perf(pacquet): parse packuments off the reactor and release the network permit before CPU work#12318
Merged
Merged
Conversation
…lease the network permit first A cold resolve parsed every packument body inline on the tokio worker that buffered it — while still holding its network-concurrency permit — and then wrote the multi-megabyte metadata mirror to disk synchronously on the same worker. High-release-cadence packages (`@fluentui/*`, `@types/node`) ship packuments of several megabytes, so a burst of them pinned every reactor worker in serde for hundreds of milliseconds at a time, stalling the sockets those workers pump and parking concurrency permits on CPU work. On a cold babylon resolve (1,476 packuments, ~520 MB) the metadata phase ran at a third of pnpm's effective throughput and the tarball pipeline sat idle behind it (a 15-second dead zone in the download timeline while the `@fluentui` subtree parsed). Both metadata fetchers now drop the request guard as soon as the body is buffered — the semaphore goes back to bounding sockets, matching the buffer-then-release shape the tarball pipeline already uses — and deserialize + persist the mirror inside one `spawn_blocking`, keeping the reactor free to pump the remaining transfers. A 304 releases the guard before the mirror disk read for the same reason. Cold babylon full installs in a Linux container improved from a ~38 s median to ~32.7 s (pnpm: ~27 s), winning 5 of 6 interleaved A/B rounds; cold `--lockfile-only` resolves dropped from 44 s to 24 s in the back-to-back pair.
Code Review by Qodo
1. Socket permit undercount on 304
|
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (3)
📜 Recent review details⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (10)
🧰 Additional context used📓 Path-based instructions (1)pacquet/**/*.rs📄 CodeRabbit inference engine (pacquet/AGENTS.md)
Files:
🧠 Learnings (4)📚 Learning: 2026-05-20T19:40:55.051ZApplied to files:
📚 Learning: 2026-05-22T00:08:44.646ZApplied to files:
📚 Learning: 2026-05-20T23:07:58.444ZApplied to files:
📚 Learning: 2026-06-06T18:58:37.156ZApplied to files:
🔇 Additional comments (3)
📝 WalkthroughWalkthroughThe PR introduces a ChangesResource lifecycle and spawn_blocking JSON parsing
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
PR Summary by Qodoperf: parse packuments off the reactor and release network permit before CPU work WalkthroughsDescription• Moves packument deserialization and mirror persistence into spawn_blocking in both fetch_full_metadata and fetch_full_metadata_cached, preventing multi-MB serde parses from pinning tokio reactor workers and stalling concurrent socket I/O. • Drops the ThrottledClientGuard (network-concurrency semaphore permit) immediately after the response body is fully buffered — before any CPU work — so the semaphore bounds sockets rather than parse time, matching the existing tarball pipeline pattern. • On 304 responses, releases the guard before the mirror disk read for the same reason. • Adds a new FetchMetadataError::ParseTask variant to surface JoinError from the blocking task (panic or runtime shutdown). DiagramsequenceDiagram
participant R as Tokio Reactor
participant S as Semaphore
participant F as Fetcher
participant B as spawn_blocking
participant D as Disk Mirror
R->>S: acquire permit
S-->>F: ThrottledClientGuard
F->>R: GET /registry/pkg (stream body)
R-->>F: raw_body buffered
F->>S: drop(client) — release permit
Note over S,R: Semaphore free; reactor pumps next sockets
F->>B: spawn_blocking(serde_json::from_str)
B->>D: save_meta (mirror write)
B-->>F: Package
High-Level AssessmentThe following are alternative approaches to this PR: 1. block_in_place instead of spawn_blocking
2. Combine guard drop + parse into a single async step without spawn_blocking
Recommendation: The PR's approach — explicit File ChangesEnhancement (1)
Other (2)
|
Contributor
Micro-Benchmark ResultsLinux |
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #12318 +/- ##
==========================================
+ Coverage 87.65% 87.71% +0.06%
==========================================
Files 288 289 +1
Lines 35475 35624 +149
==========================================
+ Hits 31095 31247 +152
+ Misses 4380 4377 -3 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Contributor
Integrated-Benchmark Report (Linux)Each scenario reports direct installs and pnpr installs. Bencher consumes pacquet@HEAD and pnpr@HEAD. Scenario: Isolated linker: fresh restore, cold cache + cold store
BENCHMARK_REPORT.json{
"results": [
{
"command": "pacquet@HEAD",
"mean": 10.017704292480001,
"stddev": 0.16067292411308276,
"median": 9.95407801108,
"user": 3.3073586999999995,
"system": 3.29330364,
"min": 9.85795805158,
"max": 10.38778808558,
"times": [
9.92033034858,
10.38778808558,
9.93648043258,
9.91833999058,
10.058451249580001,
9.93477742558,
9.992880553580001,
10.19836119758,
9.97167558958,
9.85795805158
]
},
{
"command": "pacquet@main",
"mean": 10.05747461448,
"stddev": 0.1299342290316299,
"median": 10.02964970108,
"user": 3.1967086,
"system": 3.2011163399999996,
"min": 9.97844568758,
"max": 10.41709596858,
"times": [
10.06891978858,
10.00074651658,
10.04224697958,
10.02596453858,
9.97886235658,
10.04058469658,
10.03333486358,
9.97844568758,
10.41709596858,
9.98854474858
]
},
{
"command": "pnpr@HEAD",
"mean": 5.3842999462800005,
"stddev": 0.13755358585405802,
"median": 5.35208286508,
"user": 2.5425676,
"system": 2.8733183399999995,
"min": 5.27676659358,
"max": 5.75598956258,
"times": [
5.40803522758,
5.31567597058,
5.36640672358,
5.3951475035800005,
5.2836461765800005,
5.27676659358,
5.36679216558,
5.75598956258,
5.33678053258,
5.33775900658
]
},
{
"command": "pnpr@main",
"mean": 5.40449842818,
"stddev": 0.133264656044673,
"median": 5.32742165358,
"user": 2.5429517,
"system": 2.8487267399999996,
"min": 5.2869473755800005,
"max": 5.63098520158,
"times": [
5.63098520158,
5.2869473755800005,
5.32624802258,
5.48295811758,
5.28955286058,
5.31052059858,
5.57216380758,
5.52359831258,
5.29341470058,
5.32859528458
]
}
]
}Scenario: Isolated linker: fresh restore, hot cache + hot store
BENCHMARK_REPORT.json{
"results": [
{
"command": "pacquet@HEAD",
"mean": 0.64843110006,
"stddev": 0.012855532797284895,
"median": 0.6489226985600001,
"user": 0.36625749999999996,
"system": 1.3115138799999997,
"min": 0.6305207220600001,
"max": 0.6724569330600001,
"times": [
0.6724569330600001,
0.6305207220600001,
0.64758253706,
0.65026286006,
0.6583546570600001,
0.6534967460600001,
0.6372756930600001,
0.6356680450600001,
0.65862360206,
0.6400692050600001
]
},
{
"command": "pacquet@main",
"mean": 0.6651677703600001,
"stddev": 0.03565056959686553,
"median": 0.64578285706,
"user": 0.36864449999999993,
"system": 1.31026848,
"min": 0.6373482340600001,
"max": 0.7383924080600001,
"times": [
0.6427776390600001,
0.69820455406,
0.7383924080600001,
0.6651904750600001,
0.6380965580600001,
0.64878807506,
0.7030181590600001,
0.6388038340600001,
0.6373482340600001,
0.6410577670600001
]
},
{
"command": "pnpr@HEAD",
"mean": 0.77909978656,
"stddev": 0.08147469586256419,
"median": 0.75620586106,
"user": 0.3888621,
"system": 1.32867268,
"min": 0.7358225020600001,
"max": 1.00953619306,
"times": [
0.7537868200600001,
0.7560261970600001,
0.75638552506,
0.75874824606,
0.7358225020600001,
0.7416027810600001,
1.00953619306,
0.7679393290600001,
0.75935763006,
0.7517926420600001
]
},
{
"command": "pnpr@main",
"mean": 0.7764442101600001,
"stddev": 0.04927642223606241,
"median": 0.7544804190600001,
"user": 0.39519039999999994,
"system": 1.31221738,
"min": 0.7435991270600001,
"max": 0.90547633806,
"times": [
0.75748533206,
0.7736609430600001,
0.80675769606,
0.7508382770600001,
0.7469096460600001,
0.75026090406,
0.90547633806,
0.7779783320600001,
0.7435991270600001,
0.75147550606
]
}
]
}Scenario: Isolated linker: fresh install, cold cache + cold store
BENCHMARK_REPORT.json{
"results": [
{
"command": "pacquet@HEAD",
"mean": 9.170304871119999,
"stddev": 0.023374401523629327,
"median": 9.16537366612,
"user": 3.70488122,
"system": 3.28502824,
"min": 9.131187692620001,
"max": 9.20745777562,
"times": [
9.20745777562,
9.16038210462,
9.17636382262,
9.16391298862,
9.18810480762,
9.131187692620001,
9.20049435162,
9.148288282620001,
9.16002254162,
9.16683434362
]
},
{
"command": "pacquet@main",
"mean": 9.27756225862,
"stddev": 0.04630413011140634,
"median": 9.27123243262,
"user": 3.6605168199999993,
"system": 3.2652195399999995,
"min": 9.214396754620001,
"max": 9.36318135562,
"times": [
9.232666011620001,
9.280934290620001,
9.273439917620001,
9.26902494762,
9.32562192662,
9.214396754620001,
9.36318135562,
9.31989133862,
9.24391286162,
9.25255318162
]
},
{
"command": "pnpr@HEAD",
"mean": 5.10497121992,
"stddev": 0.04733092717171625,
"median": 5.10418520762,
"user": 2.3939623199999995,
"system": 2.7781312400000004,
"min": 5.04190246662,
"max": 5.19754261562,
"times": [
5.09856333462,
5.080519929619999,
5.123783502619999,
5.1098070806199996,
5.04190246662,
5.19754261562,
5.08463188762,
5.042733141619999,
5.11939551262,
5.150832727619999
]
},
{
"command": "pnpr@main",
"mean": 5.104766965719998,
"stddev": 0.10986918684231943,
"median": 5.08045095162,
"user": 2.3825613199999998,
"system": 2.76368494,
"min": 5.02116636362,
"max": 5.3985165216199995,
"times": [
5.10698995962,
5.142216501619999,
5.073444246619999,
5.048519102619999,
5.03662251062,
5.0353945506199995,
5.08745765662,
5.3985165216199995,
5.09734224362,
5.02116636362
]
}
]
}Scenario: Isolated linker: fresh install, hot cache + hot store
BENCHMARK_REPORT.json{
"results": [
{
"command": "pacquet@HEAD",
"mean": 1.4121855616599999,
"stddev": 0.013933077579351548,
"median": 1.40885295176,
"user": 1.52979132,
"system": 1.78811608,
"min": 1.39488547326,
"max": 1.43274780226,
"times": [
1.4010979292599999,
1.39838453826,
1.4085775462599999,
1.40027107926,
1.39488547326,
1.42721844426,
1.40912835726,
1.43274780226,
1.4262752862599999,
1.42326916026
]
},
{
"command": "pacquet@main",
"mean": 1.40617268466,
"stddev": 0.011358453118770977,
"median": 1.40469467176,
"user": 1.5400773200000002,
"system": 1.78112528,
"min": 1.39380359126,
"max": 1.42392898326,
"times": [
1.39528005226,
1.42392898326,
1.39471976726,
1.41460778026,
1.42343435026,
1.40352645626,
1.3982765772599999,
1.4058628872599999,
1.39380359126,
1.40828640126
]
},
{
"command": "pnpr@HEAD",
"mean": 0.67072438896,
"stddev": 0.023739783761279734,
"median": 0.66514325176,
"user": 0.33862531999999995,
"system": 1.2619566799999997,
"min": 0.6503194572600001,
"max": 0.7323890592600001,
"times": [
0.7323890592600001,
0.6738196632600001,
0.6791604132600001,
0.66364444826,
0.6736317202600001,
0.65553247926,
0.6666420552600001,
0.6518109712600001,
0.6503194572600001,
0.6602936222600001
]
},
{
"command": "pnpr@main",
"mean": 0.6716872321599999,
"stddev": 0.038514660819486735,
"median": 0.66022310926,
"user": 0.33849142,
"system": 1.2521583799999998,
"min": 0.6429031582600001,
"max": 0.77645472226,
"times": [
0.68714502426,
0.6618315452600001,
0.6541664432600001,
0.6610052832600001,
0.65944093526,
0.6429031582600001,
0.77645472226,
0.65379351726,
0.6649766492600001,
0.6551550432600001
]
}
]
}Scenario: Isolated linker: fresh install, cold cache + hot store
BENCHMARK_REPORT.json{
"results": [
{
"command": "pacquet@HEAD",
"mean": 4.896212041420001,
"stddev": 0.03101217792385769,
"median": 4.90448953812,
"user": 1.8125289199999997,
"system": 1.9896081200000002,
"min": 4.844166609619999,
"max": 4.93746342362,
"times": [
4.91859012562,
4.844166609619999,
4.93746342362,
4.908537358619999,
4.90344073062,
4.89055172662,
4.851929290619999,
4.905538345619999,
4.92755537562,
4.87434742762
]
},
{
"command": "pacquet@main",
"mean": 5.20378988752,
"stddev": 0.030883993797042063,
"median": 5.19615913362,
"user": 1.7897567200000002,
"system": 2.1917720199999997,
"min": 5.15981348362,
"max": 5.2710631356199995,
"times": [
5.22936070862,
5.18823785862,
5.204080408619999,
5.183247402619999,
5.2710631356199995,
5.21587934062,
5.15981348362,
5.18707937762,
5.18641696062,
5.21272019862
]
},
{
"command": "pnpr@HEAD",
"mean": 0.68897308072,
"stddev": 0.018057391692706468,
"median": 0.68267650912,
"user": 0.34679301999999995,
"system": 1.2992290199999998,
"min": 0.6689919186200001,
"max": 0.7291547906200001,
"times": [
0.7291547906200001,
0.69656302962,
0.6949073896200001,
0.6808321726200001,
0.6779835136200001,
0.70469350362,
0.6831554656200001,
0.67125147062,
0.68219755262,
0.6689919186200001
]
},
{
"command": "pnpr@main",
"mean": 0.68795327802,
"stddev": 0.027047176533396746,
"median": 0.6865555696200001,
"user": 0.33963652,
"system": 1.27114192,
"min": 0.65693897162,
"max": 0.7568784596200001,
"times": [
0.7568784596200001,
0.6860325056200001,
0.65693897162,
0.67555786362,
0.68821351762,
0.67767559162,
0.6870786336200001,
0.6937233986200001,
0.69292370962,
0.66451012862
]
}
]
} |
Contributor
|
Bencher Report| Branch | pr/12318 |
| Testbed | pacquet |
Click to view all benchmark results
| Benchmark | Latency | Benchmark Result milliseconds (ms) (Result Δ%) | Upper Boundary milliseconds (ms) (Limit %) |
|---|---|---|---|
| isolated-linker.fresh-install.cold-cache.cold-store | 📈 view plot 🚷 view threshold | 9,170.30 ms(+3.52%)Baseline: 8,858.08 ms | 10,629.70 ms (86.27%) |
| isolated-linker.fresh-install.cold-cache.hot-store | 📈 view plot 🚷 view threshold | 4,896.21 ms(-2.48%)Baseline: 5,020.81 ms | 6,024.97 ms (81.27%) |
| isolated-linker.fresh-install.hot-cache.hot-store | 📈 view plot 🚷 view threshold | 1,412.19 ms(-0.82%)Baseline: 1,423.87 ms | 1,708.64 ms (82.65%) |
| isolated-linker.fresh-restore.cold-cache.cold-store | 📈 view plot 🚷 view threshold | 10,017.70 ms(-0.42%)Baseline: 10,059.54 ms | 12,071.45 ms (82.99%) |
| isolated-linker.fresh-restore.hot-cache.hot-store | 📈 view plot 🚷 view threshold | 648.43 ms(-0.06%)Baseline: 648.85 ms | 778.62 ms (83.28%) |
Contributor
|
| Branch | pr/12318 |
| Testbed | pnpr |
⚠️ WARNING: No Threshold found!Without a Threshold, no Alerts will ever be generated.
Click here to create a new Threshold
For more information, see the Threshold documentation.
To only post results if a Threshold exists, set the--ci-only-thresholdsflag.
Click to view all benchmark results
| Benchmark | Latency | milliseconds (ms) |
|---|---|---|
| isolated-linker.fresh-install.cold-cache.cold-store | 📈 view plot | 5,104.97 ms |
| isolated-linker.fresh-install.cold-cache.hot-store | 📈 view plot | 688.97 ms |
| isolated-linker.fresh-install.hot-cache.hot-store | 📈 view plot | 670.72 ms |
| isolated-linker.fresh-restore.cold-cache.cold-store | 📈 view plot | 5,384.30 ms |
| isolated-linker.fresh-restore.hot-cache.hot-store | 📈 view plot | 779.10 ms |
zkochan
added a commit
that referenced
this pull request
Jun 11, 2026
…ndexed metadata mirror (#12322) ## Why Profiling a warm babylon resolve (metadata mirrors hot, ~520 MB across 1,476 packuments) showed the dominant cost was not I/O but **hydration**: every version of every packument was deserialized into typed manifests — maps, strings, and `serde_json::Value` trees built, hashed, and dropped — even though a pick consults only the version *strings* plus the handful of manifests it actually considers. typescript ships 3,800 versions; a pick needs one. The `#[serde(flatten)]` catch-all on `PackageVersion` compounded it by routing the whole struct through serde's buffering deserializer. The same hydration cost was paid on cold resolves inside the `spawn_blocking` parses from #12318. ## What Three changes, applied in profile-driven order: **1. Lazy packument hydration** (`df6f70eb57`). `Package::versions` becomes a `PackageVersions` map whose entries hold the raw JSON fragment serde captured (`Arc<RawValue>`, shared not copied) and hydrate into `Arc<PackageVersion>` on first access, cached per slot. Key scans never hydrate; `pinned_version` hydrates only the winner; the publish-date filter moves slots without touching manifests; undecodable fragments degrade to "version absent" (matching the JS implementation's tolerance); serialization re-emits raw fragments verbatim. Picked manifests travel as `Arc<PackageVersion>`. Enables serde_json's `raw_value` feature (already a workspace dep). **2. Sharded in-memory packument cache** (`4c28c6679e`). Every resolve edge consults the shared meta cache, and its single `Mutex<HashMap>` was the top mutex-wait site in the profile; it is now the same `DashMap` shape the crate's other shared maps use. Honest note: wall time was unchanged by this alone — the post-hydration profile shows the warm resolve is critical-path-bound, not lock-bound — but the contention disappears and the cache no longer serializes workers under load. **3. Indexed on-disk metadata mirror** (`126a416ae8`, maintainer-approved cache-format break). The two-line NDJSON mirror (headers + verbatim body) becomes: ``` pacquet-meta-v1 <headers_len> <index_len>\n <headers JSON> etag, modified <index JSON> name, dist-tags, time, homepage, versions: [version, offset, len] <fragments> concatenated raw per-version JSON ``` The loader reads the file once and hands `PackageVersions` byte spans into that buffer — no structural re-scan, hydration parses a slice in place. The writer streams the registry's own bytes (fragments borrow from the lazily-parsed response body), so the **cold-install cost stays one temp-file + rename per package**. Old-format files read as cache misses and are rewritten on the next 200. A span-per-fragment `pread` variant was tried first and measured *worse* (sys 0.4 s → 4.5 s; the pick paths probe many candidate fragments per package), hence the single-buffer design. ## Measurements (warm babylon `--lockfile-only` resolve, 10-core M-series) | build | wall | notes | |---|---|---| | before this PR | ~9.1 s | malloc/free + `deserialize_any` dominate the profile | | + lazy hydration | ~7.7–8 s | parse microbench 3–3.8× (`typescript` 36.1→9.5 ms) | | + indexed mirror | **~5.5–6.7 s** | sys 0.4 s; no whole-body scan | Cold resolves keep the same hydration savings inside the parse tasks; cold write volume and syscall pattern are unchanged. ## Evaluated and deliberately not included Consolidating the install-phase thread pools (tokio + global rayon + the dedicated CAS-write pool + capped blocking threads show ~100k involuntary context switches on cold installs vs ~750 for the pnpm CLI). After the resolution fixes, repeated container A/Bs show no measurable wall-clock cost from the churn — it hides entirely behind network time — and history warns that speculative concurrency reshuffles here regress badly (see the #11903 prefetch revert). Deferred until a benchmark shows it on the critical path. ## Tests New `package_versions` unit tests (hydrate-on-demand + Arc-identity caching, undecodable-fragment-as-absent, verbatim raw round-trip incl. unknown keys, eager construction, hydration-free filtering) and rewritten `mirror` tests (headers/index/fragment round-trip, span hydration, truncation → cache miss, old-format → cache miss, atomic overwrite). Full suites green: `pacquet-registry` (23), `pacquet-resolving-npm-resolver` (235), `pacquet-package-manager` + `pacquet-cli` (768); workspace clippy `-D warnings` (pedantic set), dylint, fmt.
2 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Follow-up to #12315's investigation: the one variation where pacquet still trailed pnpm on benchmarks.vlt.sh was
cleanbabylon (34.3 s vs 30.1 s on the 2026-06-10 run).A Linux-side profile (strace + download/linkat timelines in a container) localized the gap to the resolution phase, not store-write/linking as initially suspected:
--lockfile-onlyresolves: pacquet 44 s vs pnpm 17.7 s, with pacquet at ~30% of one core — network-throughput-bound, not CPU (warm-metadata resolves: pacquet 9.8 s vs pnpm 18.7 s, i.e. pacquet's CPU side is already 2× faster).@fluentui/*subtree resolved.Root cause:
fetch_full_metadata_cachedparsed every packument body inline on the tokio worker that buffered it — while still holding its network-concurrency permit — and then wrote the multi-MB metadata mirror to disk synchronously on the same worker. High-release-cadence packages ship multi-megabyte packuments; a burst of them pinned every reactor worker in serde for hundreds of milliseconds at a time, stalling the sockets those workers pump and parking the concurrency semaphore's permits on CPU work. The tarball pipeline already avoids exactly this ("body fully buffered; release the network permit before the CPU-bound work" +spawn_blocking); the packument path never got the same treatment.What
In both metadata fetchers (
fetch_full_metadata_cached,fetch_full_metadata):spawn_blocking, keeping the reactor free to pump the remaining transfers;New
FetchMetadataError::ParseTaskvariant carries aJoinErrorfrom the blocking task.Validation
--lockfile-only44 s → 24 s in the back-to-back pair.ubuntu-24.04-armrunner,cleanvariation, pnpm and patched pacquet in the same hyperfine session (the runner's network was degraded during the run — both tools ~2× slower than the official numbers — which makes the relative result the meaningful one): babylon: pacquet 54.8 s vs pnpm 59.2 s — flipped from −14 % behind (official run) to ahead under identical conditions. astrocleanstayed within noise (pacquet already led it officially). All hyperfine runs exited 0.cargo nextest runforpacquet-resolving-npm-resolver(234),pacquet-package-manager+pacquet-cli(740) all pass; clippy-D warnings, dylint, fmt clean.Profile notes for future work (not addressed here)
The extract+link phase still shows heavy scheduler churn relative to pnpm (~100 k involuntary context switches vs ~750;
sched_yieldstorms from the rayon pools): tokio workers + the dedicated CAS-write pool + the global rayon pool + up to 2×cores blocking threads can leave ~5×cores runnable threads on the box. Wall-clock impact is currently hidden behind network time, but it's the next candidate if the gap reopens on faster runners.Written by an agent (Claude Code, claude-fable-5).
Summary by CodeRabbit
Bug Fixes
Performance