# /etc/systemd/system/passkeeper.service
#
# Install / update:
#   sudo cp scripts/passkeeper.service /etc/systemd/system/passkeeper.service
#   sudo systemctl daemon-reload
#   sudo systemctl enable passkeeper
#   sudo systemctl restart passkeeper
#   journalctl -xeu passkeeper.service
#
# Phase 5 additions vs original:
#   - Restart=on-failure: systemd restarts Gunicorn if the master exits non-zero
#   - Gunicorn settings live in gunicorn.conf.py (gthread workers, timeout
#     ordered above nginx proxy_read_timeout, /dev/shm heartbeat dir)
#   - PrivateTmp, NoNewPrivileges, ProtectSystem: basic systemd sandboxing
#   - StartLimitIntervalSec / StartLimitBurst: caps restart storm

[Unit]
Description=PassKeeper Gunicorn daemon
After=network.target mysql.service
Wants=mysql.service

# Restart policy: cap to 5 restarts in 60 s to prevent restart storms
StartLimitIntervalSec=60
StartLimitBurst=5

[Service]
User=www-data
Group=www-data
WorkingDirectory=/home/spuser/PassKeeper
EnvironmentFile=/home/spuser/PassKeeper/.env

# All tuning lives in gunicorn.conf.py (worker class, counts, timeouts,
# logging) so it is versioned with the code and documented in one place.
# Override any of it with GUNICORN_* variables in the EnvironmentFile above
# rather than editing this line.
# Absolute path — do not rely on WorkingDirectory for config lookup.
ExecStart=/home/spuser/.venv/bin/gunicorn \
          -c /home/spuser/PassKeeper/gunicorn.conf.py \
          wsgi:app

# Reload: HUP re-reads config and restarts workers in place under the
# existing master.
#
# NOT USR2. USR2 forks a *second* master that inherits the listening socket,
# and retiring the old one needs a follow-up WINCH + QUIT that this unit never
# sent. The result was two masters competing for the port with systemd $MAINPID
# tracking the stale one — a later stop/restart then signalled the wrong
# process and the socket vanished, which Nginx reports as 502.
#
# HUP does not pick up changed Python source: use `systemctl restart` for code
# deploys, reload only for config-only changes.
ExecReload=/bin/kill -s HUP $MAINPID

# NO WatchdogSec here — deliberately.
#
# WatchdogSec requires the service to send WATCHDOG=1 keepalives over the sd_notify
# socket. Gunicorn only does that when systemd exports NOTIFY_SOCKET, which happens
# only under Type=notify (+ NotifyAccess=main). This unit is Type=simple (the
# default), so no keepalive was ever sent, systemd treated the service as hung, and
# SIGKILLed it every ~30 s. Restart=on-failure then brought it back after RestartSec,
# producing a repeating window of 502s from Nginx.
#
# Hung *workers* are already handled by Gunicorn's own `timeout` in gunicorn.conf.py; a crashed
# *master* is already handled by Restart=on-failure below. The watchdog added no
# coverage, only outages.
#
# To re-enable it properly (optional), all three lines are required:
#   Type=notify
#   NotifyAccess=main
#   WatchdogSec=30s

Restart=on-failure
RestartSec=5s

# Systemd sandboxing
PrivateTmp=true
NoNewPrivileges=true
ProtectSystem=strict
ReadWritePaths=/home/spuser/PassKeeper /home/spuser/logs /home/spuser/backups

[Install]
WantedBy=multi-user.target