Skip to content

RELEASE.md

Latest commit

 

History

History
118 lines (94 loc) · 4.94 KB

RELEASE.md

File metadata and controls

118 lines (94 loc) · 4.94 KB
Raw
weft

Releases

What changed in each version of Weft, newest first.

Keeping this file. A change someone using Weft would notice gets a line under Unreleased in the commit that makes it. When a version is cut, that section is renamed to the version and its date, and a new empty Unreleased section goes above it.

Distribution name: weft-server. Import package weft, CLI weft.

Unreleased

0.1.1 — 2026-10-23

Fixed

  • Process workers: the first Ctrl+C no longer hangs the supervisor (workers are signalled over pipes, not multiprocessing.Event; #1, thanks @patx). SIGTERM to the supervisor or the thread-worker controller now drains gracefully, including --drain-delay and lifespan shutdown. A worker that outlives SIGTERM is killed and reaped; workers already started are stopped when a later one fails to start.
  • Unix sockets: every client was reset on accept; they are now served.
  • HTTP/2 and HTTP/3: request bodies stream to the app after dispatch instead of being buffered first, so early responses and --request-timeout work, and an oversized HTTP/2 body gets 413. Large HTTP/3 uploads to a slow app are no longer cut off. Static files are sent whole (they were truncated after 64 KiB). An h2 app gets http.disconnect when the client leaves after the body was read. A cleartext h2 preface split across reads is no longer taken for HTTP/1.
  • WSGI: write() no longer stalls on HTTP/2 and HTTP/3, and SERVER_PROTOCOL reports HTTP/2 there. A one-block list response no longer sends Content-Length twice.
  • Response cache: a body that does not match its Content-Length is not cached; responses are aged by their Date; Accept-Encoding variants no longer leak into each other or evict each other; requests with a body skip the cache-hit path (their body was parsed as the next request); shared slots are safe when a worker holding one pauses or dies.
  • Static files: a file that shrinks or fails mid-read resets the stream or closes the connection instead of looking complete; a precompressed file's ETag comes from that file.

Changed

  • WSGI over HTTP/1 keep-alive sends finished responses with the rest of their batch: about 1.25× requests/s at saturation and lower p99 on the minimal WSGI benchmark. ASGI unchanged.

0.1.0 — 2026-09-28

The first PyPI release.

Added

  • ASGI 3.0 HTTP (spec 2.4) on Tokio, driven from asyncio.select(). Native send / receive vectorcall objects, a copied prototype scope, interned keys, and a pre-completed awaitable when the socket is ready.
  • HTTP/1.1: keep-alive, pipelining, chunked transfer both ways, Expect: 100-continue, HEAD/204/304, smuggling checks, header timeouts, a head size limit, graceful drain.
  • WebSocket: handshake, framing, fragmentation, ping/pong, limits, queue backpressure, permessage-deflate.
  • WSGI (PEP 3333): protocol detection, native environ and start_response (including write()), streamed iterables, wsgi.file_wrapper, --wsgi-threads.
  • HTTP/2 over TLS (ALPN h2) and cleartext prior knowledge.
  • HTTP/3 over QUIC (--http3, --quic-port) and WebTransport as ASGI webtransport.*. Incoming streams and datagrams; CONNECT accept/refuse. weft.webtransport.WebTransportSession and weft.contrib (Starlette / FastAPI / Django routers, AltSvcMiddleware, IETF resumable uploads).
  • TLS via rustls: --tls-cert / --tls-key, SNI when repeated, --hsts, --redirect-http.
  • Static files (--static-dir), compression, shared response cache (--cache-size).
  • Access log (text or JSON), --health-check-path, --drain-delay, --max-connections, --forwarded-allow-ips, --request-id, --trace-context, --request-start-header, --rate-limit, unix sockets, --reload and SIGHUP rolling restart, --venv, --no-lifespan.
  • Prometheus on --metrics-port / --metrics-host.
  • Process workers and --worker-mode thread on free-threaded CPython 3.14t. Py_mod_gil = Py_MOD_GIL_NOT_USED.

Known

  • ACME is not implemented; leave issuance to Caddy.
  • Raw WSGI is slower than Peregrine's asyncio-free poller (76% at 256 connections on the contract). Raw ASGI and FastAPI match Peregrine and are 1.5–1.9× Granian 2.8.3. BENCHMARKS.md.

Publishing to PyPI

Wheels are built and uploaded by .github/workflows/release.yml on a v* tag. Version in Cargo.toml must match the tag (v0.1.0 → 0.1.0).

  1. On PyPI, add a pending trusted publisher for weft-server:
    • Owner: grepjava
    • Repository: weft
    • Workflow name: release.yml
    • Environment name: release
  2. git tag v0.1.0 && git push origin v0.1.0

The first upload creates the project. Later tags reuse the same publisher.