http: fast-forward teardown of unread messages by mcollina · Pull Request #65732 · nodejs/node

mcollina · 2026-09-02T06:53:54Z

When a response finishes and the incoming message was fully received but never read (every body-less request that the handler ignores), _dump() currently calls resume(), which schedules a resume_ tick, emits 'resume' to no listeners, and runs the flow() machinery just to reach EOF.

With this change _dump() calls read(0) instead when the message is already ended, undestroyed and fully drained: read(0) at EOF goes straight to endReadable() (idempotent since #65310), so the 'end'/'close' sequence is unchanged and still asynchronous. Incomplete messages (body still arriving when the response ends) keep the resume() path. Unread client responses dumped by _http_client benefit the same way.

For a hello-world HTTP server this removes one nextTick (7 → 6 per request) and one dead emit (9 → 8 per request).

Benchmark (i7-7700, server pinned to one core, wrk -t2 -c50 on separate physical cores, 8 interleaved rounds; CPU measured from /proc/<pid>/stat over the run):

	CPU µs/request	user µs/request
main	26.6	16.7
this PR	25.3 (−4.2%)	15.8 (−5%)

The patched binary won all 8 rounds and served ~4% more requests per round.

test/parallel/test-http-*, test-stream-*, test-net-* and test-https-* all pass; test-http-dump-req-when-res-ends covers the mid-body dump case, which keeps the previous behavior.

——

This was generated with the assistance of AI. I reviewed all the code.

Second commit (http: destroy clean incoming messages in one tick): when the message was fully received and is destroyed with no error, _destroy completes synchronously instead of deferring the callback with process.nextTick() — the deferral only exists so 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' remains asynchronous. Per-request ticks drop from 6 to 5; same test suites pass (863/863).

nodejs-github-bot · 2026-09-02T06:53:59Z

Review requested:

@nodejs/http
@nodejs/net

When a response finishes and the incoming message was fully received but never read, _dump() no longer goes through resume(): read(0) at EOF schedules the 'end' emission directly, skipping the resume_ tick, the 'resume' emit and the flow() machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 7 to 6 and CPU per request by ~4%. Signed-off-by: Matteo Collina <hello@matteocollina.com>

codecov · 2026-09-02T08:44:34Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.99%. Comparing base (7b0de5e) to head (962c042).
⚠️ Report is 92 commits behind head on main.

Additional details and impacted files

@@ Coverage Diff @@## main #65732 +/- ##
==========================================
- Coverage 90.05% 89.99% -0.06% 
==========================================
Files 754 757 +3 Lines 255722 257753 +2031 Branches 48314 48885 +571 ==========================================
+ Hits 230281 231959 +1678 - Misses 16555 16895 +340 - Partials 8886 8899 +13

Files with missing lines	Coverage Δ
lib/_http_incoming.js	`97.99% <100.00%> (+0.05%)`	⬆️

... and 94 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

When an incoming message is destroyed after being fully received with no error, complete the destroy synchronously instead of deferring the callback with process.nextTick(). The deferral only exists so that 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' is still emitted asynchronously by the stream machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 6 to 5. Signed-off-by: Matteo Collina <hello@matteocollina.com>

nodejs-github-bot · 2026-09-03T05:56:26Z

CI: https://ci.nodejs.org/job/node-test-pull-request/76918/

mcollina · 2026-09-02T06:53:54Z

When a response finishes and the incoming message was fully received but never read (every body-less request that the handler ignores), _dump() currently calls resume(), which schedules a resume_ tick, emits 'resume' to no listeners, and runs the flow() machinery just to reach EOF.

With this change _dump() calls read(0) instead when the message is already ended, undestroyed and fully drained: read(0) at EOF goes straight to endReadable() (idempotent since #65310), so the 'end'/'close' sequence is unchanged and still asynchronous. Incomplete messages (body still arriving when the response ends) keep the resume() path. Unread client responses dumped by _http_client benefit the same way.

For a hello-world HTTP server this removes one nextTick (7 → 6 per request) and one dead emit (9 → 8 per request).

Benchmark (i7-7700, server pinned to one core, wrk -t2 -c50 on separate physical cores, 8 interleaved rounds; CPU measured from /proc/<pid>/stat over the run):

	CPU µs/request	user µs/request
main	26.6	16.7
this PR	25.3 (−4.2%)	15.8 (−5%)

The patched binary won all 8 rounds and served ~4% more requests per round.

test/parallel/test-http-*, test-stream-*, test-net-* and test-https-* all pass; test-http-dump-req-when-res-ends covers the mid-body dump case, which keeps the previous behavior.

——

This was generated with the assistance of AI. I reviewed all the code.

Second commit (http: destroy clean incoming messages in one tick): when the message was fully received and is destroyed with no error, _destroy completes synchronously instead of deferring the callback with process.nextTick() — the deferral only exists so 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' remains asynchronous. Per-request ticks drop from 6 to 5; same test suites pass (863/863).

nodejs-github-bot · 2026-09-02T06:53:59Z

Review requested:

@nodejs/http
@nodejs/net

When a response finishes and the incoming message was fully received but never read, _dump() no longer goes through resume(): read(0) at EOF schedules the 'end' emission directly, skipping the resume_ tick, the 'resume' emit and the flow() machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 7 to 6 and CPU per request by ~4%. Signed-off-by: Matteo Collina <hello@matteocollina.com>

codecov · 2026-09-02T08:44:34Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.99%. Comparing base (7b0de5e) to head (962c042).
⚠️ Report is 92 commits behind head on main.

Additional details and impacted files

@@ Coverage Diff @@## main #65732 +/- ##
==========================================
- Coverage 90.05% 89.99% -0.06% 
==========================================
Files 754 757 +3 Lines 255722 257753 +2031 Branches 48314 48885 +571 ==========================================
+ Hits 230281 231959 +1678 - Misses 16555 16895 +340 - Partials 8886 8899 +13

Files with missing lines	Coverage Δ
lib/_http_incoming.js	`97.99% <100.00%> (+0.05%)`	⬆️

... and 94 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

When an incoming message is destroyed after being fully received with no error, complete the destroy synchronously instead of deferring the callback with process.nextTick(). The deferral only exists so that 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' is still emitted asynchronously by the stream machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 6 to 5. Signed-off-by: Matteo Collina <hello@matteocollina.com>

nodejs-github-bot · 2026-09-03T05:56:26Z

CI: https://ci.nodejs.org/job/node-test-pull-request/76918/

mcollina · 2026-09-02T06:53:54Z

When a response finishes and the incoming message was fully received but never read (every body-less request that the handler ignores), _dump() currently calls resume(), which schedules a resume_ tick, emits 'resume' to no listeners, and runs the flow() machinery just to reach EOF.

With this change _dump() calls read(0) instead when the message is already ended, undestroyed and fully drained: read(0) at EOF goes straight to endReadable() (idempotent since #65310), so the 'end'/'close' sequence is unchanged and still asynchronous. Incomplete messages (body still arriving when the response ends) keep the resume() path. Unread client responses dumped by _http_client benefit the same way.

For a hello-world HTTP server this removes one nextTick (7 → 6 per request) and one dead emit (9 → 8 per request).

Benchmark (i7-7700, server pinned to one core, wrk -t2 -c50 on separate physical cores, 8 interleaved rounds; CPU measured from /proc/<pid>/stat over the run):

	CPU µs/request	user µs/request
main	26.6	16.7
this PR	25.3 (−4.2%)	15.8 (−5%)

The patched binary won all 8 rounds and served ~4% more requests per round.

test/parallel/test-http-*, test-stream-*, test-net-* and test-https-* all pass; test-http-dump-req-when-res-ends covers the mid-body dump case, which keeps the previous behavior.

——

This was generated with the assistance of AI. I reviewed all the code.

Second commit (http: destroy clean incoming messages in one tick): when the message was fully received and is destroyed with no error, _destroy completes synchronously instead of deferring the callback with process.nextTick() — the deferral only exists so 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' remains asynchronous. Per-request ticks drop from 6 to 5; same test suites pass (863/863).

nodejs-github-bot · 2026-09-02T06:53:59Z

Review requested:

@nodejs/http
@nodejs/net

When a response finishes and the incoming message was fully received but never read, _dump() no longer goes through resume(): read(0) at EOF schedules the 'end' emission directly, skipping the resume_ tick, the 'resume' emit and the flow() machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 7 to 6 and CPU per request by ~4%. Signed-off-by: Matteo Collina <hello@matteocollina.com>

codecov · 2026-09-02T08:44:34Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.99%. Comparing base (7b0de5e) to head (962c042).
⚠️ Report is 92 commits behind head on main.

Additional details and impacted files

@@ Coverage Diff @@## main #65732 +/- ##
==========================================
- Coverage 90.05% 89.99% -0.06% 
==========================================
Files 754 757 +3 Lines 255722 257753 +2031 Branches 48314 48885 +571 ==========================================
+ Hits 230281 231959 +1678 - Misses 16555 16895 +340 - Partials 8886 8899 +13

Files with missing lines	Coverage Δ
lib/_http_incoming.js	`97.99% <100.00%> (+0.05%)`	⬆️

... and 94 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

When an incoming message is destroyed after being fully received with no error, complete the destroy synchronously instead of deferring the callback with process.nextTick(). The deferral only exists so that 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' is still emitted asynchronously by the stream machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 6 to 5. Signed-off-by: Matteo Collina <hello@matteocollina.com>

nodejs-github-bot · 2026-09-03T05:56:26Z

CI: https://ci.nodejs.org/job/node-test-pull-request/76918/

mcollina · 2026-09-02T06:53:54Z

When a response finishes and the incoming message was fully received but never read (every body-less request that the handler ignores), _dump() currently calls resume(), which schedules a resume_ tick, emits 'resume' to no listeners, and runs the flow() machinery just to reach EOF.

With this change _dump() calls read(0) instead when the message is already ended, undestroyed and fully drained: read(0) at EOF goes straight to endReadable() (idempotent since #65310), so the 'end'/'close' sequence is unchanged and still asynchronous. Incomplete messages (body still arriving when the response ends) keep the resume() path. Unread client responses dumped by _http_client benefit the same way.

For a hello-world HTTP server this removes one nextTick (7 → 6 per request) and one dead emit (9 → 8 per request).

Benchmark (i7-7700, server pinned to one core, wrk -t2 -c50 on separate physical cores, 8 interleaved rounds; CPU measured from /proc/<pid>/stat over the run):

	CPU µs/request	user µs/request
main	26.6	16.7
this PR	25.3 (−4.2%)	15.8 (−5%)

The patched binary won all 8 rounds and served ~4% more requests per round.

test/parallel/test-http-*, test-stream-*, test-net-* and test-https-* all pass; test-http-dump-req-when-res-ends covers the mid-body dump case, which keeps the previous behavior.

——

This was generated with the assistance of AI. I reviewed all the code.

Second commit (http: destroy clean incoming messages in one tick): when the message was fully received and is destroyed with no error, _destroy completes synchronously instead of deferring the callback with process.nextTick() — the deferral only exists so 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' remains asynchronous. Per-request ticks drop from 6 to 5; same test suites pass (863/863).

nodejs-github-bot · 2026-09-02T06:53:59Z

Review requested:

@nodejs/http
@nodejs/net

When a response finishes and the incoming message was fully received but never read, _dump() no longer goes through resume(): read(0) at EOF schedules the 'end' emission directly, skipping the resume_ tick, the 'resume' emit and the flow() machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 7 to 6 and CPU per request by ~4%. Signed-off-by: Matteo Collina <hello@matteocollina.com>

codecov · 2026-09-02T08:44:34Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.99%. Comparing base (7b0de5e) to head (962c042).
⚠️ Report is 92 commits behind head on main.

Additional details and impacted files

@@ Coverage Diff @@## main #65732 +/- ##
==========================================
- Coverage 90.05% 89.99% -0.06% 
==========================================
Files 754 757 +3 Lines 255722 257753 +2031 Branches 48314 48885 +571 ==========================================
+ Hits 230281 231959 +1678 - Misses 16555 16895 +340 - Partials 8886 8899 +13

Files with missing lines	Coverage Δ
lib/_http_incoming.js	`97.99% <100.00%> (+0.05%)`	⬆️

... and 94 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

When an incoming message is destroyed after being fully received with no error, complete the destroy synchronously instead of deferring the callback with process.nextTick(). The deferral only exists so that 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' is still emitted asynchronously by the stream machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 6 to 5. Signed-off-by: Matteo Collina <hello@matteocollina.com>

nodejs-github-bot · 2026-09-03T05:56:26Z

CI: https://ci.nodejs.org/job/node-test-pull-request/76918/

mcollina · 2026-09-02T06:53:54Z

When a response finishes and the incoming message was fully received but never read (every body-less request that the handler ignores), _dump() currently calls resume(), which schedules a resume_ tick, emits 'resume' to no listeners, and runs the flow() machinery just to reach EOF.

With this change _dump() calls read(0) instead when the message is already ended, undestroyed and fully drained: read(0) at EOF goes straight to endReadable() (idempotent since #65310), so the 'end'/'close' sequence is unchanged and still asynchronous. Incomplete messages (body still arriving when the response ends) keep the resume() path. Unread client responses dumped by _http_client benefit the same way.

For a hello-world HTTP server this removes one nextTick (7 → 6 per request) and one dead emit (9 → 8 per request).

Benchmark (i7-7700, server pinned to one core, wrk -t2 -c50 on separate physical cores, 8 interleaved rounds; CPU measured from /proc/<pid>/stat over the run):

	CPU µs/request	user µs/request
main	26.6	16.7
this PR	25.3 (−4.2%)	15.8 (−5%)

The patched binary won all 8 rounds and served ~4% more requests per round.

test/parallel/test-http-*, test-stream-*, test-net-* and test-https-* all pass; test-http-dump-req-when-res-ends covers the mid-body dump case, which keeps the previous behavior.

——

This was generated with the assistance of AI. I reviewed all the code.

Second commit (http: destroy clean incoming messages in one tick): when the message was fully received and is destroyed with no error, _destroy completes synchronously instead of deferring the callback with process.nextTick() — the deferral only exists so 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' remains asynchronous. Per-request ticks drop from 6 to 5; same test suites pass (863/863).

nodejs-github-bot · 2026-09-02T06:53:59Z

Review requested:

@nodejs/http
@nodejs/net

When a response finishes and the incoming message was fully received but never read, _dump() no longer goes through resume(): read(0) at EOF schedules the 'end' emission directly, skipping the resume_ tick, the 'resume' emit and the flow() machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 7 to 6 and CPU per request by ~4%. Signed-off-by: Matteo Collina <hello@matteocollina.com>

codecov · 2026-09-02T08:44:34Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.99%. Comparing base (7b0de5e) to head (962c042).
⚠️ Report is 92 commits behind head on main.

Additional details and impacted files

@@ Coverage Diff @@## main #65732 +/- ##
==========================================
- Coverage 90.05% 89.99% -0.06% 
==========================================
Files 754 757 +3 Lines 255722 257753 +2031 Branches 48314 48885 +571 ==========================================
+ Hits 230281 231959 +1678 - Misses 16555 16895 +340 - Partials 8886 8899 +13

Files with missing lines	Coverage Δ
lib/_http_incoming.js	`97.99% <100.00%> (+0.05%)`	⬆️

... and 94 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

When an incoming message is destroyed after being fully received with no error, complete the destroy synchronously instead of deferring the callback with process.nextTick(). The deferral only exists so that 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' is still emitted asynchronously by the stream machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 6 to 5. Signed-off-by: Matteo Collina <hello@matteocollina.com>

nodejs-github-bot · 2026-09-03T05:56:26Z

CI: https://ci.nodejs.org/job/node-test-pull-request/76918/

mcollina · 2026-09-02T06:53:54Z

When a response finishes and the incoming message was fully received but never read (every body-less request that the handler ignores), _dump() currently calls resume(), which schedules a resume_ tick, emits 'resume' to no listeners, and runs the flow() machinery just to reach EOF.

With this change _dump() calls read(0) instead when the message is already ended, undestroyed and fully drained: read(0) at EOF goes straight to endReadable() (idempotent since #65310), so the 'end'/'close' sequence is unchanged and still asynchronous. Incomplete messages (body still arriving when the response ends) keep the resume() path. Unread client responses dumped by _http_client benefit the same way.

For a hello-world HTTP server this removes one nextTick (7 → 6 per request) and one dead emit (9 → 8 per request).

Benchmark (i7-7700, server pinned to one core, wrk -t2 -c50 on separate physical cores, 8 interleaved rounds; CPU measured from /proc/<pid>/stat over the run):

	CPU µs/request	user µs/request
main	26.6	16.7
this PR	25.3 (−4.2%)	15.8 (−5%)

The patched binary won all 8 rounds and served ~4% more requests per round.

test/parallel/test-http-*, test-stream-*, test-net-* and test-https-* all pass; test-http-dump-req-when-res-ends covers the mid-body dump case, which keeps the previous behavior.

——

This was generated with the assistance of AI. I reviewed all the code.

Second commit (http: destroy clean incoming messages in one tick): when the message was fully received and is destroyed with no error, _destroy completes synchronously instead of deferring the callback with process.nextTick() — the deferral only exists so 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' remains asynchronous. Per-request ticks drop from 6 to 5; same test suites pass (863/863).

nodejs-github-bot · 2026-09-02T06:53:59Z

Review requested:

@nodejs/http
@nodejs/net

When a response finishes and the incoming message was fully received but never read, _dump() no longer goes through resume(): read(0) at EOF schedules the 'end' emission directly, skipping the resume_ tick, the 'resume' emit and the flow() machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 7 to 6 and CPU per request by ~4%. Signed-off-by: Matteo Collina <hello@matteocollina.com>

codecov · 2026-09-02T08:44:34Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.99%. Comparing base (7b0de5e) to head (962c042).
⚠️ Report is 92 commits behind head on main.

Additional details and impacted files

@@ Coverage Diff @@## main #65732 +/- ##
==========================================
- Coverage 90.05% 89.99% -0.06% 
==========================================
Files 754 757 +3 Lines 255722 257753 +2031 Branches 48314 48885 +571 ==========================================
+ Hits 230281 231959 +1678 - Misses 16555 16895 +340 - Partials 8886 8899 +13

Files with missing lines	Coverage Δ
lib/_http_incoming.js	`97.99% <100.00%> (+0.05%)`	⬆️

... and 94 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

When an incoming message is destroyed after being fully received with no error, complete the destroy synchronously instead of deferring the callback with process.nextTick(). The deferral only exists so that 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' is still emitted asynchronously by the stream machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 6 to 5. Signed-off-by: Matteo Collina <hello@matteocollina.com>

nodejs-github-bot · 2026-09-03T05:56:26Z

CI: https://ci.nodejs.org/job/node-test-pull-request/76918/

mcollina · 2026-09-02T06:53:54Z

When a response finishes and the incoming message was fully received but never read (every body-less request that the handler ignores), _dump() currently calls resume(), which schedules a resume_ tick, emits 'resume' to no listeners, and runs the flow() machinery just to reach EOF.

With this change _dump() calls read(0) instead when the message is already ended, undestroyed and fully drained: read(0) at EOF goes straight to endReadable() (idempotent since #65310), so the 'end'/'close' sequence is unchanged and still asynchronous. Incomplete messages (body still arriving when the response ends) keep the resume() path. Unread client responses dumped by _http_client benefit the same way.

For a hello-world HTTP server this removes one nextTick (7 → 6 per request) and one dead emit (9 → 8 per request).

Benchmark (i7-7700, server pinned to one core, wrk -t2 -c50 on separate physical cores, 8 interleaved rounds; CPU measured from /proc/<pid>/stat over the run):

	CPU µs/request	user µs/request
main	26.6	16.7
this PR	25.3 (−4.2%)	15.8 (−5%)

The patched binary won all 8 rounds and served ~4% more requests per round.

test/parallel/test-http-*, test-stream-*, test-net-* and test-https-* all pass; test-http-dump-req-when-res-ends covers the mid-body dump case, which keeps the previous behavior.

——

This was generated with the assistance of AI. I reviewed all the code.

Second commit (http: destroy clean incoming messages in one tick): when the message was fully received and is destroyed with no error, _destroy completes synchronously instead of deferring the callback with process.nextTick() — the deferral only exists so 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' remains asynchronous. Per-request ticks drop from 6 to 5; same test suites pass (863/863).

nodejs-github-bot · 2026-09-02T06:53:59Z

Review requested:

@nodejs/http
@nodejs/net

When a response finishes and the incoming message was fully received but never read, _dump() no longer goes through resume(): read(0) at EOF schedules the 'end' emission directly, skipping the resume_ tick, the 'resume' emit and the flow() machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 7 to 6 and CPU per request by ~4%. Signed-off-by: Matteo Collina <hello@matteocollina.com>

codecov · 2026-09-02T08:44:34Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.99%. Comparing base (7b0de5e) to head (962c042).
⚠️ Report is 92 commits behind head on main.

Additional details and impacted files

@@ Coverage Diff @@## main #65732 +/- ##
==========================================
- Coverage 90.05% 89.99% -0.06% 
==========================================
Files 754 757 +3 Lines 255722 257753 +2031 Branches 48314 48885 +571 ==========================================
+ Hits 230281 231959 +1678 - Misses 16555 16895 +340 - Partials 8886 8899 +13

Files with missing lines	Coverage Δ
lib/_http_incoming.js	`97.99% <100.00%> (+0.05%)`	⬆️

... and 94 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

When an incoming message is destroyed after being fully received with no error, complete the destroy synchronously instead of deferring the callback with process.nextTick(). The deferral only exists so that 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' is still emitted asynchronously by the stream machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 6 to 5. Signed-off-by: Matteo Collina <hello@matteocollina.com>

nodejs-github-bot · 2026-09-03T05:56:26Z

CI: https://ci.nodejs.org/job/node-test-pull-request/76918/

mcollina · 2026-09-02T06:53:54Z

When a response finishes and the incoming message was fully received but never read (every body-less request that the handler ignores), _dump() currently calls resume(), which schedules a resume_ tick, emits 'resume' to no listeners, and runs the flow() machinery just to reach EOF.

With this change _dump() calls read(0) instead when the message is already ended, undestroyed and fully drained: read(0) at EOF goes straight to endReadable() (idempotent since #65310), so the 'end'/'close' sequence is unchanged and still asynchronous. Incomplete messages (body still arriving when the response ends) keep the resume() path. Unread client responses dumped by _http_client benefit the same way.

For a hello-world HTTP server this removes one nextTick (7 → 6 per request) and one dead emit (9 → 8 per request).

Benchmark (i7-7700, server pinned to one core, wrk -t2 -c50 on separate physical cores, 8 interleaved rounds; CPU measured from /proc/<pid>/stat over the run):

	CPU µs/request	user µs/request
main	26.6	16.7
this PR	25.3 (−4.2%)	15.8 (−5%)

The patched binary won all 8 rounds and served ~4% more requests per round.

test/parallel/test-http-*, test-stream-*, test-net-* and test-https-* all pass; test-http-dump-req-when-res-ends covers the mid-body dump case, which keeps the previous behavior.

——

This was generated with the assistance of AI. I reviewed all the code.

Second commit (http: destroy clean incoming messages in one tick): when the message was fully received and is destroyed with no error, _destroy completes synchronously instead of deferring the callback with process.nextTick() — the deferral only exists so 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' remains asynchronous. Per-request ticks drop from 6 to 5; same test suites pass (863/863).

nodejs-github-bot · 2026-09-02T06:53:59Z

Review requested:

@nodejs/http
@nodejs/net

When a response finishes and the incoming message was fully received but never read, _dump() no longer goes through resume(): read(0) at EOF schedules the 'end' emission directly, skipping the resume_ tick, the 'resume' emit and the flow() machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 7 to 6 and CPU per request by ~4%. Signed-off-by: Matteo Collina <hello@matteocollina.com>

codecov · 2026-09-02T08:44:34Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.99%. Comparing base (7b0de5e) to head (962c042).
⚠️ Report is 92 commits behind head on main.

Additional details and impacted files

@@ Coverage Diff @@## main #65732 +/- ##
==========================================
- Coverage 90.05% 89.99% -0.06% 
==========================================
Files 754 757 +3 Lines 255722 257753 +2031 Branches 48314 48885 +571 ==========================================
+ Hits 230281 231959 +1678 - Misses 16555 16895 +340 - Partials 8886 8899 +13

Files with missing lines	Coverage Δ
lib/_http_incoming.js	`97.99% <100.00%> (+0.05%)`	⬆️

... and 94 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

When an incoming message is destroyed after being fully received with no error, complete the destroy synchronously instead of deferring the callback with process.nextTick(). The deferral only exists so that 'error' listeners attached right after destroy(err) still receive the error, which cannot matter when there is no error. 'close' is still emitted asynchronously by the stream machinery. Reduces the per-request nextTick count of a hello-world HTTP server from 6 to 5. Signed-off-by: Matteo Collina <hello@matteocollina.com>

nodejs-github-bot · 2026-09-03T05:56:26Z

CI: https://ci.nodejs.org/job/node-test-pull-request/76918/

nodejs-github-bot added http Issues and PRs related to the http subsystem. needs-ci PRs that need a full CI run. labels Sep 2, 2026

mcollina force-pushed the http-fast-request-teardown branch from e4f8389 to 803bd87Compare September 2, 2026 07:04

mcollina mentioned this pull request Sep 2, 2026
stream: skip unobserved 'readable' emission at EOF #65749
Open

ronag approved these changes Sep 3, 2026
View reviewed changes

mcollina added the request-ci Add this label to start a Jenkins CI on a PR. label Sep 3, 2026

github-actionsBot removed the request-ci Add this label to start a Jenkins CI on a PR. label Sep 3, 2026

jasnell approved these changes Sep 3, 2026
View reviewed changes