Skip to content

Releases

v0.5.10

Choose a tag to compare

@kaitk kaitk released this 24 Oct 08:24

HTTP/3. Goose v2026.3, goose-hash v0.0.4, goose-tls v0.0.6, and goose-compress
v0.1.4. No breaking changes.

Added:

  • HTTP/3 (RFC 9114) over QUIC version 1, written in Goose, in the optional module
    http.h3, over goose-tls v0.0.6's QUIC handshake and packet protection. It is off
    unless a listener asks (Listener.h3, beside tls), and a program that doesn't
    import the module opens no UDP socket (API). Linux's loops and
    macOS's kqueue; no later, 0-RTT, or migration yet.
  • A listener with no address binds HTTP/3 to every address, IPv6's included, as its TCP
    listener does.

v0.5.9

Choose a tag to compare

@kaitk kaitk released this 24 Oct 05:57

A smaller worker, timers on time, and WebSockets up to max_conns. Goose v2026.3,
goose-hash v0.0.4, goose-tls v0.0.5, and goose-compress v0.1.4.

Changed:

  • The loops' timers fall due by the precise monotonic clock on Linux and macOS, not
    the coarse one, which is up to a scheduler tick old: /delay/10 at 64 connections
    answered in 12.2 ms, and now in 10.2 ms
    (decisions).
  • A timer table adds its cells as timers need them, up to its capacity, rather than
    writing every cell at open, and a box's table of live keys is left as calloc
    zeroed it. A worker with max_waiting at max_conns, 16,384, held about 0.8 MiB
    before its first wait; now it holds what its waits use.
  • Config.ring_buffers is 64, down from 256: io_uring writes every provided buffer
    in turn, so each worker held 1 MiB of them at any load, and 64 serve as fast on
    Linux (decisions). On Windows, IOCP
    serves 1 to 7% fewer requests a second with 64; a program there that wants them
    back sets ring_buffers: 256.
  • A WebSocket is promised bytes of the gather budget only while it holds part of a
    message, a message in fragments or a frame split across reads, not from its upgrade
    on. Idle connections and messages that arrive whole cost no budget, so a worker holds
    as many WebSockets as max_conns allows, where the defaults stopped it at 1,021.
    ws_upgrade no longer answers 503; a message the budget can't hold closes its
    connection with 1013 (Try Again Later)
    (decisions).

Removed:

  • http::ws_admit. A program that called it before ws_switch drops the call; its
    WebSockets are then promised their bytes as ws_upgrade's are.

v0.5.8

Choose a tag to compare

@kaitk kaitk released this 23 Oct 20:19

Compression from 256 bytes, and goose-compress v0.1.4. Goose v2026.3, goose-hash
v0.0.4, goose-tls v0.0.5, and goose-compress v0.1.4.

Changed:

  • http::Compression's min_size is 256 bytes, down from 1,024: gzip saves bytes on
    JSON from about 192 bytes on, and an 888-byte answer went out raw
    (decisions). A program
    that wants the old threshold passes http::Compression { min_size: 1024, .. }.
  • goose-compress v0.1.4, whose DEFLATE level 1 matcher and symbol coder are C, with
    the same bytes: gzip at level 1 takes 12 to 15 percent less time on 4 to 8 KB JSON
    bodies.
  • The drain example stops on SIGTERM, so its io_uring listener closes.

v0.5.7

Choose a tag to compare

@kaitk kaitk released this 23 Oct 17:56

An io_uring send fix, and goose-compress v0.1.3. Goose v2026.3, goose-hash
v0.0.4, goose-tls v0.0.5, and goose-compress v0.1.3.

Changed:

  • goose-compress v0.1.3, whose CRC-32 uses the CPU's carryless multiply (PCLMULQDQ
    on x86-64, PMULL on arm64) in a C shim, and whose load32 is inlined. Gzip and
    deflate responses are faster, with the same bytes; in Plod's HttpArena entry,
    json-comp serves 20% more requests a second.

Fixed:

  • On io_uring, a send made at once, without the ring, reserves its completion before
    any byte goes out. Out of memory, it once returned an error after sending every byte;
    now it sends nothing.

v0.5.6

Choose a tag to compare

@kaitk kaitk released this 23 Oct 17:56

Fuzz targets for the parsers, and io_uring closes cancel by tag. Goose v2026.3,
goose-hash v0.0.4, goose-tls v0.0.5, and goose-compress v0.1.2.

Added:

  • content_type_of knows .map (application/json), .webmanifest
    (application/manifest+json), and .otf (font/otf).
  • fuzz/: libFuzzer targets for the request head and chunked body, HPACK, h2 frames,
    and multipart (development).

Changed:

  • On io_uring, closing a connection cancels its receive and its send in flight by
    their tags, and only when they're queued, rather than everything on its fd, which
    the kernel finds by walking its whole poll table
    (decisions).
  • goose-compress v0.1.2, whose DEFLATE, gzip, and CRC-32 are faster.

Fixed:

  • On io_uring, a file's send closed while it filled its pipe no longer leaves the
    pipe, with that piece in it, for the next file's send.

v0.5.5

Choose a tag to compare

@kaitk kaitk released this 23 Oct 04:10

Drain credits for publishers on upgraded connections. Goose v2026.3, goose-hash
v0.0.4, goose-tls v0.0.5, and goose-compress v0.1.1.

Added:

  • after_drain(out) requests a DRAINED event when an upgraded connection's output
    reaches its socket, so a publisher can keep one bounded batch in flight. Pending
    mailbox waits stay armed; closing cancels the credit. The application still bounds
    its producer queue and batches (API).

Fixed:

  • On Windows, a Config.backlog below 200 stays as asked, rather than being raised
    to 200.

v0.5.4

Choose a tag to compare

@kaitk kaitk released this 22 Oct 20:37

A blocking HTTP/1.1 and HTTPS client. Goose v2026.3, goose-hash v0.0.4, goose-tls
v0.0.5, and goose-compress v0.1.1.

Added:

  • The module http.client: HTTP/1.1 requests, plain or over TLS, on the blocking
    socket, for a lane's thread or a program without a loop. A session keeps its
    connection for the next request to the same host; http::get, post, put,
    patch, delete, and request take a block that adds header fields, a bearer
    token, or a body, http::json for a JSON one. It decodes chunked bodies, and
    doesn't follow redirects, keep cookies, or decode compressed bodies. A program that
    imports it declares goose-tls (docs/api.md, "An HTTP client").

v0.5.3

Choose a tag to compare

@kaitk kaitk released this 22 Oct 20:26

A TLS connection's context chosen per connection, for ACME's TLS-ALPN-01. Goose
v2026.3, goose-hash v0.0.4, goose-tls v0.0.5, and goose-compress v0.1.1.

Added:

  • A TLS connection's context chosen at its ClientHello, as ACME's TLS-ALPN-01 needs:
    http::tls_server(ctx, chooser) gives a worker an http::TlsServerOf<C>, and on a
    context given to tls::choose_context, the handshake calls the program's
    http::tls_hello(chooser, conn), which may move the connection with
    tls::use_context. http::TlsServer and http::tls_server(ctx) are unchanged,
    and add a check only where the handshake is advanced (docs/api.md, "HTTPS").

v0.5.2

Choose a tag to compare

@kaitk kaitk released this 22 Oct 20:26

Host names in the blocking connect, and goose-tls's faster server. Goose v2026.3,
goose-hash v0.0.4, goose-tls v0.0.5, and goose-compress v0.1.1.

Added:

  • http::connect(host, port, ms), the blocking connect for a thread without a loop,
    resolves a host name with the system's resolver, then tries its addresses in the
    resolver's order, up to 8, all within ms. The lookup itself isn't bounded by ms,
    and fails with resolve's ENOENT for a name with no address or EAGAIN for one
    that failed for now. The loop's http::connect_start still fails a name with
    ENOSYS, since resolving blocks the loop.

Changed:

  • goose-tls v0.0.5: the server's overhead cut, with one HKDF and one HMAC context
    reused through a handshake and records decrypted in place. In goose-tls's own
    measurements, 8.3% less CPU a json-like request, 2.7% a full handshake, and 11.8% a
    resumed one. No API change.

v0.5.1

Choose a tag to compare

@kaitk kaitk released this 22 Oct 20:26

h2 over TLS, goose-tls's own TLS 1.3 server, and outbound sockets released when an h2
stream ends. Goose v2026.3, goose-hash v0.0.4, goose-tls v0.0.4, and
goose-compress v0.1.1.

Added:

  • h2 over TLS on both loops, chosen by ALPN: a TLS listener whose context lists h2,
    as tls::server_files(chain, key, ["h2", "http/1.1"]), serves h2 to a client that
    offers it and HTTP/1.1 to one that doesn't, on one port. No new name; a context
    that lists only http/1.1 serves as before. h2spec passes over TLS on both loops.
    The benchmark server serves it from one context on both its TLS ports, the port + 1
    and the port + 363 (8081 and 8443 beside 8080, TLS_PORT and H2_PORT), and
    serves /static/ as /files/, for HttpArena's static profiles.
  • An idle h2 connection, h2c or over TLS, is sent GOAWAY before it closes, and over
    TLS close_notify after it.
  • http::file_tag(f), the ETag that serve_file sends, which serve_file now
    calls, so a program that answers a 304 or 416 itself can match it.
  • http::take_compressed(z, coding, c), which takes and resets an encoder from the
    worker's pool, and which begin_compressed now calls once it has written the head,
    so a program that writes its own head can stream compressed.
  • http::close_file(f), which closes a File that open_file opened and the
    program won't serve, the last of open_precompressed's parts to be public.

Changed:

  • goose-tls v0.0.4: the TLS server is goose-tls's own TLS 1.3 server in Goose over
    libcrypto, 2.9 KB per idle connection where libssl held 28.5 KB, with session
    tickets; a client offering ALPN without http/1.1 now fails with
    no_application_protocol. The kTLS handoff is gone, with the benchmark server's
    KTLS=1.

Fixed:

  • On io_uring, an h2 stream reset while it waits on http::wait_socket, or answered in
    the call that waited, cancels the socket's poll, which held a socket the handler hung
    up open until its next event or the connection's close. test/outbound/ tests both
    over h2c on every loop where wait_socket works.
  • On io_uring, an HTTP/1.1 handler that answers in the call that waited on
    http::wait_socket cancels the socket's poll too.
  • An h2 stream refused with 413, for a body past Config.max_body, calls its handler
    with CLOSED when it was owed a call, as a reset does, so a socket it holds is hung
    up. test/outbound/ tests both on every loop where wait_socket works.