Deployment
Process layout
One php omnitrix serve per host (or container) starts the OpenSwoole master, a manager, N HTTP
workers and optional CPU-pool processes. Run it under a supervisor
(systemd, s6, Kubernetes) that sends SIGTERM to stop and SIGUSR1 to reload code.
# /etc/systemd/system/omnitrix-api.service
[Service]
WorkingDirectory=/srv/api
ExecStart=/usr/bin/php omnitrix serve
ExecReload=/bin/kill -USR1 $MAINPID
KillSignal=SIGTERM
TimeoutStopSec=30
Restart=always
User=api
EnvironmentFile=/etc/omnitrix-api.env
LimitNOFILE=65536
Sizing
server.workers: number of cores available to the API (I/O-bound code needs no more).database.pool.maxis per worker: total connections = workers × max. Keep the sum below the database'smax_connectionsacross all API hosts.server.max_coroutinebounds in-flight requests per worker (503 above it).http_client.max_concurrentandpool.max_per_hostbound outbound fan-out.- CPU pool: one process per core you can spare for CPU jobs.
Shutdown and reload
- Default
server.mode = processdrains in-flight requests on SIGTERM (bounded bymax_wait_time), then closes pools inWorkerStop. Set the supervisor timeout abovemax_wait_time. - SIGUSR1 replaces workers one by one with new code; OPCache with
validate_timestamps=0needs a reload (or restart) to pick up new files.
Reverse proxy (optional)
Nginx in front terminates TLS and buffers slow clients:
upstream omnitrix { server 127.0.0.1:8080; keepalive 256; }
server {
listen 443 ssl http2;
client_max_body_size 1m; # match http.max_body_size
location / {
proxy_pass http://omnitrix;
proxy_http_version 1.1;
proxy_set_header Connection "";
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_read_timeout 35s; # above http.request_timeout
}
}
Behind a proxy, Request::remoteAddress() is the proxy; give RateLimit a key function
that reads a trusted forwarded header.
Kernel limits
Raise net.core.somaxconn and tcp_max_syn_backlog (≥ 4096) and the open-file limit. Small
values show up as 1 s SYN retransmits under bursts of new connections.
Logging
Write logs to a file or a collector that always reads; never leave stderr connected to a pipe
nobody drains (a full pipe blocks workers). Set server.log_file for OpenSwoole's own log.
Queue workers
Run php omnitrix queue:work as separate supervised processes (see Queues);
they stop gracefully on SIGTERM. Use a stop timeout above the queue's job timeout.
# /etc/systemd/system/omnitrix-worker@.service (systemctl start omnitrix-worker@1 …)
[Service]
WorkingDirectory=/srv/api
ExecStart=/usr/bin/php omnitrix queue:work
KillSignal=SIGTERM
TimeoutStopSec=90
Restart=always
User=api
EnvironmentFile=/etc/omnitrix-api.env
Releasing a new version
composer install --no-dev --optimize-autoloader
php omnitrix migrate # only migrate:fresh, migrate:reset, rollback and db:seed need --force in production
systemctl reload omnitrix-api # SIGUSR1: workers are replaced one by one
systemctl restart 'omnitrix-worker@*'
Checklist
debug => false; secrets only from the environment; TLS verification left on./metricsprotected (token middleware) or reachable only internally./health/livefor liveness,/health/readyfor readiness/load-balancer checks.- Database user with least privilege;
statement_timeoutset. - Supervisor stop timeout >
max_wait_time; file descriptor limit raised.