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
9.7 KiB
Round 17 — dedup ingest/verify hash-clone purge, CardDAV vCard TYPE tokens
Benchmark-gated, same rule as ROUND2–16: every change ships with a
BEFORE/AFTER benchmark and an equivalence/safety gate; an AFTER that doesn't
beat its BEFORE is rolled back (never applied). The roll-back rule is encoded
directly into the harness as a GATE FAIL … rollback non-zero exit, so a
regression fails CI rather than shipping.
This round targets the content-addressable dedup write path — the item
carried on the ROUND15 deferred list as "dedup_service hash-String
re-allocations" — from both ends: the streaming ingest loop that hashes and
stores every uploaded chunk, and the delta-commit verification read that
re-hashes a proposed chunk sequence. Plus a CardDAV micro-cut: the vCard
emitters allocated a throw-away upper-cased String per TYPE= token.
Measured on 4 cores / 15 GiB, no PostgreSQL needed for any Round-17 arm (release-profile counting-allocator example). Reproduce any row with:
cargo run --release --features bench --example bench_round17_micro
Summary
| # | change | key metric | before → after |
|---|---|---|---|
| D2 | Chunk-ingest (store_from_stream) allocated the 64-char hex hash String 3× per chunk (to_hex().to_string() + chunk_hashes.push(clone) + session_seen.insert(clone) — the last dropped on the spot for 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 split so a duplicate moves the hex in. |
64-chunk batch, 1-in-2 dup | 214 → 149 allocs/op (65 fewer) · 1.14× wall · set clone gone + dup manifest clone gone |
| D1 | hash_chunk_sequence (delta-commit verification) took chunks: &[(String,u64)] and fed the backend stream with chunks.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 and into_iter() it. |
64-chunk verify | 65 → 0 internal allocs/op · ~2.5 µs of clone work removed per verify |
| V1 | contact_to_vcard / generate_vcard emitted every EMAIL/TEL/ADR TYPE= token via ty.to_uppercase() — one throw-away String per token per contact. New shared fmt::push_upper writes the upper-cased chars straight into the vCard buffer. |
8 tokens/op | 13 → 5 allocs/op (8 fewer) · 1.19× wall |
Allocs/op is the deterministic primary gate (identical run to run); the wall figures are single-shot and noise-bounded (D1's AFTER arm is a near-zero-cost read, so its ratio swings 60–120× between runs — the stable fact is the ~2.5 µs / 65-alloc clone removed).
[D2] Chunk-ingest — dedup set keyed on the raw digest
The streaming ingest loop (DedupService::store_from_stream) is the hottest
write path in the system: it runs for every chunk of every upload. Per
chunk it produced the 64-char hex hash and then allocated it three times:
let hash = blake3::hash(&data).to_hex().to_string(); // A: the hex String
chunk_hashes.push(hash.clone()); // B: manifest copy (always)
if session_seen.insert(hash.clone()) { // C: dedup-set copy (always)
pending.push((hash, Bytes::from(data))); // A moved into the write batch
}
session_seen is the intra-upload dedup set (has this exact chunk already
appeared in this stream? — repeated blocks, zero-padded regions, re-chunked
near-duplicates). It was a HashSet<String>, so clone C heap-allocated a
64-byte key for every chunk — and on a duplicate, insert allocated the clone
only to drop it when the key already existed. Pure waste on the case a dedup
store exists to make cheap.
A BLAKE3 digest is [u8; 32] — Copy, no heap, and the hex is a lossless
rendering of it, so keying the set on the raw digest is behaviour-identical:
let digest = blake3::hash(&data);
let hash = digest.to_hex().to_string(); // A (once)
if session_seen.insert(*digest.as_bytes()) { // Copy key — zero heap
chunk_hashes.push(hash.clone()); // B (new chunk only)
pending.push((hash, Bytes::from(data))); // A moved
} else {
chunk_hashes.push(hash); // dup: move, no clone
}
Clone C is gone for every chunk; clone B is gone for every duplicate (it moves the hex into the manifest instead). The set also holds 32-byte inline keys instead of 64-byte heap Strings and hashes 32 bytes per membership test. Net per chunk: 3 → 2 allocs (new) / 3 → 1 (duplicate) — strictly fewer on every input, unique or duplicate.
64-chunk, 1-in-2-duplicate batch: 214 → 149 allocs/op (65 fewer), 1.14×
wall. Gate (in-harness, replica of the exact before/after loop bodies): the
observable output is byte-for-byte equal — the ordered chunk_hashes
manifest, the chunk_sizes, and the distinct write-set pending all match;
only the private set's key representation differs — plus the gate_allocs
rollback exit on any AFTER that fails to reduce allocations.
[D1] hash_chunk_sequence — take the chunk Vec by value
The delta-sync commit (delta_upload_service::commit) verifies a client's
proposed manifest by streaming the pinned chunks back out of the backend and
recomputing the whole-file BLAKE3. The caller already builds a fresh owned
Vec<(String,u64)> (request.chunks.iter().map(|c| (c.h.clone(), c.s)).collect()),
but hash_chunk_sequence took it by &[(String,u64)] and then did
futures::stream::iter(chunks.iter().cloned()) — re-cloning every chunk hash
a second time to feed the stream.
Taking chunks: Vec<(String,u64)> by value and into_iter()-ing it (the caller
drops one &) moves those Strings straight into the stream: zero internal
clones. The streamed (hash, size) pairs are byte-identical, so the recomputed
hash and every per-chunk size check are unchanged.
The section isolates exactly the clone the old signature forced (the caller's
.collect() is identical on both shapes and excluded): 65 → 0 allocs/op for
a 64-chunk manifest — ~2.5 µs of clone work removed per verify (the AFTER arm is
a near-zero-cost read, so the wall ratio is large but noisy: 60–120×). Gate: the
old internal clone is asserted to be a pure copy (moving changes nothing
observable), plus gate_allocs.
[V1] CardDAV vCard TYPE= tokens — push_upper
Both vCard emitters — carddav_adapter::contact_to_vcard (the CardDAV
REPORT/GET path) and ContactService::generate_vcard — wrote each address /
phone / email TYPE= parameter with write!(…, "{}", ty.to_uppercase()), and
str::to_uppercase() heap-allocates a fresh String for every token. A contact
with several emails/phones/addresses pays one alloc per token, per emit, on
every address-book sync.
New shared helper common::fmt::push_upper(buf, s) writes the upper-cased chars
(char::to_uppercase, so byte-identical to str::to_uppercase — including
ß → SS and dotless-i) straight into the vCard buffer; the five call sites push
the fixed prefix, the upper-cased token, and the value directly. Zero
temporaries.
8-token contact: 13 → 5 allocs/op (8 fewer), 1.19× wall. Gates:
push_upper is unit-tested byte-equal to str::to_uppercase across ASCII /
mixed / multi-char-uppercase / dotless-i inputs (fmt::tests), the section
asserts the full emitted vCard is byte-identical before/after, and the existing
carddav_adapter_test::test_contact_to_vcard_full pins the whole document.
Not shipped — deferred to a later round
Surfaced during the Round-17 audit but not landed (each wants its own decision, a Postgres fixture, or a streaming-I/O benchmark):
- Storage I/O —
encrypted_blob_backendframe size (evaluated, kept): the ROUND15 note floated 64 KiB → 256 KiB plaintext emit frames.PLAINTEXT_EMIT_SIZEis a deliberate match to the 64 KiB the unencrypted backends stream, so downstream consumers see the same backpressure shape; changing it is a behaviour change that needs a streaming-throughput A/B (TTFB + syscalls + peak RSS), not an alloc micro-bench. Left as-is pending that harness. - Storage I/O —
CachedBlobBackendwrite path (needs an fs harness): per-writecreate_dir_alleven when the shard dir exists, and inline evictionremove_fileon the reactor thread (carried from ROUND15). - Backend query-shape (needs Postgres):
music_storage_adapter::list_public_playlists1 + NCOUNT(*)fold; contact REST listings over-fetch the multi-KBvcardTEXT though theContactDtomappers never read it (wants a lite row mapper). - Backend CPU/alloc (no Postgres, next micro-pack): the two WebDAV PROPFIND
surfaces still quote
d:getetaginto a freshStringper row andformat!the per-row href per child (the CalDAV reused-buffer treatment never reached them); REST calendar-event edit re-format!s the wholeical_databody once per changed property. - Frontend (vitest-benchmarkable):
VirtualRows.offsetsprefix-sum rebuilt in full on every photos-timeline page; the dotfile filter andResourceList.itemIndexByIdre-scan the whole accumulated list per page.
Environment / methodology
cargo run --release --features bench --example bench_round17_micro— counting global allocator, no Postgres. Tunables:BENCH_ITERS(100000),BENCH_CHUNKS(64),BENCH_DUP_RATIO(2).- Each section is BEFORE (verbatim replica of the shipped-before shape) vs AFTER (verbatim replica of the shipped-after shape) with a byte/-value equivalence gate; the shipped source now matches each AFTER arm.
- Roll-back rule encoded per section:
std::process::exit(1)withGATE FAIL … rollbackif an AFTER arm fails to reduce allocations.