HTTP/2 defaults jump for throughput (
1576cb8)
Node now defaults HTTP/2 stream flow control to 4MB and the connection window to 32MB, a big increase from the prior 64KB baseline. That should materially improve throughput on higher-latency links, while keeping existing apps compatible unless they override settings.
TLS handshake callbacks are deferred off the OpenSSL stack (
d18457b)
SNICallback and OCSPRequest are now scheduled asynchronously instead of running directly from the TLS library stack, avoiding re-entrancy during handshake. This also tightens an error path in client hello handling so teardown scenarios don’t accidentally resume handshakes.
TLS client-hello parsing is replaced with native callback support (5b940dd)
Node drops the hand-rolled ClientHello parser and related plumbing in favor of modern OpenSSL/BoringSSL early ClientHello callbacks for resume-session flows. This is a substantial simplification of the TLS stack and removes a fragile internal subsystem.
QUIC gains explicit STOP_SENDING handling (13e4e3d)
A new stream.onstopsending callback exposes peer-initiated STOP_SENDING events separately from RESET_STREAM, making QUIC stream shutdown semantics clearer and more precise. The update also threads the new callback through docs, bindings, and tests.
HTTP response completion now matches Writable more closely (2ddade6)
writableFinished is now based on actual flush completion, and end() callbacks report success or the flush error instead of always hanging off 'finish'. That fixes cases where HTTP responses could signal completion too early or lose the real failure reason after write errors.
FFI multi-arg fast calls accept pointer BigInts (
2b350bb)
The multi-argument fast path now accepts BigInt pointer values the same way the single-argument path already did, instead of rejecting them during raw pointer extraction. This fixes a real compatibility gap for pointer-based FFI calls.
FFI reuses libffi call plans (
31cde9f)
On x86-64 System V, fixed-signature FFI calls now reuse precomputed libffi call plans instead of rebuilding argument placement on every invocation. That trims per-call overhead in the generic path and is backed by a new benchmark.
ReadableStream pipeTo cuts per-chunk allocations (3fc98b8)
The webstreams pipe-to implementation was refactored to reuse write/read bookkeeping instead of allocating new promise records and closures for every chunk. This is a performance-focused internal rewrite that should reduce GC pressure in steady-state streaming.
Task runner paths are handled as UTF-8 on Windows (
793a6ec)
Filesystem paths are now converted with UTF-8 instead of the active Windows code page when passed into Node and libuv interfaces. That prevents non-ASCII paths from being mangled or rejected on Windows.
Other misc changes
- npm updated to 11.19.0.
- nghttp2, nghttp3, ngtcp2, simdjson, acorn, minimatch, and googletest dependency bumps.
- SQLite test and close-behavior fixes; in-memory test database cleanup.
- Misc build/tooling updates for perfetto and nix automation.
- Documentation and grammar edits across HTTP/2, QUIC, CLI, dgram, addons, and test docs.
- Smaller FFI, test runner, and SharedArrayBuffer pointer support fixes.