[igalshilman]

restate.serve

A one-line way to serve a Restate app, as the counterpart of restate.connect:

app = restate.app([greeter])

if __name__ == "__main__":
    asyncio.run(restate.serve(app))                      # serve on $PORT or 9080
    # asyncio.run(restate.connect(app, name="greeter"))  # or connect out to Restate Cloud

• What it serves: async def serve(app, *, host="0.0.0.0", port=None) listens with the cleartext HTTP/2 that Restate connects with, on uvicorn 0.53+ with zttp. It stops within a few seconds of SIGINT or SIGTERM.
• A new serve extra: pip install "restate-sdk[serve]" brings uvicorn and zttp on Python 3.11+. Without the extra, serve raises an ImportError naming it, and ordinary installs get no new dependencies.
• Signals: the app installs its own SIGTERM hook on its first request, through install_shutdown_hook. That keeps one asyncio handler per signal and would replace uvicorn's handlers, so serve registers the server's exit through the same hook first and turns off uvicorn's own signal capture. The server and the app's hook then both run.
• README: the example is served with restate.serve; hypercorn stays as the alternative, including for Python 3.10.

Tests

tests/serve.py:

• Without the extra, serve raises an ImportError naming it (checked in a subprocess with uvicorn hidden).
• End to end, a program serving a greeter with restate.serve:
  • registers with a real Restate, which reports HTTP/2.0 for the deployment;
  • answers a call through Restate;
  • stops promptly on SIGTERM after serving those requests.
  • (Skipped where the extra isn't installed, which includes Python 3.10.)

#262 (session objects and agents) will use this in place of its own sessions.serve.

🤖 Generated with Claude Code

[pullfrog]

ℹ️ No critical issues — three small rough edges inline.

Reviewed changes

Reviewed the full PR: a new restate.serve entrypoint on uvicorn+zttp HTTP/2, a serve extra, README updates, and tests. I also checked it locally with uvicorn 0.54 and zttp 0.0.35 on Python 3.12. /discover is served over HTTP/2, and both SIGTERM and SIGINT stop the process in about 0.25s while an h2 connection is still open. The shared-hook ordering (server exit first, then the app's _on_sigterm) works as the comment describes.

• restate.serve — lazily imports uvicorn/zttp, checks for uvicorn ≥ 0.53, and serves with http="zttp", http2=True, lifespan="off" and a 3s graceful shutdown.
• Signal handling — sends uvicorn's handle_exit through install_shutdown_hook, so it shares one asyncio dispatcher with the app's SIGTERM hook, and turns off uvicorn's own capture_signals.
• serve extra — uvicorn>=0.53 and zttp>=0.0.34, both limited to Python 3.11+. uv.lock now uses uvicorn 0.38 below 3.11 and 0.54 from 3.11 on.
• README — the example now uses restate.serve, with hypercorn still listed as the alternative (including for 3.10).
• Tests — a subprocess check of the missing-extra ImportError, plus a Docker end-to-end test: register, check HTTP/2.0, invoke, then SIGTERM.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using claude-opus-5-5 ｜ 𝕏}

[github-actions]

Test Results

  8 files  ±0    8 suites  ±0   3m 36s ⏱️ -19s
 71 tests ±0   71 ✅ ±0  0 💤 ±0  0 ❌ ±0 
322 runs  ±0  322 ✅ ±0  0 💤 ±0  0 ❌ ±0 

Results for commit ba95e1b. ± Comparison against base commit 39bb159.

[pullfrog]

✅ No new issues found. All three earlier findings are fixed.

Reviewed changes

Reviewed bcbdb31, the follow-up commit since the last Pullfrog review.

• Signal fallback — server.capture_signals is now replaced only after both install_shutdown_hook calls succeed. On Windows or in a non-main thread, the first add_signal_handler call raises, so uvicorn keeps its own signal handling and no hook is left half-installed.
• port=0 — an explicit 0 now reaches uvicorn as an OS-assigned port instead of falling back to $PORT/9080. The docstring now says this.
• Shutdown test — SIGTERM is now sent while Restate still holds its HTTP/2 connections. The test waits up to 10s for the program to exit and only then stops the container in finally, so docker stop time doesn't count against the budget.

  ^{｜ View workflow run ｜ Using claude-opus-5-5 ｜ 𝕏}

[pullfrog]

✅ No new issues found. The new commit makes the end-to-end test steadier and easier to debug when it fails.

Reviewed changes

Reviewed 97fea5e, a test-only commit to tests/serve.py since the last Pullfrog review.

• Container-first ordering — RestateContainer now starts before find_free_port(). Docker's host-port mappings are then already taken and can't collide with the program's port.
• Readiness wait — a new listening() helper polls a TCP connect on 127.0.0.1:{port} (up to about 10s) and fails early with the program's output if the program exits. This replaces the 100-try registration retry loop, so registration now happens once against a server that is already listening.
• Program output capture — stdout and stderr go to tmp_path/program.log and are attached to the readiness and shutdown assertion messages. Before, they were sent to DEVNULL.

  ^{｜ View workflow run ｜ Using claude-opus-5-5 ｜ 𝕏}