rusteerio 1.1.2 → 1.1.3

package/Cargo.lock

@@ -720,7 +720,7 @@ dependencies = [
 
 [[package]]
 name = "rusteerio"
-version = "1.1.2"
+version = "1.1.3"
 dependencies = [
  "ego-tree",
  "html5ever",

package/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "rusteerio"
-version = "1.1.2"
+version = "1.1.3"
 edition = "2021"
 description = "Rust HTML5 parser and CSS selector engine for the rusteerio Node.js package"
 license = "MIT"

package/README.md

@@ -198,18 +198,18 @@ Rusteerio was faster in **all 45 cases**. Representative results:
 
 | Workload | CPUs | Cheerio | rusteerio | rusteerio |
 | --- | ---: | ---: | ---: | --- |
-| Load/query, 500 articles | 1 | 36.53 ms | 3.37 ms | 10.82× faster |
-| Load/query, 5,000 articles | 1 | 152.78 ms | 34.90 ms | 4.38× faster |
-| HTML mutations, 5,000 articles | 1 | 241.28 ms | 68.12 ms | 3.54× faster |
-| XML load/query, 5,000 articles | 1 | 65.28 ms | 15.24 ms | 4.28× faster |
-| Extract 16 documents × 5,000 articles | 1 | 2,874.89 ms | 447.12 ms | 6.43× faster |
-| Extract 16 documents × 5,000 articles | 8 | 805.16 ms | 122.93 ms | 6.55× faster |
-| Nested callback extraction, 8 × 500 articles | 8 | 54.17 ms | 33.09 ms | 1.64× faster |
-| Cold extraction, 16 × 20 articles | 8 | 15.19 ms | 3.17 ms | 4.79× faster |
-| 8 concurrent batches × 8 × 300 articles | 8 | 127.82 ms | 23.94 ms | 5.34× faster |
-| HTML stream, 50,000 elements | 1 | 1,086.17 ms | 297.04 ms | 3.66× faster |
-| XML stream, 50,000 elements | 8 | 122.06 ms | 67.44 ms | 1.81× faster |
-| Append/update/remove, 4,000 cycles | 1 | 613.95 ms | 388.99 ms | 1.58× faster |
+| Load/query, 500 articles | 1 | 41.89 ms | 3.77 ms | 11.11× faster |
+| Load/query, 5,000 articles | 1 | 184.64 ms | 41.27 ms | 4.47× faster |
+| HTML mutations, 5,000 articles | 1 | 308.55 ms | 89.40 ms | 3.45× faster |
+| XML load/query, 5,000 articles | 1 | 90.55 ms | 20.00 ms | 4.53× faster |
+| Extract 16 documents × 5,000 articles | 1 | 3,615.68 ms | 557.12 ms | 6.49× faster |
+| Extract 16 documents × 5,000 articles | 8 | 1,103.64 ms | 178.32 ms | 6.19× faster |
+| Nested callback extraction, 8 × 500 articles | 8 | 76.75 ms | 42.52 ms | 1.81× faster |
+| Cold extraction, 16 × 20 articles | 8 | 22.66 ms | 5.65 ms | 4.01× faster |
+| 8 concurrent batches × 8 × 300 articles | 8 | 153.41 ms | 26.38 ms | 5.81× faster |
+| HTML stream, 50,000 elements | 1 | 1,551.73 ms | 384.28 ms | 4.04× faster |
+| XML stream, 50,000 elements | 8 | 168.84 ms | 108.97 ms | 1.55× faster |
+| Append/update/remove, 4,000 cycles | 1 | 1,143.37 ms | 766.40 ms | 1.49× faster |
 
 Cold measurements include engine import, pool initialization and the first
 batch.
@@ -225,14 +225,14 @@ HTML. Walking the same nodes without enumerating attributes is faster in both.
 Very small pages can measure either way.
 
 **Memory.** A batch extracting 16 documents of 5,000 articles with eight CPUs
-ended at 581 MB RSS for rusteerio and 2,497 MB for Cheerio's worker pool, the
-HTML stream of 50,000 elements at 491 MB versus 666 MB, and the large XML
-load/query at 140 MB versus 147 MB. The XML stream ended higher, at 285 MB versus
-221 MB, and the append/update/remove workload about even (93 MB versus 91 MB).
-Linux builds use mimalloc, which keeps freed pages briefly for reuse; setting
-`MIMALLOC_PURGE_DELAY=0` returns them at once and lowers stream RSS, at a cost in
-speed for later documents. RSS includes initialization and resident input, and
-sampled peaks do not establish a universal memory bound.
+ended at 581 MB RSS for rusteerio and 2,523 MB for Cheerio's worker pool, the
+HTML stream of 50,000 elements at 364 MB versus 666 MB, and the large XML
+load/query at 140 MB versus 148 MB. The XML stream ended higher, at 236 MB versus
+221 MB, and the append/update/remove workload about even (94 MB versus 90 MB).
+Linux builds use mimalloc, which keeps freed pages briefly for reuse; after a
+parse that builds at least 100,000 nodes, the parsing thread returns them to the
+system. RSS includes initialization and resident input, and sampled peaks do not
+establish a universal memory bound.
 
 These results describe this VM and revision; hardware, runtime and input shape
 affect them. See [the results](bench/results/2026-10-14/) and

package/THIRD_PARTY_NOTICES.md

@@ -4,7 +4,7 @@ This file preserves third-party license and attribution notices for the Rust nat
 
 Generated by `node scripts/generate-notices.mjs` from `cargo metadata --locked`, crate license files, and the pinned supplemental sources recorded in `scripts/notices/sources.json`. It includes build dependencies and dependencies for all target platforms; a particular binary may use a subset. Identical notice texts are reproduced once and referenced below. The html5ever, markup5ever and html5gum crates contain local observer-hook and opt-in UTF-16 transport patches; their complete patched source and original notices are included in `native/vendor`. Other dependency source files are unmodified.
 
-Cargo.lock SHA-256: `b834f3bb9a4bb9e7889f20f789f80af645271bcfd0057cf4217843b17e3f5ae5`.
+Cargo.lock SHA-256: `b4e1749e9f6c31b8ecd8cb108285e93f365a4906f4ef04e8239959f549b7de1a`.
 
 ## Source availability
 

package/native/src/lib.rs

@@ -266,6 +266,7 @@ impl NativeDocument {
         if arena.nodes.len() > POOL_RELEASE_NODES {
             runtime::start_release_thread();
         }
+        runtime::return_parse_pages(arena.nodes.len());
         Self {
             html: RefCell::new(None),
             ids: RefCell::new(Vec::new()),
@@ -516,6 +517,7 @@ impl NativeDocument {
         if ids.len() > POOL_RELEASE_NODES {
             runtime::start_release_thread();
         }
+        runtime::return_parse_pages(ids.len());
         Self {
             html: RefCell::new(Some(html)),
             ids: RefCell::new(ids),

package/native/src/runtime.rs

@@ -183,6 +183,29 @@ static RELEASE_THREAD: OnceLock<mpsc::Sender<Box<dyn Send>>> = OnceLock::new();
 /// exists, so freeing large trees does not stall JavaScript. Documents large
 /// enough to be released later start it when they are created, because
 /// finalizers must not create threads.
+/// Trees with at least this many nodes return the parse's freed memory to
+/// the system once built.
+#[cfg(target_os = "linux")]
+const RETURN_PAGES_NODES: usize = 100_000;
+
+#[cfg(target_os = "linux")]
+extern "C" {
+    fn mi_collect(force: bool);
+}
+
+/// Building a large tree grows and then frees big buffers, such as a node
+/// vector that doubled several times. mimalloc keeps freed pages resident for
+/// reuse; after a large parse, return them on the parsing thread instead.
+pub(crate) fn return_parse_pages(nodes: usize) {
+    #[cfg(target_os = "linux")]
+    if nodes >= RETURN_PAGES_NODES {
+        // SAFETY: mimalloc's own collection frees only memory no block uses.
+        unsafe { mi_collect(true) };
+    }
+    #[cfg(not(target_os = "linux"))]
+    let _ = nodes;
+}
+
 pub(crate) fn start_release_thread() {
     RELEASE_THREAD.get_or_init(|| {
         let (sender, receiver) = mpsc::channel::<Box<dyn Send>>();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "rusteerio",
-  "version": "1.1.2",
+  "version": "1.1.3",
   "description": "Cheerio-compatible HTML processing with a Rust engine and parallel, non-blocking extraction",
   "license": "MIT",
   "repository": {