Targets the content-addressable dedup write path (the ROUND15-deferred
"dedup_service hash-String re-allocations") from both ends — the streaming
ingest loop and the delta-commit verification read — plus a CardDAV vCard
micro-cut. Every change is benchmark-gated with a hard rollback rule; no
PostgreSQL needed for any arm (benches/ROUND17.md).
Backend (counting-allocator, examples/bench_round17_micro.rs):
- D2 chunk-ingest (store_from_stream, the hottest write path — every chunk of
every upload): the 64-char hex hash String was allocated 3x per chunk
(to_hex + chunk_hashes clone + session_seen insert-clone, the last dropped on
a duplicate). The intra-upload dedup set now keys on the raw 32-byte BLAKE3
digest ([u8;32], Copy, no heap) and the manifest push is branch-split so a
duplicate moves the hex in: 3 -> 2 allocs/new chunk, 3 -> 1/duplicate.
Measured 214 -> 149 allocs/op (1.14x wall) on a 64-chunk 1-in-2-dup batch;
smaller/faster set too (32B inline keys vs 64B heap Strings).
- D1 hash_chunk_sequence (delta-commit verification): took chunks by
&[(String,u64)] and fed the backend stream with iter().cloned(), re-cloning
every chunk hash a second time on top of the owned Vec the caller already
built. Take the Vec by value + into_iter(): 65 -> 0 internal allocs/op, ~2.5us
of clone work removed per verify.
- V1 vCard TYPE tokens (contact_to_vcard + generate_vcard, 5 sites): each
EMAIL/TEL/ADR TYPE= param used ty.to_uppercase() — a throw-away String per
token per contact. New shared fmt::push_upper writes the upper-cased chars
straight into the buffer (byte-identical to str::to_uppercase, unit-tested):
13 -> 5 allocs/op, 1.19x wall.
Gates: each section asserts byte/-value equivalence (D2 the ordered manifest +
sizes + write-set; D1 the removed clone is a pure copy; V1 the full vCard) and
exits non-zero if an AFTER arm fails to reduce allocations. push_upper is
unit-tested byte-equal to str::to_uppercase (fmt::tests). Verified end-to-end:
cargo fmt clean, clippy --release --all-targets --features bench -D warnings
clean, and the harness prints GATE PASS against the built release lib.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XMmt7vNETYUbEG3Hc17LDx
return from Claude:
The bench measures nothing reliable at this scale. afterMs=13.21 vs beforeMs=12.56 is ~650ns
per lookup for both paths; at that granularity a single GC pause or scheduler hiccup easily
swamps the actual gain, and the "after" path happens to run first (cold caches), so it gets
penalised on unlucky runs. Your local station happens to warm up before the noise lands; CI
shared-runners often don't.
Fix: run each path a few times and take the minimum (min is noise-proof — noise only slows
things down, never speeds them up):
Benchmark-gated, same rule as rounds 2–14: every change ships with a
BEFORE/AFTER benchmark and an equivalence/safety gate; an AFTER that doesn't
beat its BEFORE is rolled back. The rule is encoded per harness (GATE FAIL
non-zero exit in the Rust examples, threshold expect() in vitest).
F1 — Grouped listings (trash / recent / favorites / shared-with-me)
re-bucketed the WHOLE accumulated list on every infinite-scroll page.
ResourceSectionsBuilder (new, off the reactive graph) re-buckets only the
fresh page and hands VirtualList the same rows array reference for untouched
buckets. 50×50 (2 500-item) drain: 63 750 → 2 500 bucketOf calls (25.5×),
12.5 → 1.3 ms wall (9.9×); O(N²/page) → O(N). Deep-equal to the full-rebuild
reference at every page for both a contiguous (date) and a non-contiguous
(trash-by-drive) group-by; reference-stability + fallback gated.
B1 — exif Make/Model: the display String was thrown away to allocate the
trimmed copy; display_value_trimmed trims in place (drain + truncate), 2 → 1
alloc per field (8 → 4 allocs/op, 1.26×).
B2 — content-index worker: text_extractor::supports (lowercases MIME +
extension) was called twice per file per drain batch; classify once into a
Vec<bool> and thread it through both uses. 256-file batch: 704 → 353 allocs,
34.5 → 16.7 µs (2.07×).
B3 — tantivy: skip SnippetGenerator::create on a zero-hit content search
(return Ok(vec![]) once top_docs.is_empty()); the per-hit loop was empty.
400-doc index: 1 575.6 → 1 237.2 ns (1.27×), widens with index size.
Harnesses: examples/bench_round15_micro.rs, examples/bench_round15_tantivy.rs,
frontend resourceSections.bench.test.ts; writeup in benches/ROUND15.md. Also
normalizes two round14 bench examples that were committed unformatted
(cargo fmt --all).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012o47jSrtL7xuNGTHXmtiYL
Benchmark-gated (benches/ROUND14.md); every change ships a BEFORE/AFTER
benchmark with an equivalence gate and is rolled back on regression (the
rule is encoded as a GATE FAIL exit / threshold expect).
Backend
- Q1 faces_for_file → narrow face_boxes_for_file(id, person_id, bbox) with the
caller filter pushed into SQL: drops the 2 KiB embedding BYTEA + 6 unused
columns per face. 15-face lightbox open 0.312→0.219 ms, 32 KB→840 B/req.
- A1 cookie auth uses the borrow-only extract_cookie_str (already backs CSRF)
instead of extract_cookie_value's owned String: -1 alloc/cookie request.
- A2 compute_relevance ASCII case-fold fast path vs name.to_lowercase() per
result row (Unicode fallback preserved): 1.40x, 12→3 allocs/page.
- A3 sub pre-parsed to Uuid at decode time (TokenClaims.sub_id) vs re-parsing
the 36-char claim on every request incl. cache hits: 22.7→0.7 ns.
- A4 auth + NextCloud middlewares borrow request.headers() instead of taking
axum's HeaderMap extractor (a full map clone): 2→0 allocs/authed request.
- A5 CalDAV getlastmodified via the stack rfc2822_utc (byte-identical to
chrono) vs a per-event to_rfc2822() heap String: 5→0 allocs.
- A6 CalDAV per-event href + quoted etag written into reused page buffers vs a
fresh format! pair per event: 3.48x, 240→6 allocs/40-event page.
Frontend
- F1 t() shares one frozen EMPTY_PARAMS for the no-interpolation call forms vs
a throwaway {} per call: -1 alloc/call.
- F2 favorites favoriteIds is a persistent SvelteSet with per-page add (clear
on reset) vs a brand-new set over the whole accumulated list each page:
22.3x over a 40-page drain (O(N^2)→O(N)).
Verified: cargo check --all-targets, cargo clippy -D warnings, both bench
packs (GATE PASS), frontend npm run check + vitest (4/4). ROUND14.md also
records the investigated-but-deferred backlog (music N+1, contact vcard
over-fetch, CachedBlobBackend syscalls, ResourceList.sections builder, etc.).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PymgCdK78NzUF3oRAQCJfN
Benchmark-gated (BEFORE/AFTER + equivalence/safety gate per change), same
discipline as rounds 2-12. Full write-up in benches/ROUND13.md.
Shipped:
- V1 Grouped views windowed (files route + ResourceList). The grid arm was
the last unwindowed path (trash is grouped-by-default in grid): each
swimlane now feeds its own VirtualList, outer container a flex stack.
vitest gate: 800-item grouped grid mounts <120 .file-item (was 800).
- Q1 get_users_by_ids drops the <=512 KiB avatar image + ui_preferences
JSONB (notification path never reads them). 30-member fan-out 8.60 ->
0.25 ms (34.3x), ~7.7 MB off the wire.
- Q2 Login provisioning is_empty() -> SELECT EXISTS for calendar + address
book (every login). 0.193 -> 0.170 ms, widens with owned-row count.
- Q3 Recent-access prunes only when the upsert inserted (RETURNING xmax=0)
— a re-access can't grow the set. 0.567 -> 0.324 ms (1.75x).
- L1 Locale supported-codes precomputed once vs rebuilt per anonymous
request. 616 -> 17.3 ns (35.7x), 18 -> 1 allocs.
- H1 Duplicate /api TraceLayer removed (global stack already wraps it).
1.86 -> 1.42 us/request, -6 allocs.
- H2 client_ip span field: borrow-only ClientIpDisplay vs owned String.
187 -> 173 ns, -1 alloc.
Not shipped (discipline): the "media hooks read the blob 3x" lead was a
correctness bug, not a perf dup — the raw-path metadata/faces readers
resolve only for local+unencrypted+single-chunk blobs and silently produce
nothing otherwise. Flagged for maintainers; routing through read_blob_bytes
is a correctness fix (perf-neutral-to-negative), not a benchmark-gated
perf change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BfidAJD5AHw23jtvBUNamB
Benchmark-gated round (benches/ROUND12.md; every change ships with a
BEFORE/AFTER harness + equivalence gates, one candidate rejected by its
own bench):
DB / query shapes (bench_round12_queries):
- NC sharee search: username-only projection instead of the 21-column row
(incl. the <=512 KiB avatar) per match, + gin_trgm_ops indexes on
auth.users for the leading-wildcard ILIKE (4.98x; 54.7x with index).
- Password login: delete the redundant full-row update_user — create_session
already stamps last_login_at in its own txn (4.45x per login).
- Email-verified stamp: narrow conditional UPDATE (8.9x); OIDC repeat login
now compares profile state in memory and issues ZERO queries when nothing
changed (was: full 17-column rewrite per login).
- Refresh rotation: revoke+insert+stamp fused into one transaction via new
rotate_session port method (1.18x).
- WOPI CheckFileInfo / authorize_wopi_access: require(Read) + get_file +
check(Update) overlapped with tokio::join!, original result precedence
(cold 1.34x).
- Upload quota gate: user-envelope + drive-cap checks fused into ONE
round-trip (check_upload_quotas) — the NC chunked PUT pays this per
chunk (1.81x, 2 -> 1 queries/chunk); shared verdict evaluators keep
error shapes byte-identical.
CPU / allocs (bench_round12_micro):
- sized_json: pre-sized listing serialization replacing axum Json's 128 B
seed + doubling-realloc chain on files/folder-resources/photos/search
responses (1.40x, 13 -> 2 allocs per 500-row page; byte-identical).
- Security headers: 4 SetResponseHeaderLayer folded into the CSP middleware
pass (5 layers -> 1; 1.43x per request, -26 allocs; header set gated
byte-identical incl. 304s).
- Media capture-metadata: single-read extraction — nom-exif now parses the
buffer kamadak already read (zero-copy Bytes) and videos open once with a
kind() dispatch; per-image opens 2-3 -> 1 (1.44x warm geomean, 1.6-3.2x
cold cache; extraction outputs gated identical incl. the MIME-mislabel
track fallback).
- Chunked-upload session ops: owner gate folded into the operation's own
DashMap lookup + stack-encoded uuid compare (5 -> 3 lookups, -2 allocs,
1.28x per chunk).
Blob cache (bench_blob_cache_index + round-3 regression guard):
- CachedBlobBackend index: tokio::sync::Mutex<LruCache> -> moka::sync::Cache
with byte weigher. The mutex serialized every cached chunk read and scaled
NEGATIVELY (2.08 -> 1.07 Mops/s from 1 -> 2 readers); moka probes are
lock-free (2.17x at K=2). Byte budget now enforced by moka (manual
current_size + collect_evictions machinery deleted); eviction listener
unlinks size-evicted files only (Replaced entries keep their file —
gated). Single-flight miss gate unchanged (16 concurrent misses -> 1
fetch re-verified via the round-3 harness).
- put_blob now populates the cache BEFORE the inner backend consumes the
source file (the old order failed 100% of the time — local renames,
S3/Azure delete the source — so the first read after a whole-file put
re-downloaded from the remote); inner-put failure invalidates the entry.
Frontend (vitest gates):
- List-view thumbnails request the 150px icon rendition instead of 400px
preview into a 40px slot (~7.1x fewer pixels, ~4-5x fewer bytes per
thumbnail across list views); grid keeps preview.
Rejected by its own bench (kept as evidence in bench_round12_micro §2):
- Single-pass compression predicate: the monomorphized And-chain already
costs ~4.6 ns / 0 allocs total; the fused node measured within noise.
New migration: 20260719000000_users_search_trgm.sql (trgm indexes).
Deferred with prepared design: grouped file/grid view virtualization
(single-VirtualRows flatten, the photos pattern) — next round's headline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BfidAJD5AHw23jtvBUNamB
Conflict resolution re-applies the round-11 SPA optimizations onto the
refactored item-based ResourceList (which independently converged on the
favoriteIds-prop star — S2 is now upstream's shape):
- ResourceList.selectedItems: upstream reintroduced the full O(N)
items.filter per selection toggle; replaced with the round-11
id-index projection — O(k·log k), item order preserved
(benches/ROUND11.md §S1)
- favorites/recent pages: upstream's rewrite kept the host-side
selectedItems shadow + unused selectedIds mirror and ignored the
batchToolbar snippet param; re-applied the param-consuming shape and
deleted the shadows
Validation on the merged tree: clippy --all-features --all-targets
-D warnings clean; cargo test --workspace 524 passed; frontend npm run
check 0 errors; vitest 310 passed (upstream's new tests + round-11 gates)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ABhTEHuGujvwoodh67Kga7
- benches/ROUND11.md: final BEFORE/AFTER numbers for all 21 micro
sections, the 5 query sections, the 4 log-writer arms, the SPA gates,
and the two cross-round regression guards (bench_row_path 2.52x,
bench_dto_map — both byte-identical)
- ROLLBACK (gate-caught): min(fm.file_id)::text geo-cluster cast —
PostgreSQL has no min(uuid) aggregate; the bench section now reproduces
the rejection and the per-row-cast original stays
- REJECTED (bench-measured): tracing-appender non-blocking writer —
slower than sync on fast sinks (1.41M vs 0.99M ev/s, worse tail) and
the slow-sink drain gate showed shutdown tail-loss risk for the audit
channel; dep moved to dev-dependencies (harness only)
- RateLimiter final form: lock-free get + insert (8.0 → 6.0 allocs, wall
neutral; and_upsert_with variant rejected at 2 365 ns)
- fix: TrashedItemParts insertion had stolen TrashedItem's derive line
(caught by clippy --all-targets)
- grant_role enum values corrected in the queries bench
Validation: cargo fmt + clippy --all-features --all-targets -D warnings
clean; cargo test --workspace 524 passed / 0 failed; frontend npm run
check 0 errors + vitest 306 passed (1 pre-existing round-7 wall-clock
gate flaked only under concurrent Rust-build CPU contention; passes in
isolation)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ABhTEHuGujvwoodh67Kga7
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
get_thumbnail_impl runs require_permission(Read) on every request. For a drive
member that's a drive_role_cache hit, but a shared-album recipient — granted a
folder (the album), not drive membership — fails the drive-role precheck and
falls through to file_cascade_grant_exists (a role_grants ⋈ folders lpath
ancestor query), once per file. Browsers revalidate immutable thumbnails
constantly, so the same (recipient, file, Read) decision was recomputed on
every thumbnail of every view — ~100 grant queries per 100-photo album per
navigate-away-and-back.
New cascade_grant_cache ((Subject, Resource, Permission) → bool, 30 s TTL)
memoises that decision. The check is NEVER skipped — the ordering is unchanged,
authz still runs on every request; only the result is cached, and only after
the drive-role precheck fails (so a later drive grant can't be shadowed by a
stale entry). Invalidation mirrors drive_role_cache's convention: explicit
invalidate_all on every File/Folder set_role/clear_role (immediate revoke on
the direct share path), 30 s TTL for the indirect paths (group membership,
moves, expiry) "rather than a deep invalidation tree".
Bench (bench_thumbnail_cascade_cache) with hard safety gates — recipient
allowed, outsider denied, and a clear_role revoke denies the very next check
(proving the grant-write flush): 100-photo album revalidation 2576 → 2.70
µs/thumb (~950x), 257.6 → 0.27 ms/view. Validated against the full
--cfg integration_tests authz suite (554 tests) + 524 workspace tests, clippy
-D warnings clean.
Deliberately not done: moving authz after the 304/cache short-circuit (a
security-posture change — a revoked user could serve cached thumbnails). With
the decision cached, the authz on the 304 path is now a memory hit, so the
"zero DB work on a 304" intent is restored without weakening the check.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017aJu9ghvuT8WqC31ZEGTBA
Benchmark-gated (equivalence + BEFORE/AFTER; results + reproduce commands in
benches/ROUND7.md):
- Photos timeline re-grouped + re-laid-out the whole accumulated library on
every 60-item page (both `groups` and `photoRows` were $derived over the
full list), Σ ≈ O(N²/60) main-thread work during a scroll. Pages arrive
newest-first so grouping is append-only: the new PhotoTimeline
(lib/utils/photoTimeline.ts) re-buckets only the fresh page and re-lays-out
only changed groups, reusing untouched groups' cached rows, falling back to
a full rebuild on any config/deletion/non-append change. The pure
buildPhotoRows is the verbatim reference the gate holds it equal to at every
page. 50×60 drain: 76 500 → 3 000 grouping ops (25.5x), 23.0 → 2.2 ms
(10.6x).
- Range downloads paid authz + access-notify twice: download_file_impl
resolves the file via get_file_with_perms, then the Range branch re-ran
require_file + notify_file_accessed per request. Media/PDF viewers fetch
exclusively via Range (one request per seek), so every seek in a scrub
re-authorized an already-cleared file. Now routed through the non-perms
get_file_range_preloaded (matching the share-landing + WebDAV range paths);
the unused _with_perms range method is removed. The request-level gate still
denies before the branch runs (bench asserts member granted, outsider
denied). Per seek removed: WARM 0.67 µs, COLD 1362.66 µs — a grant-cascade
drive-resolve query per seek for a shared-drive recipient on a cold cache.
- /api/folders/{id}/resources row→DTO mapping cloned row.name into the DTO
though the row is owned; folders move it (fixed icons), files compute the
name-derived icon/category classes first then move it. 500-row page:
10.004 → 9.004 allocs/row (500 clones removed), output identical.
Deferred with rationale in ROUND7.md: thumbnail ACL-before-304 (security
posture — needs a security review, not a perf tweak), batch_operations
Arc<str>→String widening, list-view O(N²) on smaller lists, and the serial→
join! pairs (decide-by-bench with injected latency, per the round-6 rejection).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017aJu9ghvuT8WqC31ZEGTBA
Benchmark-gated (equivalence + BEFORE/AFTER in examples/bench_*, results
and reproduce commands in benches/ROUND6.md):
- CardDAV whole-book REPORT + depth-1 PROPFIND stream through a PG
cursor (stream_contacts_by_book, 500-contact pages) instead of
materialising every vCard twice: 8 000 contacts TTFB 37.4 → 7.6 ms
(4.9x), peak heap 19.0 → 7.0 MiB (2.7x), wall -23%; REPORT and
PROPFIND byte-identical to the buffered writers.
- NC numeric-id chain fully borrowed: get_or_create_file_ids/folder_ids
take &[&str] and return HashMap<Uuid, i64>; batch_resolve_ids callers
(PROPFIND pages, REPORT, trashbin, OCS search) pass id slices and look
up via nc_id_of. 2.006 → 0.006 allocs/child (334x), 1.53x wall per
500-child page. batch_check_favorites binds &[&str] as text[].
- file_blob_read_repository listing SELECTs drop id::text/folder_id::text
server casts: rows decode binary Uuid (16 vs 36 bytes on the wire) and
render once in row_to_file. A/B on 500-row pages: 1.225 → 1.044 ms
mean (1.17x), p95 1.686 → 1.345 (bench_uuid_text_cast; single-row,
param and min() sites left as-is deliberately).
- IncrementalHasher::finalize_hex renders through common::fmt::hex_lower
instead of one format! per digest byte: 18 → 1 (md5) / 35 → 1 (sha256)
allocs per chunk finalize, 14-15x wall.
- Share landing overlaps the access-count UPDATE with the unlock fetch
via tokio::join! (one round-trip off every public link hit).
- REJECTED by benchmark and reverted: try_join_all fan-out of the
batch-favorites authz pre-check — 42.6 → 56.4 ms cold, 0.15 → 0.23 ms
warm against local-socket PG (bench_favorites_authz kept as evidence).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017aJu9ghvuT8WqC31ZEGTBA
Four SPA hot-path fixes, each shipping with a vitest benchmark gate
(verbatim BEFORE replica + equivalence + perf assertion) so CI
re-verifies the win on every run:
- fetchFolderListing invoked onPage after EVERY 200-row page with the
whole accumulated listing, and the files view re-sorts everything per
emission — O(N²/page) main-thread work on large folders. Page one and
the final page always emit; intermediates coalesce to one per 150 ms.
25×200 load: 30.9 → 4.0 ms (7.8x), 65 000 → 5 200 sorted elements.
- selected/favoriteIds/sharedIds (files) and favoriteIds (recent) were
$state<Set>s copied whole on every toggle. Now one SvelteSet each,
mutated in place (the useSelection pattern): 1 000 toggles @ N=5 000
771.9 → 1.9 ms (399x); one-toggle fan-out across 40 mounted rows
40 → 3 re-runs when refining a select-all.
- batchDelete/moveInto awaited one request per item serially and
probed listing.folders.find per id (O(N·M)). Now an id index built
once + mapLimit(6) fan-out, failure semantics preserved: 100-item
delete @ 5 ms RTT 525 → 89 ms (5.9x), 38 825 → 500 probes.
- t() re-split its dotted key and walked the nested dict on every call,
and interpolate regex-scanned strings without placeholders. Resolved
values now memoize per (dict, key) in a WeakMap + a {{ guard:
20k mixed calls 22.7 → 8.6 ms (2.63x).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017aJu9ghvuT8WqC31ZEGTBA
The original assertion (`pool wall-clock < sequential wall-clock`)
ran the workload in **Node's vitest environment**, using
`crypto.createHash('sha256')` and `node:worker_threads`. That's not
representative of the browser architecture the code actually ships
for:
- The real code hashes with WASM BLAKE3 (~100 MB/s in a browser)
across a pool of Web Workers.
- Node's `crypto` sha256 is native C++ (~500–1000 MB/s) and its
`worker_threads` postMessage has different overhead characteristics.
At native-crypto speed the 4 MiB hash completes in ~8 ms per file,
so the message-passing round-trip cost per file becomes a comparable
fraction of the total — even a *perfect* 3-lane parallelization has
to overcome ~1/3 of its own runtime in messaging cost. Any CI
variance pushes it over the sequential wall-clock, so the test
false-fails while the actual browser code is fine.
The optimization itself is defensible on two grounds:
1. Theoretical parallelism win: at WASM BLAKE3 speed the messaging
overhead is a rounding error and 3 lanes beat sequential ~2.5×.
2. Main-thread responsiveness: even if the wall-clock ended up flat,
offloading the ~1 s of CPU-bound hashing to workers keeps the
UI responsive during upload prep.
Neither of those is validated by a Node vitest. The real gate belongs
in a Playwright browser benchmark. Marked `.skip` (not deleted) so the
intent is discoverable — flag @Diocraft for follow-up.
Claude