perf(pm): make unified registry stateless by elrrrrrrr · Pull Request #3046 · utooland/utoo

elrrrrrrr · 2026-05-21T19:06:34Z

Summary

remove UnifiedRegistry ownership of PackageCache, OnceMap inflight gates, and stored-version dedupe state
keep registry responsibilities limited to provider job execution, HTTP fetch/parse/extract, persistence, and direct RegistryClient compatibility wrappers
update registry tests to assert clone/store sharing instead of cache sharing

Validation

cargo fmt
cargo check -p utoo-ruborist
cargo test -p utoo-ruborist service::registry::tests -- --nocapture
cargo test -p utoo-ruborist resolver::builder::tests::test_build -- --nocapture
cargo clippy --all-targets -- -D warnings --no-deps

Split Plan

Part of the resolver stack split from source PR #3028. This follows the demand mainloop PR so cache/inflight ownership lives only in the resolver loop.

gemini-code-assist

Code Review

This pull request refactors UnifiedRegistry to be stateless by removing the in-memory PackageCache and associated synchronization primitives, transitioning to a job-based manifest resolution model. Reviewers identified a critical bug where fetch_full_manifest fails on 304 Not Modified responses because the registry can no longer retrieve cached manifests. Further feedback notes potential performance regressions due to redundant I/O and network requests for non-exact version specs, and recommends restoring deleted unit tests to ensure logic coverage.

gemini-code-assist · 2026-05-21T19:14:20Z

    async fn fetch_full_manifest(&self, name: &str) -> Result<Arc<FullManifest>, Self::Error> {
-        match self.resolve_full_manifest(name).await? {
-            FullManifestResult::Full(manifest) => Ok(manifest),
-            FullManifestResult::NotModified => {
-                // 304 in trait context: caller doesn't have versions cache access,
-                // so we return an error indicating the manifest is unchanged.
-                Err(RegistryError(anyhow!(
-                    "No versions available for {} (304 Not Modified but no full manifest cached)",
-                    name
-                )))
-            }
+        match self
+            .execute_manifest_job(ManifestJob::Full {
+                name: name.to_string(),
+                spec: None,
+            })
+            .await?
+        {
+            ManifestJobDone::Full {
+                data:
+                    ManifestFullData::Full {
+                        manifest,
+                        speculative: _,
+                    },
+                ..
+            } => Ok(manifest),
+            ManifestJobDone::Full {
+                data: ManifestFullData::Versions(_),
+                ..
+            } => Err(RegistryError(anyhow!(
+                "No full manifest available for {} (304 Not Modified)",
+                name
+            ))),
+            ManifestJobDone::Version { .. } => Err(RegistryError(anyhow!(
+                "provider returned version manifest for full job {name}"
+            ))),
        }
    }


The implementation of fetch_full_manifest is broken when a persistent store is used and the server returns a 304 Not Modified response.

Because UnifiedRegistry is now stateless and does not maintain an in-memory cache of full manifests, it cannot satisfy the RegistryClient contract (which requires returning an Arc<FullManifest>) when the server indicates the manifest hasn't changed. Since execute_manifest_job automatically uses the ETag from the store, subsequent calls to fetch_full_manifest for the same package will consistently fail with an error until the manifest is updated on the server.

For this compatibility wrapper, it is better to bypass the ETag-based job execution and fetch a fresh manifest if one is explicitly requested, or ensure the store can provide the full manifest bytes.

async fn fetch_full_manifest(&self, name: &str) -> Result<Arc<FullManifest>, Self::Error> { let (manifest, etag) = manifest::fetch_full_manifest_fresh( &self.registry_url, name, manifest::MetadataFormat::Abbreviated, ) .await .map_err(RegistryError)?; let manifest = Arc::new(manifest); let versions = Arc::new(VersionsInfo { versions: Versions { version_list: manifest.versions.clone(), dist_tags: manifest.dist_tags.clone(), }, etag, last_updated: current_timestamp_secs(), }); self.store.store_versions(name, versions); Ok(manifest) }

gemini-code-assist · 2026-05-21T19:14:20Z

+                if deno_semver::Version::parse_from_npm(&fetch_spec).is_ok()
+                    && let Some(manifest) =
+                        self.store.load_version_manifest(&name, &fetch_spec).await
+                {
+                    let manifest = Arc::new(manifest);
+                    return Ok(ManifestJobDone::Version {
+                        name,
+                        spec,
+                        manifest,
+                    });
+                }


For semver-supporting registries, resolution requests using tags (e.g., latest) or ranges will now always trigger a network request because they are not exact versions and thus skip the ManifestStore lookup.

Previously, these were cached in memory via inflight_version and PackageCache. While the resolver loop is expected to handle its own caching, this change significantly impacts the performance of direct RegistryClient calls for these registries. If this is intended, it might be worth documenting that UnifiedRegistry should be used with an external caching layer for optimal performance.

gemini-code-assist · 2026-05-21T19:14:20Z

+                let manifest = Arc::new(
+                    manifest::parse_json_vec_off_runtime::<CoreVersionManifest>(bytes).await?,
+                );
+                self.store_version_manifest(&name, Arc::clone(&manifest));


The removal of stored_version deduplication state means that store_version_manifest is now called on every successful fetch or extraction. While the store operations are intended to be fire-and-forget, this can lead to a significant number of redundant asynchronous I/O tasks if multiple concurrent resolution jobs (e.g., for different ranges that resolve to the same version) finish around the same time.

If the caller (the resolver loop) does not handle deduplication of these side effects, consider if some form of lightweight deduplication is still needed here to protect the persistence layer.

gemini-code-assist · 2026-05-21T19:14:20Z

@@ -728,44 +488,7 @@ impl RegistryClient for UnifiedRegistry {

 #[cfg(test)]
 mod tests {


Several valuable tests were removed in this PR, including test_unified_registry_executes_extract_manifest_provider_job which verified the core logic of execute_manifest_job. Even though the registry is now stateless, it is important to maintain tests that exercise the provider job execution paths and verify that they interact correctly with the ManifestStore and network utilities.

I recommend restoring or adapting these tests to use the new stateless architecture.

github-actions · 2026-05-22T01:22:25Z

📊 pm-bench-phases · `0a84525` · linux (`ubuntu-latest`)

Workflow run — ant-design

PMs: utoo (this branch) · utoo-npm (latest published) · bun (latest)

npmjs.org

p0_full_cold

PM	wall	±σ	user	sys	RSS	pgMinor
bun	9.54s	0.29s	10.50s	10.56s	760M	348.3K
utoo-next	8.16s	0.27s	10.63s	12.18s	920M	124.8K
utoo-npm	9.01s	0.89s	11.01s	12.78s	1016M	124.6K
utoo	7.99s	0.12s	11.27s	12.39s	1005M	141.7K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	14.6K	19.0K	1.19G	6M	1.88G	1.76G	1M
utoo-next	108.0K	72.9K	1.16G	5M	1.71G	1.70G	2M
utoo-npm	130.2K	94.4K	1.16G	5M	1.71G	1.70G	2M
utoo	106.8K	61.0K	1.16G	6M	1.71G	1.70G	2M

p1_resolve

PM	wall	±σ	user	sys	RSS	pgMinor
bun	1.96s	0.10s	4.12s	1.02s	514M	158.3K
utoo-next	2.86s	0.03s	5.42s	1.71s	611M	92.9K
utoo-npm	3.08s	0.03s	5.66s	2.04s	612M	79.0K
utoo	2.41s	0.04s	6.12s	1.72s	637M	117.4K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	8.1K	4.6K	203M	3M	107M	-	1M
utoo-next	46.9K	66.3K	200M	2M	7M	3M	2M
utoo-npm	71.0K	88.2K	200M	2M	7M	3M	2M
utoo	15.6K	20.1K	202M	3M	7M	3M	3M

p3_cold_install

PM	wall	±σ	user	sys	RSS	pgMinor
bun	6.84s	0.13s	6.32s	10.06s	609M	207.0K
utoo-next	6.03s	0.07s	5.06s	10.96s	477M	63.3K
utoo-npm	6.10s	0.16s	5.12s	10.91s	498M	63.2K
utoo	5.90s	0.14s	4.96s	10.83s	490M	57.9K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	4.6K	7.0K	1019M	3M	1.76G	1.76G	1M
utoo-next	96.0K	47.4K	988M	3M	1.70G	1.70G	2M
utoo-npm	97.4K	50.4K	988M	2M	1.70G	1.70G	2M
utoo	90.8K	49.1K	988M	2M	1.70G	1.70G	2M

p4_warm_link

PM	wall	±σ	user	sys	RSS	pgMinor
bun	3.40s	0.04s	0.22s	2.38s	135M	32.6K
utoo-next	2.52s	0.07s	0.51s	3.88s	79M	18.4K
utoo-npm	2.41s	0.06s	0.53s	3.83s	79M	18.3K
utoo	2.38s	0.03s	0.53s	3.87s	79M	18.4K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	507	26	5M	33K	1.91G	1.73G	1M
utoo-next	42.5K	20.4K	2K	9K	1.70G	1.70G	2M
utoo-npm	41.2K	18.4K	2K	9K	1.70G	1.70G	2M
utoo	43.0K	20.5K	2K	25K	1.71G	1.70G	2M

npmmirror.com: no output captured.

elrrrrrrr added A-Pkg Manager Area: Package Manager benchmark Run pm-bench on PR labels May 21, 2026

gemini-code-assist Bot reviewed May 21, 2026

View reviewed changes

elrrrrrrr force-pushed the perf/pm-split-resolver-remove-cache-hook branch from ef1f15b to 3f47d7e Compare May 21, 2026 19:27

elrrrrrrr force-pushed the perf/pm-split-resolver-registry-stateless branch from e77fe0b to a1e1d86 Compare May 21, 2026 19:27

elrrrrrrr mentioned this pull request May 21, 2026

perf(pm): source resolver mainloop architecture #3028

Draft

elrrrrrrr force-pushed the perf/pm-split-resolver-remove-cache-hook branch from 3f47d7e to 6f349ab Compare May 21, 2026 22:32

elrrrrrrr force-pushed the perf/pm-split-resolver-registry-stateless branch 2 times, most recently from 2d30f46 to bdd2790 Compare May 21, 2026 23:09

elrrrrrrr force-pushed the perf/pm-split-resolver-remove-cache-hook branch from 6f349ab to 8b02fd5 Compare May 21, 2026 23:09

perf(pm): make unified registry stateless

44add1d

elrrrrrrr force-pushed the perf/pm-split-resolver-remove-cache-hook branch from 8b02fd5 to ed9ee58 Compare May 21, 2026 23:39

elrrrrrrr force-pushed the perf/pm-split-resolver-registry-stateless branch from bdd2790 to 44add1d Compare May 21, 2026 23:39

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

perf(pm): make unified registry stateless#3046

perf(pm): make unified registry stateless#3046
elrrrrrrr wants to merge 1 commit into
perf/pm-split-resolver-remove-cache-hookfrom
perf/pm-split-resolver-registry-stateless

elrrrrrrr commented May 21, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot May 21, 2026

Uh oh!

gemini-code-assist Bot May 21, 2026

Uh oh!

gemini-code-assist Bot May 21, 2026

Uh oh!

gemini-code-assist Bot May 21, 2026

Uh oh!

github-actions Bot commented May 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

		@@ -728,44 +488,7 @@ impl RegistryClient for UnifiedRegistry {

		#[cfg(test)]
		mod tests {

Conversation

elrrrrrrr commented May 21, 2026

Summary

Validation

Split Plan

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot May 21, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 21, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 21, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot May 21, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented May 22, 2026

📊 pm-bench-phases · 0a84525 · linux (ubuntu-latest)

npmjs.org

p0_full_cold

p1_resolve

p3_cold_install

p4_warm_link

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

📊 pm-bench-phases · `0a84525` · linux (`ubuntu-latest`)