Runnable, single-file demos for the true_async_server extension.

With a TrueAsync PHP build (the extension already loaded):

php examples/ws-echo-server.php
PORT=9000 php examples/ws-echo-server.php     # every example honours $PORT (default 8080)

If you built the extension locally and it is not in your php.ini:

php -d extension=true_async_server.so examples/ws-echo-server.php

Every server listens on 0.0.0.0:$PORT and logs a start line to stderr. Stop it with Ctrl+C.

Connect with websocat (open two terminals for the chat demo):

websocat ws://127.0.0.1:8080/

…or from a browser console:

const ws = new WebSocket('ws://127.0.0.1:8080/');
ws.onmessage = e => console.log('recv:', e.data);
ws.onopen    = () => ws.send('hello');

The same handlers work unchanged over wss:// (TLS) and HTTP/2 Extended CONNECT (RFC 8441).

gRPC messages are your protobuf payloads; the 5-byte length-prefix framing is handled for you. Call with a real client (grpcurl, generated stubs), or raw over h2c — a frame is 1 flag byte + 4-byte big-endian length + payload:

# unary greeter → "hello, world"
printf '\x00\x00\x00\x00\x05world' | curl --http2-prior-knowledge -s \
  -H 'content-type: application/grpc' -H 'te: trailers' --data-binary @- \
  http://127.0.0.1:8080/helloworld.Greeter/SayHello | tail -c +6; echo

# bidi echo, two stacked request frames → two reply frames
printf '\x00\x00\x00\x00\x03aaa\x00\x00\x00\x00\x02bb' | curl --http2-prior-knowledge -s \
  -H 'content-type: application/grpc' -H 'te: trailers' --data-binary @- \
  http://127.0.0.1:8080/echo.Echo/Stream | xxd

Reload needs the worker pool (setWorkers() > 1). The bootloader is re-run on replacement workers, so edits apply atomically per rotation:

php examples/reload-server.php          # prints the pool-parent pid
curl 127.0.0.1:8080/                    # version=v1 boot=<id>
echo v2 > examples/app-version.txt      # "change the app"
kill -HUP <pool-parent-pid>             # SIGHUP rotates the pool
curl 127.0.0.1:8080/                    # version=v2, fresh boot id, zero downtime

Triggers: enableReloadOnSignal() (SIGHUP, prod), enableHotReload([dirs]) (filesystem watcher, dev), or $server->reload() (programmatic).

The server has no embedded exporter. getStats() returns a plain PHP array; twenty lines in the example turn it into the Prometheus text format, and you can swap them for OpenMetrics, StatsD or anything else without touching the server.

php examples/observability-server.php
curl -s localhost:8080/metrics

docker compose -f examples/docker/observability/docker-compose.yml up -d
open http://localhost:3001          # Grafana, no login, dashboard loaded

examples/docker/observability/smoke-test.sh   # asserts every hop end to end

The stack runs in Docker but the server stays on the host — it needs your locally built extension, and the point is to scrape a real one.

Why the counter kinds matter. Run reload() while Grafana is open: tas_requests_total keeps climbing (a retiring worker's totals are inherited, so a scraper never sees a counter run backwards), while tas_conns_active_h1 drops to what is really open (a dead worker holds no connections, so its last reading is not carried forward as a phantom).

Two traps the example is built to avoid. Under setWorkers(N) the handler is copied into a worker thread, and that thread has its own function table — a function render() declared in your script does not exist there, so anything the handler calls must arrive through use() or a bootloader. And use the file log sink, not stream: a parent-opened stream resource cannot cross into a worker.