So the read-ahead depth is now a backend hint, not a constant:
- BlobStorageBackend::read_prefetch() default 1 (sequential; safe for local).
- S3 / Azure override to 8 (overlap GETs to hide TTFB).
- cached / encrypted / retry / migration delegate to the backend that serves
the bytes.
- Both CDC read paths use `self.backend.read_prefetch().max(1)`.
Net: local backend unchanged (no regression); remote reassembly ~4-8x faster.
Ordered `buffered` (not buffer_unordered) keeps chunks in sequence.
Bench (per-chunk fetch-latency model): buffered(1)->(4)/(8) = x3.9 / x7.8
@1ms, x4.0 / x8.1 @5ms, x4.0 / x8.0 @20ms. Local warm: 230ms@1 vs 227ms@4
(noise); local cold: 425ms@1 vs 585ms@4 (why local stays at 1).
https://claude.ai/code/session_01DCszkkU11LYxMEUWr4setK
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